WEBVTT
00:00:00.000 --> 00:00:04.950
As people fear manipulation, especially when people fear the longer AGI style threats, right?
00:00:04.950 --> 00:00:18.359
They want to see what did the model decided to manipulate you in their response, or maybe what instruction it chose to disregard, but also like some people are saying it would be very interesting to see where the model goes and realizes that it's done something bad
00:00:26.774 --> 00:00:30.504
Hello and welcome back to another episode of deep learning with PolyAI.
00:00:30.815 --> 00:00:35.145
The only podcast that sends you on a wild rollercoaster every time we kick off.
00:00:35.185 --> 00:00:36.045
I'm your host, Damien.
00:00:36.064 --> 00:00:37.765
I'm here as usual.
00:00:37.945 --> 00:00:40.365
Was CEO and co founder of Poly AI, Nikola Mrkšić.
00:00:40.645 --> 00:00:41.015
Hey, Nikola.
00:00:41.034 --> 00:00:41.384
How are you?
00:00:41.734 --> 00:00:42.284
Good to see you.
00:00:42.335 --> 00:00:42.755
I'm good.
00:00:42.755 --> 00:00:43.085
I'm good.
00:00:43.085 --> 00:00:44.115
Fresh off of a red eye.
00:00:44.155 --> 00:00:44.634
How are you?
00:00:44.634 --> 00:00:46.335
Yeah, you just came back from Montreal, correct?
00:00:46.484 --> 00:00:46.575
Yeah.
00:00:46.576 --> 00:00:47.104
Yeah.
00:00:47.104 --> 00:00:47.814
Oh, wonderful.
00:00:47.825 --> 00:00:49.314
Another conference in the bag.
00:00:49.335 --> 00:00:51.954
Well, a lot has happened this week while you were in Montreal.
00:00:51.954 --> 00:01:03.625
So I thought we'd do an episode on the AI news and review to talk about some of the key headlines that are making their way around some of the big names in the AI space or practitioners and deployers of AI and AI systems.
00:01:04.125 --> 00:01:08.805
We'll go through those fairly quickly because I want to finish on probably the biggest story.
00:01:08.814 --> 00:01:10.034
And that's something that happened.
00:01:10.484 --> 00:01:18.474
Well, I want to say probably about 25 minutes ago was announced, which is which is open AI rolling out a one or GPT five and.
00:01:18.575 --> 00:01:20.114
Previous circles, also strawberry.
00:01:20.364 --> 00:01:22.305
So we'll talk a little bit about your early impressions.
00:01:23.635 --> 00:01:25.594
This is what they would refer to as GPT five.
00:01:25.594 --> 00:01:26.275
So I think,
00:01:26.344 --> 00:01:26.775
Oh, good.
00:01:26.775 --> 00:01:33.155
Well, you could correct me then when we get to that point, but definitely want to save some time to talk about strawberry slash Oh one slash.
00:01:33.375 --> 00:01:34.965
Definitely not GPT five.
00:01:35.334 --> 00:01:47.045
Want to kick off by bringing up the CEO of Klarna in the news again, knew he made ripples when he said he was cutting 50 percent of his support staff when it came to some AI automation.
00:01:47.545 --> 00:02:01.224
And now he's basically telling Salesforce and Workday as their CRM and ERP systems goodbye, as he moves toward this, what seems to be a monolithic AI system, my question to you is.
00:02:01.549 --> 00:02:03.879
Is this just a lot of marketing flair?
00:02:03.879 --> 00:02:16.439
I, is it possible for one singular nebulous AI based system to operate an enterprise, or is this just something that's business as usual when it comes to furthering automation and it's got a heavy marketing spin on it.
00:02:17.530 --> 00:02:19.270
I don't doubt the marketing spin at all.
00:02:19.280 --> 00:02:23.689
I think that what was the other thing that they're not using Salesforce, they're building everything in house.
00:02:24.030 --> 00:02:27.830
Like we deal with buyers and clients who have this attitude and they do everything at house.
00:02:27.840 --> 00:02:35.919
So, I am not surprised that developers at Klarna can build a CRM for their own needs, whether that's the best use of their time, that's up to them.
00:02:36.349 --> 00:02:38.650
And I don't really have a strong opinion.
00:02:38.860 --> 00:02:40.370
Or care enough about far now.
00:02:40.400 --> 00:02:43.409
I don't think that it's actually all that relevant.
00:02:43.650 --> 00:02:46.360
It does sound like they're leaning in heavily and drinking their own Kool Aid.
00:02:46.360 --> 00:02:46.659
So
00:02:47.219 --> 00:02:57.070
is this something that theoretically, when it comes to building, is this something that probably be more ideal for a less complex, maybe a smaller midsize enterprise, That may be more approachable.
00:02:57.070 --> 00:03:08.491
It seems like a pretty daunting task even with the resources that Klarna may or may not have for an enterprise with the complexity and size that they do to just strike these fairly significant backend systems.
00:03:08.531 --> 00:03:09.320
So what was it?
00:03:09.350 --> 00:03:10.711
Salesforce and Workday
00:03:10.950 --> 00:03:11.610
for ERP.
00:03:12.360 --> 00:03:18.640
Look, given the nature of their business, which is like to Well, provide like loans for buying, moderately expensive things.
00:03:18.711 --> 00:03:18.971
Sure.
00:03:18.971 --> 00:03:26.591
I believe that they have a system of record where they can, track the products they're sharing and stuff and that they don't need Salesforce or given the nature of what they do.
00:03:26.591 --> 00:03:29.181
I'm almost surprised that they, I don't know how they were using it.
00:03:29.181 --> 00:03:32.121
So it does seem and feel a bit bombastic, right?
00:03:33.300 --> 00:03:43.161
I think, the previous announcement around automating support with with LLMs, given the nature of their client base, which is like younger tech savvy people, right?
00:03:43.221 --> 00:03:45.110
Those are people you can't push into those channels.
00:03:45.121 --> 00:03:47.001
So if you do it deliberately.
00:03:47.390 --> 00:03:48.200
Fair game, right?
00:03:48.450 --> 00:03:53.480
We speak to a lot of companies and, there are challenger banks that we work with and those that we don't, right?
00:03:53.610 --> 00:03:59.700
For example, Revolut is deeply focused on, well, decided not to have voice as part of their offering.
00:04:00.320 --> 00:04:03.401
Amazon doesn't really give you an ability to call Amazon either.
00:04:03.401 --> 00:04:12.526
And they just push you into digital channels and they're rich enough to just send you a replacement in case that it seems likely that you're not trying to You know, get a free item.
00:04:12.586 --> 00:04:15.336
So, these are, I think, design decisions at the top end.
00:04:15.346 --> 00:04:15.996
And I think that.
00:04:16.586 --> 00:04:20.326
They clearly did well enough for the first announcement that they're getting hooked on it.
00:04:20.326 --> 00:04:21.415
So good for them.
00:04:21.435 --> 00:04:24.865
If rebuilding an ERP on the CRM is the best use of their time.
00:04:25.896 --> 00:04:26.625
I'm happy for them.
00:04:27.216 --> 00:04:27.516
Yeah.
00:04:28.115 --> 00:04:28.396
All right.
00:04:28.435 --> 00:04:28.935
Fair enough.
00:04:29.125 --> 00:04:30.326
Klarna, you're in the past.
00:04:30.466 --> 00:04:32.196
Apple intelligence coming up next.
00:04:32.216 --> 00:04:41.415
Obviously the big Apple product showcase that occurred in the past week or so announcing basically an entire device ecosystem built around.
00:04:41.750 --> 00:04:47.670
Apple intelligence, which is, which was previously billed as they're on device AI approach.
00:04:47.680 --> 00:05:01.011
And it, there was a little bit of blending with, their previous iCloud offering and now moving into Apple intelligence, which because of the compute needs to happen on devices, which is why they've been built from the ground up.
00:05:01.560 --> 00:05:06.690
They have this private cloud system that looks to blend the two.
00:05:07.391 --> 00:05:25.221
I'm curious with iCloud being their pivotal go to and trying to onboard all Apple users into iCloud and now moving more, I would say intimately to the devices itself with Apple intelligence is this a future where those two approaches, those two go to markets like live together, peacefully.
00:05:25.711 --> 00:05:26.060
I think
00:05:26.060 --> 00:05:26.401
so.
00:05:26.451 --> 00:05:34.120
I think that it is inevitable with larger models and these compute needs and that you will have part of it on device.
00:05:34.190 --> 00:05:36.331
That was always going to be the case in one form or another.
00:05:36.331 --> 00:05:37.540
And you can trade it off, right?
00:05:37.940 --> 00:05:41.281
Previously, you could process speech recognition online or do it on the phone.
00:05:41.740 --> 00:05:45.440
And both have implications in battery life, performance, size of model.
00:05:45.841 --> 00:05:46.440
They still do.
00:05:46.940 --> 00:05:55.930
I think like with more and more features that are based on LLMs you're going to need like a faster phone with a better CPU, GPU, right?
00:05:57.161 --> 00:05:58.331
You're going to need more storage.
00:05:58.331 --> 00:06:04.430
I don't know about you, but my 256 gigabyte iPhone 14 right now feels like a piece of ancient history that I tore up.
00:06:05.771 --> 00:06:07.860
Honestly, I am out of space constantly.
00:06:07.860 --> 00:06:09.901
And the next one I buy is definitely going to be like a terabyte.
00:06:10.451 --> 00:06:13.360
So, I think it's a good way for them to.
00:06:13.591 --> 00:06:15.651
Create a need for people to upgrade their phones.
00:06:15.971 --> 00:06:16.891
I can't wait to get one.
00:06:16.930 --> 00:06:18.310
So I'm excited about everything.
00:06:18.321 --> 00:06:20.521
And, they're pushing a lot of it into Siri.
