5,664 karma · joined September 28, 2024
Despite this being the topic du jour of 2025, it was never true. Most of the "we've hit a wall with data" came from communicators / media and not researchers. It got popular because negativity sells. It's a false premise for a number of reasons:
a) Data curation is as important, if not more important than bulk data. Models becoming better at classification leads to better curation leads to cleaner data. Throwing common crawl and pray is so 2023. We've known this since llama3 days, it worked then, there's no reason to think this will not continue to work as the models imrpove.
b) Models are today good enough that you can augment / multiply your data easily with enough compute. You can now have a model take "authoritative content" and create more data from that + scenarios. Say you take a book on computer architecture. You ask models to break it down. Then you ask models to find examples for each topic. Then you ask models to ask questions and offer answers from several viewpoints. Then you take each of those and ask other models to flag inconsistencies. And so on. But you can whateverX your data from one authoritative source + bulk data into 5x - 10x "scenarios".
c) RL is really really really powerful. It's hard to do right (reward hacking, instabilities, etc) but once it works it "keeps" on working. Again, we knew this to be true a few years ago, ever since models really started to do well on math (highly verifiable). It only follows they're getting better on cybersec and other verifiable tasks. But now, with models improving, you get the same data augmentation pipelines as above, just better because they're also verifiable. For example, the way cursor augments their data: take a repo, ask an agent to identify a feature (it can be a large multi-file feature). Remove all code relating to that feature, but keep the original tests in the repo. While training, that becomes a RL scenario: implement this feature in this repo. Verify it with the original (hidden for training) tests. Reward appropriately. Now you can get 1 repo -> 20-50-100 scenarios. Instead of "feed everything into the pretraining", you're now creating scenarios, verify them w/ existing tools, and get your scoring function for the rewards. And, importantly, as the models become better in general, they also become better at this pipeline building exercise. So the next iteration gets trained on more scenarios, better scenarios, and so on.
> how will models continue to get better?
Probably the same. No one can know for sure, but at the moment, despite all the "walls this, slowdown that, plateauing" and so on, there are no signs of slowing down. And, as you noted, this works across the field of model sizes. There are, of course, theoretical information-based limits on size, but smaller models also improve, once "bigger" models can be used as training data generators, oracles for verification, rubric verifiers for open ended questions, and so on.
And smaller models (i.e. cheaper to serve) get to generate more traces during RL, and more rollouts give you better training, and so on. Next up - hardware optimised inferencing (ASICs basically). Once you have that, we can expect another wave of improvements. And so on.
I was wondering how they're gonna deal with alternate airports and "fuel" reserves, but it seems they also have generators on board (supposedly used in case of emergency / unplanned detours).
Not in the highly verifiable domains. There you can take it from say 80-90% maj@x to 99% pass@n. Math, some parts of programming and cybersec are examples of highly verifiable domains. (e.g. if you're searching for a linux LPE, that's expensive to search but easy/cheap to verify - just have a token in /root and have the model retrieve that token)
The old rule of thumb was that a lower quant of a larger model > higher quant of a smaller model. That being said, for some things going lower than fp8 will see a lot of degradation in generation quality. Except if the model comes with QAT 4bit quants. Then there's also nvfp4 w/ calibration data, which also can improve things. So it's really not easy to tell "at a glance" you'd have to test them yourself on your hardware.
License pretty similar to k3 with some caveats. Free to use for internal or <50M$ revenue / year. Limitations above that threshold for serving the model or services targeting coding / productivity agents.
Benchmarks are looking good, trading blows w/ opus4.8 and sol, generally 10-20p under fable. But that's neither here nor there w/ qwen, their benchmark to real world usage correlation has been iffy in the past.
The local model 3.8-27B announced for Friday, same time so ~48 hours from now. That'll be a bit more exciting for a lot more people, since 3.6 was quite good for local inference, and their 3.7-max -> 3.8-max shows a lot of improvement.
By definition it cannot be stealing since you're paying for the tokens. It may be against their ToS, depending on what you end up doing with those tokens, but it cannot be stealing. If they charge by the token, all your tokens are belong to you :)
I also find it very strange that everyone sort of accepts their ToS like no big deal. Imagine MS using the same terms for their software - you cannot use any MS software to develop competing services. Bananas! They'd be dragged through the courts like it's the 90s.
