WEBVTT
00:00:00.160 --> 00:00:06.000
I asked ChatGPT's Astra to make me a cool cybunk scene in 3D software Blender.
00:00:06.559 --> 00:00:08.080
This is what it built.
00:00:08.720 --> 00:00:11.439
Then, Nick, I rebuilt our website with it.
00:00:11.599 --> 00:00:19.359
Then I took an already AI built course after a Canvas cartridge and had the file inside Canvas, able to learn from it in about 10 minutes.
00:00:19.600 --> 00:00:20.800
The rubrics are set up.
00:00:21.039 --> 00:00:22.719
The quizzes were built perfectly.
00:00:22.879 --> 00:00:29.280
There was an authentic assessment that ran through the whole course and had me building up a portfolio to show at a job interview.
00:00:29.440 --> 00:00:35.759
And somewhere in there, while using OpenAI's latest model, I got a feeling that I haven't had with AI for a while.
00:00:36.079 --> 00:00:37.759
Holy smokes, Batman.
00:00:37.920 --> 00:00:39.439
What else can this thing do?
00:00:39.759 --> 00:00:40.719
Yeah, wow.
00:00:40.880 --> 00:00:43.119
Okay, I had a look at that blender demo.
00:00:43.280 --> 00:00:44.960
We we also need to talk about Fable though.
00:00:45.039 --> 00:00:47.920
I feel you're straying, and Astra's getting a lot of your attention.
00:00:48.240 --> 00:00:55.280
Well, Nvidia's Jensen Huang shouting out that Astra is probably AGI hasn't helped all the the the fanfare.
00:00:55.359 --> 00:01:01.039
But for our last episode of the season, we've each bought things that made us stop and say, have you seen this cool thing?
00:01:01.280 --> 00:01:13.599
Things that we built, things that other people are making, the fact that the state of the art is now trumped in about seven days, and later a game whose characters run on a simulation built from a fruit flies neural wiring.
00:01:13.760 --> 00:01:14.400
Yes, really.
00:01:15.760 --> 00:01:17.439
You need to explain that.
00:01:18.000 --> 00:01:18.799
Maybe we will.
00:01:18.879 --> 00:01:19.519
At some point.
00:01:19.840 --> 00:01:21.280
Let's get to the title.
00:01:30.959 --> 00:01:33.599
Nick, it is our last episode of the season.
00:01:33.680 --> 00:01:34.640
So I hope you're well.
00:01:34.799 --> 00:01:35.840
Are you doing good?
00:01:36.239 --> 00:01:37.200
Doing great, man.
00:01:37.280 --> 00:01:44.560
Like uh really, really yeah, bitter, bit of sweet, but I think a really good note to go off on, especially as sunlight is off on holiday for a while.
00:01:44.799 --> 00:01:46.400
A little bit of a jort.
00:01:46.640 --> 00:01:47.200
A bit of a jort.
00:01:47.359 --> 00:01:50.799
Well, we're back before the end of the year, but we thought we'd finish the season with a bit of a show and tell.
00:01:51.040 --> 00:01:57.840
We spend plenty of time picking about these technologies and what they mean, what could go wrong, what universities should do about them.
00:01:58.000 --> 00:02:04.799
So today we're giving ourselves a bit of room to enjoy the fact that some of this stuff is absolutely extraordinary.
00:02:05.120 --> 00:02:09.120
My last few weeks with Astra has given me, wow, this technology is so powerful.
00:02:09.280 --> 00:02:23.439
That feeling again, and sure, the pendulum swings to existential dread because of all the things it can do and what that means for future economic possibilities, or when safety researchers leave certain frontier labs and said that frontier labs were in an arms race to bring the end of humankind as we know it.
00:02:23.680 --> 00:02:26.479
But hey, let's try keep it light.
00:02:26.800 --> 00:02:28.879
Doomsday Dale is not here today.
00:02:29.199 --> 00:02:36.879
So I'll start with Astra, because we've been out for a few weeks now, and I'm still getting that urge to just ask it increasingly unreasonable questions.
00:02:37.439 --> 00:02:40.560
Was that a so such as a very detailed blender brief?
00:02:40.800 --> 00:02:46.240
I want to be clear, my very detailed render blender brief, and you can watch on YouTube for the actual clip that it made.
00:02:46.400 --> 00:02:48.560
It was make me a cool cyberpunk thing.
00:02:48.800 --> 00:02:55.680
There was clear artistic vision here, but I watched as it scanned my computer to see if a 3D software blender was installed.
00:02:55.840 --> 00:03:00.639
It set up the render settings, the camera, it rigged up a character that could move about the world.
00:03:00.800 --> 00:03:05.360
All the things that are just not that fun when you're building out cool 3D scenes.
00:03:05.599 --> 00:03:06.879
I was shooketh.
00:03:07.120 --> 00:03:15.360
Also, I used this wireframe as like a skeleton and took it into AI video generation platform C Dance and applied textures to it, again, completely guided by Astra.
00:03:15.680 --> 00:03:20.639
It was pure flow state, and God, it tickled that part of my brain for creating something cool.
00:03:20.719 --> 00:03:22.479
It just happened so fast.
00:03:23.120 --> 00:03:26.400
I don't think I have the deep, deep pockets and resources that you do.
00:03:26.479 --> 00:03:29.599
I've I have played a bit, I've dabbled, I've dabbled is the best I can do.
00:03:29.680 --> 00:03:30.800
And I'm absolutely blown away.
00:03:30.879 --> 00:03:33.919
Some of the demos that I've seen online are extraordinary.
00:03:34.080 --> 00:03:36.479
Your cyberpunk video, that is sick.
00:03:36.639 --> 00:03:41.919
I would pay money and line up and sit in dark rooms surrounded by strangers, iMacs or otherwise to watch that.
00:03:42.080 --> 00:03:43.199
That is amazing.
00:03:43.439 --> 00:03:47.520
It does look cool, and again, kind of want to build out the world a little bit more and again kind of flesh this out.
00:03:47.599 --> 00:03:49.039
And this is what these tools kind of offer.
00:03:49.120 --> 00:03:52.400
But Nick, the spread of what Astra can do is what I loved.
00:03:52.479 --> 00:03:56.159
So our podcast website has for a long time been sorely lacking.
00:03:56.319 --> 00:04:00.080
Uh your HTML is broken, was one of the pieces of feedback that I received.
00:04:00.159 --> 00:04:03.840
And I was far too ashamed to admit that it was just my personal credit taste there.
00:04:04.000 --> 00:04:09.520
So I know this isn't purely Astra example, but what I did note is that I didn't have to go back and forth.
00:04:09.759 --> 00:04:12.560
I just had the thing after a single prompt.
00:04:12.800 --> 00:04:18.079
And as for building out learning content, it's been something that I've tried on previous models, and the results were okay.
00:04:18.240 --> 00:04:25.360
But with Astra, we built out learning materials, the lesson flow, the assessments, even lesson plans and how this might translate to a face-to-face environment.
00:04:25.519 --> 00:04:28.560
And then after it actually drained up its own LMS environment.
00:04:28.720 --> 00:04:31.759
So my push was like, how do you take this into the canvas environment?
00:04:32.000 --> 00:04:34.000
It made its own LMS.
00:04:34.399 --> 00:04:35.040
It did.
00:04:35.600 --> 00:04:38.240
And I was like, I don't want that, I already have one of those.
00:04:38.560 --> 00:04:39.759
Oh my god.
00:04:39.920 --> 00:04:42.160
All right, so that's insane.
00:04:42.639 --> 00:04:49.439
The the the lag of so that was where the time was spent to make making from scratch a learning management system.
00:04:49.680 --> 00:04:50.079
Correct.
00:04:50.160 --> 00:04:54.240
Oh my god and I was like, hey, thanks so much for doing this and taking my content and putting it to it.
00:04:54.319 --> 00:04:54.720
But I have this.
00:04:54.800 --> 00:04:56.480
Can you just put into a Canvas cartridge?
00:04:56.720 --> 00:04:57.439
And it was great.
00:04:57.519 --> 00:05:01.439
The quick bit was kind of packaging it all up, that existing material into Canvas.
00:05:01.600 --> 00:05:03.759
That's something that would take me like a week to do.
00:05:04.079 --> 00:05:06.399
It's almost just boring, busy work.
00:05:06.639 --> 00:05:09.279
And the level of quality I received was great.
00:05:09.439 --> 00:05:22.879
So once I had imported it, I expected things like broken headings, rubrics that need to be wired up again, broken links, all the things that anyone who's built any of these tools and used AI for the last couple of years kind of expects, like to get halfway there, but you're still gonna need to be a work.
00:05:23.120 --> 00:05:25.360
But no, this one was perfect.
00:05:25.519 --> 00:05:38.959
I mean, sure, the proof is in the classroom about whether this is good, is high quality, but we're now in a world where you can pump in your content and get a very decent learning experience, including images, diagrams, in a really quick amount of time.
00:05:39.120 --> 00:05:40.160
This is crazy.
00:05:40.480 --> 00:05:45.360
I know a file export is a really difficult thing to make sexy on YouTube, but extrapolate, yeah.
00:05:45.600 --> 00:05:45.839
Right.
00:05:46.000 --> 00:05:53.839
Yeah, so which is why we opened up on the amazing city streets of your your cyberpunk uh to sing my incredibly strong editorial judgment.
00:05:54.000 --> 00:05:57.279
But across these three examples, I kept having that same reaction.
00:05:57.439 --> 00:05:59.439
Hang on, I can just ask to do this as well.
00:05:59.600 --> 00:06:07.519
And that's different from how I felt with a lot of these incremental improvements we've seen since I guess even GPT-5 and I guess was it Opus 4?
00:06:07.680 --> 00:06:13.759
Sure, maybe it could do some of these cool things, but after a whole bunch of tinkering and heartache, this kind of removed that struggle.
00:06:13.920 --> 00:06:16.800
This had me looking around at other things that I wanted to make.
00:06:16.959 --> 00:06:22.879
That app, that game, it is just the coolest little toolbox, and what's more, it's actually quite an economical model.
00:06:23.040 --> 00:06:30.720
I spent the greater part of this year using Claude, but I ride my usage meter the whole time, terrified that Fayette was going to eat everything on one of its endless floor paths.
00:06:31.040 --> 00:06:40.319
Astra's new model, I don't actually have that same fear, and the fact that the lowest end thinking of Astra, so like Astra Low, I think it is, majorly outforms the last sweeter models.
00:06:40.480 --> 00:06:42.160
It's pretty good bang for buck.
00:06:42.480 --> 00:06:43.519
Full call out though.
00:06:43.600 --> 00:06:46.000
I am waiting for my limits to reset tomorrow.
00:06:46.160 --> 00:06:50.800
So before I completely disappear into my own experiments, Nick, what have you got?
00:06:50.959 --> 00:06:51.680
What have you bought?
00:06:51.839 --> 00:06:55.279
Start with one of something that you really want to show me.
00:06:55.839 --> 00:06:58.560
You called Dibbs on Astro.
00:06:58.800 --> 00:07:05.839
And so what I got is I got to use that as an excuse to go down a little rabbit hole of mine, which is around robotics and world models.
00:07:05.920 --> 00:07:08.560
So I want to give you a before first so you feel the after, right?
00:07:08.639 --> 00:07:15.439
So July this year, one of the best robotics labs in the world taught a robot to turn a sock inside out.
00:07:15.759 --> 00:07:21.519
Nick, may I remind you that this episode is to keep our audience very engaged?
00:07:21.839 --> 00:07:23.360
You have bought me a sock.
00:07:25.040 --> 00:07:27.040
Master has bought me a sock.
00:07:27.279 --> 00:07:29.439
No, no, but here's the interesting thing, right?
00:07:29.519 --> 00:07:35.040
So to do this, a human set of the controls and puppeted the robot through the task 176 times, right?
00:07:35.120 --> 00:07:38.879
So eight hours, one sock, same lap, same month peeling an orange.
00:07:38.959 --> 00:07:40.319
They went through so many Mandarins.
00:07:40.480 --> 00:07:48.480
The researchers joke that the local grocer was looking at a takeover when it worked 52% of the time, way slower than you, right?
00:07:48.879 --> 00:07:51.439
But that was July, and now it's August.
00:07:51.600 --> 00:08:02.160
Someone stands in a kitchen with a phone at eye level, films themselves potting a plant 10 minutes, hand the robot the video, and the robot pots the robot pots the plant.
00:08:03.199 --> 00:08:03.759
Hallelujah.
00:08:06.160 --> 00:08:10.000
One oh no exactly.
00:08:10.319 --> 00:08:10.959
Exactly.
00:08:11.199 --> 00:08:14.800
So this is no no no no, but thinks you this, because there are other fun things with it as well.
00:08:14.879 --> 00:08:21.360
So this is a company called Skilled, August, one video of a person, not a robot, but a human film from their own point of view.
00:08:21.600 --> 00:08:22.639
Robot does task.
00:08:22.800 --> 00:08:31.120
Things that are never seen, like potting a plant, pancakes, pour over coffee, assembling that annoying IKEA furniture up to 10 minutes, dozens of steps.
00:08:31.439 --> 00:08:32.720
It's just trained on the video.
00:08:32.960 --> 00:08:34.960
No, and that's the crazy thing, right?
00:08:35.039 --> 00:08:37.759
So nothing is trained, the weights don't change.
00:08:37.919 --> 00:08:39.279
It's kind of like a recipe.
00:08:39.440 --> 00:08:42.960
It reads the video the way that you or I would read a recipe and it does it.
00:08:43.120 --> 00:08:46.960
You know, you know how historically you would paste an example historically.
00:08:47.279 --> 00:08:51.120
Two months ago, you would paste an example into a chatbot and say, write me one like this.
00:08:51.360 --> 00:08:52.080
Same trick.
00:08:52.240 --> 00:08:56.639
The example is a video, and the thing copies, the thing it copies is a set of movements.
00:08:56.799 --> 00:08:58.559
The video is the instruction.
00:08:58.720 --> 00:09:05.120
And now this sounds like a really odd way to kick off a discussion about this, but this lends itself to something called world models.
00:09:05.279 --> 00:09:11.840
And I want to nerd out feverishly with those with you in just a minute because it is, I think it's the one after Astra.