00:06:20.680 --> 00:06:21.600
Someone who used to work there.
00:06:21.600 --> 00:06:25.401
I'm really excited to see all the ways in which Siri will improve.
00:06:25.781 --> 00:06:27.500
Big believer in Apple in general.
00:06:27.500 --> 00:06:32.060
I think that no company really treats its product with as much like.
00:06:32.896 --> 00:06:35.216
Respect and love and care as they do.
00:06:35.365 --> 00:06:35.675
Yeah.
00:06:36.036 --> 00:06:45.875
So, like when it comes to privacy and designing the right approach around giving you all that while giving you the best that you can have in terms of these features.
00:06:46.341 --> 00:06:47.120
I'm really excited.
00:06:47.120 --> 00:06:50.031
The whole demo was like explaining images, right?
00:06:50.071 --> 00:06:54.430
Search over one place where I would still go out at the Apple ecosystem always was Google photos.
00:06:54.430 --> 00:06:55.500
I don't know about you, right?
00:06:55.990 --> 00:07:04.430
But if you're looking for like expenses, and you have 50 things to submit, like Google photos tends to be better at searching, right?
00:07:04.680 --> 00:07:05.141
Oh, sure.
00:07:05.290 --> 00:07:05.560
Yeah.
00:07:05.750 --> 00:07:12.050
If I'm dealing with like paperwork and I've taken a photo, I know I have, but I can't find it like Google photos have.
00:07:12.730 --> 00:07:17.911
To this point, been a lot better at like vaguely describing the thing and it being found in my photos.
00:07:17.911 --> 00:07:21.761
So I'm excited for Apple to catch up with that without giving all my data to Google.
00:07:21.781 --> 00:07:23.021
So, it's cool.
00:07:23.081 --> 00:07:23.790
I'm really excited.
00:07:23.790 --> 00:07:26.019
Yeah.
00:07:26.100 --> 00:07:27.040
Oh, I definitely will.
00:07:27.071 --> 00:07:30.040
I'm typically on a three a three model cycle.
00:07:30.321 --> 00:07:31.540
But I'm sitting on a.
00:07:32.130 --> 00:07:34.370
I think at 13, so I'm ready to move up.
00:07:34.761 --> 00:07:36.420
It just go so happen to coincide.
00:07:36.620 --> 00:07:50.091
The privacy thing is a huge issue and that's always been one of the flagship tenants of Apple is, allowing for the utmost of consumer privacy, even going so far as to blocking, federal and governmental intrusion for some devices.
00:07:50.531 --> 00:07:52.290
And this definitely plays into that.
00:07:52.815 --> 00:08:01.826
What does it say when they're pushing the safety and security of on device privacy again, going back to the cloud before, are they basically damning their previous efforts?
00:08:03.505 --> 00:08:11.516
I'm just saying like they've always billed that iCloud was very private and secure as well, but they're saying you get added security when you move to on device.
00:08:11.615 --> 00:08:14.295
If you're saying that, are you basically, I
00:08:14.295 --> 00:08:16.795
don't think that's what they're saying.
00:08:16.795 --> 00:08:23.596
I think for the family of things and tasks that you're going to need to be on device for it is safe.
00:08:23.605 --> 00:08:24.795
I think maybe.
00:08:25.415 --> 00:08:31.555
What it may be is a bit of a dig at other services where, if you're using chat GPT, they can use your data to train, right?
00:08:31.656 --> 00:08:39.796
I'm not exactly sure what's in the terms and conditions of Android devices with Gemini and stuff, but I feel like it probably is angled more in that direction.
00:08:40.441 --> 00:08:44.071
Like war than specifically against iCloud per se.
00:08:45.230 --> 00:08:45.890
Yeah, that's fair.
00:08:45.931 --> 00:08:46.390
That's fair.
00:08:46.390 --> 00:08:56.390
So, so arguably saying of these two behemoths of the two like dominant forces and, portable consumer devices, like we're going to be more secure and more respectful of your private data.
00:08:56.971 --> 00:08:59.539
Yeah, I think that's always a fair assumption when it comes to this too.
00:08:59.811 --> 00:09:00.681
Yeah, I agree.
00:09:01.140 --> 00:09:12.870
Speaking of your history with Siri you were among the teams that you were among one of the first teams to pioneer what has become largely attributed as the first consumer voice assistant to reach.
00:09:12.870 --> 00:09:14.041
We were only
00:09:14.041 --> 00:09:14.821
acquired into
00:09:14.821 --> 00:09:15.211
Siri.
00:09:15.211 --> 00:09:15.921
You know what?
00:09:15.921 --> 00:09:17.041
We're the badge with pride.
00:09:17.041 --> 00:09:17.280
Nickel.
00:09:17.280 --> 00:09:19.510
I'm still going to tell everybody you were on one of the first teams.
00:09:19.591 --> 00:09:21.490
But you obviously had an intimate.
00:09:21.600 --> 00:09:26.400
Involvement with the research and development that went into what became Siri.
00:09:27.630 --> 00:09:33.640
This is the first probably truly transformational iteration of the voice assistant since its inception.
00:09:33.640 --> 00:09:37.451
So what are you looking forward to seeing what this next generation of Siri brings?
00:09:38.201 --> 00:09:39.211
It's a tough one, right?
00:09:39.260 --> 00:09:44.341
I can't really talk much about the history of Siri while I was there, but it was the first.
00:09:44.860 --> 00:09:50.260
Of them, which meant that, um, the mistakes that could have been made were made first date.
00:09:50.311 --> 00:09:54.890
I think just maybe to redirect to a company that it didn't work out where I can talk a bit more freely.
00:09:56.081 --> 00:10:01.010
No, I think when Amazon would Alexa hit the gold vein of like smart speakers in homes
00:10:01.010 --> 00:10:01.980
and
00:10:01.980 --> 00:10:08.806
then went into a big push there, I think they actually built a system clunkier than Siri, cause they were really like mass producing it.
00:10:09.015 --> 00:10:15.416
It felt the Soviets have 14, 000 tanks and then they lose the cold war without a bullet fire because their economy crumbles.
00:10:15.416 --> 00:10:17.355
Like Alexa was built very quickly.
00:10:17.441 --> 00:10:18.265
To be complex.
00:10:18.265 --> 00:10:19.639
And it was very complex.
00:10:19.640 --> 00:10:22.760
So the technical debt of those systems is hard to address.
00:10:22.760 --> 00:10:30.740
And now when you basically change the paradigm so much around, LLMs doing what you previously had to have very complex ontologies for.
00:10:30.740 --> 00:10:35.850
So, think of like different applications and domains and things that have to be built up one by one.
00:10:35.900 --> 00:10:37.841
And very complex controllable things.
00:10:37.841 --> 00:10:40.061
And, with an LLM, you can't really relinquish control.
00:10:40.061 --> 00:10:51.206
So, I think it really remains to be seen what's, what's Will have been, it will not be much like our generative deployments have not been a straight up, like here's a prompt and run an LLM.
00:10:51.206 --> 00:11:00.716
If you have different requirements, both like in terms of controllability, understanding safety, et cetera, you can't just go and throw an LLM at it.
00:11:00.865 --> 00:11:06.245
So I think what I'm really interested in seeing is just like how LMS will have been used to change the output, right?
00:11:06.245 --> 00:11:10.535
I think the general conversational ability, and it depends, right?
00:11:10.586 --> 00:11:16.296
Apple's always been very opinionated around, what's your, you shouldn't say where, how far to go in different places.
00:11:16.306 --> 00:11:17.475
It's a hard thing to get right.
00:11:17.846 --> 00:11:30.056
Cause you're like done if you're doing them, if you don't, in terms of trying to go on like a philosophical debate with someone or taking a joke too far and stuff, so you can do more than they will dare to do simply because of who they are.
00:11:30.056 --> 00:11:32.091
They can do a lot more than they will dare to do.
00:11:32.140 --> 00:11:33.701
So it'll be exciting nonetheless.
00:11:33.730 --> 00:11:50.671
But I think historically the only step that you can take is up when it comes to home, voice assistants or smart speakers, historically, all those limitations, as far as when it hits the consumer level, it causes a lot of frustration because of some of those safeguards that are put in place.
00:11:50.681 --> 00:11:51.671
So maybe.
00:11:51.686 --> 00:11:54.265
I think it's people are very critical of all of them.
00:11:54.316 --> 00:12:00.926
Apart from a very small minority that really loves them the way they are in their power users and know exactly how to use them.
00:12:00.956 --> 00:12:01.416
Exactly.
00:12:01.436 --> 00:12:01.625
But the
00:12:01.625 --> 00:12:12.696
truth is, I think the real thing, and the reason why it would probably I created a company that's focused on enterprise use cases is that I think there's a stronger need there because there are repeatable things that happen at scale over the phone, right?
00:12:13.046 --> 00:12:17.875
The problem with consumer voice assistance is that like everyone imagines a different thing.
00:12:18.201 --> 00:12:26.681
We all have this fuzzy understanding of what this future Jarvis, mega, Jarvis style, mega powerful voice assistant will be.
00:12:26.701 --> 00:12:28.971
What do you really need it to do?
00:12:29.730 --> 00:12:30.610
Exactly.
00:12:30.650 --> 00:12:31.791
That's the thing is that I've always,
00:12:31.811 --> 00:12:36.461
Like whenever the inevitable end of, I've had this conversation many times in the last few weeks, right?
00:12:36.931 --> 00:12:39.571
The inevitable conclusion is, well, she know everything about me.
00:12:39.801 --> 00:12:47.750
And then as it books a trip for me and my family, she know that, in our itinerary, my life, my, my wife loves Thai food and I love Japanese food.
00:12:47.750 --> 00:12:49.691
So come on, man, you wouldn't trust.
