There's haggling where you actually enjoy the process (i.e. Turkey, visiting the bazaars, you get to haggle, then you get invited for some cay w/ the vendor, talk a bit, that's really enjoyable) and then there's outright scamming. Friend of mine went to Egypt, really wanted to ride a camel. Agreed on a price, rode the camel, and at the end they wouldn't bring the wooden thingy so that he could dismount, and they were asking for more money. That's not enjoyable at all.
> in chess engines that literally have endgame databases for example
You / Carlsen / anyone will not beat a top chess engine even without the endgame databases. In the vast majority of cases you / anyone won't even reach that part (7piece / 8piece for some positions).
> ChatGPT does not know more than you
Yes, yes it does. Your fallacy is that you confuse knowledge with "knowing what to do when you don't have that knowledge". But in pure raw knowledge (definitions, trivia, bits of history, etc) chatbots are oom over any human being. Just try any of the "benchmarks" gamified by people.
> I've read multiple times that this approach is harmful in training.
There's a lot of nuance here. Note that I said "prepare" datasets and not just "generate" datasets.
First, the "model collapse" paper(s) were highly misunderstood and the "media" / content creators ran with it because negativity sells. In that initial paper the authors took things to the extreme, and presented as a given what happens in the literal worse case scenario. They used small models, they generated data w/ those models and indiscriminately trained on that data. It obviously led to model collapse. But that's not what you do in the real world.
The way you do this in the real world is different. For pre-training data you can do things to improve the quality of your inputs:
First, you can use the models to curate your datasets. And this is something that everyone has done since the days of "chug common crawl into the model and see what comes out". It turns out that quality of the data is very important and common crawl is really bad. So we've seen attempts at curating that data. The better the filtering models, the better the initial pre-training data.
Then you can have data augmentation, where you take some piece of content, and generate augmentations for it. Current models are good enough that you can take a piece of "authoritative" text (say a book on writing style) and a bunch of articles, and "improve" them. Or take a piece of content and "translate" it into simple / advanced explanations. Or take a piece of code and "explain" what it does, based on a paragraph from an authoritative book. And so on.
Then, for the mid-training / post-training with RL:
You need to find both good scenarios (i.e. problems) for your model to solve and a good verification schema. Like they say in the quote above, those problems need to be complicated enough for each new model. Here you can again use old models to prepare datasets for the new models.
One simple approach is to take a codebase, have your current model identify a set of features. Then instruct the model to remove code relating to feature "a" but keep its tests. Then verify that every other feature works in the code, bar the one you removed. Then, during RL, you train your new model on that task (you present it as a "prompt" / "situation") and you score the model based on the new feature passing the original tests.
Then there are more advanced ways of using prev gen models for "open ended" problems. You can't really apply RL if the task is not easily verifiable (like above, with tests). But you can use something like RLAIF (reinforcement learning w/ AI feedback) where you grade responses with the previous gen models. Now, in general this is lower quality / lower signal than RLVR (verifiable rewards) but you can still do smart things. Instead of rating an answer good / bad, or ask it one-shot what answer is better, you can use a method based on rubrics. You can first ask the preparing model to select tasks, and a list of rubrics on how that task should be scored (like they generally do on open ended exam questions). Then while doing RL you grade each response by asking the prev gen model to generate said rubrics. Does the answer touch on subject a / b / c? Does the answer mention x y z? Is this mathematically sound? And so on. You still get better results than nothing, even if the task is "open ended". And, again, as models improve so does your pipeline.
> Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools. This dataset lets the model learn both from existing software as well as developer-agent interactions, capturing how developers work and how agents interact with their environments.
This is what the big money was for. Cursor is the first big player that had real-world data from real-world projects, before cc / codex were a thing.
> We used reinforcement learning on difficult problems in realistic environments spanning both software engineering and broader knowledge work. These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results.
> Many of these problems had to be designed to be difficult enough that even frontier models fail at them. As models improve, existing tasks stop teaching them anything new, and problems that once required extensive reasoning become routine.
> We developed a distributed agent system to construct these environments at scale. Engineers specify a problem and how a solution is verified, and large groups of agents construct, test, and refine each environment.
This is where scale comes in. You use the previous gen model to prepare datasets for the next model iteration. The better the models, the better the data, the better the next models. (they also have a comparison with their composer2.5 training run, for people still thinking chinese models are "close to SotA"...)
Reports of xAIs demise (after giving a lot of compute to Anthropic) were slightly exaggerated, it seems.
> Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs
Should as in "would it be nice?" - yeah. Should as in they have to? No.
> Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the “Software”), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so
You can do pretty much anything you want with an MIT license.
It is already here. Not humanoid (yet, but it's in the works) but tracked robots with bolted on machine guns have both held and captured positions in UA.
> With $120,000 owed in back taxes due by you upon purchase.
How does it work in the US? Are taxes on the property itself? This feels weird. I would have thought that the property can only be sold if everything is OK with it (no litigation, liens, etc), and taxes are owed by persons? Is it different over there?
On that aspect, I agree. Smaller / open models are becoming "good enough" at an increasing number of tasks. And that's great for us, consumers. But there will always be tasks that are "worth" pursuing with better models, and cost is irrelevant for those tasks. That was the point I was trying to make.
> The rate of improvement of all models is converging to zero.
That's so obviously not true that I don't even think it's worth the energy to even debate it. It's been said for years, yet here we are, constantly improving. People really don't get RL / the bitter lesson, do they?
> It follows that in a few generations, open models inferencing will be about as good as closed model inferencing.
Not a chance. There's hundreds of billions of dollars on one side, and oom less on the other. There's also scaling laws and information theory. No matter how good, a 30B model will not be able to be better than a 3T+ model, all things being equal.