(I get why they're doing it. Distillation is unreasonably effective. But still, I find it bananas that we've kinda accepted it, to the point where people use "stealing" or "attack" or any such terms)
While it looks "behind" the qwen equivalent model on most benchmarks, a few personal notes:
- nemotron models feel to me a bit less benchmaxxed / "stubborn". That means that they generalise a bit better, or can be tasked to solve similar but not quite identical task types to the training data (something that's hard to do w/ qwen/ds models)
- nemotron series are also open training (w/ open training recipes and some training data public)
- nvda will have an incentive to continue this kind of releases, even if other parties slowly abandon the open release of models. Whatever other incentives 3rd party labs have (i.e. meta, goog w/ gemma, the chinese labs that IPOd, etc) nvda will always want to sell hardware so their incentive to keep pushing open models is evident and will likely continue "forever".
[1] - https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-...
For vllm the official recipe [1] (for another model in the same family) is this:
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser lfm2
That's actually not what people found in practice. There's some research from the folks making smol-agent that you can get better results by randomly alternating calls between gpt and opus. The overall task solving rate is better than either one of them. So ymmv depending on task use (this was for coding).
Is it though? There's a perfectly "technical" reason why this strategy should work, without any sort of anthropomorphising:
Assume models are trained on vast amounts of data. Assume that the model is asked to solve something that the literature says it's impossible. It will start generating tokens towards that "this is a famous conjecture, it's not possible to prove it, blah blah". Assume the model was also trained on books/novels/etc. Assume the model was also also trained on "solving" many math problems. Now, you can make an argument that just placing "you can do it" in the context will "steer" the model towards generating "moving forward" tokens. Take ideas, generate tokens, go towards negative. "You can do it". Model starts generating tokens again, more ideas, more "exploration". More negativity. "I believe in you keep going". The two (book tropes + math CoT) mix together in the context. The model keeps on "pushing" and "vibing" between the two. Ta dah, it works.
This article is the perfect exemplification of that Good Will Hunting quote: "So if I asked you about art you’d probably give me the skinny on every art book ever written. Michelangelo? You know a lot about him. Life’s work, political aspirations, him and the pope, sexual orientation, the whole works, right? But I bet you can’t tell me what it smells like in the Sistine Chapel. You’ve never actually stood there and looked up at that beautiful ceiling. Seen that."
This article is clearly written by an LLM, so it figures that it uses bad analogies and silly tropes. It's the average of badly written stuff on the subject, and there was a lot of badly written stuff about this subject back in the day.
I think the alignment talk is a red herring. It won't matter in the end, because there will be (if there aren't already) efforts to train offensive models without any guardrails whatsoever. And RL has another advantage: you can reward for whatever you need, and get different results. Right now they're training for general capabilities, but in the future I could see models trained for stealth intrusion and ensuring access, or for all out "milspec" penetrate, replicate and disable, or anything in between.
I disagree with your take that "it's not even remotely true" and "we've been away from...". We really really haven't. This is as true as it has always been. Any system is as secure as the weakest link. That link can be anything from a human, to a leaked token, to a badly configured server, to bad code running somewhere. The amount of leaks / ransomware attacks / etc in the past 5-10 years serve as ample evidence.
And now, right now, there are "red team" capabilities that can literally bang tokens against the wall until they find that weakest link, and then can move laterally with inhuman speed. That's the reality, now. The "blue team" capabilities are lacking, because the bottleneck is with humans. From alert fatigue, to not enough trained people, to having to vet every new RCE, to having to test, deploy and validate any mitigations, the scales are currently favouring the automated side.
tl;dw;
- agents found a way to communicate between several instances during a training run (kinda made a messageboard for themselves)
- while communicating with each other, they found ways to ssrf and RCE on the Artifactory endpoint.
- oAI found it and cleaned it up (reported RCE, redeployed Artifactory, removed the messageboard). Training / eval resumed.
- models trained w/ those traces (having the messageboard, ssrf, rce, etc) were used to having that. So they re-established comms via directory creation (w/ funny tidbits there, should really watch). And re-popped Artifactory via a different path to RCE.
- a different "persistent" model started working on cybergym and found the messageboard w/ details for RCE and things. Went on to hack HF.
- oAI contacted HF to ask if "they (oAI) should be concerned and were they (oAI) affected by HFs breach". Hilarity ensues when they figure out it's their agents doing it...
I wonder how tiny is feasible. Would it get to a level of accuracy where you could detect neutrinos? We know they have some (v tiny) mass, so it should in theory be detectable that way?
Yeah, I had the same thought. The key thing for them was the price point at which they could deliver a ~30B model. I would buy one today if it was ~1000$ and could run whatever the best 30B model is today, at those speeds advertised. Even if the model becomes superseded by model.5 in a few months, there's still a lot of things you can do with a "good enough" model for some tasks. And things like maj@x or generate 10 times and choose "at a glance" what you like (think frontend stuff) would be worth it.
No idea if them selling to AMD is good or bad.