00:09:12.159 --> 00:09:20.240
I think it's the next stage when we get past like just text and video and voice and images, like those amazing ones that you just showed us.
00:09:20.320 --> 00:09:23.679
I'm sort of I'm looking forward to 27, very futurist of me.
00:09:23.840 --> 00:09:32.639
But this is this stuff is really, is really crazy that we've now got machines that are now learning from an input as complex as a video, kind of like we do.
00:09:32.960 --> 00:09:33.519
Ooh.
00:09:33.919 --> 00:09:39.279
Nick, you you've you've got me interested in socks and the world models.
00:09:39.360 --> 00:09:43.679
But before we do that, I want to talk about what other people have been doing with all these cool models.
00:09:44.480 --> 00:09:46.399
First, look at what people are making.
00:09:46.559 --> 00:09:56.559
So Ethan Moloch, often read uh often mentioned on the show, he's been using Astra to make an explorable reconstruction of the library of Alexandria in a surrounding setting.
00:09:56.720 --> 00:10:02.960
So you can walk around, you can listen to a narrator tour, you can open texts within the library to read.
00:10:03.120 --> 00:10:06.320
He's published this project as well, so you can actually go out there and play with it yourself.
00:10:06.639 --> 00:10:12.240
I don't want to throw any shade, but you rebuilt our website and he rebuilt Alexandria.
00:10:12.559 --> 00:10:17.120
I would argue that we've both made important contributions to the internet in the last week or so.
00:10:17.360 --> 00:10:25.600
But look at the combination here, because you've got a place to explore, you've got things to interact with, you've got material to read, and the project does acknowledge the historic historical uncertainty.
00:10:25.840 --> 00:10:36.799
Its documentation separates evidence from reconstruction and make clear that layout is actually an interpretation, which gives it an interesting way to teach and an interesting teaching question.
00:10:36.960 --> 00:10:40.000
Imagine asking students to explore something like this and then challenge it.
00:10:40.080 --> 00:10:40.799
What can we support?
00:10:40.879 --> 00:10:42.000
What has somebody inferred?
00:10:42.080 --> 00:10:44.639
What has been added to make the experience work?
00:10:45.919 --> 00:10:47.200
And then there's interesting things.
00:10:47.360 --> 00:10:50.799
Do you have them use the environment or or make their own?
00:10:51.279 --> 00:10:51.919
100%.
00:10:53.039 --> 00:10:55.600
Both sound interesting to me, but think about what this opens up.
00:10:55.679 --> 00:11:03.120
Let's stop talking about cheating, but let's start talking about, hey, you can get people to build entire worlds, especially making a small part of something, and I have to justify my choices.
00:11:03.279 --> 00:11:07.519
You'd have to explain why you put a broom there, why that object belongs, what evidence you've used.
00:11:07.759 --> 00:11:10.559
But even before we get to teaching, I love the ambition of the request.
00:11:10.799 --> 00:11:13.440
Could you build me somewhere that I can run around?
00:11:13.600 --> 00:11:15.440
And you just can with Astra.
00:11:15.600 --> 00:11:21.039
And there's another one that I want to talk about, because it's probably closer to the thing that freaked me out about Astra a little bit.
00:11:21.279 --> 00:11:31.360
Max Weinbeck gave it images of Apple Park, which is famously where the iPhone and everything else is kind of built, and basically said to Astra, rebuild this in Blender.
00:11:31.759 --> 00:11:32.720
And it did.
00:11:34.000 --> 00:11:36.399
Wow, that's ridiculous.
00:11:37.440 --> 00:11:41.360
And it's a beautiful, beautiful building, and it builds every single environment.
00:11:41.440 --> 00:11:43.200
There's a huge amount of um images.
00:11:43.279 --> 00:11:47.440
I would highly recommend you kind of look at what it how it kind of approached it.
00:11:47.600 --> 00:11:50.000
Use something like 6.27 billion tokens.
00:11:50.159 --> 00:12:12.799
It responded 30,000 times in that in that model, spun up 489 different subagents, crazy, and had to look at a bunch of two-dimensional images, understand what they're dealing with with a real three-dimensional place, work out the geometry, the relationships between the buildings, the material, the landscape, the lighting, then actually operate Blender, which is a notoriously hard piece of software, to produce something.
00:12:13.039 --> 00:12:14.799
Max described the result as absurd.
00:12:15.039 --> 00:12:19.279
I think that's a reasonably fair assessment, and this connects directly to what I was saying about my cyberpunk scene.
00:12:19.360 --> 00:12:23.360
I gave it an incredibly sophisticated prompt, make me a cool cyberpunk scene.
00:12:23.759 --> 00:12:24.879
Masterclass.
00:12:25.200 --> 00:12:25.759
Correct.
00:12:26.159 --> 00:12:31.519
But Max is effectively saying, here is a real place, work backwards from the images and reconstruct it.
00:12:31.679 --> 00:12:35.519
And that gets me because we've had models that can look at the image for a long time.
00:12:35.679 --> 00:12:43.919
They can tell you what's in Apple Park, they can describe the architecture, they can probably give you an incredibly detailed explanation of how you might model it, but this is different experientially.
00:12:44.080 --> 00:12:46.960
You give it images and then there's Apple Park models sitting in Blender.
00:12:47.440 --> 00:12:51.360
It was doing things like scanning public records for the type of tree that it should model.
00:12:51.519 --> 00:12:55.519
He ran different experiments to ensure that the glass rendering matched reality.
00:12:55.679 --> 00:13:01.360
That's a judgment of a professional and the problem solver of someone who doesn't want to go to San Francisco but knows that the information is somewhere.
00:13:01.679 --> 00:13:03.600
That's astounding.
00:13:03.759 --> 00:13:06.720
I mean, the the token usage just for one.
00:13:07.279 --> 00:13:11.279
And the the refraction just as a single problem.
00:13:11.519 --> 00:13:12.879
And problem's not the right word.
00:13:12.960 --> 00:13:14.720
Like, I mean, those are so complex.
00:13:14.799 --> 00:13:26.480
And I'm just to channel the cynic of of anyone who might be this thing, uh obviously there would be the that's astounding, but there would there would maybe be the caveat and architect, 3D artists would find all sorts of things wrong with it, yeah.
00:13:26.720 --> 00:13:27.120
Probably.
00:13:27.200 --> 00:13:32.000
And I'm not claiming it has reverse engineered Apple's architectural drawings from some photographs.
00:13:32.399 --> 00:13:36.720
The dimensions, the dimensions are probably wrong, details can be invented.
00:13:36.799 --> 00:13:42.720
Someone who actually does this professionally may judge this as is going to judge the model far better than I can.
00:13:42.960 --> 00:13:46.080
Also, it kind of gave up on things when there was an information vacuum.
00:13:46.159 --> 00:13:49.679
But still, it did make the thing, and that is the Astra feeling for me.
00:13:49.759 --> 00:13:56.720
And this is where a point where another five like little points on a benchmark, we keep seeing all these benchmark graphics, they actually don't matter anymore.
00:13:57.039 --> 00:13:57.440
Yeah.
00:13:57.679 --> 00:14:02.320
I mean, presumably then the the next question is what happens, those those information gaps things?
00:14:02.480 --> 00:14:06.399
What happens when you give it somewhere less documented than Apple Park?
00:14:06.639 --> 00:14:06.960
Correct.
00:14:07.200 --> 00:14:12.399
Or your own building, your laboratory, a historical location where you have enough reference material, a classroom, a house, etc.
00:14:12.559 --> 00:14:12.799
etc.
00:14:13.279 --> 00:14:16.720
What else could I give this thing and see whether it can do it?
00:14:16.879 --> 00:14:20.159
Perhaps it would build a track within Ableton.
00:14:20.320 --> 00:14:24.000
And this one gives our audio listeners something as well, assuming it plays.
00:14:24.480 --> 00:14:42.000
Uh Pietro Shennaro, who after building Atlantis and a Beyblade PS1 game, which looks addictive as hell, and a video to code demonstration, connected Astra to music platform Ableton through a software integration, and he says it created the synth instruments, the musical parts, and the arrangement for this track.
00:14:42.240 --> 00:14:45.840
I'm not a music guy, someone who knows music.
00:14:45.919 --> 00:14:48.799
I love the music, but I can listen to this and go, this is cool.
00:14:48.879 --> 00:14:49.840
This is a cool track.
00:14:50.240 --> 00:14:52.559
It's got that kind of lo-fi vibe, etc.
00:14:52.720 --> 00:14:52.960
etc.
00:14:53.200 --> 00:14:55.279
But again, really, really cool.
00:14:55.519 --> 00:15:00.960
You could say things like, keep that bass line, make the opening stranger, give me something completely excessive when the drums come in.
00:15:01.200 --> 00:15:06.320
And that's the part that I want to play with and can play with now with things like Astra.
00:15:07.039 --> 00:15:12.080
Yeah, okay, so I'm guessing we're coming back next season with a 12-minute uh theme tune.
00:15:12.159 --> 00:15:12.480
Yes.
00:15:12.720 --> 00:15:13.679
And it will be art.
00:15:13.840 --> 00:15:18.720
It'll be a concoction between me and Astra with a beautiful cyberpunk rendering.
00:15:19.120 --> 00:15:21.679
This is what I mean about the range though, historical environments.
00:15:21.759 --> 00:15:27.519
An editable, editable locomotive, a music production project, walking around Apple Park.
00:15:27.759 --> 00:15:31.279
Each example has its own tools and setup, but there's work around the model.
00:15:31.360 --> 00:15:35.120
But the variety of jobs people are attempting with it is just what keeps capturing me.
00:15:35.519 --> 00:15:41.759
Alright, Nick, I'm assuming you wanna you wanna talk about about Fable about Bad Anthropic for a bit.
00:15:42.240 --> 00:15:46.559
I mean, before you get your Astra T-shirt out, like I mean, I think I think we should, yeah?
00:15:46.879 --> 00:15:51.840
Well, I think Fable 5, particularly 5.1, deserves a proper place in this conversation.
00:15:52.000 --> 00:16:01.200
At Astra's launch, artificial analysis had Fable 5.1 ahead in both its intelligence index and its coding agent index in the configuration they tested.
00:16:01.360 --> 00:16:06.879
So on balance, Fable 5.1 is apparently better than Astra.
00:16:07.919 --> 00:16:13.360
Which I mean, yeah, which on paper is is is amazing, right?
00:16:13.519 --> 00:16:16.000
But I I was just listening to the Ableton thing there.
00:16:16.080 --> 00:16:17.279
It just came through on my end.
00:16:17.360 --> 00:16:25.200
I mean, these things like the the cyberpunk thing, the Apple Park, the the the this this response then, why am I blown away?
00:16:25.279 --> 00:16:27.279
This makes this response quite interesting, yeah.
00:16:27.600 --> 00:16:30.960
Well, Fable has repeatedly given me that feeling of wow, that is really smart.
00:16:31.120 --> 00:16:33.679
Also, thankfully, the writing style is not horrible.
00:16:33.919 --> 00:16:37.519
Fable 5, like OG one was horrendous.
00:16:37.679 --> 00:16:38.879
I've said it once and I'll say it again.
00:16:38.960 --> 00:16:41.440
I'll not hang out with Fable 5.0 on weekends.
00:16:41.519 --> 00:16:43.679
It was the worst, they've kind of fixed it in 5.1.
00:16:43.840 --> 00:16:47.919
But Astra has given me a very physical reaction to the breadth of what I can get done.
00:16:48.000 --> 00:16:49.440
And I'm describing my own experience here.
00:16:49.519 --> 00:16:51.600
I haven't conducted comprehensive comparisons.
00:16:51.759 --> 00:16:55.360
I'm assuming Fable is capable of substantial practical work too.
00:16:55.600 --> 00:16:59.360
Actually, the example I wanted to bring kind of brings that up quite well.
00:16:59.840 --> 00:17:14.319
Anthropic reports that Fable 5.1 used historical radar images from NASA's Magellan mission alongside an existing elevation map to train a neural network and produce a higher resolution elevation map covering roughly a third of Venus.
00:17:14.960 --> 00:17:16.000
No shade.
00:17:16.240 --> 00:17:19.119
Again, a fairly strong answer to your website.
00:17:19.359 --> 00:17:20.799
That's very brutal.
00:17:20.880 --> 00:17:23.839
But we've established a bit of a hierarchy, I'd imagine.
00:17:23.920 --> 00:17:27.039
We've got my website, we've got Alexandra, we've got another planet.
00:17:27.279 --> 00:17:34.000
We can let other anthropic says that new map reveals details around two to three kilometers.
00:17:34.160 --> 00:17:34.880
Just think about that.
00:17:34.960 --> 00:17:35.119
Wow.
00:17:35.359 --> 00:17:36.799
Think about how far the innerness is away.
00:17:36.880 --> 00:17:44.319
I don't have that number on me, but we can see, we can understand to two to three kilometres compared with 10 to 20 kilometers previously.
00:17:44.400 --> 00:17:45.200
That's their published result.
00:17:45.440 --> 00:17:47.519
So we should keep it all attributed to Anthropic.
00:17:47.680 --> 00:17:50.559
But as an example of something to attempt with these tools, it really is extraordinary.
00:17:50.640 --> 00:17:56.880
It makes you wonder what's sitting in existing data sets that somebody hasn't had the time yet or the resource or the right method to actually investigate.
00:17:57.200 --> 00:18:03.839
I think that's a great qualifier because it adds nuance, it complicates the idea that this one is the clever one and the other one does things, yeah.
00:18:04.319 --> 00:18:04.880
Correct.
00:18:05.359 --> 00:18:06.720
They can both do things.
00:18:06.880 --> 00:18:09.680
The difference I'm trying to describe is how they've landed for me.
00:18:09.759 --> 00:18:11.279
And it's just kind of the tool you want to play with.
00:18:11.359 --> 00:18:15.920
With Astra, I kept moving between very different kinds of work and finding something useful on the other side.
00:18:16.079 --> 00:18:19.920
And I suspect that's why it has occupied so much of my attention at this stage.
00:18:21.680 --> 00:18:22.559
That's sick.
