WEBVTT
00:00:02.560 --> 00:00:09.919
One of the more important AI models you've probably never heard of and likely won't use just got released.
00:00:10.080 --> 00:00:17.280
This may get slept on, but Inkling is a huge release from Thinking Machines Lab.
00:00:17.600 --> 00:00:25.920
So if you haven't heard of Thinking Machines, they're led by Mira Marathi, the former CTO of OpenAI.
00:00:26.160 --> 00:00:32.560
So why is Inkling maybe the most important AI model you probably won't use?
00:00:33.039 --> 00:00:45.520
Because it's now the best open source model from an American company, and Thinking Machines is betting on the future of AI, fine-tuning as a service.
00:00:45.840 --> 00:00:54.799
The model itself, though, it's multimodal, it's agentic, it's customizable, and it's available through enterprise distribution on day one.
00:00:55.039 --> 00:01:00.399
Inkling doesn't need to beat every Chinese model or Claude Fable to be relevant.
00:01:00.560 --> 00:01:12.000
It only needs to pique the interest of a few cost-conscious enterprises to not only be extremely profitable, but to also help shift the conversation around enterprise AI.
00:01:12.239 --> 00:01:22.799
Frontier models are becoming so powerful they can actually fine-tune smaller, practical models without much iteration or without even much expertise.
00:01:23.599 --> 00:01:27.680
So, has a new category re-emerged?
00:01:27.920 --> 00:01:29.519
Is fine-tuning back?
00:01:29.840 --> 00:01:36.079
And what might think machines' first major product mean for the broader AI competition?
00:01:36.480 --> 00:01:38.400
Well, here's the big picture.
00:01:38.640 --> 00:01:42.079
Inkling could change how enterprises buy AI.
00:01:42.319 --> 00:01:46.560
So this was just released Wednesday, so hours ago.
00:01:46.879 --> 00:01:51.760
And it is not the smartest model, but it is strategically important.
00:01:51.920 --> 00:01:58.799
I think it gives companies a credible American open weight model alternative to Chinese models.
00:01:58.879 --> 00:02:06.079
And inkling could be the best general purpose open model because it is multimodal, right?
00:02:06.159 --> 00:02:09.840
A lot of the open source, open weight Chinese models aren't.
00:02:10.000 --> 00:02:11.599
Many of them are text only.
00:02:11.759 --> 00:02:14.719
So Inkling, it is multimodal, it is a gentic.
00:02:15.439 --> 00:02:26.319
And as workflows in the real world lag so far behind model capabilities, I do think that there's a real market for bespoke middle-of-the-pack AI.
00:02:27.039 --> 00:02:30.159
So on today's show, here's what you're gonna learn.
00:02:30.240 --> 00:02:34.319
You're gonna learn why an open model matters, even if you're never gonna deploy it.
00:02:34.479 --> 00:02:40.400
You're gonna understand how Inkling reopens enterprise AI options beyond just Chinese open models.
00:02:40.560 --> 00:02:45.840
You're gonna know why frontier models could make fine-tuning practical beyond research teams.
00:02:46.000 --> 00:02:52.159
And you're gonna know how model shopping is gonna change budgets, vendors, and competitive leverage.
00:02:52.319 --> 00:02:53.520
All right, let's get to it.
00:02:53.680 --> 00:02:55.360
Welcome to Everyday AI.
00:02:55.520 --> 00:02:57.360
If you're new here, my name is Jordan Wilson.
00:02:57.439 --> 00:02:58.240
We do this every day.
00:02:58.400 --> 00:03:08.000
It's your daily live stream podcast and free daily newsletter helping business leaders like you and me keep up with the nonstop AI updates because my gosh, can't take an hour off.
00:03:08.159 --> 00:03:10.319
I help you decide what's important.
00:03:10.479 --> 00:03:14.000
I tell you how to use all this information to grow your company and your career.
00:03:14.080 --> 00:03:17.520
So it starts here with the unedited, unscripted live stream podcast.
00:03:17.599 --> 00:03:21.360
But make sure you go to our website at your everydayai.com.
00:03:21.439 --> 00:03:22.319
That's your cheat code.
00:03:22.479 --> 00:03:27.919
We're gonna be not just recapping the highlights from today's show, but go sign up for our free daily newsletter.
00:03:28.000 --> 00:03:32.879
We're gonna be re uh giving you everything else that you need to know that's happening in the world of AI today.
00:03:33.039 --> 00:03:36.879
Because yeah, it's one of those things you gotta like go and work it out.
00:03:36.960 --> 00:03:37.599
It's a muscle.
00:03:37.759 --> 00:03:38.879
Use it every single day.
00:03:39.120 --> 00:03:42.560
Uh, and yeah, all the AI news will be in our newsletter.
00:03:42.719 --> 00:03:43.759
All right, let's get into it.
00:03:43.919 --> 00:03:45.599
Live stream audience, good to see you.
00:03:45.680 --> 00:03:47.039
Uh, Adam joining us from St.
00:03:47.199 --> 00:03:47.280
St.
00:03:47.439 --> 00:03:54.000
Louis, Jose from Santiago, uh, Angie joining from Montana, Amiko, Tokyo, Brian, what's up, Brian?
00:03:54.080 --> 00:03:55.199
Joining from Minnesota.
00:03:55.360 --> 00:04:03.919
Uh, so uh, let's talk about maybe the most important AI model you'll probably never use.
00:04:04.240 --> 00:04:12.159
So there's a lot of different factors that have been compounding over the last, I would say, three months.
00:04:12.400 --> 00:04:26.160
And to put it in a very short uh summary, frontier models are probably for the most part too much for many enterprises, right?
00:04:26.319 --> 00:04:34.160
I'll say this I, you know, probably the Fortune 500, they can squeeze as much juice and uh, you know, make it worth the cost.
00:04:34.240 --> 00:04:43.920
But I'd say for many companies, especially those probably in the Fortune 5 or like the Fortune 501 to the 5000, right?
00:04:44.000 --> 00:04:45.600
So this isn't everyone.
00:04:45.759 --> 00:04:59.120
Um, you know, I don't think this changes the equation for every single company, every single business out there, because so many people are still gonna want, you know, their Chat GPT Enterprise, their, you know, Gemini, their Claude, their co-pilot.
00:04:59.199 --> 00:05:07.759
They want to make things easy, and they're not necessarily, you know, looking at, okay, how do open models or fine-tuning models change our strategy?
00:05:07.920 --> 00:05:10.240
But for so many in our audience, it does.
00:05:10.480 --> 00:05:13.040
And this is a really big deal.
00:05:13.279 --> 00:05:16.000
So here's a little bit more about the model itself.
00:05:16.240 --> 00:05:20.959
So, model, uh, the model is from Thinking Machines and it's called Inkling.
00:05:21.279 --> 00:05:26.079
It is a 975 billion parameter model, and it's multimodal.
00:05:26.160 --> 00:05:28.079
So, text, images, and audio.
00:05:28.240 --> 00:05:45.759
Uh, what is interesting as well, uh, previously Thinking Machines did demo kind of a similar dual purpose or uh you know two-way street um audio model as well, so that can kind of hear and listen as well as talk at the same time, like OpenAI's new GPT live.
00:05:45.920 --> 00:05:56.000
So Inkling is led by former OpenAI CTO Miro Marathi, and its quick arrival raises, I think, the next big question, which is how good is it?
00:05:56.160 --> 00:05:58.560
Well, here's from the company themselves from their release.
00:05:58.720 --> 00:06:07.439
They say our model called Inkling is a mixture of experts transformer with 975 billion total parameters, 41 billion active.
00:06:07.519 --> 00:06:10.720
It supports a context window of up to 1 million tokens.
00:06:10.879 --> 00:06:16.000
It was pre-trained on 45 trillion tokens of text, images, audio, and video.
00:06:16.079 --> 00:06:19.040
It is the first in a family of models of different sizes.
00:06:19.120 --> 00:06:34.480
Uh, alongside it, we are sharing a preview of Inkling Small, a lighter weight model with 12 billion active parameters, trained with a smaller recipe that achieves strong performance with even lower costs and latency.
00:06:34.720 --> 00:06:42.560
Inkling reasons natively over text, images, and audio and balances costs with performance through efficient and controllable thinking effort.
00:06:42.879 --> 00:06:48.800
We trained it to be a broad balanced foundation model, strong across many domains, flexible enough to adapt.
00:06:49.040 --> 00:06:54.160
Inkling is not the strongest over model overall model available today, open or closed.
00:06:54.319 --> 00:07:04.879
Instead, a combination of qualities makes it a good open weights base for customization, multimodal capabilities, efficient thinking, and availability on Tinker for fine tuning.
00:07:05.120 --> 00:07:08.560
Inkling is just the start, our first release in a model family.