00:12:49.931 --> 00:12:54.490
Like a human being to do right, you're probably going right.
00:12:54.490 --> 00:12:56.441
So people have a lot of nebulous.
00:12:56.816 --> 00:13:02.235
Desires that kind of like point towards I want this or that, and they'll never be happy with the consumer voice.
00:13:02.245 --> 00:13:04.806
So I think it's just, it's difficult, right?
00:13:04.836 --> 00:13:08.355
Like it might be an improvement that is like 10 X and you'll still have people.
00:13:08.855 --> 00:13:14.135
Who are expecting the Jarvis, expecting the C3PO, expecting the fully interactive, like 100 percent of the
00:13:14.135 --> 00:13:14.905
integrations.
00:13:14.916 --> 00:13:16.235
And then we go back to.
00:13:16.625 --> 00:13:18.576
The classic what would you trust it to do?
00:13:18.586 --> 00:13:19.355
What permissions?
00:13:19.416 --> 00:13:19.676
Yeah.
00:13:20.186 --> 00:13:21.485
How angry will you be when it gets?
00:13:21.566 --> 00:13:21.885
It's the
00:13:21.885 --> 00:13:24.796
same people that are like, but I don't want it to know any of my information too.
00:13:24.846 --> 00:13:28.855
It's like you should know me, but I'm not telling you anything about me and I'm preventing you from, yeah.
00:13:28.855 --> 00:13:32.785
Well, having settled that one day, when like one of these companies goes there.
00:13:33.306 --> 00:13:35.416
For the people that want both Apple will still be the best.
00:13:35.495 --> 00:13:39.515
I sound like a, an Apple app, but absolutely.
00:13:41.597 --> 00:13:42.268
To the core.
00:13:42.327 --> 00:13:42.817
Thank you.
00:13:42.837 --> 00:13:43.437
Dad joke.
00:13:43.988 --> 00:13:44.278
All right.
00:13:44.307 --> 00:13:44.557
Cool.
00:13:44.577 --> 00:13:47.357
Well, excited to see what's in store for the next iteration.
00:13:47.357 --> 00:13:49.607
And I'm glad that you and I are going to get the same model phone.
00:13:49.687 --> 00:13:51.727
We can sit here and geek out over it together.
00:13:52.028 --> 00:13:52.307
Okay.
00:13:52.408 --> 00:13:56.337
Moving on Sam Altman backs XO watt.
00:13:56.628 --> 00:13:59.557
It's a solar storage company.
00:13:59.577 --> 00:14:24.677
They raise a 20 million seed round, obviously clean storage, minimal carbon footprint impact, climate friendly energy solutions have always been top of mind, well, at least in the past decade or so more so when the compute Strain has hit because of these LLMs that are out there using tremendous amounts of processing power.
00:14:25.217 --> 00:14:30.878
Is he buying into his own pockets a little bit with this move?
00:14:31.207 --> 00:14:32.087
Look, can you blame him?
00:14:32.138 --> 00:14:32.687
I dunno.
00:14:32.788 --> 00:14:34.947
The truth is yeah, like we need a lot more computer.
00:14:34.947 --> 00:14:41.288
I think the fact that we need so much more compute now probably makes the development of a lot of these technologies a greater priority.
00:14:41.378 --> 00:14:42.727
You could just do nuclear, right?
00:14:42.727 --> 00:14:52.418
I think I think there are videos of Elon circulating where he's talking about, we could like power everything we have, we need on the planet with enough solar and the right battery technology.
00:14:52.768 --> 00:14:55.067
So yeah, I think people are excited and they'll work on it, right?
00:14:55.168 --> 00:14:58.758
Like whether it's solar, nuclear, depending on how you feel about it, right?
00:14:58.847 --> 00:15:02.148
I think everyone's like gearing up for like more and more, right?
00:15:02.447 --> 00:15:03.298
In terms of compute.
00:15:03.298 --> 00:15:04.807
So like energy is a
00:15:05.673 --> 00:15:13.832
But we've also talked about how there's also a race to reduce the resources by building efficiencies and, progressive models, who's going to win.
00:15:14.273 --> 00:15:15.482
No there's no winning, right?
00:15:15.482 --> 00:15:24.482
Like in, in building like you can always increase capacity, model size, training, quality of chips, speed, to train yet a bigger model.
00:15:24.602 --> 00:15:30.092
Of course I think innovation in terms of like high quality synthetic data, better architectures will.
00:15:31.278 --> 00:15:33.577
Like this, like two axis along which you can progress.
00:15:33.587 --> 00:15:35.538
And they both amplify the end outcome.
00:15:35.587 --> 00:15:37.128
I remember being at a talk where.
00:15:37.707 --> 00:15:49.268
Ilya Sutskever talked about Hey, transformers just sped things up by a few years because with hardware developing LSTM cells three, four years later would have achieved the GPT 3 moment.
00:15:49.278 --> 00:15:56.827
So, yeah, if you have another, and I think there are many people proposing all kinds of architectures, some of which are pretty strong as alternatives to transformers.
00:15:56.837 --> 00:15:59.957
So that's just people will work and attack on both fronts.
00:15:59.977 --> 00:16:02.927
And, ideally we just continue to make progress from both.
00:16:04.057 --> 00:16:09.577
So everybody wins with this play, maybe a bit of a conflict of interest.
00:16:09.648 --> 00:16:11.548
We'll leave that to to progression down the road.
00:16:11.548 --> 00:16:11.927
Not our business.
00:16:11.927 --> 00:16:12.615
All
00:16:12.615 --> 00:16:13.302
right.
00:16:13.302 --> 00:16:14.677
Open AI.
00:16:14.677 --> 00:16:14.748
Can
00:16:14.748 --> 00:16:15.217
we buy in?
00:16:15.217 --> 00:16:15.988
That's my question.
00:16:17.748 --> 00:16:19.658
Is there a Kickstarter open AI?
00:16:19.717 --> 00:16:21.097
Oh, one just announced.
00:16:21.332 --> 00:16:31.673
Like literally moments ago this is the much anticipated code name, strawberry reasoning model and thinking model that opening eyes bringing to market.
00:16:31.903 --> 00:16:34.273
I know you've had a chance to tinker with it a little bit.
00:16:34.373 --> 00:16:36.123
What are your early impressions?
00:16:36.513 --> 00:16:37.013
It's phenomenal.
00:16:37.712 --> 00:16:39.643
Just maybe like to explain to everyone.
00:16:39.643 --> 00:16:40.253
Basically.
00:16:40.658 --> 00:16:46.038
I think if you've got a plus subscription, you already have access where like you ask a model.
00:16:46.077 --> 00:16:52.057
And now, instead of just like blurting out a response, it does the chain of thought reasoning where basically it.
00:16:52.807 --> 00:16:57.138
Well, they decided not to show the full work, but think of if you're solving a math problem, right?
00:16:57.447 --> 00:16:58.057
On an exam,
00:16:58.327 --> 00:16:58.607
right?
00:16:59.077 --> 00:17:04.268
And you're writing and you're doing a lot of work and you go down some way down like one approach.
00:17:04.317 --> 00:17:05.097
It doesn't really work.
00:17:05.097 --> 00:17:05.948
You scratch that.
00:17:06.538 --> 00:17:08.657
You move these two pieces of paper aside.
00:17:08.778 --> 00:17:09.778
You try another thing.
00:17:09.867 --> 00:17:10.627
You get there.
00:17:11.087 --> 00:17:18.647
Then maybe if you've got time, you like write it up properly, like one, a four page and looks elegant, like you're like a genius just got there linearly.
00:17:18.768 --> 00:17:19.718
That's what's happening now.
00:17:19.718 --> 00:17:26.028
So if you ask it to like I asked it for like an M& A strategy for poly, which of our competitors should require, et cetera.
00:17:26.317 --> 00:17:32.913
And it really it shows Oh, look at competitors, looking at synergies, looking at technologies, looking at market coverage, and it only shows you this high level thing.
00:17:34.002 --> 00:17:34.363
Thanks.
00:17:34.413 --> 00:17:39.982
And really, this is test time compute, we're basically, if you ask me a question, I need to respond immediately.
00:17:40.413 --> 00:17:49.313
That's like, how good am I at being like quick witted and thinking, but I could also go away, which we never do for this podcast, but Hey and think about something for a while.
00:17:49.373 --> 00:17:54.883
Think of the best response, do some research, take a chain of thoughts down like one Avenue.
00:17:55.353 --> 00:17:57.323
And see that maybe it's not that interesting, right?
00:17:57.323 --> 00:17:59.742
And then I go back, backtrack, try something else.
00:18:00.192 --> 00:18:01.603
And then, see where the whole thing ends up.
00:18:01.603 --> 00:18:02.823
And maybe I'll just show you the work.
00:18:02.823 --> 00:18:04.603
So at the moment, they're not showing the whole work.
00:18:04.833 --> 00:18:07.913
It's actually very interesting to, to see how the model got there.
00:18:07.913 --> 00:18:16.192
And I think that people are excited about this because As people fear manipulation, especially when people fear the longer AGI style threats, right?
00:18:16.192 --> 00:18:36.423
They want to see what did the model decided to manipulate you in their response, or maybe what instruction it chose to disregard, but also like some people are saying it would be very interesting to see where the model goes and realizes that it's done something bad, goes back, decides not to do it, but then they say that they decided not to show it because that would actually.
00:18:37.292 --> 00:18:43.823
Basically give you a well visualize something that might be like in a violation of the policies, right?
00:18:44.303 --> 00:18:44.593
Sure.
00:18:45.042 --> 00:18:47.103
So, cause it's it's Transparency
00:18:47.103 --> 00:18:59.813
could actually be a little bit counterproductive in that it shows something that would violate safeguards or whatever, and ultimately probabilistically we moved away from this, but we're showing it to you.