00:18:22.720 --> 00:18:23.359
That's sick.
00:18:23.519 --> 00:18:24.720
I hadn't seen the Venus one.
00:18:24.880 --> 00:18:29.039
Like, I mean, just in terms of those applications, as you say, from vast data sets.
00:18:29.279 --> 00:18:38.720
Um, shout out to the laggard, maybe in this in the stage, some of the Google Earth stuff that's coming out in terms of like weather mapping and everything, is just like an absolute bananas.
00:18:39.119 --> 00:18:46.480
And these real-world applications to vast libraries of corpora of text that we've had, we've just been sitting on for forever.
00:18:46.720 --> 00:18:58.960
And we we we take them and turn them into something that you do not need a PhD in five weeks of review time or five months or five years, you know, that you can instead and engage with and enjoy because some of these things.
00:18:59.200 --> 00:19:02.000
Venus, a topographical map of Venus, is just cool.
00:19:02.160 --> 00:19:04.559
Eight-year-old me is like, whoa, that's sick.
00:19:04.720 --> 00:19:10.720
I want to like to see it at the same lenses like Google Maps to be able to do that on Venus, another like incredible.
00:19:11.039 --> 00:19:17.759
So let's talk about AGI for a little bit because because this is this is what's been around since these new worlds have come out.
00:19:17.920 --> 00:19:19.519
We may be here at AGI.
00:19:19.680 --> 00:19:21.759
And I want to say something that will probably annoy people.
00:19:22.000 --> 00:19:26.720
At this stage, the AGI definition doesn't really matter to how seriously I take this technology.
00:19:26.880 --> 00:19:29.920
The line of what machines can actually do keeps moving.
00:19:30.079 --> 00:19:36.079
So let's look at the mass example recently that's got a bit of controversy, solves super challenging Earth problems.
00:19:36.319 --> 00:19:37.680
We'll ignore the controversy.
00:19:37.839 --> 00:19:46.960
OpenAI threw thousands of Asians at these problems in parallel, pursuing different approaches, following promising leads, and then using Astra itself to help formalize and verify the results.
00:19:47.119 --> 00:19:51.359
It's a pretty extraordinary way of attacking a problem humans have been wrestling with for decades.
00:19:51.519 --> 00:19:55.680
And the model that started doing it was trained on August 28th.
00:19:55.920 --> 00:19:58.400
Astra came out on September 3rd.
00:19:58.559 --> 00:20:02.720
So we're looking at Astra right now, saying, wow, look at what this thing can do.
00:20:02.880 --> 00:20:12.160
And the next thing, the next state-of-the-art model is already in the building solving complex math problems, doing stuff like solving super hard math problems with one million dollar bounties.
00:20:12.319 --> 00:20:15.200
So maybe the question isn't less have we reached HEI?
00:20:15.359 --> 00:20:20.160
Maybe it's how quickly these list of things we thought machines couldn't do disappears.
00:20:20.640 --> 00:20:24.079
Because right now, I feel like that list keeps getting shorter.
00:20:24.319 --> 00:20:26.720
We could spend the whole episode trying to decide whether it qualifies.
00:20:26.799 --> 00:20:29.359
I'm more interested in what happens when I give it a job.
00:20:29.519 --> 00:20:30.160
Can it do it?
00:20:30.240 --> 00:20:31.279
Where does it fall over?
00:20:31.519 --> 00:20:32.799
How much checking does it need?
00:20:32.960 --> 00:20:35.039
What happens when I let it work across several steps?
00:20:35.119 --> 00:20:37.759
Those questions affect what I'm gonna do tomorrow.
00:20:38.559 --> 00:20:51.119
Just to shout out Gary Marcus here, like I mean, perennial skeptic, like he he points out that there's quite a difference between what a human, so artificial general intelligence is roughly equated to what a human can do, right?
00:20:51.440 --> 00:20:52.559
But which human?
00:20:52.799 --> 00:20:56.880
Are we talking like the drunk tired guy at a bar face down at 4 a.m.
00:20:57.119 --> 00:21:00.160
on an insert your local street where people do things of that nature?
00:21:00.400 --> 00:21:06.960
Or are we talking your Einstein level physicist in the specific area of expertise, or perhaps Einstein doing something he's terrible at?
00:21:07.119 --> 00:21:07.759
You know what I mean?
00:21:07.839 --> 00:21:12.079
So these are very soft uh measures and metrics and everything else.
00:21:12.240 --> 00:21:23.119
So I think I think he's probably and I I get the feeling from you that you're comfortable leaving the label unresolved, that there's maybe a hype element marketing in there as well.
00:21:23.440 --> 00:21:24.400
Incredibly comfortable.
00:21:24.559 --> 00:21:31.200
I personally feel we will cross the Rubicon, enter the singularity, pick your term, and we'll look back and go, oh, was that it?
00:21:31.440 --> 00:21:34.160
As the next magical thing happens, or perhaps terrifying thing.
00:21:34.319 --> 00:21:37.119
There's people doing research can keep working on those precise definitions.
00:21:37.279 --> 00:21:43.119
I don't need a verdict on AGI before I'm allowed to look at this thing and go, wow, this is a very powerful technology.
00:21:43.440 --> 00:21:46.799
And Astra has also given me a little bit of the fear.
00:21:47.599 --> 00:21:49.279
The fear, yeah.
00:21:49.920 --> 00:21:52.160
Um I don't know if you felt it recently.
00:21:52.559 --> 00:21:56.319
I've been I've been fearless recently, but Astra it's crapped back in.
00:21:56.799 --> 00:21:59.680
The moment when those implications start to catch up with you, you're enjoying something.
00:22:00.240 --> 00:22:02.160
You're asking it to make a thing, it produces a result.
00:22:02.640 --> 00:22:12.640
And then you stop for a second because you realize you had an assumption about how difficult that task was, or how many different skills you'd need, or how much help you'd have to organise before you could even attempt that thing.
00:22:12.880 --> 00:22:14.720
But suddenly you're looking at the output.
00:22:14.880 --> 00:22:16.079
There is excitement in that.
00:22:16.160 --> 00:22:16.480
That's awesome.
00:22:16.559 --> 00:22:18.240
That's the whole thing that's kind of unlocked this last week.
00:22:18.319 --> 00:22:21.920
But there's also a little jolt of, ooh, I need to reconsider this whole thing a lot.
00:22:22.079 --> 00:22:28.559
I came to the office this week and generally said, I don't know if I want to talk about some of the things that I'm doing because I did it so fast without any struggle.
00:22:28.720 --> 00:22:29.680
I am uncomfortable.
00:22:30.000 --> 00:22:33.119
100% and very hard on the fear thing.
00:22:33.279 --> 00:22:35.119
Like I think this is part of the complexity.
00:22:35.200 --> 00:22:38.640
And just to, I've got a question for you on just a sec, but just an observation.
00:22:38.799 --> 00:22:40.480
This is a layered space.
00:22:40.720 --> 00:23:06.720
Because I mean, inherent to what you just said is the second order consequences of third and everything else for all that, the the hype around all these things, the anthropic, well, we know we're early mid-September, you know, in terms of you mentioned at the beginning lab whistleblowers for lack of another term, or people talking about greater than 10% chances of X, Y, and Z concerning, but there are intermediate steps around what AGI means and what it means for us as economic participants and society and everything else.
00:23:06.799 --> 00:23:11.119
Like um, I I a hundred percent hear you on that discomfort as well.
00:23:11.359 --> 00:23:18.319
So, do you think though that comes from the individual thing or the aggregate of how many different things that you've been trying?
00:23:18.400 --> 00:23:20.079
The the the broader package?
00:23:20.400 --> 00:23:31.680
I think the range has been the thing for uh these releases that have made me a bit kind of aware and how little effort I feel I'm putting in to get good results, impactful results.
00:23:31.839 --> 00:23:34.400
One impressive result is easy for me to put in a box.
00:23:34.559 --> 00:23:40.079
It's also highly likely with enough know-how and the right harness, previous models could probably have done something similar.
00:23:40.240 --> 00:23:48.000
But I'm talking about simplicity, which gives a layman access to great power versus the bloke and his 15 Mac minis and all these custom prompts.
00:23:48.079 --> 00:23:49.519
Like that's not needed anymore.
00:23:49.759 --> 00:23:58.240
And when I moved from our website to a course package, to Blender, to scraping public data to build out a research base and more and more, then I'm looking at what other people have done in completely different software.
00:23:58.319 --> 00:24:01.759
I start wondering where else my expectations are completely out of date.
00:24:01.920 --> 00:24:07.440
And the next thought is what could somebody who really knows this field do with the same tool?
00:24:07.599 --> 00:24:14.160
Or on the flip side, how much should I put into this to ensure there is some productive struggle that feels like I've earned the outcome?
00:24:15.200 --> 00:24:17.039
There's a there's a flip side to that as well.
00:24:17.519 --> 00:24:28.160
Like the the the perception of false mastery, the illusion rather of sort of of false mastery, that is you can also produce something impressive looking in an area where you're poorly equipped to judge it.
00:24:28.480 --> 00:24:28.960
Correct.
00:24:29.119 --> 00:24:29.359
Absolutely.
00:24:29.440 --> 00:24:32.480
And I think it's a the big fallacy of all these AI tools more broadly.
00:24:32.559 --> 00:24:35.119
I can judge whether our website communicates what I want to communicate.
00:24:35.279 --> 00:24:43.039
I know I can spend more time on lighting to convey the story and my cool cyberpunk setting, but I don't need an AGI verdict to take this seriously.
00:24:43.200 --> 00:24:48.960
I've got the work in front of me I would previously have struggled with, both productively and non-productively, to get to the same results.
00:24:49.200 --> 00:24:52.000
I've I've I brought us down a little bit, so let's go to word models.
00:24:52.319 --> 00:24:54.640
You've spent this episode building places, right?
00:24:54.720 --> 00:24:55.920
I think that's highly apropos.
00:24:56.000 --> 00:25:00.640
So uh Cyberpunk Street, Molox Alexandria, the Apple Park from Photos.
00:25:00.880 --> 00:25:03.759
Astro gets you there with Blender with an existing tool.
00:25:03.920 --> 00:25:05.759
Uh I got a version for you without Blender.
00:25:06.079 --> 00:25:14.319
And uh as I kind of said earlier, I think I think it's the next point, and this is not just me, this is people like Faith A Lee, commonly known as the godmother of AI.
00:25:14.400 --> 00:25:18.480
Um, she's got a company called World Labs, they have a product called Marble.
00:25:18.640 --> 00:25:20.480
Um, Google has one called Genie.
00:25:20.880 --> 00:25:24.400
Basically, if people aren't aware of this, you type a description of a room.
00:25:24.480 --> 00:25:31.599
Um, so think a hospital ward, a courtroom, RPD, thank you, on stage now, airport hangar or factory floor.
00:25:31.759 --> 00:25:32.400
Yeah.
00:25:33.200 --> 00:25:35.359
And it makes stuff.
00:25:35.759 --> 00:25:37.359
It makes the whole world.
00:25:37.680 --> 00:25:38.079
Yeah.
00:25:38.240 --> 00:25:45.759
So these are these are 3D environments that stay put, not some Will Smith Spaghetti Index video that morphs when you turn your head, right?
00:25:45.920 --> 00:25:48.559
This is a room that's still there when you turn back.
00:25:48.720 --> 00:25:52.880
You walk through it as Dale's doing now, you can edit it, export it into a game engine.
00:25:52.960 --> 00:25:57.440
So here we're talking Unreal or Unity with solid walls and lighting.
00:25:57.680 --> 00:25:59.359
So what do it cost though?
00:25:59.599 --> 00:26:00.240
Yeah.
00:26:00.559 --> 00:26:01.279
Good question.
00:26:01.440 --> 00:26:06.559
I mean, a dollar, dollar twenty American here in this case, rough draft for cents.
00:26:06.720 --> 00:26:12.640
Um, so think again, uh, educational applications or nursing, been doing a lot of talk with people about that recently.
00:26:12.880 --> 00:26:18.160
Make a ward for a lot of coffee, like make three for a coffee, maybe.
00:26:18.400 --> 00:26:20.720
And I mean, I'm not I'm not reading that off a website, right?
00:26:20.799 --> 00:26:22.000
So we we're building these.
00:26:22.079 --> 00:26:27.920
So I was in a talk recently, I had a marble export dropped into Unreal on my machine.
00:26:28.079 --> 00:26:33.200
Collision fields, the mesh and all of that on, so you can't just walk through walls, it's a walkable space.
00:26:33.440 --> 00:26:38.000
And um, me walking around a room that didn't exist the day before.
00:26:38.240 --> 00:26:42.240
I think that the thing that this really brings for me though is what happens to simulation.
00:26:42.319 --> 00:26:49.359
And you've kind of alluded to this as well with the Blender thing, but you've still got that really high threshold of using Blender, or maybe you don't, you know, like I mean, even knowing it exists.
00:26:49.440 --> 00:26:56.319
But what happens to simulation as a pedagogy or experiential learning when simulation stops requiring a budget?
00:26:56.640 --> 00:27:01.920
Because Alexandria, your Alexandria example, these these this is the answer starting to arrive.
00:27:02.079 --> 00:27:05.599
This is a a place to explore, these are things to touch.
00:27:05.759 --> 00:27:12.160
A teacher, an educational experience with students uh being asked what's evidence and what's reconstruction.
00:27:12.400 --> 00:27:16.480
Now for a down payment and a coffee, you know, it's it's amazing.
00:27:17.200 --> 00:27:20.240
And I got an email this week that got me so excited.
00:27:20.400 --> 00:27:32.319
It's the the next version of the World Labs products Atlas, and it said, Yep, the wait list, they're starting to activate it, and we'll start to see kind of get access to the next model from Stern 14th, which is crazy.
00:27:32.640 --> 00:27:33.839
Yeah, yeah, 100%.
00:27:34.400 --> 00:27:46.240
Because what you've got there is simulation, demonstration, execution, if you like, in terms of like I mean, putting on my this is a great idea, I'd like to try this, uh, questions or approaches I used to get from CCs and PMs.