00:07:08.639 --> 00:07:09.920
We will continue to build on.
00:07:10.079 --> 00:07:13.519
We want to make customization accessible for more use cases.
00:07:13.600 --> 00:07:17.519
So Inkling is available for fine tuning on Tinker today.
00:07:17.680 --> 00:07:26.160
Uh, picking the right base model to fine-tune is a qualitative judgment that combines measurable benchmarks with a unique feel of a model that comes from playing with it.
00:07:26.319 --> 00:07:34.480
To enable the latter, we're adding the Inkling Playground in the Tinker Council, a developer-facing interface for chatting with Inkling.
00:07:34.560 --> 00:07:39.279
To show what customization means in practice, we asked Inkling to fine-tune itself using Tinker.
00:07:39.360 --> 00:07:42.959
The model wrote its own fine-tuning job, ran it, and evaluated the result.
00:07:43.439 --> 00:07:47.120
So, long story, kind of short, right?
00:07:47.279 --> 00:07:50.240
If you want to go use Tinkling, uh Tinkling, right?
00:07:50.319 --> 00:07:52.319
And I don't, I don't know why.
00:07:52.560 --> 00:07:59.199
The combination of Thinking Machines Lab and Tinker, it just doesn't roll off the ton, uh, right.
00:07:59.439 --> 00:08:08.240
But if you want to go try it out, so they do have like a dev council where you can go quote unquote chat, which is interesting that they just didn't release a chat version.
00:08:08.480 --> 00:08:14.000
Um, anyways, this is, I think, a really big deal.
00:08:14.240 --> 00:08:16.560
And here's one of the reasons why.
00:08:17.040 --> 00:08:18.560
Uh, yes, benchmarks.
00:08:18.639 --> 00:08:23.360
So, podcast audience is showing the artificial analysis intelligence index.
00:08:23.439 --> 00:08:31.600
Uh, but this is uh shaded here by our uh closed proprietary and our open weights models.
00:08:31.759 --> 00:08:37.440
So let me zoom way out for our very non-technical audience, or if you're very new to AI.
00:08:37.600 --> 00:08:38.240
What's the difference?
00:08:38.320 --> 00:08:40.320
What's proprietary open source, right?
00:08:40.480 --> 00:08:44.559
Proprietary are models that you can't really modify them to the core.
00:08:44.720 --> 00:08:54.320
You can add custom instructions to them and you know change their behavior that way, but you can't really change the foundations of these proprietary models, right?
00:08:54.399 --> 00:08:59.919
Those are the uh the Clauds, the GPTs, the Geminis, etc., right?
00:09:00.080 --> 00:09:04.080
Then you have these open source or open weight models.
00:09:04.399 --> 00:09:10.240
And these are ones uh that for the most part China has been absolutely dominating on.
00:09:10.399 --> 00:09:18.480
And they've been doing it through distillation, which is uh, you know, not exactly uh uh something the American labs are happy with.
00:09:18.639 --> 00:09:30.399
But the you know, the Chinese labs are essentially, you know, stealing or borrowing whatever you might call it, uh, the work of the uh American labs and making their own versions of these models.
00:09:30.639 --> 00:09:32.159
And then they serve those, right?
00:09:32.240 --> 00:09:41.440
So you can if if you're a big company and if you have the the server racks, you can use these models and the open uh open source, open weight models.
00:09:41.679 --> 00:09:50.000
If you have the infrastructure, you can download them and run them 24-7 and not really pay any additional costs if you have that capacity.
00:09:50.240 --> 00:09:55.200
So that is the um the allure of these open source models.
00:09:55.360 --> 00:10:01.440
Or as consumer hardware becomes more capable, being able to run some of these locally, all right.
00:10:01.600 --> 00:10:15.039
So some of these models, not necessarily the ones I'm showing here uh on screen, but you know, some of these open source models, if you do have a very expensive, very beefy uh, you know, machine, you can run a slower version of them.
00:10:15.120 --> 00:10:19.360
So that's kind of the premise and the difference between these proprietary and open source models.
00:10:19.440 --> 00:10:27.279
But here's why I think it's interesting because Inkling, where it came at and the uh where it landed on the artificial intelligence index, uh 41.
00:10:27.440 --> 00:10:36.080
So right now your uh leaders are uh Fable 5 with a 60 and GPT-5.6 soul with a 59.
00:10:36.159 --> 00:10:41.279
All right, so a 41, you know, seems like okay, that's a pretty big drop-off, middle of the pack, right?
00:10:41.519 --> 00:10:54.080
True, but when you put into context that, you know, go back about seven-ish months, uh, the leaders of the pack at that time, well, it was GPT-5.2 uh with a 42.
00:10:54.240 --> 00:11:05.360
So uh the numbers change, all right, because the benchmarks that go into this, uh the artificial uh artificial analysis intelligence index, those uh benchmarks get updated.
00:11:05.440 --> 00:11:15.120
But it's essentially, you know, about a dozen or so different benchmarks that are always updated and rotated, uh, that tell you how good is a model compared to you know the most important factors.
00:11:15.279 --> 00:11:27.840
And Inkling, again, only being about seven months behind the frontier is actually pretty impressive for the first release from a company that we didn't really know what they were working on, right?
00:11:27.919 --> 00:11:34.399
We really didn't start hearing from Thinking Machines Lab for their first like year that they launched, right?
00:11:34.480 --> 00:11:39.120
So they launched, I think it was uh quarter one or quarter two of 2025.
00:11:39.279 --> 00:11:54.399
We didn't really hear anything from them for a year, and then we heard they're working on you know Tinker, and then we heard that they were making their own model for fine-tuning, and then we saw this um you know, this bi-directional voice model, but we didn't actually see the inkling model until well, less than 24 hours ago.
00:11:54.559 --> 00:12:03.120
But if you put it like that, it is about seven months behind the frontier, but it is an American model, which is actually important, and it's multimodal.
00:12:03.200 --> 00:12:08.240
Those two things alone, I think, have the potential to reshape what's possible.
00:12:08.480 --> 00:12:20.639
Because my thought is right, that model right there, it's not gonna, you know, if your team is AI native, if you have, you know, agents running, if you have workflow set, you're not gonna be able to swip.
00:12:20.879 --> 00:12:22.320
Like, let me just be honest, right?
00:12:22.399 --> 00:12:26.879
You can't just, you know, click copy and paste uh and put inkling in there.
00:12:27.039 --> 00:12:37.039
But for those companies that are still finding their footing or larger enterprise companies that are looking to uh, you know, chunk off a big piece of their workflows to something maybe more affordable.
00:12:37.440 --> 00:12:41.600
Inkling is actually not a bad option, uh right.
00:12:42.159 --> 00:12:45.039
So the real product though is Tinker.
00:12:45.360 --> 00:12:49.279
It is they are trying to turn fine-tuning into a service.
00:12:49.360 --> 00:12:55.200
So Inkling, yes, it can be downloaded, but even those compressed versions need 600 gigabytes of memory.
00:12:55.279 --> 00:13:04.399
So yeah, you're probably not running this or using this unless you are a large enterprise organization uh with your own, you know, GPU server infrastructure.
00:13:04.559 --> 00:13:17.600
So Tinker, though, manages the training so companies can customize the models without actually owning the GPUs, and the business model essentially turns that openness into what I think could be the first strategic reset.
00:13:17.919 --> 00:13:22.639
So uh let's talk about those potential strategic resets.
00:13:22.879 --> 00:13:28.799
Uh, reset number one, American open weights could reopen procurement.
00:13:29.120 --> 00:13:35.519
So many big enterprises haven't been able to touch some of these open source Chinese models.
00:13:35.600 --> 00:13:41.039
Well, be because of the current, you know, China-US relationship, right?
00:13:41.200 --> 00:13:45.519
Especially for those companies that do business with the government, right?
00:13:45.600 --> 00:14:08.480
And we've seen the US government get uh much more involved here recently, um between you know, export controls, um, but also a huge thing here is that we've also seen reports in the past week that China may actually, which is, I don't know, funny or interesting, right?
00:14:08.639 --> 00:14:14.000
That China may shut down uh its models to other countries, right?
00:14:14.080 --> 00:14:28.639
So even though they are distilling from US companies, we've seen reports they may not let, you know, who knows how that uh will be set up, but they may not quote unquote allow overseas companies to use their open source models.
00:14:28.720 --> 00:14:36.320
So I don't know how open source that actually makes them, and especially if they're just distilling them from US labs, anyways, but that's beside the points.
00:14:36.480 --> 00:14:44.159
But I think for so many enterprises, they haven't been able to look at the open source category yet because of that reason, right?
00:14:44.320 --> 00:14:56.559
If you're a company that has big government contracts, if you are using uh Chinese open source models, that's gonna put you in a sticky situation, or your uh RFP might be DOA, right?