00:18:59.823 --> 00:19:03.272
So you are now getting an odd impression that this is something that's being processed.
00:19:04.167 --> 00:19:04.327
Well,
00:19:04.327 --> 00:19:04.577
yeah.
00:19:04.577 --> 00:19:14.938
I think the more meaningful one, is just competitive advantage because if they show the whole thing, it really shows the inner workings of the reinforcement learning.
00:19:14.968 --> 00:19:16.038
Cause this is like test time.
00:19:16.038 --> 00:19:17.837
Reinforcement learning goes back and forth.
00:19:18.228 --> 00:19:19.998
It trades time for compute there.
00:19:20.198 --> 00:19:25.087
The other interesting thing that people should be aware of is now you're not just charged for the tokens produced.
00:19:25.498 --> 00:19:27.637
You're charged for all that scrap paper produced.
00:19:28.218 --> 00:19:30.567
As you really, yes.
00:19:30.657 --> 00:19:36.807
So I'm not, I haven't gotten to the bottom of just like exactly how, they're going to tell you what's been used and their limitations.
00:19:36.807 --> 00:19:38.458
Maybe their configuration options.
00:19:38.478 --> 00:19:39.367
I've not seen them.
00:19:39.887 --> 00:19:45.167
But you'll be charged for more than it's like saying, Hey, if you're working with lawyers, go do this for me.
00:19:45.167 --> 00:19:46.538
Well, how long did you spend on it?
00:19:46.667 --> 00:19:49.067
Like four hours or eight hours.
00:19:49.807 --> 00:19:51.117
You're not buying the actual output.
00:19:51.127 --> 00:19:53.178
You're buying the time and similar process here.
00:19:53.288 --> 00:19:53.788
Exactly.
00:19:53.867 --> 00:19:57.758
So here there are these invisible tokens that you can't see, but you will be charged for it.
00:19:57.768 --> 00:20:00.077
So we'll see how the market reacts to that.
00:20:00.077 --> 00:20:01.887
But like the model is incredible, right?
00:20:01.928 --> 00:20:02.208
Yeah.
00:20:02.307 --> 00:20:08.317
But you have the stats I saw was that it performs at the 49th percentile of the IOI international Olympiad in informatics.
00:20:10.327 --> 00:20:12.778
I spent like much of my high school trying to qualify for that.
00:20:12.807 --> 00:20:14.807
And I I forgot to okay.
00:20:14.978 --> 00:20:17.097
Revealing childhood trauma is part of the podcast.
00:20:17.488 --> 00:20:23.458
I forgot to print an answer module or something, which puts me down from being second to fifth.
00:20:23.458 --> 00:20:24.557
And I never got to the IOI.
00:20:25.298 --> 00:20:27.258
But now it can be, it's fine.
00:20:27.288 --> 00:20:29.028
Cause now we're doing an
00:20:29.028 --> 00:20:29.938
entire episode on
00:20:29.938 --> 00:20:30.157
that
00:20:30.157 --> 00:20:30.448
moment.
00:20:30.468 --> 00:20:31.377
We need to
00:20:31.377 --> 00:20:31.948
dissect it.
00:20:32.188 --> 00:20:32.428
Yeah.
00:20:32.428 --> 00:20:32.597
Yeah.
00:20:32.597 --> 00:20:32.778
Yeah.
00:20:32.778 --> 00:20:33.028
Yeah.
00:20:33.087 --> 00:20:33.647
Counseling.
00:20:33.647 --> 00:20:44.327
But, um, at this point, an AI model is like roughly half among the pack, like the best four people from every country in the world sit down and code, but these are high school students, right?
00:20:44.337 --> 00:20:45.807
So up to the age of 18 ish.
00:20:46.278 --> 00:20:48.708
We don't have an AI model that in 16, 17 seconds.
00:20:49.198 --> 00:20:55.607
Can blurt out like code solution for things that these kids sit for five hours and solve these three problems.
00:20:55.657 --> 00:21:02.288
So, if someone had told me that even as two years ago, I'm like GPT 3, 3.
00:21:02.288 --> 00:21:04.518
5, it wasn't clear that it was going to go that fast.
00:21:04.817 --> 00:21:15.357
Although coding and the other thing that if you're prepared for these kinds of competitions, uh, you realize after a while is that There is a lot of like rote learning of things.
00:21:15.688 --> 00:21:25.988
What people do at those competitions is, there's 30 minutes to the start, everyone sits down in front of a workstation, not your workstation, because you can't use your resources and people just start writing scripts, right?
00:21:26.347 --> 00:21:31.038
Like they're like, I'm going to write, this like maximum flow algorithm, because I'm probably going to need it in one of the problems.
00:21:31.038 --> 00:21:35.417
Or I'm going to write like whatever, a complex graph algorithm, because I might just need it.
00:21:35.417 --> 00:21:42.708
So, the truth is when you train on these things and these solutions, it is still very, well, it's fascinating.
00:21:42.708 --> 00:21:44.057
The language model can fit it.
00:21:45.067 --> 00:21:47.087
Together in a way that it will work.
00:21:47.097 --> 00:21:56.847
It is incredible that it can do it at that level, but it's also not surprising that it can produce these things because it's probably seen a lot of them in training.
00:21:57.018 --> 00:22:09.657
It might've seen so many micro variations of these problems that it just goes someone who's lived forever and seen like 10, 000 versions of the problem, I think that if you're setting those tasks, like as the commission of.
00:22:10.117 --> 00:22:13.438
Like IOI, like you run out of ideas, right?
00:22:13.597 --> 00:22:14.157
So it's hard.
00:22:14.208 --> 00:22:15.877
It's exhausted at some point.
00:22:15.877 --> 00:22:16.188
So.
00:22:17.282 --> 00:22:20.663
It's not surprising, but it's also extremely surprising that you can do it.
00:22:20.712 --> 00:22:21.792
It's it's incredible.
00:22:21.792 --> 00:22:26.423
But like back to this chain of thought reasoning the number of these steps and sequences is increasing.
00:22:26.462 --> 00:22:28.803
And like this is not GPT 5, right?
00:22:28.803 --> 00:22:30.133
This is basically the 4.
00:22:30.623 --> 00:22:31.673
5 moment, right?
00:22:31.673 --> 00:22:31.742
I
00:22:31.772 --> 00:22:32.423
digress then.
00:22:32.492 --> 00:22:33.182
I digress.
00:22:33.732 --> 00:22:36.682
Well, they've decided not 5 for what it's worth, right?
00:22:36.863 --> 00:22:37.782
I don't even think it's GPT.
00:22:37.782 --> 00:22:39.198
I think it's OpenAI's 01.
00:22:39.557 --> 00:22:39.817
Yeah.
00:22:39.948 --> 00:22:40.807
Oh, yeah.
00:22:41.167 --> 00:22:41.508
I don't know.
00:22:41.557 --> 00:22:42.708
But it's it's really cool.
00:22:42.807 --> 00:22:46.048
Like it, like showing that, like both sequence of reason.
00:22:46.048 --> 00:22:46.798
I love that visibility.
00:22:46.798 --> 00:22:46.807
I
00:22:46.837 --> 00:22:50.478
think it's a step towards that responsible AI movement as well, which is.
00:22:50.478 --> 00:22:51.038
No, people are
00:22:51.038 --> 00:22:52.958
attacking it because it doesn't show the full thing.
00:22:53.067 --> 00:22:53.327
Sure.
00:22:53.337 --> 00:22:54.518
Because they would love to see it.
00:22:54.627 --> 00:22:55.758
But Haters gonna hate.
00:22:56.768 --> 00:22:57.238
Absolutely.
00:22:57.258 --> 00:23:03.587
Look, I think that there's been a lot of um, some would say that like the two, three companies have separated the top.
00:23:03.617 --> 00:23:03.987
That's.
00:23:04.347 --> 00:23:04.917
Probably true.
00:23:04.968 --> 00:23:05.167
Yeah.
00:23:05.198 --> 00:23:08.137
But a lot of, we're saying that OpenAI has lost its edge.
00:23:08.887 --> 00:23:09.557
Not today.
00:23:10.067 --> 00:23:10.637
Not today.
00:23:11.788 --> 00:23:12.087
All right.
00:23:12.218 --> 00:23:12.617
Awesome.
00:23:12.917 --> 00:23:21.018
Well, thanks for recapping some big happenings in in the AI news world, and it was a pleasure as always to chat with you, I'm going to look into the international Olympiad of informatics,
00:23:21.018 --> 00:23:24.457
but
00:23:24.458 --> 00:23:26.228
that was a nice little glimpse into your past.
00:23:26.238 --> 00:23:26.597
It's almost
00:23:26.887 --> 00:23:27.147
funny.
00:23:27.157 --> 00:23:28.317
Did they name it after?
00:23:28.528 --> 00:23:32.678
Like the visa that all tech entrepreneurs go for when they go to America.
00:23:33.087 --> 00:23:33.577
I don't know.
00:23:33.698 --> 00:23:34.127
Did they,
00:23:34.377 --> 00:23:36.538
which came first, which was the first IOI?
00:23:36.657 --> 00:23:37.397
Anyway, but I appreciate it.
00:23:37.397 --> 00:23:37.968
I like to do it.
00:23:37.968 --> 00:23:38.446
I do one.
00:23:38.497 --> 00:23:38.856
Oh, one.
00:23:38.856 --> 00:23:39.106
Yeah.
00:23:39.217 --> 00:23:39.487
Okay.
00:23:40.027 --> 00:23:43.906
Well, it was a pleasure as always a lot of fun to chat with you and hopefully we'll catch you next week.
00:23:43.957 --> 00:23:44.676
Thanks for joining us.
00:23:45.007 --> 00:23:46.156
See you, Damien.