00:27:46.400 --> 00:27:55.440
I'm like, okay, uh, that's quite an undertaking because this used to need a lab, a budget, a team, building their environment, collecting the demos, training the thing.
00:27:55.680 --> 00:27:59.759
Now it's uh all right, just smash that up while I have a coffee.
00:28:00.000 --> 00:28:04.000
And um, that's kind of where the sock problem went.
00:28:04.160 --> 00:28:04.799
You know what I mean?
00:28:04.880 --> 00:28:09.519
It used to be how many demos, now it's the demo, the demo, right?
00:28:09.759 --> 00:28:10.000
Yes.
00:28:10.160 --> 00:28:17.039
If if one good video does the work of of hundreds, one bad video does the work of hundreds of bad ones, I guess, as well.
00:28:17.119 --> 00:28:20.960
So the whole game is the quality, who's holding the camera, whether they actually know how to hold the thing.
00:28:21.039 --> 00:28:32.559
I guess sort of linking back up again to what you were saying around actual mastery and taste as it's been described and other things, whether whether it is done the way that you'd want a hundred copies of, if you like.
00:28:32.880 --> 00:28:34.559
But that's that's a teacher.
00:28:34.799 --> 00:28:35.759
Yeah, a hundred percent.
00:28:35.920 --> 00:28:36.160
Yeah.
00:28:36.319 --> 00:28:38.000
Uh and and that's what I try, right?
00:28:38.160 --> 00:28:40.799
So every skills lab used to be too expensive to build.
00:28:40.960 --> 00:28:43.200
So nursing, moot courts, factory floor.
00:28:43.359 --> 00:28:45.759
I mean, a lab that you'd never let a first hair touch.
00:28:45.920 --> 00:28:48.720
So cheap to build and free to break.
00:28:48.880 --> 00:28:56.000
I have a big long list of people who are keen to get their hands dirty with this thing, everything from aviation to logistics to build worlds.
00:28:56.079 --> 00:28:57.599
Because how fun is that, mate?
00:28:57.839 --> 00:28:58.720
Inception.
00:28:59.119 --> 00:29:00.720
You you're you're an olden man.
00:29:00.799 --> 00:29:02.160
You know, am I am I am I wrong?
00:29:02.319 --> 00:29:02.480
Of course.
00:29:03.039 --> 00:29:03.200
Of course.
00:29:03.359 --> 00:29:03.839
Yeah, yeah, yeah.
00:29:04.319 --> 00:29:06.480
So you can you can play the inception.
00:29:06.559 --> 00:29:08.079
Do you want to flip a city dale?
00:29:08.240 --> 00:29:09.119
How cool would that be?
00:29:09.279 --> 00:29:15.839
Unreal physics, and these are experiences which are permanent, which are shared, which you can put in the hands around the eyes.
00:29:15.920 --> 00:29:19.200
I mean, in terms of this, or create immersive spaces.
00:29:19.440 --> 00:29:21.440
Are two spec on demand.
00:29:21.599 --> 00:29:25.680
And so what I see from that, and the robots saying I didn't really build this up enough.
00:29:25.920 --> 00:29:29.359
This is the next point after this, where the machines learn these kind of things.
00:29:29.599 --> 00:29:32.400
That is the fear, and I feel that's strong there, 100%.
00:29:32.960 --> 00:29:41.359
But that the the wonder of this, we could go to that that Venus scan that you said and walk, walk on the surface of another world.
00:29:41.519 --> 00:29:42.960
How cool is that?
00:29:43.200 --> 00:29:45.599
And it's all about show don't tell.
00:29:45.920 --> 00:29:46.400
100%.
00:29:46.960 --> 00:29:47.839
Show don't tell.
00:29:48.079 --> 00:29:53.440
Can you please uh not to not to get away from this because I love it, but you promised me a fly.
00:29:53.680 --> 00:29:54.799
I want to hear about this fly.
00:29:54.880 --> 00:29:55.599
I don't know enough.
00:29:56.000 --> 00:30:02.160
Seems seems harsh to to spin from building the entire world to talking about a fly.
00:30:02.319 --> 00:30:06.079
What a way, what a way to go out on our last season, our last episode.
00:30:06.319 --> 00:30:06.880
But this is cool.
00:30:07.200 --> 00:30:07.599
Never give me a few.
00:30:07.759 --> 00:30:08.480
This is a cool rabbit hole.
00:30:08.640 --> 00:30:09.839
We need to get really excited.
00:30:09.920 --> 00:30:16.720
So there's a clip of a fly brain model playing beat saber, which is crazy.
00:30:16.799 --> 00:30:20.799
So the creator explained that the version showed that they've been fitted to a gameplay route.
00:30:20.880 --> 00:30:22.640
So we should be careful what this actually demonstrates.
00:30:22.720 --> 00:30:27.200
But the example I really want to talk about is the fly escape room.
00:30:27.599 --> 00:30:34.160
You arrange objects inside a house, you release the flies, and you try to get them to find a way outside.
00:30:34.319 --> 00:30:42.480
And the fly's movement is driven by simulated neural activity using connections drawn from a real fruit fly nervous system map.
00:30:42.880 --> 00:30:45.519
The fly, the fly is an MPC.
00:30:45.759 --> 00:30:47.680
A non-player character, yes, that's right.
00:30:47.759 --> 00:30:57.440
And a computer controlled character whose behavior is influenced by a model built around this biological wiring, which is a fairly extraordinary sentence to say about just this kind of browser game.
00:30:57.519 --> 00:30:58.640
This is all built on Astra.
00:30:59.680 --> 00:31:00.880
Oh my god.
00:31:01.119 --> 00:31:04.480
Okay, so where does the wiring come from?
00:31:04.799 --> 00:31:16.640
I'm assuming only because you want to map a flyer this weekend, but researchers at HHMI, Genilia, Google Research, and their collaborators map the male fruit flies brain and central nervous brain and central nervous system.
00:31:16.799 --> 00:31:23.680
The full data set contains roughly 166,000 neurons, about 125 million synaptic connections.
00:31:24.000 --> 00:31:30.160
AI helped reconstruct the neural structures from microscope images with human experts checking and annotating the results.
00:31:30.240 --> 00:31:32.000
That map is called a connectome.
00:31:32.319 --> 00:31:34.480
I'm getting nerdy here, but it's cool.
00:31:34.720 --> 00:31:37.680
Essentially, it records which neurons connect to which.
00:31:38.640 --> 00:31:42.720
Oh my god, this is the first step to upload me to the cloud.
00:31:43.599 --> 00:31:48.160
And can I say, Cyber, it's kind of fun that someone just like, yeah, you know what?
00:31:48.240 --> 00:31:52.240
The first best use case for this is let's let's make an escape room.
00:31:52.559 --> 00:31:54.079
I love when the internet's fun.
00:31:54.160 --> 00:31:56.160
And the internet has been very fun recently.
00:31:56.319 --> 00:31:58.160
It's a very, very, very internet response.
00:31:58.400 --> 00:31:58.880
Absolutely.
00:31:59.279 --> 00:32:00.640
A very scientific resource.
00:32:00.799 --> 00:32:02.400
We live in strange, wonderful times.
00:32:02.559 --> 00:32:03.599
But there's an important detail here.
00:32:03.680 --> 00:32:10.799
The game's documentation says it uses a selected part of the connectome, its neuron equations, the sensory inputs, and the movement rules, they're models.
00:32:10.960 --> 00:32:15.119
And the author explicitly cautions that it hasn't been validated as a digital replica of a living fly.
00:32:15.519 --> 00:32:19.200
Look a bit boring, because for me, this is just replicating hundreds of flies.
00:32:19.680 --> 00:32:22.480
But even at this level, I find it so fascinating.
00:32:22.960 --> 00:32:31.200
You can actually follow a digital fly, inspect its simulated neural activity while it explores, because we've just spent the episode talking about cool tools and how we can build worlds.
00:32:31.440 --> 00:32:38.480
Now we're looking at an experiment that takes a biological connection data and uses it to influence the behavior of something inside a virtual environment.
00:32:38.640 --> 00:32:41.839
Nick, I'd seen this fly park cars recently.
00:32:42.160 --> 00:32:43.759
How are we not in a simulation?
00:32:43.920 --> 00:32:49.839
No, this is the bit this is this is genuinely the that first step and the practical, how you make that happen.
00:32:50.000 --> 00:32:50.799
There we go, here we go.
00:32:50.880 --> 00:32:53.920
Here's the fly's the fly driving a car.
00:32:57.519 --> 00:32:58.160
Oh my goodness.
00:32:58.400 --> 00:33:03.200
Simple sacrocy AI sop, an insane, insane science.
00:33:05.759 --> 00:33:06.559
It's sound good.
00:33:06.960 --> 00:33:07.759
It's sound good.
00:33:09.359 --> 00:33:12.079
What is a fly expressing that it wants to honk a horn?
00:33:12.480 --> 00:33:15.359
What lived experience do they make sound?
00:33:17.599 --> 00:33:18.720
What am I meaning?
00:33:18.799 --> 00:33:19.839
Why am I doing this?
00:33:20.079 --> 00:33:20.720
Sorry, God.
00:33:22.559 --> 00:33:23.680
Fly Pacito.
00:33:23.839 --> 00:33:24.240
Oh my god.
00:33:24.400 --> 00:33:27.599
But imagine the possibilities for virtual animals or small ecosystems.
00:33:27.680 --> 00:33:33.440
How far you could get by giving the inhabitants different biologically informed models and seeing how they interact.
00:33:33.680 --> 00:33:38.400
I'm projecting it beyond this particular game, but that's the question it kind of puts in our head, Nick.
00:33:38.799 --> 00:33:42.640
Yeah, I I guess how much yeah, but this is what I was asking about the horn, right?
00:33:42.720 --> 00:33:47.440
So how much comes from the fly itself and how much from the rules the developer put around it.
00:33:48.000 --> 00:33:48.400
100%.
00:33:48.799 --> 00:33:54.480
And the way I'm thinking about it, it's kind of like treating the fly like a graphics card and seeing how it react and et cetera, et cetera.
00:33:54.640 --> 00:33:56.720
But there's an educational experience here as well.
00:33:56.880 --> 00:34:03.440
Here's a model, here's what went into it, here's what it does, and how would you investigate whether its behavior tells you anything useful about the system it represents?
00:34:04.240 --> 00:34:06.160
And how do you get the flies to leave the house?
00:34:08.320 --> 00:34:09.760
That's the the modeling.
00:34:09.840 --> 00:34:14.559
So that's the billions of dollars of compute that Chen Senwang's building his his fortune on.
00:34:14.880 --> 00:34:16.719
I think that's a good place to finish the season.
00:34:20.159 --> 00:34:25.199
Can we just take a second to acknowledge that that is the beginning of a simulated reality?
00:34:25.360 --> 00:34:28.079
And all of that, we live in the Sims.
00:34:28.400 --> 00:34:30.880
That that is the that is the end point, you know what I mean?
00:34:31.119 --> 00:34:33.199
And it's it's simulations the whole way down.
00:34:34.159 --> 00:34:35.360
It starts with flies.
00:34:39.840 --> 00:34:41.679
I am crying inside.
00:34:41.840 --> 00:34:43.199
But wondrously as well.
00:34:43.360 --> 00:34:48.800
Like, I mean, I could fly in that world, upload me to the fly to the cloud where maybe I'm gonna fly.
00:34:48.960 --> 00:34:51.840
I'll tell you why I'm beeping, because this guy won't move.
00:34:52.159 --> 00:34:58.639
And I think that's a great place to finish the season because underneath all the model comparisons, I've got a fairly simple reaction to what we've been showing.
00:34:58.800 --> 00:35:00.880
These are things I'm willing to try right now.
00:35:01.039 --> 00:35:02.400
Maybe not the fly.
00:35:02.559 --> 00:35:05.360
And I suspect there's still plenty of things that I'm treating as too difficult.
00:35:05.440 --> 00:35:06.719
Maybe the computer can't do that.
00:35:06.800 --> 00:35:09.519
That I need to go back and test those assumptions.
00:35:10.800 --> 00:35:12.239
It had the inception thing.
00:35:12.639 --> 00:35:14.480
Oh, so Nick, what are you gonna do over the break?
00:35:14.559 --> 00:35:16.079
Are you gonna but what's what's the plan?
00:35:16.159 --> 00:35:17.280
You've got you got a bit of time off?
00:35:18.000 --> 00:35:19.280
No time off for me, sucker.
00:35:19.519 --> 00:35:21.039
No, I'm obviously still at work, mate.
00:35:21.119 --> 00:35:23.519
Like uh unlike some of jaunting around.
00:35:23.599 --> 00:35:30.159
But um, my big plan for for for the spare time, the the though I'll miss you, of course, inception style world building.
00:35:30.400 --> 00:35:34.880
I mean, uh like I said, aviation, logistics, film, games, business, so much cool stuff.
00:35:35.119 --> 00:35:47.280
Plan is to export it, get someone who's actually an expert on these things to get in there, walk around it, shoot holes in it, tell me what's wrong, and then we're gonna try and get some students in there, if we can, maybe next semester.
00:35:47.440 --> 00:35:52.239
Because I reckon if it survives the aviation mechanic, right, it goes in the course.
00:35:52.400 --> 00:35:53.840
Uh, what about yourself?
00:35:54.159 --> 00:35:57.039
I am trying to keep the fear under control.
00:35:57.199 --> 00:35:58.079
That's the main thing.
00:35:58.320 --> 00:36:00.000
So I guess just go play with these tools.
00:36:00.079 --> 00:36:01.760
Go see if you can build out a fly.
00:36:01.920 --> 00:36:04.480
But for everyone else, thank you so much for listening to this season.
00:36:04.559 --> 00:36:05.519
We'll be back soon.
00:36:05.679 --> 00:36:06.079
I promise.
00:36:06.159 --> 00:36:07.599
It's only a short hiatus.
00:36:07.840 --> 00:36:09.519
Thank you for status intelligence.
00:36:09.760 --> 00:36:12.880
You can find us on YouTube, you can find us on all your favorite podcast feeds.
00:36:13.119 --> 00:36:20.320
Until then, stay curious, stay intelligent, and stay a fly.
00:36:20.559 --> 00:36:22.159
Um I'm not sure.