00:14:56.639 --> 00:15:02.480
You might not even be considered if you are a company that has been using uh Chinese open source models.
00:15:02.639 --> 00:15:04.399
So that's a big unlock.
00:15:04.799 --> 00:15:05.440
All right.
00:15:06.320 --> 00:15:20.879
And I think that many companies have also just wanted to use open source models, but the whole fact that these are Chinese models, and you know, we've seen reports on, you know, can you actually trust, right?
00:15:21.039 --> 00:15:31.840
Like what's kind of the uh the the messaging that may be coming out of this that may not be in line with companies that have stronger, you know, American values or stronger Western values.
00:15:32.000 --> 00:15:47.679
I'm not gonna get into that, but you know, there's obviously many different reasons, not just geopolitical reasons, that so many companies here in the US haven't really been able to uh, you know, convince their board or convince, you know, anyone to go down the open source route just because it is all Chinese models.
00:15:48.960 --> 00:15:52.240
So here's potential reset too.
00:15:52.480 --> 00:15:57.440
Um, the model overhang makes this customization very timely.
00:15:57.759 --> 00:16:05.039
Uh so I've talked about this a lot over the past like three months, but I think we're now at this point where there's a model overhang.
00:16:05.120 --> 00:16:08.240
Um, it's a little bit different than the capabilities gap that I talked about.
00:16:08.399 --> 00:16:25.279
There's the great anthropic study uh from uh it seems like it was from so long ago, but it was only from a couple of months ago, their uh labor index report, uh, you know, that essentially showed the capabilities um of these models and then what companies were actually using them for, right?
00:16:25.440 --> 00:16:30.480
They uh anonymized, I think it was 400,000 uh agentic chats and they mapped it all out.
00:16:30.559 --> 00:16:42.000
And essentially what they said is these models are so capable, but you know, maybe you know, enterprises are using, you know, on average, about 10 to 20 percent of the model capabilities.
00:16:42.159 --> 00:16:54.720
So essentially these models right now, the frontier models, are way more powerful than most than the average company actually needs, or even has the actual capabilities to take advantage of.
00:16:54.879 --> 00:17:06.960
So that kind of adoption gap makes fine-tuning middle-of-the-pack models for stable repeated work, maybe an actual new and intriguing area of AI.
00:17:07.119 --> 00:17:29.759
You know, couple that with the fact that we have seen this whiplash, the um token maxing to you know, value maxing or token efficiency whiplash, where you know, earlier in, you know, from December 2025 to I would say March 2026, you know, enterprise companies were like, yes, you know, we've been all in on AI, so go use as many tokens as you can, right?
00:17:29.839 --> 00:17:33.119
And then it's like, wait, these token costs are getting higher and higher.
00:17:33.279 --> 00:17:38.240
And, you know, certain uh model providers aren't token efficient, right?
00:17:38.319 --> 00:17:41.119
Uh the data says that is anthropic, right?
00:17:41.200 --> 00:17:44.319
Um, so all of a sudden these companies have these huge API bills.
00:17:44.480 --> 00:17:52.960
So now we've seen uh, you know, in the May, June, July, this whiplash of companies being like, wait, we have to start reining spend in.
00:17:53.599 --> 00:18:04.640
And part of it is, well, they're realizing that we don't need, you know, a fable five type model for you know a hundred employees rewriting their emails.
00:18:04.720 --> 00:18:11.839
Uh, and that's what I think these middle-of-the-pack bespoke customized models could actually be a big thing.
00:18:12.319 --> 00:18:13.599
And why is that?
00:18:13.759 --> 00:18:22.559
Well, fine-tuning could return because now we actually have frontier models that are smart enough to teach, right?
00:18:22.799 --> 00:18:24.480
Which we haven't really had before.
00:18:24.640 --> 00:18:28.319
I actually have you know uh a couple exciting examples that I'm gonna go over here.
00:18:28.400 --> 00:18:30.559
But you know, this the concept of fine-tuning, right?
00:18:30.720 --> 00:18:32.400
So this is the teacher-student model.
00:18:32.559 --> 00:18:35.519
This is essentially right, and we've also got reports.
00:18:35.599 --> 00:18:46.640
If you read our newsletter every single day, we've gotten now some sniffs of real, like, okay, we're at this point of recursive self-improvement where you know models are actually, according to research, right?
00:18:46.720 --> 00:18:51.519
There's actually a very interesting paper from earlier this week that, yeah, the models are actually improving themselves.
00:18:51.599 --> 00:18:56.240
Um, right, but that also lends itself to the big, big models.
00:18:56.400 --> 00:19:13.119
Um, I think maybe GBD56 Soul is the first one I've seen uh in mass that we've seen actual real examples from uh that can teach and train smaller, much, much, much smaller versions of themselves to be used for many different purposes, right?
00:19:13.200 --> 00:19:19.119
Because these frontier models can generate data evaluations, code, foul uh failure analysis uh analysis.
00:19:19.359 --> 00:19:29.359
So I think it is not just you know, tinker this fine-tuning as a service from thinking machines and uh their new model.
00:19:29.680 --> 00:19:37.039
It's the combination of that plus this new tier of frontier models that make fine-tuning a reality, right?
00:19:37.119 --> 00:19:40.400
Because it has to become common language, right?
00:19:40.480 --> 00:19:42.000
And I think we're gonna see that.
00:19:42.160 --> 00:19:49.599
It's like I could right now, and I have an example that I think is a really awesome one, uh, but I think I forgot to include the the photo of it in my slide.
00:19:49.759 --> 00:19:56.400
Um, but you can, if you have a decent enough machine, uh, even a local one, right?
00:19:56.480 --> 00:20:07.759
Like a Mac Studio, you can start building very, very small models yourself with hardly any idea of what the heck you're doing for very small purposes.
00:20:08.160 --> 00:20:08.559
All right.
00:20:08.640 --> 00:20:18.240
So three of these more recent examples of this concept of, you know, well, in this case, GPT-5.Soul uh being in the fine-tuned models.
00:20:18.319 --> 00:20:27.279
So OpenAI's uh Jason Luai uh said that GPT-56 Soul actually helped post-train OpenAI's small model, GPT-5.
00:20:28.480 --> 00:20:33.039
Completing the work estimated to require two researchers, roughly two weeks.
00:20:33.200 --> 00:20:37.759
Uh, this is the one I forgot to put the screenshot in uh on my on my slides here.
00:20:37.839 --> 00:20:41.279
Uh, but Pietro uh Sherano, hopefully I got the name right.
00:20:41.359 --> 00:20:47.440
Uh so he's uh an CEO of uh of an AI company, but he said just a fun little project that he worked on.
00:20:47.680 --> 00:20:53.359
He used GPT-5.6 to build a small local model from his iMessage history.
00:20:53.519 --> 00:20:57.519
So essentially, I think he had it, you know, work overnight or something like that.
00:20:57.680 --> 00:21:02.400
And it literally read every single message in his uh Mac history.
00:21:02.559 --> 00:21:05.279
Uh, you know, so on your Mac you can access your iMessage.
00:21:05.359 --> 00:21:08.160
If you're you know one of our green bubble friends, you're like, what does that mean?
00:21:08.240 --> 00:21:08.400
Right.
00:21:08.559 --> 00:21:11.519
So on my computer, I have my text messages.
00:21:11.599 --> 00:21:13.200
So he just let it go.
00:21:13.359 --> 00:21:24.319
It read everything, it understood everything, ran some models, some evaluations, and all of a sudden he had a literal uh small language model that GBT56 made for him, trained on that.
00:21:24.400 --> 00:21:27.440
So this isn't like a custom GPT, right?
00:21:27.519 --> 00:21:32.720
It's like, oh, I'm using the big model uh to um give it instructions.
00:21:32.880 --> 00:21:35.759
No, it is a separate model itself, right?
00:21:35.920 --> 00:21:36.960
Absolutely crazy.
00:21:37.039 --> 00:21:39.359
Uh then NVIDIA used Codacs.
00:21:39.440 --> 00:21:42.559
They just put a blog post out uh about this, I think on Tuesday.
00:21:42.640 --> 00:21:57.839
Uh NVIDIA just used Kodaks and GVD5.6 to post train its Cosmos 3 nano model, reportedly improving accuracy in it from 454 to 93 in one day using two prompts.
00:21:58.079 --> 00:21:59.759
All right, I'll I'll I'll make sure to.
00:22:00.079 --> 00:22:02.319
Reshare that in today's uh newsletter.
00:22:02.480 --> 00:22:08.160
I think we did put it in Tuesdays or yesterday's, but uh since I'm mentioning it on the show, I'll make sure to put it in there.
00:22:08.319 --> 00:22:16.799
But here's the reality two prompts, these are off the shelf creating or improving existing models with today's frontier technology.