00:00:00.000 --> 00:00:04.950
As people fear manipulation, especially when people fear the longer AGI style threats, right?
00:00:04.950 --> 00:00:18.359
They want to see what did the model decided to manipulate you in their response, or maybe what instruction it chose to disregard, but also like some people are saying it would be very interesting to see where the model goes and realizes that it's done something bad
00:00:26.774 --> 00:00:30.504
Hello and welcome back to another episode of deep learning with PolyAI.
00:00:30.815 --> 00:00:35.145
The only podcast that sends you on a wild rollercoaster every time we kick off.
00:00:35.185 --> 00:00:36.045
I'm your host, Damien.
00:00:36.064 --> 00:00:37.765
I'm here as usual.
00:00:37.945 --> 00:00:40.365
Was CEO and co founder of Poly AI, Nikola Mrkšić.
00:00:40.645 --> 00:00:41.015
Hey, Nikola.
00:00:41.034 --> 00:00:41.384
How are you?
00:00:41.734 --> 00:00:42.284
Good to see you.
00:00:42.335 --> 00:00:42.755
I'm good.
00:00:42.755 --> 00:00:43.085
I'm good.
00:00:43.085 --> 00:00:44.115
Fresh off of a red eye.
00:00:44.155 --> 00:00:44.634
How are you?
00:00:44.634 --> 00:00:46.335
Yeah, you just came back from Montreal, correct?
00:00:46.484 --> 00:00:46.575
Yeah.
00:00:46.576 --> 00:00:47.104
Yeah.
00:00:47.104 --> 00:00:47.814
Oh, wonderful.
00:00:47.825 --> 00:00:49.314
Another conference in the bag.
00:00:49.335 --> 00:00:51.954
Well, a lot has happened this week while you were in Montreal.
00:00:51.954 --> 00:01:03.625
So I thought we'd do an episode on the AI news and review to talk about some of the key headlines that are making their way around some of the big names in the AI space or practitioners and deployers of AI and AI systems.
00:01:04.125 --> 00:01:08.805
We'll go through those fairly quickly because I want to finish on probably the biggest story.
00:01:08.814 --> 00:01:10.034
And that's something that happened.
00:01:10.484 --> 00:01:18.474
Well, I want to say probably about 25 minutes ago was announced, which is which is open AI rolling out a one or GPT five and.
00:01:18.575 --> 00:01:20.114
Previous circles, also strawberry.
00:01:20.364 --> 00:01:22.305
So we'll talk a little bit about your early impressions.
00:01:23.635 --> 00:01:25.594
This is what they would refer to as GPT five.
00:01:25.594 --> 00:01:26.275
So I think,
00:01:26.344 --> 00:01:26.775
Oh, good.
00:01:26.775 --> 00:01:33.155
Well, you could correct me then when we get to that point, but definitely want to save some time to talk about strawberry slash Oh one slash.
00:01:33.375 --> 00:01:34.965
Definitely not GPT five.
00:01:35.334 --> 00:01:47.045
Want to kick off by bringing up the CEO of Klarna in the news again, knew he made ripples when he said he was cutting 50 percent of his support staff when it came to some AI automation.
00:01:47.545 --> 00:02:01.224
And now he's basically telling Salesforce and Workday as their CRM and ERP systems goodbye, as he moves toward this, what seems to be a monolithic AI system, my question to you is.
00:02:01.549 --> 00:02:03.879
Is this just a lot of marketing flair?
00:02:03.879 --> 00:02:16.439
I, is it possible for one singular nebulous AI based system to operate an enterprise, or is this just something that's business as usual when it comes to furthering automation and it's got a heavy marketing spin on it.
00:02:17.530 --> 00:02:19.270
I don't doubt the marketing spin at all.
00:02:19.280 --> 00:02:23.689
I think that what was the other thing that they're not using Salesforce, they're building everything in house.
00:02:24.030 --> 00:02:27.830
Like we deal with buyers and clients who have this attitude and they do everything at house.
00:02:27.840 --> 00:02:35.919
So, I am not surprised that developers at Klarna can build a CRM for their own needs, whether that's the best use of their time, that's up to them.
00:02:36.349 --> 00:02:38.650
And I don't really have a strong opinion.
00:02:38.860 --> 00:02:40.370
Or care enough about far now.
00:02:40.400 --> 00:02:43.409
I don't think that it's actually all that relevant.
00:02:43.650 --> 00:02:46.360
It does sound like they're leaning in heavily and drinking their own Kool Aid.
00:02:46.360 --> 00:02:46.659
So
00:02:47.219 --> 00:02:57.070
is this something that theoretically, when it comes to building, is this something that probably be more ideal for a less complex, maybe a smaller midsize enterprise, That may be more approachable.
00:02:57.070 --> 00:03:08.491
It seems like a pretty daunting task even with the resources that Klarna may or may not have for an enterprise with the complexity and size that they do to just strike these fairly significant backend systems.
00:03:08.531 --> 00:03:09.320
So what was it?
00:03:09.350 --> 00:03:10.711
Salesforce and Workday
00:03:10.950 --> 00:03:11.610
for ERP.
00:03:12.360 --> 00:03:18.640
Look, given the nature of their business, which is like to Well, provide like loans for buying, moderately expensive things.
00:03:18.711 --> 00:03:18.971
Sure.
00:03:18.971 --> 00:03:26.591
I believe that they have a system of record where they can, track the products they're sharing and stuff and that they don't need Salesforce or given the nature of what they do.
00:03:26.591 --> 00:03:29.181
I'm almost surprised that they, I don't know how they were using it.
00:03:29.181 --> 00:03:32.121
So it does seem and feel a bit bombastic, right?
00:03:33.300 --> 00:03:43.161
I think, the previous announcement around automating support with with LLMs, given the nature of their client base, which is like younger tech savvy people, right?
00:03:43.221 --> 00:03:45.110
Those are people you can't push into those channels.
00:03:45.121 --> 00:03:47.001
So if you do it deliberately.
00:03:47.390 --> 00:03:48.200
Fair game, right?
00:03:48.450 --> 00:03:53.480
We speak to a lot of companies and, there are challenger banks that we work with and those that we don't, right?
00:03:53.610 --> 00:03:59.700
For example, Revolut is deeply focused on, well, decided not to have voice as part of their offering.
00:04:00.320 --> 00:04:03.401
Amazon doesn't really give you an ability to call Amazon either.
00:04:03.401 --> 00:04:12.526
And they just push you into digital channels and they're rich enough to just send you a replacement in case that it seems likely that you're not trying to You know, get a free item.
00:04:12.586 --> 00:04:15.336
So, these are, I think, design decisions at the top end.
00:04:15.346 --> 00:04:15.996
And I think that.
00:04:16.586 --> 00:04:20.326
They clearly did well enough for the first announcement that they're getting hooked on it.
00:04:20.326 --> 00:04:21.415
So good for them.
00:04:21.435 --> 00:04:24.865
If rebuilding an ERP on the CRM is the best use of their time.
00:04:25.896 --> 00:04:26.625
I'm happy for them.
00:04:27.216 --> 00:04:27.516
Yeah.
00:04:28.115 --> 00:04:28.396
All right.
00:04:28.435 --> 00:04:28.935
Fair enough.
00:04:29.125 --> 00:04:30.326
Klarna, you're in the past.
00:04:30.466 --> 00:04:32.196
Apple intelligence coming up next.
00:04:32.216 --> 00:04:41.415
Obviously the big Apple product showcase that occurred in the past week or so announcing basically an entire device ecosystem built around.
00:04:41.750 --> 00:04:47.670
Apple intelligence, which is, which was previously billed as they're on device AI approach.
00:04:47.680 --> 00:05:01.011
And it, there was a little bit of blending with, their previous iCloud offering and now moving into Apple intelligence, which because of the compute needs to happen on devices, which is why they've been built from the ground up.
00:05:01.560 --> 00:05:06.690
They have this private cloud system that looks to blend the two.
00:05:07.391 --> 00:05:25.221
I'm curious with iCloud being their pivotal go to and trying to onboard all Apple users into iCloud and now moving more, I would say intimately to the devices itself with Apple intelligence is this a future where those two approaches, those two go to markets like live together, peacefully.
00:05:25.711 --> 00:05:26.060
I think
00:05:26.060 --> 00:05:26.401
so.
00:05:26.451 --> 00:05:34.120
I think that it is inevitable with larger models and these compute needs and that you will have part of it on device.
00:05:34.190 --> 00:05:36.331
That was always going to be the case in one form or another.
00:05:36.331 --> 00:05:37.540
And you can trade it off, right?
00:05:37.940 --> 00:05:41.281
Previously, you could process speech recognition online or do it on the phone.
00:05:41.740 --> 00:05:45.440
And both have implications in battery life, performance, size of model.
00:05:45.841 --> 00:05:46.440
They still do.
00:05:46.940 --> 00:05:55.930
I think like with more and more features that are based on LLMs you're going to need like a faster phone with a better CPU, GPU, right?
00:05:57.161 --> 00:05:58.331
You're going to need more storage.
00:05:58.331 --> 00:06:04.430
I don't know about you, but my 256 gigabyte iPhone 14 right now feels like a piece of ancient history that I tore up.
00:06:05.771 --> 00:06:07.860
Honestly, I am out of space constantly.
00:06:07.860 --> 00:06:09.901
And the next one I buy is definitely going to be like a terabyte.
00:06:10.451 --> 00:06:13.360
So, I think it's a good way for them to.
00:06:13.591 --> 00:06:15.651
Create a need for people to upgrade their phones.
00:06:15.971 --> 00:06:16.891
I can't wait to get one.
00:06:16.930 --> 00:06:18.310
So I'm excited about everything.
00:06:18.321 --> 00:06:20.521
And, they're pushing a lot of it into Siri.