00:00:00.160 --> 00:00:06.000
I asked ChatGPT's Astra to make me a cool cybunk scene in 3D software Blender.
00:00:06.559 --> 00:00:08.080
This is what it built.
00:00:08.720 --> 00:00:11.439
Then, Nick, I rebuilt our website with it.
00:00:11.599 --> 00:00:19.359
Then I took an already AI built course after a Canvas cartridge and had the file inside Canvas, able to learn from it in about 10 minutes.
00:00:19.600 --> 00:00:20.800
The rubrics are set up.
00:00:21.039 --> 00:00:22.719
The quizzes were built perfectly.
00:00:22.879 --> 00:00:29.280
There was an authentic assessment that ran through the whole course and had me building up a portfolio to show at a job interview.
00:00:29.440 --> 00:00:35.759
And somewhere in there, while using OpenAI's latest model, I got a feeling that I haven't had with AI for a while.
00:00:36.079 --> 00:00:37.759
Holy smokes, Batman.
00:00:37.920 --> 00:00:39.439
What else can this thing do?
00:00:39.759 --> 00:00:40.719
Yeah, wow.
00:00:40.880 --> 00:00:43.119
Okay, I had a look at that blender demo.
00:00:43.280 --> 00:00:44.960
We we also need to talk about Fable though.
00:00:45.039 --> 00:00:47.920
I feel you're straying, and Astra's getting a lot of your attention.
00:00:48.240 --> 00:00:55.280
Well, Nvidia's Jensen Huang shouting out that Astra is probably AGI hasn't helped all the the the fanfare.
00:00:55.359 --> 00:01:01.039
But for our last episode of the season, we've each bought things that made us stop and say, have you seen this cool thing?
00:01:01.280 --> 00:01:13.599
Things that we built, things that other people are making, the fact that the state of the art is now trumped in about seven days, and later a game whose characters run on a simulation built from a fruit flies neural wiring.
00:01:13.760 --> 00:01:14.400
Yes, really.
00:01:15.760 --> 00:01:17.439
You need to explain that.
00:01:18.000 --> 00:01:18.799
Maybe we will.
00:01:18.879 --> 00:01:19.519
At some point.
00:01:19.840 --> 00:01:21.280
Let's get to the title.
00:01:30.959 --> 00:01:33.599
Nick, it is our last episode of the season.
00:01:33.680 --> 00:01:34.640
So I hope you're well.
00:01:34.799 --> 00:01:35.840
Are you doing good?
00:01:36.239 --> 00:01:37.200
Doing great, man.
00:01:37.280 --> 00:01:44.560
Like uh really, really yeah, bitter, bit of sweet, but I think a really good note to go off on, especially as sunlight is off on holiday for a while.
00:01:44.799 --> 00:01:46.400
A little bit of a jort.
00:01:46.640 --> 00:01:47.200
A bit of a jort.
00:01:47.359 --> 00:01:50.799
Well, we're back before the end of the year, but we thought we'd finish the season with a bit of a show and tell.
00:01:51.040 --> 00:01:57.840
We spend plenty of time picking about these technologies and what they mean, what could go wrong, what universities should do about them.
00:01:58.000 --> 00:02:04.799
So today we're giving ourselves a bit of room to enjoy the fact that some of this stuff is absolutely extraordinary.
00:02:05.120 --> 00:02:09.120
My last few weeks with Astra has given me, wow, this technology is so powerful.
00:02:09.280 --> 00:02:23.439
That feeling again, and sure, the pendulum swings to existential dread because of all the things it can do and what that means for future economic possibilities, or when safety researchers leave certain frontier labs and said that frontier labs were in an arms race to bring the end of humankind as we know it.
00:02:23.680 --> 00:02:26.479
But hey, let's try keep it light.
00:02:26.800 --> 00:02:28.879
Doomsday Dale is not here today.
00:02:29.199 --> 00:02:36.879
So I'll start with Astra, because we've been out for a few weeks now, and I'm still getting that urge to just ask it increasingly unreasonable questions.
00:02:37.439 --> 00:02:40.560
Was that a so such as a very detailed blender brief?
00:02:40.800 --> 00:02:46.240
I want to be clear, my very detailed render blender brief, and you can watch on YouTube for the actual clip that it made.
00:02:46.400 --> 00:02:48.560
It was make me a cool cyberpunk thing.
00:02:48.800 --> 00:02:55.680
There was clear artistic vision here, but I watched as it scanned my computer to see if a 3D software blender was installed.
00:02:55.840 --> 00:03:00.639
It set up the render settings, the camera, it rigged up a character that could move about the world.
00:03:00.800 --> 00:03:05.360
All the things that are just not that fun when you're building out cool 3D scenes.
00:03:05.599 --> 00:03:06.879
I was shooketh.
00:03:07.120 --> 00:03:15.360
Also, I used this wireframe as like a skeleton and took it into AI video generation platform C Dance and applied textures to it, again, completely guided by Astra.
00:03:15.680 --> 00:03:20.639
It was pure flow state, and God, it tickled that part of my brain for creating something cool.
00:03:20.719 --> 00:03:22.479
It just happened so fast.
00:03:23.120 --> 00:03:26.400
I don't think I have the deep, deep pockets and resources that you do.
00:03:26.479 --> 00:03:29.599
I've I have played a bit, I've dabbled, I've dabbled is the best I can do.
00:03:29.680 --> 00:03:30.800
And I'm absolutely blown away.
00:03:30.879 --> 00:03:33.919
Some of the demos that I've seen online are extraordinary.
00:03:34.080 --> 00:03:36.479
Your cyberpunk video, that is sick.
00:03:36.639 --> 00:03:41.919
I would pay money and line up and sit in dark rooms surrounded by strangers, iMacs or otherwise to watch that.
00:03:42.080 --> 00:03:43.199
That is amazing.
00:03:43.439 --> 00:03:47.520
It does look cool, and again, kind of want to build out the world a little bit more and again kind of flesh this out.
00:03:47.599 --> 00:03:49.039
And this is what these tools kind of offer.
00:03:49.120 --> 00:03:52.400
But Nick, the spread of what Astra can do is what I loved.
00:03:52.479 --> 00:03:56.159
So our podcast website has for a long time been sorely lacking.
00:03:56.319 --> 00:04:00.080
Uh your HTML is broken, was one of the pieces of feedback that I received.
00:04:00.159 --> 00:04:03.840
And I was far too ashamed to admit that it was just my personal credit taste there.
00:04:04.000 --> 00:04:09.520
So I know this isn't purely Astra example, but what I did note is that I didn't have to go back and forth.
00:04:09.759 --> 00:04:12.560
I just had the thing after a single prompt.
00:04:12.800 --> 00:04:18.079
And as for building out learning content, it's been something that I've tried on previous models, and the results were okay.
00:04:18.240 --> 00:04:25.360
But with Astra, we built out learning materials, the lesson flow, the assessments, even lesson plans and how this might translate to a face-to-face environment.
00:04:25.519 --> 00:04:28.560
And then after it actually drained up its own LMS environment.
00:04:28.720 --> 00:04:31.759
So my push was like, how do you take this into the canvas environment?
00:04:32.000 --> 00:04:34.000
It made its own LMS.
00:04:34.399 --> 00:04:35.040
It did.
00:04:35.600 --> 00:04:38.240
And I was like, I don't want that, I already have one of those.
00:04:38.560 --> 00:04:39.759
Oh my god.
00:04:39.920 --> 00:04:42.160
All right, so that's insane.
00:04:42.639 --> 00:04:49.439
The the the lag of so that was where the time was spent to make making from scratch a learning management system.
00:04:49.680 --> 00:04:50.079
Correct.
00:04:50.160 --> 00:04:54.240
Oh my god and I was like, hey, thanks so much for doing this and taking my content and putting it to it.
00:04:54.319 --> 00:04:54.720
But I have this.
00:04:54.800 --> 00:04:56.480
Can you just put into a Canvas cartridge?
00:04:56.720 --> 00:04:57.439
And it was great.
00:04:57.519 --> 00:05:01.439
The quick bit was kind of packaging it all up, that existing material into Canvas.
00:05:01.600 --> 00:05:03.759
That's something that would take me like a week to do.
00:05:04.079 --> 00:05:06.399
It's almost just boring, busy work.
00:05:06.639 --> 00:05:09.279
And the level of quality I received was great.
00:05:09.439 --> 00:05:22.879
So once I had imported it, I expected things like broken headings, rubrics that need to be wired up again, broken links, all the things that anyone who's built any of these tools and used AI for the last couple of years kind of expects, like to get halfway there, but you're still gonna need to be a work.
00:05:23.120 --> 00:05:25.360
But no, this one was perfect.
00:05:25.519 --> 00:05:38.959
I mean, sure, the proof is in the classroom about whether this is good, is high quality, but we're now in a world where you can pump in your content and get a very decent learning experience, including images, diagrams, in a really quick amount of time.
00:05:39.120 --> 00:05:40.160
This is crazy.
00:05:40.480 --> 00:05:45.360
I know a file export is a really difficult thing to make sexy on YouTube, but extrapolate, yeah.
00:05:45.600 --> 00:05:45.839
Right.
00:05:46.000 --> 00:05:53.839
Yeah, so which is why we opened up on the amazing city streets of your your cyberpunk uh to sing my incredibly strong editorial judgment.
00:05:54.000 --> 00:05:57.279
But across these three examples, I kept having that same reaction.
00:05:57.439 --> 00:05:59.439
Hang on, I can just ask to do this as well.
00:05:59.600 --> 00:06:07.519
And that's different from how I felt with a lot of these incremental improvements we've seen since I guess even GPT-5 and I guess was it Opus 4?
00:06:07.680 --> 00:06:13.759
Sure, maybe it could do some of these cool things, but after a whole bunch of tinkering and heartache, this kind of removed that struggle.
00:06:13.920 --> 00:06:16.800
This had me looking around at other things that I wanted to make.
00:06:16.959 --> 00:06:22.879
That app, that game, it is just the coolest little toolbox, and what's more, it's actually quite an economical model.
00:06:23.040 --> 00:06:30.720
I spent the greater part of this year using Claude, but I ride my usage meter the whole time, terrified that Fayette was going to eat everything on one of its endless floor paths.
00:06:31.040 --> 00:06:40.319
Astra's new model, I don't actually have that same fear, and the fact that the lowest end thinking of Astra, so like Astra Low, I think it is, majorly outforms the last sweeter models.
00:06:40.480 --> 00:06:42.160
It's pretty good bang for buck.
00:06:42.480 --> 00:06:43.519
Full call out though.
00:06:43.600 --> 00:06:46.000
I am waiting for my limits to reset tomorrow.
00:06:46.160 --> 00:06:50.800
So before I completely disappear into my own experiments, Nick, what have you got?
00:06:50.959 --> 00:06:51.680
What have you bought?
00:06:51.839 --> 00:06:55.279
Start with one of something that you really want to show me.
00:06:55.839 --> 00:06:58.560
You called Dibbs on Astro.
00:06:58.800 --> 00:07:05.839
And so what I got is I got to use that as an excuse to go down a little rabbit hole of mine, which is around robotics and world models.
00:07:05.920 --> 00:07:08.560
So I want to give you a before first so you feel the after, right?
00:07:08.639 --> 00:07:15.439
So July this year, one of the best robotics labs in the world taught a robot to turn a sock inside out.
00:07:15.759 --> 00:07:21.519
Nick, may I remind you that this episode is to keep our audience very engaged?
00:07:21.839 --> 00:07:23.360
You have bought me a sock.
00:07:25.040 --> 00:07:27.040
Master has bought me a sock.
00:07:27.279 --> 00:07:29.439
No, no, but here's the interesting thing, right?
00:07:29.519 --> 00:07:35.040
So to do this, a human set of the controls and puppeted the robot through the task 176 times, right?
00:07:35.120 --> 00:07:38.879
So eight hours, one sock, same lap, same month peeling an orange.
00:07:38.959 --> 00:07:40.319
They went through so many Mandarins.
00:07:40.480 --> 00:07:48.480
The researchers joke that the local grocer was looking at a takeover when it worked 52% of the time, way slower than you, right?
00:07:48.879 --> 00:07:51.439
But that was July, and now it's August.
00:07:51.600 --> 00:08:02.160
Someone stands in a kitchen with a phone at eye level, films themselves potting a plant 10 minutes, hand the robot the video, and the robot pots the robot pots the plant.
00:08:03.199 --> 00:08:03.759
Hallelujah.
00:08:06.160 --> 00:08:10.000
One oh no exactly.
00:08:10.319 --> 00:08:10.959
Exactly.
00:08:11.199 --> 00:08:14.800
So this is no no no no, but thinks you this, because there are other fun things with it as well.
00:08:14.879 --> 00:08:21.360
So this is a company called Skilled, August, one video of a person, not a robot, but a human film from their own point of view.
00:08:21.600 --> 00:08:22.639
Robot does task.
00:08:22.800 --> 00:08:31.120
Things that are never seen, like potting a plant, pancakes, pour over coffee, assembling that annoying IKEA furniture up to 10 minutes, dozens of steps.
00:08:31.439 --> 00:08:32.720
It's just trained on the video.
00:08:32.960 --> 00:08:34.960
No, and that's the crazy thing, right?
00:08:35.039 --> 00:08:37.759
So nothing is trained, the weights don't change.
00:08:37.919 --> 00:08:39.279
It's kind of like a recipe.
00:08:39.440 --> 00:08:42.960
It reads the video the way that you or I would read a recipe and it does it.
00:08:43.120 --> 00:08:46.960
You know, you know how historically you would paste an example historically.
00:08:47.279 --> 00:08:51.120
Two months ago, you would paste an example into a chatbot and say, write me one like this.
00:08:51.360 --> 00:08:52.080
Same trick.
00:08:52.240 --> 00:08:56.639
The example is a video, and the thing copies, the thing it copies is a set of movements.
00:08:56.799 --> 00:08:58.559
The video is the instruction.
00:08:58.720 --> 00:09:05.120
And now this sounds like a really odd way to kick off a discussion about this, but this lends itself to something called world models.
00:09:05.279 --> 00:09:11.840
And I want to nerd out feverishly with those with you in just a minute because it is, I think it's the one after Astra.