00:22:16.960 --> 00:22:19.039
So it's not just inkling, right?
00:22:19.200 --> 00:22:35.839
It is more of a reflection of the combination of these very powerful frontier models that can literally build, run the evaluations, run the testing, run the QA, can literally build large small language models on their own, uh, and also uh train you know other variations of themselves.
00:22:36.160 --> 00:22:43.759
So, one example though, that we got from Thinking Machines, they did share uh recently one use case about Bridgewater, right?
00:22:43.839 --> 00:22:45.920
So using their Tinker service, right?
00:22:46.000 --> 00:22:52.720
So this fine-tuning as a service, uh Bridgewater customized Quen, uh, which is a Chinese open source model.
00:22:52.880 --> 00:23:11.519
So this is, you know, obviously before Inkling was uh available, or at least before they were ready to use it and talk about it, uh, they did come out with this customer use case of Bridgewater customizing uh the Chinese open source Quen model through uh Thinking Machines Tinker platform for recurring financial judgment tasks.
00:23:11.759 --> 00:23:20.720
So it report uh it reportedly beat the best tested frontier model while costing 13.8 times less.
00:23:21.119 --> 00:23:30.400
So in this case, specialized judgment beat the frontier defaults, and that kind of explains this potential model shopping shift.
00:23:31.279 --> 00:23:35.519
So let me just put it out here, y'all.
00:23:35.839 --> 00:23:39.920
Not one to be like, hey, I told you guys this a long time ago.
00:23:40.160 --> 00:23:47.680
Maybe I was just a little too early or a little too weird, but I think two years ago in my AI prediction series, right?
00:23:47.839 --> 00:23:59.599
I talked about this very thing and the uh the rise of potentially seeing thousands of small language models created by the large language models.
00:23:59.839 --> 00:24:09.119
And although we're still maybe not there because uh, you know, there is this whole thing called, you know, compute is scarce, uh, right?
00:24:09.359 --> 00:24:16.720
Where you necessarily you can't have uh anthropic and open AI and Google and Microsoft and Meta and Grok.
00:24:17.039 --> 00:24:27.839
They can't necessarily, you know, pull uh gigawatts of compute to create thousands of these bespoke uh middle tier and small language models, but that's still the reality.
00:24:27.920 --> 00:24:43.119
And I think that this here, uh, the combination of a GPT-5.6 soul uh being able to create small language models and uh post-train its own versions and the combination of thinking machines uh coming up with this as a service.
00:24:43.359 --> 00:24:51.119
I think there's finally my uh I forgot if it was late 2023 or late 2024 prediction could be coming true.
00:24:51.200 --> 00:25:02.880
Where I think we are eventually and very soon going to see large language models create hundreds or thousands of versions of small language models uh because they're cheaper.
00:25:03.039 --> 00:25:16.240
And there's finally the appetite to care about it because nine months ago, people didn't necessarily care because we were still right, we were still on this, you know, this richie rich blank check AI, right?
00:25:16.319 --> 00:25:18.079
It for$20 a month.
00:25:18.240 --> 00:25:21.440
You couldn't, most 90% of employees, right?
00:25:21.519 --> 00:25:28.160
If you had a team or a business plan paying$20 to$50 a month per seat, most employees couldn't get through that.
00:25:28.240 --> 00:25:32.559
So companies didn't necessarily care about being efficient with their AI.
00:25:32.720 --> 00:25:44.880
The fact of right, the the concept of a small language model or a a medium-sized model fine-tuned for specific tasks didn't necessarily matter because the big model could still do it.
00:25:44.960 --> 00:25:49.759
And everyone essentially, we were playing with monopoly money until about four months ago.
00:25:49.920 --> 00:25:56.319
So the whole concept of small language models or fine-tuned models literally didn't matter because it was free money.
00:25:56.559 --> 00:26:00.160
Now, because of the whiplash, this is more important than ever.
00:26:00.319 --> 00:26:04.000
So, as we wrap, here's what I want to talk about.
00:26:05.839 --> 00:26:07.599
Reset free, right?
00:26:07.920 --> 00:26:09.359
Model shopping is now big.
00:26:09.440 --> 00:26:11.920
And this is now, I think, part of the executive playbook.
00:26:12.000 --> 00:26:14.079
And we saw Microsoft, right?
00:26:14.240 --> 00:26:26.400
Microsoft is reportedly gonna start moving uh away from open AI and anthropic models, not moving away, but they're gonna start mixing in their own models as well as reportedly deep seek models.
00:26:26.480 --> 00:26:38.160
So even the biggest enterprise customers are starting to realize that for a majority of day-to-day tasks, you might not need the number one model in the world to do a big chunk of the work.
00:26:38.400 --> 00:26:48.640
And I think now it's more of this concept of maybe renting frontier intelligence for those uh, you know, for the ambiguous and risky uh and and changing work.
00:26:49.519 --> 00:26:57.119
But I think the future, which I've talked about, is using the right model for the right purpose at the right time.
00:26:57.279 --> 00:26:59.440
And eventually that will become automatic, right?
00:26:59.519 --> 00:27:00.960
I've actually built some things like that.
00:27:01.119 --> 00:27:06.640
I'm like, I'm like, I wonder why no one else is doing this because it's not like especially since GPT 5.6 came out.
00:27:06.799 --> 00:27:09.920
Like I built skills that essentially model route, right?
00:27:10.000 --> 00:27:17.759
I can use Fable inside of Codex, I can use GPT 5.6 uh inside of Claude Desktop, right?
00:27:17.839 --> 00:27:22.000
If you have a little bit of skill and enough patience, um, you can do those things, right?
00:27:22.079 --> 00:27:27.759
And I've done the same things where um, not that I ever hit my usage, but just a practice, right?
00:27:27.920 --> 00:27:32.880
Where I have I've built in model routing within codecs or you know, chat GPT work.
00:27:33.039 --> 00:27:43.839
So if you are a cost conscious um, you know, company, this is the future of working with multiple models, and you're not just gonna throw, you know, every single request at one big model.
00:27:44.880 --> 00:27:46.880
So here's the new AI playbook.
00:27:46.960 --> 00:27:54.240
You own the workflow, but you're probably just gonna rent the frontier or just exclusively use the frontier for those specific purposes.
00:27:54.400 --> 00:28:02.160
So you have to score every workflow by the stakes, the volume, the privacy, and how much proprietary judgment actually differentiates it.
00:28:02.319 --> 00:28:07.599
So you have to start to default to more of this economical model, right?
00:28:07.680 --> 00:28:18.880
Like you have to start mapping out that workflow and saying, okay, maybe we send the 20% to you know, if you especially if you're using it via the API, to that frontier model, right?
00:28:19.039 --> 00:28:23.759
But I mean, my gosh, talk about middle, middle and lower tier models.
00:28:23.920 --> 00:28:31.920
I mean, open AIs, uh, Terra and Luna, when it comes to a cost efficiency, are legit off the charts, right?
00:28:32.000 --> 00:28:38.480
So if I'm advising companies, you know, I'm saying like you should be using these models for the most part, right?
00:28:38.640 --> 00:28:48.079
The the the Luna model on like max setting is probably more than enough for almost 90% of the work that you would do.
00:28:48.240 --> 00:29:04.880
So it is kind of shifting away from this one model for all purposes to understanding, which I know is is confusing because it is easy to hit that easy button, but we can't just hit that easy button every single time because that button is going to start to cost more and more money as it requires more and more power.
00:29:05.039 --> 00:29:05.599
Uh, right.
00:29:05.759 --> 00:29:15.440
So you have to default to those sometimes more economical models, route by difficulty, and then fine-tune potentially stable, repeated, measurable work.
00:29:15.599 --> 00:29:22.960
And you have to stop asking which model is the smartest and start asking which model is enough for each job.
00:29:23.119 --> 00:29:24.400
All right, that's a wrap.
00:29:24.480 --> 00:29:26.079
I hope this one was helpful.
00:29:26.240 --> 00:29:32.720
The pretty exciting release, maybe not just for the model itself, but more for what it represents.
00:29:32.880 --> 00:29:34.400
So I hope this was helpful.
00:29:34.480 --> 00:29:36.640
If so, please let me know about it.
00:29:36.720 --> 00:29:38.960
Go sign up for our free daily newsletter.
00:29:39.039 --> 00:29:43.680
Drop me a line when you get that automated uh, you know, email, uh, welcome email.
00:29:43.839 --> 00:29:49.680
But also, if you could do me a favor, go subscribe to the podcast on Spotify or Apple Podcasts.
00:29:49.759 --> 00:29:50.720
Appreciate you tuning in.
00:29:50.880 --> 00:29:53.920
We'll see you back tomorrow and every day for more everyday AI.
00:29:54.000 --> 00:29:55.119
Thanks, y'all.