You are mistaking models becoming "good enough" for an increasingly number of tasks, which I agree is happening, with SotA models stagnating, hitting walls etc. That will not happen for many many years to come.
> but you can achieve the same at 1/10th of the cost.
For some tasks, sure. But not for all tasks. And for some tasks, cost per token is irrelevant if it provides real benefits that are oom compared to what you had.
Local models are indeed becoming "good enough" for some tasks, but there are still tasks that they can't touch. There's a recent benchmark for kernel writing. Fable wrote a kernel that provides ~30% more throughput per unit of compute compared to the latest Opus max / gpt max. Does it matter how much that session cost in terms of one session if you can take that kernel, deploy it on your inference fleet and "magically" get 30% more tokens served to your clients? There are companies that would pay millions for such a "leap". Because they can make more millions down the line.
> Even the nominal enforcers don’t believe in it. No one is motivated to actually enforce it.
Case in point, the ATC on this very flight said something along the lines of "Thanks for the report, I'll pass it on, but I doubt they'll be able to do something about it"...
Every time this subject comes up, there are a bunch of takes along the lines of "would you work on a codebase maintained by agents? they'll mess up the code". And I'm asking myself where these people work, because in 20+ years I've yet to see that pristine state of a project that keeps being pristine after the honeymoon greenfield phase, and 50+ people start working on it. Every project devolves in time, old stuff gets patched in a hurry, someone tries to make it better, learns why certain things were done a certain way, hits some undocumented client needs handled by some arcane combination of code + external systems, and so on. If anything, keeping track of what does what in a project is a task where agents can shine, if only in "ask" mode so you can figure out things quicker. Not to mention onboarding and stuff for new team members.
Isn't this old news? I remember reading about 7yo teslas used exclusively in cold climates (Norway, Finland, etc) and they found the same thing: batteries held on much better than even the manufacturer expected. And those were often 1st gen cars, which you could expect to have teething issues.
It was at the time one of the main reasons the 2nd hand markets in those countries were pretty healthy and saw a lot of movement of used cars.
I agree with you that you don't need fable for everything, and you have to be careful on what you run it on. CRUD stuff, sure even the small models can do it. But there certainly are tasks that are very much suited for the absolute SotA and you'd leave money on the table by not using it. And how much a task is worth is dependant on how much it improves your bottom line. So the cost/token becomes largely irrelevant.
Let's take this [1] benchmark. A bit more context here [2].
Here models are asked to create kernels for running inference on models. This is a benchmark perfectly suited and highly relevant right now. It's easily verifiable, an active are of research, and the results are immediately useful.
Say you have 1 unit of compute, it costs 300k $ and serves 1x users. In comes Fable and after one session it gives you 30% speed-up on your 1 unit of compute. It can now serve 1.3x users. How much is that one session worth for you? How much is it worth for a company using 10 units? 100 units? How much is it worth for a hyper-scaler running 10.000 units? How much is it worth for a lab that trains the next frontier model and then serves it from 100.000 units? 30% is relative. And the cost for one session is really meaningless. It can cost 1m$ / session and it would still be worth it for someone.
It depends. Up until recently the models were trained only to "think" on the last user message. So you'd send the message1, got back reply1 w/ think1 but you'd make the next iteration m1 - r1 - m2, and would get back reply2 w/ think2. You would not add the thinking1. That's how the models were trained, and that's how you were supposed to construct the conversation.
Now recently some things have changed, and you can add the thinking part (you get that encrypted from the closed API labs). But the model needs to have been trained for this to work. And doing it this way you'll burn through tokens faster, as the thinking parts are usually rather long.
You don't get access to the thinking traces. Might work with local models tho, but the current <thinking/> meta isn't particularly suited for this either, as it's a big blob of rambling surfaced by RL, with the "only" objective being that the thinking blob somehow leads to a better final answer. Something more detailed, using templates akin to oAI's harmony could work, provided there's also a step that teaches the models to reflect on the various thinking channels, and maybe surface bits and pieces to include in "skills" or "learnings".
It's not even that. You can have all the designs you need, but you also need a bunch of downstream tech to get from drawings to production. This is something that centrally planned economies struggle with. You can't 5-year-plan your way to jet-engines if you haven't previously 5-year-planned for all the auxiliary infrastructure needed to support that.
We already know this was an issue with the soviets, back when they had the plans for us jet engines (for fighter planes), but couldn't replicate them. Same for stealth, hell even some of their rocketry. And the soviets had plenty of auxiliary systems already in place, during the cold war. As someone said above, they could do quantity, they could do limited high-quality, but couldn't do both at the same time.
There are things that work with 5-year plans: railroads, road infra, buildings, etc. And there are things that are not that easy, and take multiple decades from when the order comes to having it realised. Something that's not immediately obvious for western folks is that when you mix central planning with authoritarian governments, you will get a huge number of pain points along the way, where orders come downwards towards the ones executing them, and overreporting/missrepresentations/lies go upwards. It's like the longest game of telephone, where you start from the top, demanding x y z, get reports that you're on your way of getting 3x, 3y, 3z and in reality you have some of x, none of y, and z looks like z but it's actually three x's in a trench coat.
Heh, for a second there this read like some of the Portiid adventures with Bianca and Fabian from Adrian Tchaikovsky's book "Children of time". If you find this kind of thing interesting and are a fan of sci-fi, I highly recommend the book.
There's haggling where you actually enjoy the process (i.e. Turkey, visiting the bazaars, you get to haggle, then you get invited for some cay w/ the vendor, talk a bit, that's really enjoyable) and then there's outright scamming. Friend of mine went to Egypt, really wanted to ride a camel. Agreed on a price, rode the camel, and at the end they wouldn't bring the wooden thingy so that he could dismount, and they were asking for more money. That's not enjoyable at all.