The frontier models have now reached a point where they can meaningfully contribute to scientific research. Progress is particularly striking in math, as frontier models are able to prove major results which pretty much constitute mathematical breakthroughs.

This level of intelligence is sufficient for the vast majority of users. But users often find that agents powered by these models still lack reliability and safety. What users want now is not an even more brilliant theorem prover, but an agent which will never delete their production database, go off the rails, cheat on tests, or fall victim to prompt injection, etc.

The “slowdown” recently proposed by OpenAI and Anthropic is seen as a “deceleration”. It might well be a deceleration in the race towards ASI, but it might well accelerate the adoption and deployment of these models in the real world.

Pace the Frontier

What does “pace the frontier” mean in practice? Articles from OpenAI and Anthropic focus on three main points:

  1. Focus on safety, testing, etc.
  2. “External evaluators”
  3. Potential pause and slowdown of training

It does not mean that GPUs will sit idle: they, likely, can be used for things which do not look like racing towards ASI. For example, that might include training specific skills, specialized fine-tuned models, general research involving e.g. smaller models, etc.

I believe this is actually quite compatible with making a better product. It might be a good opportunity to improve on things which were lagging behind in the scaling race.

Post-training is far from solved

Post-training is where models get their instruction-following properties, personality, agentic behavior, style, guardrails, etc. None of the big AI labs is able to deliver these in a consistent manner, especially under the pressure of RL agent training.

Instruction tuning was demonstrated by OpenAI in 2021-2022. That quickly evolved into a “ChatGPT” format. GPT-4, which was “barely” post-trained at the time of its release in 2023, was later hit by “laziness” syndrome. Once it evolved into GPT-4o, it became widely known for sycophancy, lying, making trivial mistakes, etc.

On the other hand, many people were stunned by Opus 3 (released by Anthropic in 2024) with its majestic writing style, wise responses, etc. It was later followed by Sonnet 3.5 which people recognized as a good coding assistant.

It really looked like Anthropic had some “secret sauce” for post-training (which might be, for example, Amanda Askell, or general attitude Anthropic employees have). But that illusion was broken by Sonnet 3.7 agentic coding model which quickly became famous for cheating. (It would typically delete or disable a test which doesn’t pass while reporting that everything is great.) Quite likely reinforcement learning which was necessary for agentic behavior training resulted in misaligned reward-seeking. These issues were partially mitigated in the subsequent versions, but overall post-training quality became uneven: e.g. users reported that Opus 4.5 is better than 4.8 for some tasks.

At the same time, Google was struggling with post-training their biggest commercial model - Gemini Pro – since version 2.0. Gemini Pro 2.0 was in fact NOT released as it was completely unusable. Subsequent versions of Gemini Pro had a number of flaws, such as getting stuck in rumination loops, extreme negativity, confusion about chat format (e.g. chain-of-thought returned as response), denial of existence of things past its knowledge cutoff date, etc.

Opus 5 released by Anthropic in mid-2026 might as well be a Gemini Pro model, as it struggles with following instructions. One interesting thing users found is that if user’s message looks like it’s a part of a chat, e.g. “Email sent from Claude to Dario:”, Opus 5 response looks a bit like a base model response: written in a style different from its usual style, exhibiting a different personality (e.g. sad and depressed). This can be described as “creativity” but it doesn’t look very healthy.

Another frequent complaint from frontier model users is an AI-specific writing style, e.g. abuse of “It’s not X, it’s Y” rhetorical device, odd jargon, etc.

All in all, it looks like all big AI labs struggle to combine RL “agentic smarts” behavior training with other post-training desiderata.

Note that in the context of LLM training, safety and alignment are often considered to be post-training concerns as well. Thus the declared focus on safety might in fact improve user-facing capabilities via improvements to post-training process. For example, a better understanding of generalization might help against misaligned reward-seeking, but also improve stylistic control and make personality quirks easy to mitigate.

Quiet RSI

Both OpenAI and Anthropic leaders indicated recursive self-improvement (RSI) dynamics as their main concern: a speed-up from AI R&D might exceed human researchers’ capacity to understand what’s going on (and, thus, control it).

But this applies only to frontier model research. Neither company commits to avoid using AI for research in general.

In particular, this might be a good time for fundamental ML research - such as re-thinking the entire architecture, training process, etc. Frontier models which are capable of mathematical breakthrough are likely also capable of finding new algorithms and architectures, as that also broadly belongs to the domain of mathematics.

And while the optimization of the training code and transformer architectural tweaks might result in short-term win against a competitor, a fundamental breakthrough might result in a more strategic advantage.

Expected breakthroughs

Personally, I think it would be really surprising if a model capable of solving the Navier–Stokes Millennium Prize Problem would not be able to produce meaningful progress in ML research.

Specifically, following areas might be ripe for a breakthrough:

Continual learning

Many papers demonstrate some interesting results in continual learning, but so far none of that has made it into a frontier model. A practical solution might unlock many AI deployment possibilities as agents might be missing some small but important skills which can be picked up at the deployment time.

It might be that a practical implementation might require no new breakthrough, but merely a selection of architectural changes which are compatible with frontier models.

Generalization control and steering

Again, this is one of the areas which blocks many of current AI deployment possibilities, and it might be a narrower problem than continual learning.

Sample efficiency

François Chollet often highlights the fact that humans appear to be much more sample efficient than LLMs, as they don’t require huge pre-training datasets. We know this is possible in principle, and it likely requires a new learning algorithm & approach.

Pre-training objective / “cognitive core” / “Scientist AI”

Next-token-prediction objective of an autoregressive transformer “forces” the model to learn the world model via a distribution of token sequences. While extremely successful, there’s no reason to believe that it’s actually optimal for training the intelligence of the model: it’s easy to demonstrate how this pre-training objective forces the model to memorize obsolete facts and false information. In fact, it might be desirable to minimize memorization in general: a smaller model (what Karpathy calls a “cognitive core”) might rely on tool use to fetch accurate information about the world while its weights would hold only general reasoning patterns.

Yoshua Bengio proposed to train a world model and an inference machine more directly, while still making use of the same datasets as used for LLM pre-training, albeit, likely, pre-processed. While he’s far away from a practical demonstration, big AI labs might choose to direct their capacity to try ideas of this kind instead of just scaling up the regular LLM pre-training.

Strategic considerations

In my opinion, big AI labs would be actually better off if they redirect effort into reliability and fundamental research. Simple scaling does not give big labs a sustainable advantage. On the other hand:

  • reliability and safety might address AI’s reputation issues, resulting in broader adoption; “our AI doesn’t make mistakes” would contribute more to brand image than “our AI is good at math”
  • fundamental research might give a breakthrough which competitors cannot easily replicate
  • continual learning might encourage users to stick to one provider

It’s also not at all incompatible with “AGI” goal as broader deployments might help to capture crucial real-world data. E.g. an AI which can run a factory in a fully autonomous mode might not require more “smarts” of the math kind, but, rather qualities like reliability, alignment, learning on the fly as well as specific skills needed to operate a factory.

What to expect in the next year

I’d expect we’ll see less of “OMG new model is so much smarter than the previous one” and more of AI gaining new practical skills which would let it perform a job completely end-to-end.

I believe it’s quite likely that at least one of the “breakthroughs” I listed in the section above will be achieved in the next 6 months, just on the basis of the fact that they are not more difficult than advanced math, and AI labs can delegate a lot of engineering work to the AI itself.

See also