00:00:02.560 --> 00:00:09.919
One of the more important AI models you've probably never heard of and likely won't use just got released.
00:00:10.080 --> 00:00:17.280
This may get slept on, but Inkling is a huge release from Thinking Machines Lab.
00:00:17.600 --> 00:00:25.920
So if you haven't heard of Thinking Machines, they're led by Mira Marathi, the former CTO of OpenAI.
00:00:26.160 --> 00:00:32.560
So why is Inkling maybe the most important AI model you probably won't use?
00:00:33.039 --> 00:00:45.520
Because it's now the best open source model from an American company, and Thinking Machines is betting on the future of AI, fine-tuning as a service.
00:00:45.840 --> 00:00:54.799
The model itself, though, it's multimodal, it's agentic, it's customizable, and it's available through enterprise distribution on day one.
00:00:55.039 --> 00:01:00.399
Inkling doesn't need to beat every Chinese model or Claude Fable to be relevant.
00:01:00.560 --> 00:01:12.000
It only needs to pique the interest of a few cost-conscious enterprises to not only be extremely profitable, but to also help shift the conversation around enterprise AI.
00:01:12.239 --> 00:01:22.799
Frontier models are becoming so powerful they can actually fine-tune smaller, practical models without much iteration or without even much expertise.
00:01:23.599 --> 00:01:27.680
So, has a new category re-emerged?
00:01:27.920 --> 00:01:29.519
Is fine-tuning back?
00:01:29.840 --> 00:01:36.079
And what might think machines' first major product mean for the broader AI competition?
00:01:36.480 --> 00:01:38.400
Well, here's the big picture.
00:01:38.640 --> 00:01:42.079
Inkling could change how enterprises buy AI.
00:01:42.319 --> 00:01:46.560
So this was just released Wednesday, so hours ago.
00:01:46.879 --> 00:01:51.760
And it is not the smartest model, but it is strategically important.
00:01:51.920 --> 00:01:58.799
I think it gives companies a credible American open weight model alternative to Chinese models.
00:01:58.879 --> 00:02:06.079
And inkling could be the best general purpose open model because it is multimodal, right?
00:02:06.159 --> 00:02:09.840
A lot of the open source, open weight Chinese models aren't.
00:02:10.000 --> 00:02:11.599
Many of them are text only.
00:02:11.759 --> 00:02:14.719
So Inkling, it is multimodal, it is a gentic.
00:02:15.439 --> 00:02:26.319
And as workflows in the real world lag so far behind model capabilities, I do think that there's a real market for bespoke middle-of-the-pack AI.
00:02:27.039 --> 00:02:30.159
So on today's show, here's what you're gonna learn.
00:02:30.240 --> 00:02:34.319
You're gonna learn why an open model matters, even if you're never gonna deploy it.
00:02:34.479 --> 00:02:40.400
You're gonna understand how Inkling reopens enterprise AI options beyond just Chinese open models.
00:02:40.560 --> 00:02:45.840
You're gonna know why frontier models could make fine-tuning practical beyond research teams.
00:02:46.000 --> 00:02:52.159
And you're gonna know how model shopping is gonna change budgets, vendors, and competitive leverage.
00:02:52.319 --> 00:02:53.520
All right, let's get to it.
00:02:53.680 --> 00:02:55.360
Welcome to Everyday AI.
00:02:55.520 --> 00:02:57.360
If you're new here, my name is Jordan Wilson.
00:02:57.439 --> 00:02:58.240
We do this every day.
00:02:58.400 --> 00:03:08.000
It's your daily live stream podcast and free daily newsletter helping business leaders like you and me keep up with the nonstop AI updates because my gosh, can't take an hour off.
00:03:08.159 --> 00:03:10.319
I help you decide what's important.
00:03:10.479 --> 00:03:14.000
I tell you how to use all this information to grow your company and your career.
00:03:14.080 --> 00:03:17.520
So it starts here with the unedited, unscripted live stream podcast.
00:03:17.599 --> 00:03:21.360
But make sure you go to our website at your everydayai.com.
00:03:21.439 --> 00:03:22.319
That's your cheat code.
00:03:22.479 --> 00:03:27.919
We're gonna be not just recapping the highlights from today's show, but go sign up for our free daily newsletter.
00:03:28.000 --> 00:03:32.879
We're gonna be re uh giving you everything else that you need to know that's happening in the world of AI today.
00:03:33.039 --> 00:03:36.879
Because yeah, it's one of those things you gotta like go and work it out.
00:03:36.960 --> 00:03:37.599
It's a muscle.
00:03:37.759 --> 00:03:38.879
Use it every single day.
00:03:39.120 --> 00:03:42.560
Uh, and yeah, all the AI news will be in our newsletter.
00:03:42.719 --> 00:03:43.759
All right, let's get into it.
00:03:43.919 --> 00:03:45.599
Live stream audience, good to see you.
00:03:45.680 --> 00:03:47.039
Uh, Adam joining us from St.
00:03:47.199 --> 00:03:47.280
St.
00:03:47.439 --> 00:03:54.000
Louis, Jose from Santiago, uh, Angie joining from Montana, Amiko, Tokyo, Brian, what's up, Brian?
00:03:54.080 --> 00:03:55.199
Joining from Minnesota.
00:03:55.360 --> 00:04:03.919
Uh, so uh, let's talk about maybe the most important AI model you'll probably never use.
00:04:04.240 --> 00:04:12.159
So there's a lot of different factors that have been compounding over the last, I would say, three months.
00:04:12.400 --> 00:04:26.160
And to put it in a very short uh summary, frontier models are probably for the most part too much for many enterprises, right?
00:04:26.319 --> 00:04:34.160
I'll say this I, you know, probably the Fortune 500, they can squeeze as much juice and uh, you know, make it worth the cost.
00:04:34.240 --> 00:04:43.920
But I'd say for many companies, especially those probably in the Fortune 5 or like the Fortune 501 to the 5000, right?
00:04:44.000 --> 00:04:45.600
So this isn't everyone.
00:04:45.759 --> 00:04:59.120
Um, you know, I don't think this changes the equation for every single company, every single business out there, because so many people are still gonna want, you know, their Chat GPT Enterprise, their, you know, Gemini, their Claude, their co-pilot.
00:04:59.199 --> 00:05:07.759
They want to make things easy, and they're not necessarily, you know, looking at, okay, how do open models or fine-tuning models change our strategy?
00:05:07.920 --> 00:05:10.240
But for so many in our audience, it does.
00:05:10.480 --> 00:05:13.040
And this is a really big deal.
00:05:13.279 --> 00:05:16.000
So here's a little bit more about the model itself.
00:05:16.240 --> 00:05:20.959
So, model, uh, the model is from Thinking Machines and it's called Inkling.
00:05:21.279 --> 00:05:26.079
It is a 975 billion parameter model, and it's multimodal.
00:05:26.160 --> 00:05:28.079
So, text, images, and audio.
00:05:28.240 --> 00:05:45.759
Uh, what is interesting as well, uh, previously Thinking Machines did demo kind of a similar dual purpose or uh you know two-way street um audio model as well, so that can kind of hear and listen as well as talk at the same time, like OpenAI's new GPT live.
00:05:45.920 --> 00:05:56.000
So Inkling is led by former OpenAI CTO Miro Marathi, and its quick arrival raises, I think, the next big question, which is how good is it?
00:05:56.160 --> 00:05:58.560
Well, here's from the company themselves from their release.
00:05:58.720 --> 00:06:07.439
They say our model called Inkling is a mixture of experts transformer with 975 billion total parameters, 41 billion active.
00:06:07.519 --> 00:06:10.720
It supports a context window of up to 1 million tokens.
00:06:10.879 --> 00:06:16.000
It was pre-trained on 45 trillion tokens of text, images, audio, and video.
00:06:16.079 --> 00:06:19.040
It is the first in a family of models of different sizes.
00:06:19.120 --> 00:06:34.480
Uh, alongside it, we are sharing a preview of Inkling Small, a lighter weight model with 12 billion active parameters, trained with a smaller recipe that achieves strong performance with even lower costs and latency.
00:06:34.720 --> 00:06:42.560
Inkling reasons natively over text, images, and audio and balances costs with performance through efficient and controllable thinking effort.
00:06:42.879 --> 00:06:48.800
We trained it to be a broad balanced foundation model, strong across many domains, flexible enough to adapt.
00:06:49.040 --> 00:06:54.160
Inkling is not the strongest over model overall model available today, open or closed.
00:06:54.319 --> 00:07:04.879
Instead, a combination of qualities makes it a good open weights base for customization, multimodal capabilities, efficient thinking, and availability on Tinker for fine tuning.
00:07:05.120 --> 00:07:08.560
Inkling is just the start, our first release in a model family.