00:09:12.159 --> 00:09:20.240
I think it's the next stage when we get past like just text and video and voice and images, like those amazing ones that you just showed us.
00:09:20.320 --> 00:09:23.679
I'm sort of I'm looking forward to 27, very futurist of me.
00:09:23.840 --> 00:09:32.639
But this is this stuff is really, is really crazy that we've now got machines that are now learning from an input as complex as a video, kind of like we do.
00:09:32.960 --> 00:09:33.519
Ooh.
00:09:33.919 --> 00:09:39.279
Nick, you you've you've got me interested in socks and the world models.
00:09:39.360 --> 00:09:43.679
But before we do that, I want to talk about what other people have been doing with all these cool models.
00:09:44.480 --> 00:09:46.399
First, look at what people are making.
00:09:46.559 --> 00:09:56.559
So Ethan Moloch, often read uh often mentioned on the show, he's been using Astra to make an explorable reconstruction of the library of Alexandria in a surrounding setting.
00:09:56.720 --> 00:10:02.960
So you can walk around, you can listen to a narrator tour, you can open texts within the library to read.
00:10:03.120 --> 00:10:06.320
He's published this project as well, so you can actually go out there and play with it yourself.
00:10:06.639 --> 00:10:12.240
I don't want to throw any shade, but you rebuilt our website and he rebuilt Alexandria.
00:10:12.559 --> 00:10:17.120
I would argue that we've both made important contributions to the internet in the last week or so.
00:10:17.360 --> 00:10:25.600
But look at the combination here, because you've got a place to explore, you've got things to interact with, you've got material to read, and the project does acknowledge the historic historical uncertainty.
00:10:25.840 --> 00:10:36.799
Its documentation separates evidence from reconstruction and make clear that layout is actually an interpretation, which gives it an interesting way to teach and an interesting teaching question.
00:10:36.960 --> 00:10:40.000
Imagine asking students to explore something like this and then challenge it.
00:10:40.080 --> 00:10:40.799
What can we support?
00:10:40.879 --> 00:10:42.000
What has somebody inferred?
00:10:42.080 --> 00:10:44.639
What has been added to make the experience work?
00:10:45.919 --> 00:10:47.200
And then there's interesting things.
00:10:47.360 --> 00:10:50.799
Do you have them use the environment or or make their own?
00:10:51.279 --> 00:10:51.919
100%.
00:10:53.039 --> 00:10:55.600
Both sound interesting to me, but think about what this opens up.
00:10:55.679 --> 00:11:03.120
Let's stop talking about cheating, but let's start talking about, hey, you can get people to build entire worlds, especially making a small part of something, and I have to justify my choices.
00:11:03.279 --> 00:11:07.519
You'd have to explain why you put a broom there, why that object belongs, what evidence you've used.
00:11:07.759 --> 00:11:10.559
But even before we get to teaching, I love the ambition of the request.
00:11:10.799 --> 00:11:13.440
Could you build me somewhere that I can run around?
00:11:13.600 --> 00:11:15.440
And you just can with Astra.
00:11:15.600 --> 00:11:21.039
And there's another one that I want to talk about, because it's probably closer to the thing that freaked me out about Astra a little bit.
00:11:21.279 --> 00:11:31.360
Max Weinbeck gave it images of Apple Park, which is famously where the iPhone and everything else is kind of built, and basically said to Astra, rebuild this in Blender.
00:11:31.759 --> 00:11:32.720
And it did.
00:11:34.000 --> 00:11:36.399
Wow, that's ridiculous.
00:11:37.440 --> 00:11:41.360
And it's a beautiful, beautiful building, and it builds every single environment.
00:11:41.440 --> 00:11:43.200
There's a huge amount of um images.
00:11:43.279 --> 00:11:47.440
I would highly recommend you kind of look at what it how it kind of approached it.
00:11:47.600 --> 00:11:50.000
Use something like 6.27 billion tokens.
00:11:50.159 --> 00:12:12.799
It responded 30,000 times in that in that model, spun up 489 different subagents, crazy, and had to look at a bunch of two-dimensional images, understand what they're dealing with with a real three-dimensional place, work out the geometry, the relationships between the buildings, the material, the landscape, the lighting, then actually operate Blender, which is a notoriously hard piece of software, to produce something.
00:12:13.039 --> 00:12:14.799
Max described the result as absurd.
00:12:15.039 --> 00:12:19.279
I think that's a reasonably fair assessment, and this connects directly to what I was saying about my cyberpunk scene.
00:12:19.360 --> 00:12:23.360
I gave it an incredibly sophisticated prompt, make me a cool cyberpunk scene.
00:12:23.759 --> 00:12:24.879
Masterclass.
00:12:25.200 --> 00:12:25.759
Correct.
00:12:26.159 --> 00:12:31.519
But Max is effectively saying, here is a real place, work backwards from the images and reconstruct it.
00:12:31.679 --> 00:12:35.519
And that gets me because we've had models that can look at the image for a long time.
00:12:35.679 --> 00:12:43.919
They can tell you what's in Apple Park, they can describe the architecture, they can probably give you an incredibly detailed explanation of how you might model it, but this is different experientially.
00:12:44.080 --> 00:12:46.960
You give it images and then there's Apple Park models sitting in Blender.
00:12:47.440 --> 00:12:51.360
It was doing things like scanning public records for the type of tree that it should model.
00:12:51.519 --> 00:12:55.519
He ran different experiments to ensure that the glass rendering matched reality.
00:12:55.679 --> 00:13:01.360
That's a judgment of a professional and the problem solver of someone who doesn't want to go to San Francisco but knows that the information is somewhere.
00:13:01.679 --> 00:13:03.600
That's astounding.
00:13:03.759 --> 00:13:06.720
I mean, the the token usage just for one.
00:13:07.279 --> 00:13:11.279
And the the refraction just as a single problem.
00:13:11.519 --> 00:13:12.879
And problem's not the right word.
00:13:12.960 --> 00:13:14.720
Like, I mean, those are so complex.
00:13:14.799 --> 00:13:26.480
And I'm just to channel the cynic of of anyone who might be this thing, uh obviously there would be the that's astounding, but there would there would maybe be the caveat and architect, 3D artists would find all sorts of things wrong with it, yeah.
00:13:26.720 --> 00:13:27.120
Probably.
00:13:27.200 --> 00:13:32.000
And I'm not claiming it has reverse engineered Apple's architectural drawings from some photographs.
00:13:32.399 --> 00:13:36.720
The dimensions, the dimensions are probably wrong, details can be invented.
00:13:36.799 --> 00:13:42.720
Someone who actually does this professionally may judge this as is going to judge the model far better than I can.
00:13:42.960 --> 00:13:46.080
Also, it kind of gave up on things when there was an information vacuum.
00:13:46.159 --> 00:13:49.679
But still, it did make the thing, and that is the Astra feeling for me.
00:13:49.759 --> 00:13:56.720
And this is where a point where another five like little points on a benchmark, we keep seeing all these benchmark graphics, they actually don't matter anymore.
00:13:57.039 --> 00:13:57.440
Yeah.
00:13:57.679 --> 00:14:02.320
I mean, presumably then the the next question is what happens, those those information gaps things?
00:14:02.480 --> 00:14:06.399
What happens when you give it somewhere less documented than Apple Park?
00:14:06.639 --> 00:14:06.960
Correct.
00:14:07.200 --> 00:14:12.399
Or your own building, your laboratory, a historical location where you have enough reference material, a classroom, a house, etc.
00:14:12.559 --> 00:14:12.799
etc.
00:14:13.279 --> 00:14:16.720
What else could I give this thing and see whether it can do it?
00:14:16.879 --> 00:14:20.159
Perhaps it would build a track within Ableton.
00:14:20.320 --> 00:14:24.000
And this one gives our audio listeners something as well, assuming it plays.
00:14:24.480 --> 00:14:42.000
Uh Pietro Shennaro, who after building Atlantis and a Beyblade PS1 game, which looks addictive as hell, and a video to code demonstration, connected Astra to music platform Ableton through a software integration, and he says it created the synth instruments, the musical parts, and the arrangement for this track.
00:14:42.240 --> 00:14:45.840
I'm not a music guy, someone who knows music.
00:14:45.919 --> 00:14:48.799
I love the music, but I can listen to this and go, this is cool.
00:14:48.879 --> 00:14:49.840
This is a cool track.
00:14:50.240 --> 00:14:52.559
It's got that kind of lo-fi vibe, etc.
00:14:52.720 --> 00:14:52.960
etc.
00:14:53.200 --> 00:14:55.279
But again, really, really cool.
00:14:55.519 --> 00:15:00.960
You could say things like, keep that bass line, make the opening stranger, give me something completely excessive when the drums come in.
00:15:01.200 --> 00:15:06.320
And that's the part that I want to play with and can play with now with things like Astra.
00:15:07.039 --> 00:15:12.080
Yeah, okay, so I'm guessing we're coming back next season with a 12-minute uh theme tune.
00:15:12.159 --> 00:15:12.480
Yes.
00:15:12.720 --> 00:15:13.679
And it will be art.
00:15:13.840 --> 00:15:18.720
It'll be a concoction between me and Astra with a beautiful cyberpunk rendering.
00:15:19.120 --> 00:15:21.679
This is what I mean about the range though, historical environments.
00:15:21.759 --> 00:15:27.519
An editable, editable locomotive, a music production project, walking around Apple Park.
00:15:27.759 --> 00:15:31.279
Each example has its own tools and setup, but there's work around the model.
00:15:31.360 --> 00:15:35.120
But the variety of jobs people are attempting with it is just what keeps capturing me.
00:15:35.519 --> 00:15:41.759
Alright, Nick, I'm assuming you wanna you wanna talk about about Fable about Bad Anthropic for a bit.
00:15:42.240 --> 00:15:46.559
I mean, before you get your Astra T-shirt out, like I mean, I think I think we should, yeah?
00:15:46.879 --> 00:15:51.840
Well, I think Fable 5, particularly 5.1, deserves a proper place in this conversation.
00:15:52.000 --> 00:16:01.200
At Astra's launch, artificial analysis had Fable 5.1 ahead in both its intelligence index and its coding agent index in the configuration they tested.
00:16:01.360 --> 00:16:06.879
So on balance, Fable 5.1 is apparently better than Astra.
00:16:07.919 --> 00:16:13.360
Which I mean, yeah, which on paper is is is amazing, right?
00:16:13.519 --> 00:16:16.000
But I I was just listening to the Ableton thing there.
00:16:16.080 --> 00:16:17.279
It just came through on my end.
00:16:17.360 --> 00:16:25.200
I mean, these things like the the cyberpunk thing, the Apple Park, the the the this this response then, why am I blown away?
00:16:25.279 --> 00:16:27.279
This makes this response quite interesting, yeah.
00:16:27.600 --> 00:16:30.960
Well, Fable has repeatedly given me that feeling of wow, that is really smart.
00:16:31.120 --> 00:16:33.679
Also, thankfully, the writing style is not horrible.
00:16:33.919 --> 00:16:37.519
Fable 5, like OG one was horrendous.
00:16:37.679 --> 00:16:38.879
I've said it once and I'll say it again.
00:16:38.960 --> 00:16:41.440
I'll not hang out with Fable 5.0 on weekends.
00:16:41.519 --> 00:16:43.679
It was the worst, they've kind of fixed it in 5.1.
00:16:43.840 --> 00:16:47.919
But Astra has given me a very physical reaction to the breadth of what I can get done.
00:16:48.000 --> 00:16:49.440
And I'm describing my own experience here.
00:16:49.519 --> 00:16:51.600
I haven't conducted comprehensive comparisons.
00:16:51.759 --> 00:16:55.360
I'm assuming Fable is capable of substantial practical work too.
00:16:55.600 --> 00:16:59.360
Actually, the example I wanted to bring kind of brings that up quite well.
00:16:59.840 --> 00:17:14.319
Anthropic reports that Fable 5.1 used historical radar images from NASA's Magellan mission alongside an existing elevation map to train a neural network and produce a higher resolution elevation map covering roughly a third of Venus.
00:17:14.960 --> 00:17:16.000
No shade.
00:17:16.240 --> 00:17:19.119
Again, a fairly strong answer to your website.
00:17:19.359 --> 00:17:20.799
That's very brutal.
00:17:20.880 --> 00:17:23.839
But we've established a bit of a hierarchy, I'd imagine.
00:17:23.920 --> 00:17:27.039
We've got my website, we've got Alexandra, we've got another planet.
00:17:27.279 --> 00:17:34.000
We can let other anthropic says that new map reveals details around two to three kilometers.
00:17:34.160 --> 00:17:34.880
Just think about that.
00:17:34.960 --> 00:17:35.119
Wow.
00:17:35.359 --> 00:17:36.799
Think about how far the innerness is away.
00:17:36.880 --> 00:17:44.319
I don't have that number on me, but we can see, we can understand to two to three kilometres compared with 10 to 20 kilometers previously.
00:17:44.400 --> 00:17:45.200
That's their published result.
00:17:45.440 --> 00:17:47.519
So we should keep it all attributed to Anthropic.
00:17:47.680 --> 00:17:50.559
But as an example of something to attempt with these tools, it really is extraordinary.
00:17:50.640 --> 00:17:56.880
It makes you wonder what's sitting in existing data sets that somebody hasn't had the time yet or the resource or the right method to actually investigate.
00:17:57.200 --> 00:18:03.839
I think that's a great qualifier because it adds nuance, it complicates the idea that this one is the clever one and the other one does things, yeah.
00:18:04.319 --> 00:18:04.880
Correct.
00:18:05.359 --> 00:18:06.720
They can both do things.
00:18:06.880 --> 00:18:09.680
The difference I'm trying to describe is how they've landed for me.
00:18:09.759 --> 00:18:11.279
And it's just kind of the tool you want to play with.
00:18:11.359 --> 00:18:15.920
With Astra, I kept moving between very different kinds of work and finding something useful on the other side.
00:18:16.079 --> 00:18:19.920
And I suspect that's why it has occupied so much of my attention at this stage.
00:18:21.680 --> 00:18:22.559
That's sick.