00:06:20.680 --> 00:06:21.600
Someone who used to work there.
00:06:21.600 --> 00:06:25.401
I'm really excited to see all the ways in which Siri will improve.
00:06:25.781 --> 00:06:27.500
Big believer in Apple in general.
00:06:27.500 --> 00:06:32.060
I think that no company really treats its product with as much like.
00:06:32.896 --> 00:06:35.216
Respect and love and care as they do.
00:06:35.365 --> 00:06:35.675
Yeah.
00:06:36.036 --> 00:06:45.875
So, like when it comes to privacy and designing the right approach around giving you all that while giving you the best that you can have in terms of these features.
00:06:46.341 --> 00:06:47.120
I'm really excited.
00:06:47.120 --> 00:06:50.031
The whole demo was like explaining images, right?
00:06:50.071 --> 00:06:54.430
Search over one place where I would still go out at the Apple ecosystem always was Google photos.
00:06:54.430 --> 00:06:55.500
I don't know about you, right?
00:06:55.990 --> 00:07:04.430
But if you're looking for like expenses, and you have 50 things to submit, like Google photos tends to be better at searching, right?
00:07:04.680 --> 00:07:05.141
Oh, sure.
00:07:05.290 --> 00:07:05.560
Yeah.
00:07:05.750 --> 00:07:12.050
If I'm dealing with like paperwork and I've taken a photo, I know I have, but I can't find it like Google photos have.
00:07:12.730 --> 00:07:17.911
To this point, been a lot better at like vaguely describing the thing and it being found in my photos.
00:07:17.911 --> 00:07:21.761
So I'm excited for Apple to catch up with that without giving all my data to Google.
00:07:21.781 --> 00:07:23.021
So, it's cool.
00:07:23.081 --> 00:07:23.790
I'm really excited.
00:07:23.790 --> 00:07:26.019
Yeah.
00:07:26.100 --> 00:07:27.040
Oh, I definitely will.
00:07:27.071 --> 00:07:30.040
I'm typically on a three a three model cycle.
00:07:30.321 --> 00:07:31.540
But I'm sitting on a.
00:07:32.130 --> 00:07:34.370
I think at 13, so I'm ready to move up.
00:07:34.761 --> 00:07:36.420
It just go so happen to coincide.
00:07:36.620 --> 00:07:50.091
The privacy thing is a huge issue and that's always been one of the flagship tenants of Apple is, allowing for the utmost of consumer privacy, even going so far as to blocking, federal and governmental intrusion for some devices.
00:07:50.531 --> 00:07:52.290
And this definitely plays into that.
00:07:52.815 --> 00:08:01.826
What does it say when they're pushing the safety and security of on device privacy again, going back to the cloud before, are they basically damning their previous efforts?
00:08:03.505 --> 00:08:11.516
I'm just saying like they've always billed that iCloud was very private and secure as well, but they're saying you get added security when you move to on device.
00:08:11.615 --> 00:08:14.295
If you're saying that, are you basically, I
00:08:14.295 --> 00:08:16.795
don't think that's what they're saying.
00:08:16.795 --> 00:08:23.596
I think for the family of things and tasks that you're going to need to be on device for it is safe.
00:08:23.605 --> 00:08:24.795
I think maybe.
00:08:25.415 --> 00:08:31.555
What it may be is a bit of a dig at other services where, if you're using chat GPT, they can use your data to train, right?
00:08:31.656 --> 00:08:39.796
I'm not exactly sure what's in the terms and conditions of Android devices with Gemini and stuff, but I feel like it probably is angled more in that direction.
00:08:40.441 --> 00:08:44.071
Like war than specifically against iCloud per se.
00:08:45.230 --> 00:08:45.890
Yeah, that's fair.
00:08:45.931 --> 00:08:46.390
That's fair.
00:08:46.390 --> 00:08:56.390
So, so arguably saying of these two behemoths of the two like dominant forces and, portable consumer devices, like we're going to be more secure and more respectful of your private data.
00:08:56.971 --> 00:08:59.539
Yeah, I think that's always a fair assumption when it comes to this too.
00:08:59.811 --> 00:09:00.681
Yeah, I agree.
00:09:01.140 --> 00:09:12.870
Speaking of your history with Siri you were among the teams that you were among one of the first teams to pioneer what has become largely attributed as the first consumer voice assistant to reach.
00:09:12.870 --> 00:09:14.041
We were only
00:09:14.041 --> 00:09:14.821
acquired into
00:09:14.821 --> 00:09:15.211
Siri.
00:09:15.211 --> 00:09:15.921
You know what?
00:09:15.921 --> 00:09:17.041
We're the badge with pride.
00:09:17.041 --> 00:09:17.280
Nickel.
00:09:17.280 --> 00:09:19.510
I'm still going to tell everybody you were on one of the first teams.
00:09:19.591 --> 00:09:21.490
But you obviously had an intimate.
00:09:21.600 --> 00:09:26.400
Involvement with the research and development that went into what became Siri.
00:09:27.630 --> 00:09:33.640
This is the first probably truly transformational iteration of the voice assistant since its inception.
00:09:33.640 --> 00:09:37.451
So what are you looking forward to seeing what this next generation of Siri brings?
00:09:38.201 --> 00:09:39.211
It's a tough one, right?
00:09:39.260 --> 00:09:44.341
I can't really talk much about the history of Siri while I was there, but it was the first.
00:09:44.860 --> 00:09:50.260
Of them, which meant that, um, the mistakes that could have been made were made first date.
00:09:50.311 --> 00:09:54.890
I think just maybe to redirect to a company that it didn't work out where I can talk a bit more freely.
00:09:56.081 --> 00:10:01.010
No, I think when Amazon would Alexa hit the gold vein of like smart speakers in homes
00:10:01.010 --> 00:10:01.980
and
00:10:01.980 --> 00:10:08.806
then went into a big push there, I think they actually built a system clunkier than Siri, cause they were really like mass producing it.
00:10:09.015 --> 00:10:15.416
It felt the Soviets have 14, 000 tanks and then they lose the cold war without a bullet fire because their economy crumbles.
00:10:15.416 --> 00:10:17.355
Like Alexa was built very quickly.
00:10:17.441 --> 00:10:18.265
To be complex.
00:10:18.265 --> 00:10:19.639
And it was very complex.
00:10:19.640 --> 00:10:22.760
So the technical debt of those systems is hard to address.
00:10:22.760 --> 00:10:30.740
And now when you basically change the paradigm so much around, LLMs doing what you previously had to have very complex ontologies for.
00:10:30.740 --> 00:10:35.850
So, think of like different applications and domains and things that have to be built up one by one.
00:10:35.900 --> 00:10:37.841
And very complex controllable things.
00:10:37.841 --> 00:10:40.061
And, with an LLM, you can't really relinquish control.
00:10:40.061 --> 00:10:51.206
So, I think it really remains to be seen what's, what's Will have been, it will not be much like our generative deployments have not been a straight up, like here's a prompt and run an LLM.
00:10:51.206 --> 00:11:00.716
If you have different requirements, both like in terms of controllability, understanding safety, et cetera, you can't just go and throw an LLM at it.
00:11:00.865 --> 00:11:06.245
So I think what I'm really interested in seeing is just like how LMS will have been used to change the output, right?
00:11:06.245 --> 00:11:10.535
I think the general conversational ability, and it depends, right?
00:11:10.586 --> 00:11:16.296
Apple's always been very opinionated around, what's your, you shouldn't say where, how far to go in different places.
00:11:16.306 --> 00:11:17.475
It's a hard thing to get right.
00:11:17.846 --> 00:11:30.056
Cause you're like done if you're doing them, if you don't, in terms of trying to go on like a philosophical debate with someone or taking a joke too far and stuff, so you can do more than they will dare to do simply because of who they are.
00:11:30.056 --> 00:11:32.091
They can do a lot more than they will dare to do.
00:11:32.140 --> 00:11:33.701
So it'll be exciting nonetheless.
00:11:33.730 --> 00:11:50.671
But I think historically the only step that you can take is up when it comes to home, voice assistants or smart speakers, historically, all those limitations, as far as when it hits the consumer level, it causes a lot of frustration because of some of those safeguards that are put in place.
00:11:50.681 --> 00:11:51.671
So maybe.
00:11:51.686 --> 00:11:54.265
I think it's people are very critical of all of them.
00:11:54.316 --> 00:12:00.926
Apart from a very small minority that really loves them the way they are in their power users and know exactly how to use them.
00:12:00.956 --> 00:12:01.416
Exactly.
00:12:01.436 --> 00:12:01.625
But the
00:12:01.625 --> 00:12:12.696
truth is, I think the real thing, and the reason why it would probably I created a company that's focused on enterprise use cases is that I think there's a stronger need there because there are repeatable things that happen at scale over the phone, right?
00:12:13.046 --> 00:12:17.875
The problem with consumer voice assistance is that like everyone imagines a different thing.
00:12:18.201 --> 00:12:26.681
We all have this fuzzy understanding of what this future Jarvis, mega, Jarvis style, mega powerful voice assistant will be.
00:12:26.701 --> 00:12:28.971
What do you really need it to do?
00:12:29.730 --> 00:12:30.610
Exactly.
00:12:30.650 --> 00:12:31.791
That's the thing is that I've always,
00:12:31.811 --> 00:12:36.461
Like whenever the inevitable end of, I've had this conversation many times in the last few weeks, right?
00:12:36.931 --> 00:12:39.571
The inevitable conclusion is, well, she know everything about me.
00:12:39.801 --> 00:12:47.750
And then as it books a trip for me and my family, she know that, in our itinerary, my life, my, my wife loves Thai food and I love Japanese food.
00:12:47.750 --> 00:12:49.691
So come on, man, you wouldn't trust.