00:07:08.639 --> 00:07:09.920
We will continue to build on.
00:07:10.079 --> 00:07:13.519
We want to make customization accessible for more use cases.
00:07:13.600 --> 00:07:17.519
So Inkling is available for fine tuning on Tinker today.
00:07:17.680 --> 00:07:26.160
Uh, picking the right base model to fine-tune is a qualitative judgment that combines measurable benchmarks with a unique feel of a model that comes from playing with it.
00:07:26.319 --> 00:07:34.480
To enable the latter, we're adding the Inkling Playground in the Tinker Council, a developer-facing interface for chatting with Inkling.
00:07:34.560 --> 00:07:39.279
To show what customization means in practice, we asked Inkling to fine-tune itself using Tinker.
00:07:39.360 --> 00:07:42.959
The model wrote its own fine-tuning job, ran it, and evaluated the result.
00:07:43.439 --> 00:07:47.120
So, long story, kind of short, right?
00:07:47.279 --> 00:07:50.240
If you want to go use Tinkling, uh Tinkling, right?
00:07:50.319 --> 00:07:52.319
And I don't, I don't know why.
00:07:52.560 --> 00:07:59.199
The combination of Thinking Machines Lab and Tinker, it just doesn't roll off the ton, uh, right.
00:07:59.439 --> 00:08:08.240
But if you want to go try it out, so they do have like a dev council where you can go quote unquote chat, which is interesting that they just didn't release a chat version.
00:08:08.480 --> 00:08:14.000
Um, anyways, this is, I think, a really big deal.
00:08:14.240 --> 00:08:16.560
And here's one of the reasons why.
00:08:17.040 --> 00:08:18.560
Uh, yes, benchmarks.
00:08:18.639 --> 00:08:23.360
So, podcast audience is showing the artificial analysis intelligence index.
00:08:23.439 --> 00:08:31.600
Uh, but this is uh shaded here by our uh closed proprietary and our open weights models.
00:08:31.759 --> 00:08:37.440
So let me zoom way out for our very non-technical audience, or if you're very new to AI.
00:08:37.600 --> 00:08:38.240
What's the difference?
00:08:38.320 --> 00:08:40.320
What's proprietary open source, right?
00:08:40.480 --> 00:08:44.559
Proprietary are models that you can't really modify them to the core.
00:08:44.720 --> 00:08:54.320
You can add custom instructions to them and you know change their behavior that way, but you can't really change the foundations of these proprietary models, right?
00:08:54.399 --> 00:08:59.919
Those are the uh the Clauds, the GPTs, the Geminis, etc., right?
00:09:00.080 --> 00:09:04.080
Then you have these open source or open weight models.
00:09:04.399 --> 00:09:10.240
And these are ones uh that for the most part China has been absolutely dominating on.
00:09:10.399 --> 00:09:18.480
And they've been doing it through distillation, which is uh, you know, not exactly uh uh something the American labs are happy with.
00:09:18.639 --> 00:09:30.399
But the you know, the Chinese labs are essentially, you know, stealing or borrowing whatever you might call it, uh, the work of the uh American labs and making their own versions of these models.
00:09:30.639 --> 00:09:32.159
And then they serve those, right?
00:09:32.240 --> 00:09:41.440
So you can if if you're a big company and if you have the the server racks, you can use these models and the open uh open source, open weight models.
00:09:41.679 --> 00:09:50.000
If you have the infrastructure, you can download them and run them 24-7 and not really pay any additional costs if you have that capacity.
00:09:50.240 --> 00:09:55.200
So that is the um the allure of these open source models.
00:09:55.360 --> 00:10:01.440
Or as consumer hardware becomes more capable, being able to run some of these locally, all right.
00:10:01.600 --> 00:10:15.039
So some of these models, not necessarily the ones I'm showing here uh on screen, but you know, some of these open source models, if you do have a very expensive, very beefy uh, you know, machine, you can run a slower version of them.
00:10:15.120 --> 00:10:19.360
So that's kind of the premise and the difference between these proprietary and open source models.
00:10:19.440 --> 00:10:27.279
But here's why I think it's interesting because Inkling, where it came at and the uh where it landed on the artificial intelligence index, uh 41.
00:10:27.440 --> 00:10:36.080
So right now your uh leaders are uh Fable 5 with a 60 and GPT-5.6 soul with a 59.
00:10:36.159 --> 00:10:41.279
All right, so a 41, you know, seems like okay, that's a pretty big drop-off, middle of the pack, right?
00:10:41.519 --> 00:10:54.080
True, but when you put into context that, you know, go back about seven-ish months, uh, the leaders of the pack at that time, well, it was GPT-5.2 uh with a 42.
00:10:54.240 --> 00:11:05.360
So uh the numbers change, all right, because the benchmarks that go into this, uh the artificial uh artificial analysis intelligence index, those uh benchmarks get updated.
00:11:05.440 --> 00:11:15.120
But it's essentially, you know, about a dozen or so different benchmarks that are always updated and rotated, uh, that tell you how good is a model compared to you know the most important factors.
00:11:15.279 --> 00:11:27.840
And Inkling, again, only being about seven months behind the frontier is actually pretty impressive for the first release from a company that we didn't really know what they were working on, right?
00:11:27.919 --> 00:11:34.399
We really didn't start hearing from Thinking Machines Lab for their first like year that they launched, right?
00:11:34.480 --> 00:11:39.120
So they launched, I think it was uh quarter one or quarter two of 2025.
00:11:39.279 --> 00:11:54.399
We didn't really hear anything from them for a year, and then we heard they're working on you know Tinker, and then we heard that they were making their own model for fine-tuning, and then we saw this um you know, this bi-directional voice model, but we didn't actually see the inkling model until well, less than 24 hours ago.
00:11:54.559 --> 00:12:03.120
But if you put it like that, it is about seven months behind the frontier, but it is an American model, which is actually important, and it's multimodal.
00:12:03.200 --> 00:12:08.240
Those two things alone, I think, have the potential to reshape what's possible.
00:12:08.480 --> 00:12:20.639
Because my thought is right, that model right there, it's not gonna, you know, if your team is AI native, if you have, you know, agents running, if you have workflow set, you're not gonna be able to swip.
00:12:20.879 --> 00:12:22.320
Like, let me just be honest, right?
00:12:22.399 --> 00:12:26.879
You can't just, you know, click copy and paste uh and put inkling in there.
00:12:27.039 --> 00:12:37.039
But for those companies that are still finding their footing or larger enterprise companies that are looking to uh, you know, chunk off a big piece of their workflows to something maybe more affordable.
00:12:37.440 --> 00:12:41.600
Inkling is actually not a bad option, uh right.
00:12:42.159 --> 00:12:45.039
So the real product though is Tinker.
00:12:45.360 --> 00:12:49.279
It is they are trying to turn fine-tuning into a service.
00:12:49.360 --> 00:12:55.200
So Inkling, yes, it can be downloaded, but even those compressed versions need 600 gigabytes of memory.
00:12:55.279 --> 00:13:04.399
So yeah, you're probably not running this or using this unless you are a large enterprise organization uh with your own, you know, GPU server infrastructure.
00:13:04.559 --> 00:13:17.600
So Tinker, though, manages the training so companies can customize the models without actually owning the GPUs, and the business model essentially turns that openness into what I think could be the first strategic reset.
00:13:17.919 --> 00:13:22.639
So uh let's talk about those potential strategic resets.
00:13:22.879 --> 00:13:28.799
Uh, reset number one, American open weights could reopen procurement.
00:13:29.120 --> 00:13:35.519
So many big enterprises haven't been able to touch some of these open source Chinese models.
00:13:35.600 --> 00:13:41.039
Well, be because of the current, you know, China-US relationship, right?
00:13:41.200 --> 00:13:45.519
Especially for those companies that do business with the government, right?
00:13:45.600 --> 00:14:08.480
And we've seen the US government get uh much more involved here recently, um between you know, export controls, um, but also a huge thing here is that we've also seen reports in the past week that China may actually, which is, I don't know, funny or interesting, right?
00:14:08.639 --> 00:14:14.000
That China may shut down uh its models to other countries, right?
00:14:14.080 --> 00:14:28.639
So even though they are distilling from US companies, we've seen reports they may not let, you know, who knows how that uh will be set up, but they may not quote unquote allow overseas companies to use their open source models.
00:14:28.720 --> 00:14:36.320
So I don't know how open source that actually makes them, and especially if they're just distilling them from US labs, anyways, but that's beside the points.
00:14:36.480 --> 00:14:44.159
But I think for so many enterprises, they haven't been able to look at the open source category yet because of that reason, right?
00:14:44.320 --> 00:14:56.559
If you're a company that has big government contracts, if you are using uh Chinese open source models, that's gonna put you in a sticky situation, or your uh RFP might be DOA, right?