00:18:22.720 --> 00:18:23.359
That's sick.
00:18:23.519 --> 00:18:24.720
I hadn't seen the Venus one.
00:18:24.880 --> 00:18:29.039
Like, I mean, just in terms of those applications, as you say, from vast data sets.
00:18:29.279 --> 00:18:38.720
Um, shout out to the laggard, maybe in this in the stage, some of the Google Earth stuff that's coming out in terms of like weather mapping and everything, is just like an absolute bananas.
00:18:39.119 --> 00:18:46.480
And these real-world applications to vast libraries of corpora of text that we've had, we've just been sitting on for forever.
00:18:46.720 --> 00:18:58.960
And we we we take them and turn them into something that you do not need a PhD in five weeks of review time or five months or five years, you know, that you can instead and engage with and enjoy because some of these things.
00:18:59.200 --> 00:19:02.000
Venus, a topographical map of Venus, is just cool.
00:19:02.160 --> 00:19:04.559
Eight-year-old me is like, whoa, that's sick.
00:19:04.720 --> 00:19:10.720
I want to like to see it at the same lenses like Google Maps to be able to do that on Venus, another like incredible.
00:19:11.039 --> 00:19:17.759
So let's talk about AGI for a little bit because because this is this is what's been around since these new worlds have come out.
00:19:17.920 --> 00:19:19.519
We may be here at AGI.
00:19:19.680 --> 00:19:21.759
And I want to say something that will probably annoy people.
00:19:22.000 --> 00:19:26.720
At this stage, the AGI definition doesn't really matter to how seriously I take this technology.
00:19:26.880 --> 00:19:29.920
The line of what machines can actually do keeps moving.
00:19:30.079 --> 00:19:36.079
So let's look at the mass example recently that's got a bit of controversy, solves super challenging Earth problems.
00:19:36.319 --> 00:19:37.680
We'll ignore the controversy.
00:19:37.839 --> 00:19:46.960
OpenAI threw thousands of Asians at these problems in parallel, pursuing different approaches, following promising leads, and then using Astra itself to help formalize and verify the results.
00:19:47.119 --> 00:19:51.359
It's a pretty extraordinary way of attacking a problem humans have been wrestling with for decades.
00:19:51.519 --> 00:19:55.680
And the model that started doing it was trained on August 28th.
00:19:55.920 --> 00:19:58.400
Astra came out on September 3rd.
00:19:58.559 --> 00:20:02.720
So we're looking at Astra right now, saying, wow, look at what this thing can do.
00:20:02.880 --> 00:20:12.160
And the next thing, the next state-of-the-art model is already in the building solving complex math problems, doing stuff like solving super hard math problems with one million dollar bounties.
00:20:12.319 --> 00:20:15.200
So maybe the question isn't less have we reached HEI?
00:20:15.359 --> 00:20:20.160
Maybe it's how quickly these list of things we thought machines couldn't do disappears.
00:20:20.640 --> 00:20:24.079
Because right now, I feel like that list keeps getting shorter.
00:20:24.319 --> 00:20:26.720
We could spend the whole episode trying to decide whether it qualifies.
00:20:26.799 --> 00:20:29.359
I'm more interested in what happens when I give it a job.
00:20:29.519 --> 00:20:30.160
Can it do it?
00:20:30.240 --> 00:20:31.279
Where does it fall over?
00:20:31.519 --> 00:20:32.799
How much checking does it need?
00:20:32.960 --> 00:20:35.039
What happens when I let it work across several steps?
00:20:35.119 --> 00:20:37.759
Those questions affect what I'm gonna do tomorrow.
00:20:38.559 --> 00:20:51.119
Just to shout out Gary Marcus here, like I mean, perennial skeptic, like he he points out that there's quite a difference between what a human, so artificial general intelligence is roughly equated to what a human can do, right?
00:20:51.440 --> 00:20:52.559
But which human?
00:20:52.799 --> 00:20:56.880
Are we talking like the drunk tired guy at a bar face down at 4 a.m.
00:20:57.119 --> 00:21:00.160
on an insert your local street where people do things of that nature?
00:21:00.400 --> 00:21:06.960
Or are we talking your Einstein level physicist in the specific area of expertise, or perhaps Einstein doing something he's terrible at?
00:21:07.119 --> 00:21:07.759
You know what I mean?
00:21:07.839 --> 00:21:12.079
So these are very soft uh measures and metrics and everything else.
00:21:12.240 --> 00:21:23.119
So I think I think he's probably and I I get the feeling from you that you're comfortable leaving the label unresolved, that there's maybe a hype element marketing in there as well.
00:21:23.440 --> 00:21:24.400
Incredibly comfortable.
00:21:24.559 --> 00:21:31.200
I personally feel we will cross the Rubicon, enter the singularity, pick your term, and we'll look back and go, oh, was that it?
00:21:31.440 --> 00:21:34.160
As the next magical thing happens, or perhaps terrifying thing.
00:21:34.319 --> 00:21:37.119
There's people doing research can keep working on those precise definitions.
00:21:37.279 --> 00:21:43.119
I don't need a verdict on AGI before I'm allowed to look at this thing and go, wow, this is a very powerful technology.
00:21:43.440 --> 00:21:46.799
And Astra has also given me a little bit of the fear.
00:21:47.599 --> 00:21:49.279
The fear, yeah.
00:21:49.920 --> 00:21:52.160
Um I don't know if you felt it recently.
00:21:52.559 --> 00:21:56.319
I've been I've been fearless recently, but Astra it's crapped back in.
00:21:56.799 --> 00:21:59.680
The moment when those implications start to catch up with you, you're enjoying something.
00:22:00.240 --> 00:22:02.160
You're asking it to make a thing, it produces a result.
00:22:02.640 --> 00:22:12.640
And then you stop for a second because you realize you had an assumption about how difficult that task was, or how many different skills you'd need, or how much help you'd have to organise before you could even attempt that thing.
00:22:12.880 --> 00:22:14.720
But suddenly you're looking at the output.
00:22:14.880 --> 00:22:16.079
There is excitement in that.
00:22:16.160 --> 00:22:16.480
That's awesome.
00:22:16.559 --> 00:22:18.240
That's the whole thing that's kind of unlocked this last week.
00:22:18.319 --> 00:22:21.920
But there's also a little jolt of, ooh, I need to reconsider this whole thing a lot.
00:22:22.079 --> 00:22:28.559
I came to the office this week and generally said, I don't know if I want to talk about some of the things that I'm doing because I did it so fast without any struggle.
00:22:28.720 --> 00:22:29.680
I am uncomfortable.
00:22:30.000 --> 00:22:33.119
100% and very hard on the fear thing.
00:22:33.279 --> 00:22:35.119
Like I think this is part of the complexity.
00:22:35.200 --> 00:22:38.640
And just to, I've got a question for you on just a sec, but just an observation.
00:22:38.799 --> 00:22:40.480
This is a layered space.
00:22:40.720 --> 00:23:06.720
Because I mean, inherent to what you just said is the second order consequences of third and everything else for all that, the the hype around all these things, the anthropic, well, we know we're early mid-September, you know, in terms of you mentioned at the beginning lab whistleblowers for lack of another term, or people talking about greater than 10% chances of X, Y, and Z concerning, but there are intermediate steps around what AGI means and what it means for us as economic participants and society and everything else.
00:23:06.799 --> 00:23:11.119
Like um, I I a hundred percent hear you on that discomfort as well.
00:23:11.359 --> 00:23:18.319
So, do you think though that comes from the individual thing or the aggregate of how many different things that you've been trying?
00:23:18.400 --> 00:23:20.079
The the the broader package?
00:23:20.400 --> 00:23:31.680
I think the range has been the thing for uh these releases that have made me a bit kind of aware and how little effort I feel I'm putting in to get good results, impactful results.
00:23:31.839 --> 00:23:34.400
One impressive result is easy for me to put in a box.
00:23:34.559 --> 00:23:40.079
It's also highly likely with enough know-how and the right harness, previous models could probably have done something similar.
00:23:40.240 --> 00:23:48.000
But I'm talking about simplicity, which gives a layman access to great power versus the bloke and his 15 Mac minis and all these custom prompts.
00:23:48.079 --> 00:23:49.519
Like that's not needed anymore.
00:23:49.759 --> 00:23:58.240
And when I moved from our website to a course package, to Blender, to scraping public data to build out a research base and more and more, then I'm looking at what other people have done in completely different software.
00:23:58.319 --> 00:24:01.759
I start wondering where else my expectations are completely out of date.
00:24:01.920 --> 00:24:07.440
And the next thought is what could somebody who really knows this field do with the same tool?
00:24:07.599 --> 00:24:14.160
Or on the flip side, how much should I put into this to ensure there is some productive struggle that feels like I've earned the outcome?
00:24:15.200 --> 00:24:17.039
There's a there's a flip side to that as well.
00:24:17.519 --> 00:24:28.160
Like the the the perception of false mastery, the illusion rather of sort of of false mastery, that is you can also produce something impressive looking in an area where you're poorly equipped to judge it.
00:24:28.480 --> 00:24:28.960
Correct.
00:24:29.119 --> 00:24:29.359
Absolutely.
00:24:29.440 --> 00:24:32.480
And I think it's a the big fallacy of all these AI tools more broadly.
00:24:32.559 --> 00:24:35.119
I can judge whether our website communicates what I want to communicate.
00:24:35.279 --> 00:24:43.039
I know I can spend more time on lighting to convey the story and my cool cyberpunk setting, but I don't need an AGI verdict to take this seriously.
00:24:43.200 --> 00:24:48.960
I've got the work in front of me I would previously have struggled with, both productively and non-productively, to get to the same results.
00:24:49.200 --> 00:24:52.000
I've I've I brought us down a little bit, so let's go to word models.
00:24:52.319 --> 00:24:54.640
You've spent this episode building places, right?
00:24:54.720 --> 00:24:55.920
I think that's highly apropos.
00:24:56.000 --> 00:25:00.640
So uh Cyberpunk Street, Molox Alexandria, the Apple Park from Photos.
00:25:00.880 --> 00:25:03.759
Astro gets you there with Blender with an existing tool.
00:25:03.920 --> 00:25:05.759
Uh I got a version for you without Blender.
00:25:06.079 --> 00:25:14.319
And uh as I kind of said earlier, I think I think it's the next point, and this is not just me, this is people like Faith A Lee, commonly known as the godmother of AI.
00:25:14.400 --> 00:25:18.480
Um, she's got a company called World Labs, they have a product called Marble.
00:25:18.640 --> 00:25:20.480
Um, Google has one called Genie.
00:25:20.880 --> 00:25:24.400
Basically, if people aren't aware of this, you type a description of a room.
00:25:24.480 --> 00:25:31.599
Um, so think a hospital ward, a courtroom, RPD, thank you, on stage now, airport hangar or factory floor.
00:25:31.759 --> 00:25:32.400
Yeah.
00:25:33.200 --> 00:25:35.359
And it makes stuff.
00:25:35.759 --> 00:25:37.359
It makes the whole world.
00:25:37.680 --> 00:25:38.079
Yeah.
00:25:38.240 --> 00:25:45.759
So these are these are 3D environments that stay put, not some Will Smith Spaghetti Index video that morphs when you turn your head, right?
00:25:45.920 --> 00:25:48.559
This is a room that's still there when you turn back.
00:25:48.720 --> 00:25:52.880
You walk through it as Dale's doing now, you can edit it, export it into a game engine.
00:25:52.960 --> 00:25:57.440
So here we're talking Unreal or Unity with solid walls and lighting.
00:25:57.680 --> 00:25:59.359
So what do it cost though?
00:25:59.599 --> 00:26:00.240
Yeah.
00:26:00.559 --> 00:26:01.279
Good question.
00:26:01.440 --> 00:26:06.559
I mean, a dollar, dollar twenty American here in this case, rough draft for cents.
00:26:06.720 --> 00:26:12.640
Um, so think again, uh, educational applications or nursing, been doing a lot of talk with people about that recently.
00:26:12.880 --> 00:26:18.160
Make a ward for a lot of coffee, like make three for a coffee, maybe.
00:26:18.400 --> 00:26:20.720
And I mean, I'm not I'm not reading that off a website, right?
00:26:20.799 --> 00:26:22.000
So we we're building these.
00:26:22.079 --> 00:26:27.920
So I was in a talk recently, I had a marble export dropped into Unreal on my machine.
00:26:28.079 --> 00:26:33.200
Collision fields, the mesh and all of that on, so you can't just walk through walls, it's a walkable space.
00:26:33.440 --> 00:26:38.000
And um, me walking around a room that didn't exist the day before.
00:26:38.240 --> 00:26:42.240
I think that the thing that this really brings for me though is what happens to simulation.
00:26:42.319 --> 00:26:49.359
And you've kind of alluded to this as well with the Blender thing, but you've still got that really high threshold of using Blender, or maybe you don't, you know, like I mean, even knowing it exists.
00:26:49.440 --> 00:26:56.319
But what happens to simulation as a pedagogy or experiential learning when simulation stops requiring a budget?
00:26:56.640 --> 00:27:01.920
Because Alexandria, your Alexandria example, these these this is the answer starting to arrive.
00:27:02.079 --> 00:27:05.599
This is a a place to explore, these are things to touch.
00:27:05.759 --> 00:27:12.160
A teacher, an educational experience with students uh being asked what's evidence and what's reconstruction.
00:27:12.400 --> 00:27:16.480
Now for a down payment and a coffee, you know, it's it's amazing.
00:27:17.200 --> 00:27:20.240
And I got an email this week that got me so excited.
00:27:20.400 --> 00:27:32.319
It's the the next version of the World Labs products Atlas, and it said, Yep, the wait list, they're starting to activate it, and we'll start to see kind of get access to the next model from Stern 14th, which is crazy.
00:27:32.640 --> 00:27:33.839
Yeah, yeah, 100%.
00:27:34.400 --> 00:27:46.240
Because what you've got there is simulation, demonstration, execution, if you like, in terms of like I mean, putting on my this is a great idea, I'd like to try this, uh, questions or approaches I used to get from CCs and PMs.