00:12:49.931 --> 00:12:54.490
Like a human being to do right, you're probably going right.
00:12:54.490 --> 00:12:56.441
So people have a lot of nebulous.
00:12:56.816 --> 00:13:02.235
Desires that kind of like point towards I want this or that, and they'll never be happy with the consumer voice.
00:13:02.245 --> 00:13:04.806
So I think it's just, it's difficult, right?
00:13:04.836 --> 00:13:08.355
Like it might be an improvement that is like 10 X and you'll still have people.
00:13:08.855 --> 00:13:14.135
Who are expecting the Jarvis, expecting the C3PO, expecting the fully interactive, like 100 percent of the
00:13:14.135 --> 00:13:14.905
integrations.
00:13:14.916 --> 00:13:16.235
And then we go back to.
00:13:16.625 --> 00:13:18.576
The classic what would you trust it to do?
00:13:18.586 --> 00:13:19.355
What permissions?
00:13:19.416 --> 00:13:19.676
Yeah.
00:13:20.186 --> 00:13:21.485
How angry will you be when it gets?
00:13:21.566 --> 00:13:21.885
It's the
00:13:21.885 --> 00:13:24.796
same people that are like, but I don't want it to know any of my information too.
00:13:24.846 --> 00:13:28.855
It's like you should know me, but I'm not telling you anything about me and I'm preventing you from, yeah.
00:13:28.855 --> 00:13:32.785
Well, having settled that one day, when like one of these companies goes there.
00:13:33.306 --> 00:13:35.416
For the people that want both Apple will still be the best.
00:13:35.495 --> 00:13:39.515
I sound like a, an Apple app, but absolutely.
00:13:41.597 --> 00:13:42.268
To the core.
00:13:42.327 --> 00:13:42.817
Thank you.
00:13:42.837 --> 00:13:43.437
Dad joke.
00:13:43.988 --> 00:13:44.278
All right.
00:13:44.307 --> 00:13:44.557
Cool.
00:13:44.577 --> 00:13:47.357
Well, excited to see what's in store for the next iteration.
00:13:47.357 --> 00:13:49.607
And I'm glad that you and I are going to get the same model phone.
00:13:49.687 --> 00:13:51.727
We can sit here and geek out over it together.
00:13:52.028 --> 00:13:52.307
Okay.
00:13:52.408 --> 00:13:56.337
Moving on Sam Altman backs XO watt.
00:13:56.628 --> 00:13:59.557
It's a solar storage company.
00:13:59.577 --> 00:14:24.677
They raise a 20 million seed round, obviously clean storage, minimal carbon footprint impact, climate friendly energy solutions have always been top of mind, well, at least in the past decade or so more so when the compute Strain has hit because of these LLMs that are out there using tremendous amounts of processing power.
00:14:25.217 --> 00:14:30.878
Is he buying into his own pockets a little bit with this move?
00:14:31.207 --> 00:14:32.087
Look, can you blame him?
00:14:32.138 --> 00:14:32.687
I dunno.
00:14:32.788 --> 00:14:34.947
The truth is yeah, like we need a lot more computer.
00:14:34.947 --> 00:14:41.288
I think the fact that we need so much more compute now probably makes the development of a lot of these technologies a greater priority.
00:14:41.378 --> 00:14:42.727
You could just do nuclear, right?
00:14:42.727 --> 00:14:52.418
I think I think there are videos of Elon circulating where he's talking about, we could like power everything we have, we need on the planet with enough solar and the right battery technology.
00:14:52.768 --> 00:14:55.067
So yeah, I think people are excited and they'll work on it, right?
00:14:55.168 --> 00:14:58.758
Like whether it's solar, nuclear, depending on how you feel about it, right?
00:14:58.847 --> 00:15:02.148
I think everyone's like gearing up for like more and more, right?
00:15:02.447 --> 00:15:03.298
In terms of compute.
00:15:03.298 --> 00:15:04.807
So like energy is a
00:15:05.673 --> 00:15:13.832
But we've also talked about how there's also a race to reduce the resources by building efficiencies and, progressive models, who's going to win.
00:15:14.273 --> 00:15:15.482
No there's no winning, right?
00:15:15.482 --> 00:15:24.482
Like in, in building like you can always increase capacity, model size, training, quality of chips, speed, to train yet a bigger model.
00:15:24.602 --> 00:15:30.092
Of course I think innovation in terms of like high quality synthetic data, better architectures will.
00:15:31.278 --> 00:15:33.577
Like this, like two axis along which you can progress.
00:15:33.587 --> 00:15:35.538
And they both amplify the end outcome.
00:15:35.587 --> 00:15:37.128
I remember being at a talk where.
00:15:37.707 --> 00:15:49.268
Ilya Sutskever talked about Hey, transformers just sped things up by a few years because with hardware developing LSTM cells three, four years later would have achieved the GPT 3 moment.
00:15:49.278 --> 00:15:56.827
So, yeah, if you have another, and I think there are many people proposing all kinds of architectures, some of which are pretty strong as alternatives to transformers.
00:15:56.837 --> 00:15:59.957
So that's just people will work and attack on both fronts.
00:15:59.977 --> 00:16:02.927
And, ideally we just continue to make progress from both.
00:16:04.057 --> 00:16:09.577
So everybody wins with this play, maybe a bit of a conflict of interest.
00:16:09.648 --> 00:16:11.548
We'll leave that to to progression down the road.
00:16:11.548 --> 00:16:11.927
Not our business.
00:16:11.927 --> 00:16:12.615
All
00:16:12.615 --> 00:16:13.302
right.
00:16:13.302 --> 00:16:14.677
Open AI.
00:16:14.677 --> 00:16:14.748
Can
00:16:14.748 --> 00:16:15.217
we buy in?
00:16:15.217 --> 00:16:15.988
That's my question.
00:16:17.748 --> 00:16:19.658
Is there a Kickstarter open AI?
00:16:19.717 --> 00:16:21.097
Oh, one just announced.
00:16:21.332 --> 00:16:31.673
Like literally moments ago this is the much anticipated code name, strawberry reasoning model and thinking model that opening eyes bringing to market.
00:16:31.903 --> 00:16:34.273
I know you've had a chance to tinker with it a little bit.
00:16:34.373 --> 00:16:36.123
What are your early impressions?
00:16:36.513 --> 00:16:37.013
It's phenomenal.
00:16:37.712 --> 00:16:39.643
Just maybe like to explain to everyone.
00:16:39.643 --> 00:16:40.253
Basically.
00:16:40.658 --> 00:16:46.038
I think if you've got a plus subscription, you already have access where like you ask a model.
00:16:46.077 --> 00:16:52.057
And now, instead of just like blurting out a response, it does the chain of thought reasoning where basically it.
00:16:52.807 --> 00:16:57.138
Well, they decided not to show the full work, but think of if you're solving a math problem, right?
00:16:57.447 --> 00:16:58.057
On an exam,
00:16:58.327 --> 00:16:58.607
right?
00:16:59.077 --> 00:17:04.268
And you're writing and you're doing a lot of work and you go down some way down like one approach.
00:17:04.317 --> 00:17:05.097
It doesn't really work.
00:17:05.097 --> 00:17:05.948
You scratch that.
00:17:06.538 --> 00:17:08.657
You move these two pieces of paper aside.
00:17:08.778 --> 00:17:09.778
You try another thing.
00:17:09.867 --> 00:17:10.627
You get there.
00:17:11.087 --> 00:17:18.647
Then maybe if you've got time, you like write it up properly, like one, a four page and looks elegant, like you're like a genius just got there linearly.
00:17:18.768 --> 00:17:19.718
That's what's happening now.
00:17:19.718 --> 00:17:26.028
So if you ask it to like I asked it for like an M& A strategy for poly, which of our competitors should require, et cetera.
00:17:26.317 --> 00:17:32.913
And it really it shows Oh, look at competitors, looking at synergies, looking at technologies, looking at market coverage, and it only shows you this high level thing.
00:17:34.002 --> 00:17:34.363
Thanks.
00:17:34.413 --> 00:17:39.982
And really, this is test time compute, we're basically, if you ask me a question, I need to respond immediately.
00:17:40.413 --> 00:17:49.313
That's like, how good am I at being like quick witted and thinking, but I could also go away, which we never do for this podcast, but Hey and think about something for a while.
00:17:49.373 --> 00:17:54.883
Think of the best response, do some research, take a chain of thoughts down like one Avenue.
00:17:55.353 --> 00:17:57.323
And see that maybe it's not that interesting, right?
00:17:57.323 --> 00:17:59.742
And then I go back, backtrack, try something else.
00:18:00.192 --> 00:18:01.603
And then, see where the whole thing ends up.
00:18:01.603 --> 00:18:02.823
And maybe I'll just show you the work.
00:18:02.823 --> 00:18:04.603
So at the moment, they're not showing the whole work.
00:18:04.833 --> 00:18:07.913
It's actually very interesting to, to see how the model got there.
00:18:07.913 --> 00:18:16.192
And I think that people are excited about this because As people fear manipulation, especially when people fear the longer AGI style threats, right?
00:18:16.192 --> 00:18:36.423
They want to see what did the model decided to manipulate you in their response, or maybe what instruction it chose to disregard, but also like some people are saying it would be very interesting to see where the model goes and realizes that it's done something bad, goes back, decides not to do it, but then they say that they decided not to show it because that would actually.
00:18:37.292 --> 00:18:43.823
Basically give you a well visualize something that might be like in a violation of the policies, right?
00:18:44.303 --> 00:18:44.593
Sure.
00:18:45.042 --> 00:18:47.103
So, cause it's it's Transparency
00:18:47.103 --> 00:18:59.813
could actually be a little bit counterproductive in that it shows something that would violate safeguards or whatever, and ultimately probabilistically we moved away from this, but we're showing it to you.