00:14:56.639 --> 00:15:02.480
You might not even be considered if you are a company that has been using uh Chinese open source models.
00:15:02.639 --> 00:15:04.399
So that's a big unlock.
00:15:04.799 --> 00:15:05.440
All right.
00:15:06.320 --> 00:15:20.879
And I think that many companies have also just wanted to use open source models, but the whole fact that these are Chinese models, and you know, we've seen reports on, you know, can you actually trust, right?
00:15:21.039 --> 00:15:31.840
Like what's kind of the uh the the messaging that may be coming out of this that may not be in line with companies that have stronger, you know, American values or stronger Western values.
00:15:32.000 --> 00:15:47.679
I'm not gonna get into that, but you know, there's obviously many different reasons, not just geopolitical reasons, that so many companies here in the US haven't really been able to uh, you know, convince their board or convince, you know, anyone to go down the open source route just because it is all Chinese models.
00:15:48.960 --> 00:15:52.240
So here's potential reset too.
00:15:52.480 --> 00:15:57.440
Um, the model overhang makes this customization very timely.
00:15:57.759 --> 00:16:05.039
Uh so I've talked about this a lot over the past like three months, but I think we're now at this point where there's a model overhang.
00:16:05.120 --> 00:16:08.240
Um, it's a little bit different than the capabilities gap that I talked about.
00:16:08.399 --> 00:16:25.279
There's the great anthropic study uh from uh it seems like it was from so long ago, but it was only from a couple of months ago, their uh labor index report, uh, you know, that essentially showed the capabilities um of these models and then what companies were actually using them for, right?
00:16:25.440 --> 00:16:30.480
They uh anonymized, I think it was 400,000 uh agentic chats and they mapped it all out.
00:16:30.559 --> 00:16:42.000
And essentially what they said is these models are so capable, but you know, maybe you know, enterprises are using, you know, on average, about 10 to 20 percent of the model capabilities.
00:16:42.159 --> 00:16:54.720
So essentially these models right now, the frontier models, are way more powerful than most than the average company actually needs, or even has the actual capabilities to take advantage of.
00:16:54.879 --> 00:17:06.960
So that kind of adoption gap makes fine-tuning middle-of-the-pack models for stable repeated work, maybe an actual new and intriguing area of AI.
00:17:07.119 --> 00:17:29.759
You know, couple that with the fact that we have seen this whiplash, the um token maxing to you know, value maxing or token efficiency whiplash, where you know, earlier in, you know, from December 2025 to I would say March 2026, you know, enterprise companies were like, yes, you know, we've been all in on AI, so go use as many tokens as you can, right?
00:17:29.839 --> 00:17:33.119
And then it's like, wait, these token costs are getting higher and higher.
00:17:33.279 --> 00:17:38.240
And, you know, certain uh model providers aren't token efficient, right?
00:17:38.319 --> 00:17:41.119
Uh the data says that is anthropic, right?
00:17:41.200 --> 00:17:44.319
Um, so all of a sudden these companies have these huge API bills.
00:17:44.480 --> 00:17:52.960
So now we've seen uh, you know, in the May, June, July, this whiplash of companies being like, wait, we have to start reining spend in.
00:17:53.599 --> 00:18:04.640
And part of it is, well, they're realizing that we don't need, you know, a fable five type model for you know a hundred employees rewriting their emails.
00:18:04.720 --> 00:18:11.839
Uh, and that's what I think these middle-of-the-pack bespoke customized models could actually be a big thing.
00:18:12.319 --> 00:18:13.599
And why is that?
00:18:13.759 --> 00:18:22.559
Well, fine-tuning could return because now we actually have frontier models that are smart enough to teach, right?
00:18:22.799 --> 00:18:24.480
Which we haven't really had before.
00:18:24.640 --> 00:18:28.319
I actually have you know uh a couple exciting examples that I'm gonna go over here.
00:18:28.400 --> 00:18:30.559
But you know, this the concept of fine-tuning, right?
00:18:30.720 --> 00:18:32.400
So this is the teacher-student model.
00:18:32.559 --> 00:18:35.519
This is essentially right, and we've also got reports.
00:18:35.599 --> 00:18:46.640
If you read our newsletter every single day, we've gotten now some sniffs of real, like, okay, we're at this point of recursive self-improvement where you know models are actually, according to research, right?
00:18:46.720 --> 00:18:51.519
There's actually a very interesting paper from earlier this week that, yeah, the models are actually improving themselves.
00:18:51.599 --> 00:18:56.240
Um, right, but that also lends itself to the big, big models.
00:18:56.400 --> 00:19:13.119
Um, I think maybe GBD56 Soul is the first one I've seen uh in mass that we've seen actual real examples from uh that can teach and train smaller, much, much, much smaller versions of themselves to be used for many different purposes, right?
00:19:13.200 --> 00:19:19.119
Because these frontier models can generate data evaluations, code, foul uh failure analysis uh analysis.
00:19:19.359 --> 00:19:29.359
So I think it is not just you know, tinker this fine-tuning as a service from thinking machines and uh their new model.
00:19:29.680 --> 00:19:37.039
It's the combination of that plus this new tier of frontier models that make fine-tuning a reality, right?
00:19:37.119 --> 00:19:40.400
Because it has to become common language, right?
00:19:40.480 --> 00:19:42.000
And I think we're gonna see that.
00:19:42.160 --> 00:19:49.599
It's like I could right now, and I have an example that I think is a really awesome one, uh, but I think I forgot to include the the photo of it in my slide.
00:19:49.759 --> 00:19:56.400
Um, but you can, if you have a decent enough machine, uh, even a local one, right?
00:19:56.480 --> 00:20:07.759
Like a Mac Studio, you can start building very, very small models yourself with hardly any idea of what the heck you're doing for very small purposes.
00:20:08.160 --> 00:20:08.559
All right.
00:20:08.640 --> 00:20:18.240
So three of these more recent examples of this concept of, you know, well, in this case, GPT-5.Soul uh being in the fine-tuned models.
00:20:18.319 --> 00:20:27.279
So OpenAI's uh Jason Luai uh said that GPT-56 Soul actually helped post-train OpenAI's small model, GPT-5.
00:20:28.480 --> 00:20:33.039
Completing the work estimated to require two researchers, roughly two weeks.
00:20:33.200 --> 00:20:37.759
Uh, this is the one I forgot to put the screenshot in uh on my on my slides here.
00:20:37.839 --> 00:20:41.279
Uh, but Pietro uh Sherano, hopefully I got the name right.
00:20:41.359 --> 00:20:47.440
Uh so he's uh an CEO of uh of an AI company, but he said just a fun little project that he worked on.
00:20:47.680 --> 00:20:53.359
He used GPT-5.6 to build a small local model from his iMessage history.
00:20:53.519 --> 00:20:57.519
So essentially, I think he had it, you know, work overnight or something like that.
00:20:57.680 --> 00:21:02.400
And it literally read every single message in his uh Mac history.
00:21:02.559 --> 00:21:05.279
Uh, you know, so on your Mac you can access your iMessage.
00:21:05.359 --> 00:21:08.160
If you're you know one of our green bubble friends, you're like, what does that mean?
00:21:08.240 --> 00:21:08.400
Right.
00:21:08.559 --> 00:21:11.519
So on my computer, I have my text messages.
00:21:11.599 --> 00:21:13.200
So he just let it go.
00:21:13.359 --> 00:21:24.319
It read everything, it understood everything, ran some models, some evaluations, and all of a sudden he had a literal uh small language model that GBT56 made for him, trained on that.
00:21:24.400 --> 00:21:27.440
So this isn't like a custom GPT, right?
00:21:27.519 --> 00:21:32.720
It's like, oh, I'm using the big model uh to um give it instructions.
00:21:32.880 --> 00:21:35.759
No, it is a separate model itself, right?
00:21:35.920 --> 00:21:36.960
Absolutely crazy.
00:21:37.039 --> 00:21:39.359
Uh then NVIDIA used Codacs.
00:21:39.440 --> 00:21:42.559
They just put a blog post out uh about this, I think on Tuesday.
00:21:42.640 --> 00:21:57.839
Uh NVIDIA just used Kodaks and GVD5.6 to post train its Cosmos 3 nano model, reportedly improving accuracy in it from 454 to 93 in one day using two prompts.
00:21:58.079 --> 00:21:59.759
All right, I'll I'll I'll make sure to.
00:22:00.079 --> 00:22:02.319
Reshare that in today's uh newsletter.
00:22:02.480 --> 00:22:08.160
I think we did put it in Tuesdays or yesterday's, but uh since I'm mentioning it on the show, I'll make sure to put it in there.
00:22:08.319 --> 00:22:16.799
But here's the reality two prompts, these are off the shelf creating or improving existing models with today's frontier technology.