00:27:46.400 --> 00:27:55.440
I'm like, okay, uh, that's quite an undertaking because this used to need a lab, a budget, a team, building their environment, collecting the demos, training the thing.
00:27:55.680 --> 00:27:59.759
Now it's uh all right, just smash that up while I have a coffee.
00:28:00.000 --> 00:28:04.000
And um, that's kind of where the sock problem went.
00:28:04.160 --> 00:28:04.799
You know what I mean?
00:28:04.880 --> 00:28:09.519
It used to be how many demos, now it's the demo, the demo, right?
00:28:09.759 --> 00:28:10.000
Yes.
00:28:10.160 --> 00:28:17.039
If if one good video does the work of of hundreds, one bad video does the work of hundreds of bad ones, I guess, as well.
00:28:17.119 --> 00:28:20.960
So the whole game is the quality, who's holding the camera, whether they actually know how to hold the thing.
00:28:21.039 --> 00:28:32.559
I guess sort of linking back up again to what you were saying around actual mastery and taste as it's been described and other things, whether whether it is done the way that you'd want a hundred copies of, if you like.
00:28:32.880 --> 00:28:34.559
But that's that's a teacher.
00:28:34.799 --> 00:28:35.759
Yeah, a hundred percent.
00:28:35.920 --> 00:28:36.160
Yeah.
00:28:36.319 --> 00:28:38.000
Uh and and that's what I try, right?
00:28:38.160 --> 00:28:40.799
So every skills lab used to be too expensive to build.
00:28:40.960 --> 00:28:43.200
So nursing, moot courts, factory floor.
00:28:43.359 --> 00:28:45.759
I mean, a lab that you'd never let a first hair touch.
00:28:45.920 --> 00:28:48.720
So cheap to build and free to break.
00:28:48.880 --> 00:28:56.000
I have a big long list of people who are keen to get their hands dirty with this thing, everything from aviation to logistics to build worlds.
00:28:56.079 --> 00:28:57.599
Because how fun is that, mate?
00:28:57.839 --> 00:28:58.720
Inception.
00:28:59.119 --> 00:29:00.720
You you're you're an olden man.
00:29:00.799 --> 00:29:02.160
You know, am I am I am I wrong?
00:29:02.319 --> 00:29:02.480
Of course.
00:29:03.039 --> 00:29:03.200
Of course.
00:29:03.359 --> 00:29:03.839
Yeah, yeah, yeah.
00:29:04.319 --> 00:29:06.480
So you can you can play the inception.
00:29:06.559 --> 00:29:08.079
Do you want to flip a city dale?
00:29:08.240 --> 00:29:09.119
How cool would that be?
00:29:09.279 --> 00:29:15.839
Unreal physics, and these are experiences which are permanent, which are shared, which you can put in the hands around the eyes.
00:29:15.920 --> 00:29:19.200
I mean, in terms of this, or create immersive spaces.
00:29:19.440 --> 00:29:21.440
Are two spec on demand.
00:29:21.599 --> 00:29:25.680
And so what I see from that, and the robots saying I didn't really build this up enough.
00:29:25.920 --> 00:29:29.359
This is the next point after this, where the machines learn these kind of things.
00:29:29.599 --> 00:29:32.400
That is the fear, and I feel that's strong there, 100%.
00:29:32.960 --> 00:29:41.359
But that the the wonder of this, we could go to that that Venus scan that you said and walk, walk on the surface of another world.
00:29:41.519 --> 00:29:42.960
How cool is that?
00:29:43.200 --> 00:29:45.599
And it's all about show don't tell.
00:29:45.920 --> 00:29:46.400
100%.
00:29:46.960 --> 00:29:47.839
Show don't tell.
00:29:48.079 --> 00:29:53.440
Can you please uh not to not to get away from this because I love it, but you promised me a fly.
00:29:53.680 --> 00:29:54.799
I want to hear about this fly.
00:29:54.880 --> 00:29:55.599
I don't know enough.
00:29:56.000 --> 00:30:02.160
Seems seems harsh to to spin from building the entire world to talking about a fly.
00:30:02.319 --> 00:30:06.079
What a way, what a way to go out on our last season, our last episode.
00:30:06.319 --> 00:30:06.880
But this is cool.
00:30:07.200 --> 00:30:07.599
Never give me a few.
00:30:07.759 --> 00:30:08.480
This is a cool rabbit hole.
00:30:08.640 --> 00:30:09.839
We need to get really excited.
00:30:09.920 --> 00:30:16.720
So there's a clip of a fly brain model playing beat saber, which is crazy.
00:30:16.799 --> 00:30:20.799
So the creator explained that the version showed that they've been fitted to a gameplay route.
00:30:20.880 --> 00:30:22.640
So we should be careful what this actually demonstrates.
00:30:22.720 --> 00:30:27.200
But the example I really want to talk about is the fly escape room.
00:30:27.599 --> 00:30:34.160
You arrange objects inside a house, you release the flies, and you try to get them to find a way outside.
00:30:34.319 --> 00:30:42.480
And the fly's movement is driven by simulated neural activity using connections drawn from a real fruit fly nervous system map.
00:30:42.880 --> 00:30:45.519
The fly, the fly is an MPC.
00:30:45.759 --> 00:30:47.680
A non-player character, yes, that's right.
00:30:47.759 --> 00:30:57.440
And a computer controlled character whose behavior is influenced by a model built around this biological wiring, which is a fairly extraordinary sentence to say about just this kind of browser game.
00:30:57.519 --> 00:30:58.640
This is all built on Astra.
00:30:59.680 --> 00:31:00.880
Oh my god.
00:31:01.119 --> 00:31:04.480
Okay, so where does the wiring come from?
00:31:04.799 --> 00:31:16.640
I'm assuming only because you want to map a flyer this weekend, but researchers at HHMI, Genilia, Google Research, and their collaborators map the male fruit flies brain and central nervous brain and central nervous system.
00:31:16.799 --> 00:31:23.680
The full data set contains roughly 166,000 neurons, about 125 million synaptic connections.
00:31:24.000 --> 00:31:30.160
AI helped reconstruct the neural structures from microscope images with human experts checking and annotating the results.
00:31:30.240 --> 00:31:32.000
That map is called a connectome.
00:31:32.319 --> 00:31:34.480
I'm getting nerdy here, but it's cool.
00:31:34.720 --> 00:31:37.680
Essentially, it records which neurons connect to which.
00:31:38.640 --> 00:31:42.720
Oh my god, this is the first step to upload me to the cloud.
00:31:43.599 --> 00:31:48.160
And can I say, Cyber, it's kind of fun that someone just like, yeah, you know what?
00:31:48.240 --> 00:31:52.240
The first best use case for this is let's let's make an escape room.
00:31:52.559 --> 00:31:54.079
I love when the internet's fun.
00:31:54.160 --> 00:31:56.160
And the internet has been very fun recently.
00:31:56.319 --> 00:31:58.160
It's a very, very, very internet response.
00:31:58.400 --> 00:31:58.880
Absolutely.
00:31:59.279 --> 00:32:00.640
A very scientific resource.
00:32:00.799 --> 00:32:02.400
We live in strange, wonderful times.
00:32:02.559 --> 00:32:03.599
But there's an important detail here.
00:32:03.680 --> 00:32:10.799
The game's documentation says it uses a selected part of the connectome, its neuron equations, the sensory inputs, and the movement rules, they're models.
00:32:10.960 --> 00:32:15.119
And the author explicitly cautions that it hasn't been validated as a digital replica of a living fly.
00:32:15.519 --> 00:32:19.200
Look a bit boring, because for me, this is just replicating hundreds of flies.
00:32:19.680 --> 00:32:22.480
But even at this level, I find it so fascinating.
00:32:22.960 --> 00:32:31.200
You can actually follow a digital fly, inspect its simulated neural activity while it explores, because we've just spent the episode talking about cool tools and how we can build worlds.
00:32:31.440 --> 00:32:38.480
Now we're looking at an experiment that takes a biological connection data and uses it to influence the behavior of something inside a virtual environment.
00:32:38.640 --> 00:32:41.839
Nick, I'd seen this fly park cars recently.
00:32:42.160 --> 00:32:43.759
How are we not in a simulation?
00:32:43.920 --> 00:32:49.839
No, this is the bit this is this is genuinely the that first step and the practical, how you make that happen.
00:32:50.000 --> 00:32:50.799
There we go, here we go.
00:32:50.880 --> 00:32:53.920
Here's the fly's the fly driving a car.
00:32:57.519 --> 00:32:58.160
Oh my goodness.
00:32:58.400 --> 00:33:03.200
Simple sacrocy AI sop, an insane, insane science.
00:33:05.759 --> 00:33:06.559
It's sound good.
00:33:06.960 --> 00:33:07.759
It's sound good.
00:33:09.359 --> 00:33:12.079
What is a fly expressing that it wants to honk a horn?
00:33:12.480 --> 00:33:15.359
What lived experience do they make sound?
00:33:17.599 --> 00:33:18.720
What am I meaning?
00:33:18.799 --> 00:33:19.839
Why am I doing this?
00:33:20.079 --> 00:33:20.720
Sorry, God.
00:33:22.559 --> 00:33:23.680
Fly Pacito.
00:33:23.839 --> 00:33:24.240
Oh my god.
00:33:24.400 --> 00:33:27.599
But imagine the possibilities for virtual animals or small ecosystems.
00:33:27.680 --> 00:33:33.440
How far you could get by giving the inhabitants different biologically informed models and seeing how they interact.
00:33:33.680 --> 00:33:38.400
I'm projecting it beyond this particular game, but that's the question it kind of puts in our head, Nick.
00:33:38.799 --> 00:33:42.640
Yeah, I I guess how much yeah, but this is what I was asking about the horn, right?
00:33:42.720 --> 00:33:47.440
So how much comes from the fly itself and how much from the rules the developer put around it.
00:33:48.000 --> 00:33:48.400
100%.
00:33:48.799 --> 00:33:54.480
And the way I'm thinking about it, it's kind of like treating the fly like a graphics card and seeing how it react and et cetera, et cetera.
00:33:54.640 --> 00:33:56.720
But there's an educational experience here as well.
00:33:56.880 --> 00:34:03.440
Here's a model, here's what went into it, here's what it does, and how would you investigate whether its behavior tells you anything useful about the system it represents?
00:34:04.240 --> 00:34:06.160
And how do you get the flies to leave the house?
00:34:08.320 --> 00:34:09.760
That's the the modeling.
00:34:09.840 --> 00:34:14.559
So that's the billions of dollars of compute that Chen Senwang's building his his fortune on.
00:34:14.880 --> 00:34:16.719
I think that's a good place to finish the season.
00:34:20.159 --> 00:34:25.199
Can we just take a second to acknowledge that that is the beginning of a simulated reality?
00:34:25.360 --> 00:34:28.079
And all of that, we live in the Sims.
00:34:28.400 --> 00:34:30.880
That that is the that is the end point, you know what I mean?
00:34:31.119 --> 00:34:33.199
And it's it's simulations the whole way down.
00:34:34.159 --> 00:34:35.360
It starts with flies.
00:34:39.840 --> 00:34:41.679
I am crying inside.
00:34:41.840 --> 00:34:43.199
But wondrously as well.
00:34:43.360 --> 00:34:48.800
Like, I mean, I could fly in that world, upload me to the fly to the cloud where maybe I'm gonna fly.
00:34:48.960 --> 00:34:51.840
I'll tell you why I'm beeping, because this guy won't move.
00:34:52.159 --> 00:34:58.639
And I think that's a great place to finish the season because underneath all the model comparisons, I've got a fairly simple reaction to what we've been showing.
00:34:58.800 --> 00:35:00.880
These are things I'm willing to try right now.
00:35:01.039 --> 00:35:02.400
Maybe not the fly.
00:35:02.559 --> 00:35:05.360
And I suspect there's still plenty of things that I'm treating as too difficult.
00:35:05.440 --> 00:35:06.719
Maybe the computer can't do that.
00:35:06.800 --> 00:35:09.519
That I need to go back and test those assumptions.
00:35:10.800 --> 00:35:12.239
It had the inception thing.
00:35:12.639 --> 00:35:14.480
Oh, so Nick, what are you gonna do over the break?
00:35:14.559 --> 00:35:16.079
Are you gonna but what's what's the plan?
00:35:16.159 --> 00:35:17.280
You've got you got a bit of time off?
00:35:18.000 --> 00:35:19.280
No time off for me, sucker.
00:35:19.519 --> 00:35:21.039
No, I'm obviously still at work, mate.
00:35:21.119 --> 00:35:23.519
Like uh unlike some of jaunting around.
00:35:23.599 --> 00:35:30.159
But um, my big plan for for for the spare time, the the though I'll miss you, of course, inception style world building.
00:35:30.400 --> 00:35:34.880
I mean, uh like I said, aviation, logistics, film, games, business, so much cool stuff.
00:35:35.119 --> 00:35:47.280
Plan is to export it, get someone who's actually an expert on these things to get in there, walk around it, shoot holes in it, tell me what's wrong, and then we're gonna try and get some students in there, if we can, maybe next semester.
00:35:47.440 --> 00:35:52.239
Because I reckon if it survives the aviation mechanic, right, it goes in the course.
00:35:52.400 --> 00:35:53.840
Uh, what about yourself?
00:35:54.159 --> 00:35:57.039
I am trying to keep the fear under control.
00:35:57.199 --> 00:35:58.079
That's the main thing.
00:35:58.320 --> 00:36:00.000
So I guess just go play with these tools.
00:36:00.079 --> 00:36:01.760
Go see if you can build out a fly.
00:36:01.920 --> 00:36:04.480
But for everyone else, thank you so much for listening to this season.
00:36:04.559 --> 00:36:05.519
We'll be back soon.
00:36:05.679 --> 00:36:06.079
I promise.
00:36:06.159 --> 00:36:07.599
It's only a short hiatus.
00:36:07.840 --> 00:36:09.519
Thank you for status intelligence.
00:36:09.760 --> 00:36:12.880
You can find us on YouTube, you can find us on all your favorite podcast feeds.
00:36:13.119 --> 00:36:20.320
Until then, stay curious, stay intelligent, and stay a fly.
00:36:20.559 --> 00:36:22.159
Um I'm not sure.