00:18:59.823 --> 00:19:03.272
So you are now getting an odd impression that this is something that's being processed.
00:19:04.167 --> 00:19:04.327
Well,
00:19:04.327 --> 00:19:04.577
yeah.
00:19:04.577 --> 00:19:14.938
I think the more meaningful one, is just competitive advantage because if they show the whole thing, it really shows the inner workings of the reinforcement learning.
00:19:14.968 --> 00:19:16.038
Cause this is like test time.
00:19:16.038 --> 00:19:17.837
Reinforcement learning goes back and forth.
00:19:18.228 --> 00:19:19.998
It trades time for compute there.
00:19:20.198 --> 00:19:25.087
The other interesting thing that people should be aware of is now you're not just charged for the tokens produced.
00:19:25.498 --> 00:19:27.637
You're charged for all that scrap paper produced.
00:19:28.218 --> 00:19:30.567
As you really, yes.
00:19:30.657 --> 00:19:36.807
So I'm not, I haven't gotten to the bottom of just like exactly how, they're going to tell you what's been used and their limitations.
00:19:36.807 --> 00:19:38.458
Maybe their configuration options.
00:19:38.478 --> 00:19:39.367
I've not seen them.
00:19:39.887 --> 00:19:45.167
But you'll be charged for more than it's like saying, Hey, if you're working with lawyers, go do this for me.
00:19:45.167 --> 00:19:46.538
Well, how long did you spend on it?
00:19:46.667 --> 00:19:49.067
Like four hours or eight hours.
00:19:49.807 --> 00:19:51.117
You're not buying the actual output.
00:19:51.127 --> 00:19:53.178
You're buying the time and similar process here.
00:19:53.288 --> 00:19:53.788
Exactly.
00:19:53.867 --> 00:19:57.758
So here there are these invisible tokens that you can't see, but you will be charged for it.
00:19:57.768 --> 00:20:00.077
So we'll see how the market reacts to that.
00:20:00.077 --> 00:20:01.887
But like the model is incredible, right?
00:20:01.928 --> 00:20:02.208
Yeah.
00:20:02.307 --> 00:20:08.317
But you have the stats I saw was that it performs at the 49th percentile of the IOI international Olympiad in informatics.
00:20:10.327 --> 00:20:12.778
I spent like much of my high school trying to qualify for that.
00:20:12.807 --> 00:20:14.807
And I I forgot to okay.
00:20:14.978 --> 00:20:17.097
Revealing childhood trauma is part of the podcast.
00:20:17.488 --> 00:20:23.458
I forgot to print an answer module or something, which puts me down from being second to fifth.
00:20:23.458 --> 00:20:24.557
And I never got to the IOI.
00:20:25.298 --> 00:20:27.258
But now it can be, it's fine.
00:20:27.288 --> 00:20:29.028
Cause now we're doing an
00:20:29.028 --> 00:20:29.938
entire episode on
00:20:29.938 --> 00:20:30.157
that
00:20:30.157 --> 00:20:30.448
moment.
00:20:30.468 --> 00:20:31.377
We need to
00:20:31.377 --> 00:20:31.948
dissect it.
00:20:32.188 --> 00:20:32.428
Yeah.
00:20:32.428 --> 00:20:32.597
Yeah.
00:20:32.597 --> 00:20:32.778
Yeah.
00:20:32.778 --> 00:20:33.028
Yeah.
00:20:33.087 --> 00:20:33.647
Counseling.
00:20:33.647 --> 00:20:44.327
But, um, at this point, an AI model is like roughly half among the pack, like the best four people from every country in the world sit down and code, but these are high school students, right?
00:20:44.337 --> 00:20:45.807
So up to the age of 18 ish.
00:20:46.278 --> 00:20:48.708
We don't have an AI model that in 16, 17 seconds.
00:20:49.198 --> 00:20:55.607
Can blurt out like code solution for things that these kids sit for five hours and solve these three problems.
00:20:55.657 --> 00:21:02.288
So, if someone had told me that even as two years ago, I'm like GPT 3, 3.
00:21:02.288 --> 00:21:04.518
5, it wasn't clear that it was going to go that fast.
00:21:04.817 --> 00:21:15.357
Although coding and the other thing that if you're prepared for these kinds of competitions, uh, you realize after a while is that There is a lot of like rote learning of things.
00:21:15.688 --> 00:21:25.988
What people do at those competitions is, there's 30 minutes to the start, everyone sits down in front of a workstation, not your workstation, because you can't use your resources and people just start writing scripts, right?
00:21:26.347 --> 00:21:31.038
Like they're like, I'm going to write, this like maximum flow algorithm, because I'm probably going to need it in one of the problems.
00:21:31.038 --> 00:21:35.417
Or I'm going to write like whatever, a complex graph algorithm, because I might just need it.
00:21:35.417 --> 00:21:42.708
So, the truth is when you train on these things and these solutions, it is still very, well, it's fascinating.
00:21:42.708 --> 00:21:44.057
The language model can fit it.
00:21:45.067 --> 00:21:47.087
Together in a way that it will work.
00:21:47.097 --> 00:21:56.847
It is incredible that it can do it at that level, but it's also not surprising that it can produce these things because it's probably seen a lot of them in training.
00:21:57.018 --> 00:22:09.657
It might've seen so many micro variations of these problems that it just goes someone who's lived forever and seen like 10, 000 versions of the problem, I think that if you're setting those tasks, like as the commission of.
00:22:10.117 --> 00:22:13.438
Like IOI, like you run out of ideas, right?
00:22:13.597 --> 00:22:14.157
So it's hard.
00:22:14.208 --> 00:22:15.877
It's exhausted at some point.
00:22:15.877 --> 00:22:16.188
So.
00:22:17.282 --> 00:22:20.663
It's not surprising, but it's also extremely surprising that you can do it.
00:22:20.712 --> 00:22:21.792
It's it's incredible.
00:22:21.792 --> 00:22:26.423
But like back to this chain of thought reasoning the number of these steps and sequences is increasing.
00:22:26.462 --> 00:22:28.803
And like this is not GPT 5, right?
00:22:28.803 --> 00:22:30.133
This is basically the 4.
00:22:30.623 --> 00:22:31.673
5 moment, right?
00:22:31.673 --> 00:22:31.742
I
00:22:31.772 --> 00:22:32.423
digress then.
00:22:32.492 --> 00:22:33.182
I digress.
00:22:33.732 --> 00:22:36.682
Well, they've decided not 5 for what it's worth, right?
00:22:36.863 --> 00:22:37.782
I don't even think it's GPT.
00:22:37.782 --> 00:22:39.198
I think it's OpenAI's 01.
00:22:39.557 --> 00:22:39.817
Yeah.
00:22:39.948 --> 00:22:40.807
Oh, yeah.
00:22:41.167 --> 00:22:41.508
I don't know.
00:22:41.557 --> 00:22:42.708
But it's it's really cool.
00:22:42.807 --> 00:22:46.048
Like it, like showing that, like both sequence of reason.
00:22:46.048 --> 00:22:46.798
I love that visibility.
00:22:46.798 --> 00:22:46.807
I
00:22:46.837 --> 00:22:50.478
think it's a step towards that responsible AI movement as well, which is.
00:22:50.478 --> 00:22:51.038
No, people are
00:22:51.038 --> 00:22:52.958
attacking it because it doesn't show the full thing.
00:22:53.067 --> 00:22:53.327
Sure.
00:22:53.337 --> 00:22:54.518
Because they would love to see it.
00:22:54.627 --> 00:22:55.758
But Haters gonna hate.
00:22:56.768 --> 00:22:57.238
Absolutely.
00:22:57.258 --> 00:23:03.587
Look, I think that there's been a lot of um, some would say that like the two, three companies have separated the top.
00:23:03.617 --> 00:23:03.987
That's.
00:23:04.347 --> 00:23:04.917
Probably true.
00:23:04.968 --> 00:23:05.167
Yeah.
00:23:05.198 --> 00:23:08.137
But a lot of, we're saying that OpenAI has lost its edge.
00:23:08.887 --> 00:23:09.557
Not today.
00:23:10.067 --> 00:23:10.637
Not today.
00:23:11.788 --> 00:23:12.087
All right.
00:23:12.218 --> 00:23:12.617
Awesome.
00:23:12.917 --> 00:23:21.018
Well, thanks for recapping some big happenings in in the AI news world, and it was a pleasure as always to chat with you, I'm going to look into the international Olympiad of informatics,
00:23:21.018 --> 00:23:24.457
but
00:23:24.458 --> 00:23:26.228
that was a nice little glimpse into your past.
00:23:26.238 --> 00:23:26.597
It's almost
00:23:26.887 --> 00:23:27.147
funny.
00:23:27.157 --> 00:23:28.317
Did they name it after?
00:23:28.528 --> 00:23:32.678
Like the visa that all tech entrepreneurs go for when they go to America.
00:23:33.087 --> 00:23:33.577
I don't know.
00:23:33.698 --> 00:23:34.127
Did they,
00:23:34.377 --> 00:23:36.538
which came first, which was the first IOI?
00:23:36.657 --> 00:23:37.397
Anyway, but I appreciate it.
00:23:37.397 --> 00:23:37.968
I like to do it.
00:23:37.968 --> 00:23:38.446
I do one.
00:23:38.497 --> 00:23:38.856
Oh, one.
00:23:38.856 --> 00:23:39.106
Yeah.
00:23:39.217 --> 00:23:39.487
Okay.
00:23:40.027 --> 00:23:43.906
Well, it was a pleasure as always a lot of fun to chat with you and hopefully we'll catch you next week.
00:23:43.957 --> 00:23:44.676
Thanks for joining us.
00:23:45.007 --> 00:23:46.156
See you, Damien.