00:22:16.960 --> 00:22:19.039
So it's not just inkling, right?
00:22:19.200 --> 00:22:35.839
It is more of a reflection of the combination of these very powerful frontier models that can literally build, run the evaluations, run the testing, run the QA, can literally build large small language models on their own, uh, and also uh train you know other variations of themselves.
00:22:36.160 --> 00:22:43.759
So, one example though, that we got from Thinking Machines, they did share uh recently one use case about Bridgewater, right?
00:22:43.839 --> 00:22:45.920
So using their Tinker service, right?
00:22:46.000 --> 00:22:52.720
So this fine-tuning as a service, uh Bridgewater customized Quen, uh, which is a Chinese open source model.
00:22:52.880 --> 00:23:11.519
So this is, you know, obviously before Inkling was uh available, or at least before they were ready to use it and talk about it, uh, they did come out with this customer use case of Bridgewater customizing uh the Chinese open source Quen model through uh Thinking Machines Tinker platform for recurring financial judgment tasks.
00:23:11.759 --> 00:23:20.720
So it report uh it reportedly beat the best tested frontier model while costing 13.8 times less.
00:23:21.119 --> 00:23:30.400
So in this case, specialized judgment beat the frontier defaults, and that kind of explains this potential model shopping shift.
00:23:31.279 --> 00:23:35.519
So let me just put it out here, y'all.
00:23:35.839 --> 00:23:39.920
Not one to be like, hey, I told you guys this a long time ago.
00:23:40.160 --> 00:23:47.680
Maybe I was just a little too early or a little too weird, but I think two years ago in my AI prediction series, right?
00:23:47.839 --> 00:23:59.599
I talked about this very thing and the uh the rise of potentially seeing thousands of small language models created by the large language models.
00:23:59.839 --> 00:24:09.119
And although we're still maybe not there because uh, you know, there is this whole thing called, you know, compute is scarce, uh, right?
00:24:09.359 --> 00:24:16.720
Where you necessarily you can't have uh anthropic and open AI and Google and Microsoft and Meta and Grok.
00:24:17.039 --> 00:24:27.839
They can't necessarily, you know, pull uh gigawatts of compute to create thousands of these bespoke uh middle tier and small language models, but that's still the reality.
00:24:27.920 --> 00:24:43.119
And I think that this here, uh, the combination of a GPT-5.6 soul uh being able to create small language models and uh post-train its own versions and the combination of thinking machines uh coming up with this as a service.
00:24:43.359 --> 00:24:51.119
I think there's finally my uh I forgot if it was late 2023 or late 2024 prediction could be coming true.
00:24:51.200 --> 00:25:02.880
Where I think we are eventually and very soon going to see large language models create hundreds or thousands of versions of small language models uh because they're cheaper.
00:25:03.039 --> 00:25:16.240
And there's finally the appetite to care about it because nine months ago, people didn't necessarily care because we were still right, we were still on this, you know, this richie rich blank check AI, right?
00:25:16.319 --> 00:25:18.079
It for$20 a month.
00:25:18.240 --> 00:25:21.440
You couldn't, most 90% of employees, right?
00:25:21.519 --> 00:25:28.160
If you had a team or a business plan paying$20 to$50 a month per seat, most employees couldn't get through that.
00:25:28.240 --> 00:25:32.559
So companies didn't necessarily care about being efficient with their AI.
00:25:32.720 --> 00:25:44.880
The fact of right, the the concept of a small language model or a a medium-sized model fine-tuned for specific tasks didn't necessarily matter because the big model could still do it.
00:25:44.960 --> 00:25:49.759
And everyone essentially, we were playing with monopoly money until about four months ago.
00:25:49.920 --> 00:25:56.319
So the whole concept of small language models or fine-tuned models literally didn't matter because it was free money.
00:25:56.559 --> 00:26:00.160
Now, because of the whiplash, this is more important than ever.
00:26:00.319 --> 00:26:04.000
So, as we wrap, here's what I want to talk about.
00:26:05.839 --> 00:26:07.599
Reset free, right?
00:26:07.920 --> 00:26:09.359
Model shopping is now big.
00:26:09.440 --> 00:26:11.920
And this is now, I think, part of the executive playbook.
00:26:12.000 --> 00:26:14.079
And we saw Microsoft, right?
00:26:14.240 --> 00:26:26.400
Microsoft is reportedly gonna start moving uh away from open AI and anthropic models, not moving away, but they're gonna start mixing in their own models as well as reportedly deep seek models.
00:26:26.480 --> 00:26:38.160
So even the biggest enterprise customers are starting to realize that for a majority of day-to-day tasks, you might not need the number one model in the world to do a big chunk of the work.
00:26:38.400 --> 00:26:48.640
And I think now it's more of this concept of maybe renting frontier intelligence for those uh, you know, for the ambiguous and risky uh and and changing work.
00:26:49.519 --> 00:26:57.119
But I think the future, which I've talked about, is using the right model for the right purpose at the right time.
00:26:57.279 --> 00:26:59.440
And eventually that will become automatic, right?
00:26:59.519 --> 00:27:00.960
I've actually built some things like that.
00:27:01.119 --> 00:27:06.640
I'm like, I'm like, I wonder why no one else is doing this because it's not like especially since GPT 5.6 came out.
00:27:06.799 --> 00:27:09.920
Like I built skills that essentially model route, right?
00:27:10.000 --> 00:27:17.759
I can use Fable inside of Codex, I can use GPT 5.6 uh inside of Claude Desktop, right?
00:27:17.839 --> 00:27:22.000
If you have a little bit of skill and enough patience, um, you can do those things, right?
00:27:22.079 --> 00:27:27.759
And I've done the same things where um, not that I ever hit my usage, but just a practice, right?
00:27:27.920 --> 00:27:32.880
Where I have I've built in model routing within codecs or you know, chat GPT work.
00:27:33.039 --> 00:27:43.839
So if you are a cost conscious um, you know, company, this is the future of working with multiple models, and you're not just gonna throw, you know, every single request at one big model.
00:27:44.880 --> 00:27:46.880
So here's the new AI playbook.
00:27:46.960 --> 00:27:54.240
You own the workflow, but you're probably just gonna rent the frontier or just exclusively use the frontier for those specific purposes.
00:27:54.400 --> 00:28:02.160
So you have to score every workflow by the stakes, the volume, the privacy, and how much proprietary judgment actually differentiates it.
00:28:02.319 --> 00:28:07.599
So you have to start to default to more of this economical model, right?
00:28:07.680 --> 00:28:18.880
Like you have to start mapping out that workflow and saying, okay, maybe we send the 20% to you know, if you especially if you're using it via the API, to that frontier model, right?
00:28:19.039 --> 00:28:23.759
But I mean, my gosh, talk about middle, middle and lower tier models.
00:28:23.920 --> 00:28:31.920
I mean, open AIs, uh, Terra and Luna, when it comes to a cost efficiency, are legit off the charts, right?
00:28:32.000 --> 00:28:38.480
So if I'm advising companies, you know, I'm saying like you should be using these models for the most part, right?
00:28:38.640 --> 00:28:48.079
The the the Luna model on like max setting is probably more than enough for almost 90% of the work that you would do.
00:28:48.240 --> 00:29:04.880
So it is kind of shifting away from this one model for all purposes to understanding, which I know is is confusing because it is easy to hit that easy button, but we can't just hit that easy button every single time because that button is going to start to cost more and more money as it requires more and more power.
00:29:05.039 --> 00:29:05.599
Uh, right.
00:29:05.759 --> 00:29:15.440
So you have to default to those sometimes more economical models, route by difficulty, and then fine-tune potentially stable, repeated, measurable work.
00:29:15.599 --> 00:29:22.960
And you have to stop asking which model is the smartest and start asking which model is enough for each job.
00:29:23.119 --> 00:29:24.400
All right, that's a wrap.
00:29:24.480 --> 00:29:26.079
I hope this one was helpful.
00:29:26.240 --> 00:29:32.720
The pretty exciting release, maybe not just for the model itself, but more for what it represents.
00:29:32.880 --> 00:29:34.400
So I hope this was helpful.
00:29:34.480 --> 00:29:36.640
If so, please let me know about it.
00:29:36.720 --> 00:29:38.960
Go sign up for our free daily newsletter.
00:29:39.039 --> 00:29:43.680
Drop me a line when you get that automated uh, you know, email, uh, welcome email.
00:29:43.839 --> 00:29:49.680
But also, if you could do me a favor, go subscribe to the podcast on Spotify or Apple Podcasts.
00:29:49.759 --> 00:29:50.720
Appreciate you tuning in.
00:29:50.880 --> 00:29:53.920
We'll see you back tomorrow and every day for more everyday AI.
00:29:54.000 --> 00:29:55.119
Thanks, y'all.
