Bienvenue sur Import AI, une newsletter sur la recherche en IA. Import AI fonctionne sur arXiv et sur les commentaires des lecteurs. Si vous souhaitez soutenir cela, veuillez vous abonner.

Les systèmes d’IA sont sur le point de commencer à se construire eux-mêmes. Qu'est-ce que cela signifie?

J'écris cet article parce que lorsque je regarde toutes les informations accessibles au public, j'arrive à contrecœur à l'idée qu'il y a une chance probable (60 % +) que la R&D sur l'IA sans implication humaine - un système d'IA suffisamment puissant pour qu'il puisse de manière plausible construire de manière autonome son propre successeur - ait lieu d'ici la fin de 2028.
C'est une grosse affaire.
Je ne sais pas comment comprendre cela.
C’est un point de vue réticent, car les implications sont si importantes que je me sens éclipsé, et je ne suis pas sûr que la société soit prête pour le type de changements qu’implique la réalisation d’une R&D automatisée sur l’IA.
Je crois maintenant que nous vivons à une époque où la recherche sur l’IA sera automatisée de bout en bout. Si cela se produit, nous franchirons un Rubicon et nous mènerons à un avenir presque impossible à prévoir. Nous en reparlerons plus tard.

Le but de cet essai est d’énumérer les raisons pour lesquelles je pense que le décollage vers une R&D entièrement automatisée sur l’IA est en train de se produire. Je discuterai de certaines des conséquences de cela, mais je prévois surtout de consacrer la majorité de cet essai à discuter des preuves de cette croyance, et je passerai la majeure partie de 2026 à étudier les implications.

In terms of timing, I don’t expect this to happen in 2026. But I think we could see an example of a “model end-to-end trains it successor” within a year or two - certainly a proof-of-concept at the non-frontier model stage, though frontier models may be harder (they’re a lot more expensive and are the product of a lot of humans working extremely hard).
My reasoning for this stems primarily from public information: papers on arXiv, bioRxiv, and NBER, as well as observing the products being deployed into the world by the frontier companies. From this data I arrive at the conclusion that all the pieces are in place for automating the production of today’s AI systems - the engineering components of AI development. And if scaling trends continue, we should prepare for models to get creative enough that they may be able to substitute for human researchers at having creative ideas for novel research paths, thus pushing forward the frontier themselves, as well as refining what is already known.

Upfront caveat
For much of this piece I’m going to try to assemble a mosaic view of AI progress out of things that have happened with many individual benchmarks. As anyone who studies benchmarks knows, all benchmarks have some idiosyncratic flaws. The important thing to me is the aggregate trend which emerges through looking at all of these datapoints together, and you should assume that I am aware of the drawbacks of each individual datapoint.

Now, let’s go through some of the evidence together.

The coding singularity - capabilities over time:
AI systems are instantiated via software and software is made out of code.

AI systems have revolutionized the production of code. This has happened due to two related trends: AI systems have gotten better at writing complicated real-world code, and AI systems have gotten much better at chaining together many linear coding tasks (e.g, writing code, then testing it) independent of human oversight.

Two things that exemplify this trend are SWE-Bench and the METR time horizons plot.

Solving real-world software engineering problems:
SWE-Bench is a widely used coding test which evaluates how well AI systems can solve real world GitHub issues. When SWE-Bench launched in late 2023 the best score at the time was Claude 2 which had an overall success rate of ~2%. Claude Mythos Preview gets 93.9%, effectively saturating the benchmark. (All benchmarks have some amount of noise inherent to them, so there’s usually a point where you score high enough that you are running into the limitations of the benchmark itself rather than your method - for instance, about 6% of the labels in the ImageNet validation set are wrong or ambiguous).
SWE-Bench is a reliable proxy for the general issue of coding competency and the impact of AI on software engineering. The vast majority of people I meet at frontier labs and around Silicon Valley now code entirely through AI systems. Increasingly, they use AI systems to write the tests and check the code as well. In other words, AI systems have gotten good enough to automate a major component of AI R&D, speeding up all the humans that work on it.

Mesurer la capacité d’un système d’IA à accomplir des tâches qui prennent beaucoup de temps :
METR crée un graphique qui nous renseigne sur la complexité des tâches que les IA peuvent accomplir, mesurée par le nombre d'heures qu'un humain qualifié prendrait pour les accomplir. La mesure clé ici est celle qui vous indique l’horizon temporel approximatif sur lequel les systèmes d’IA peuvent être fiables à 50 % pour un ensemble de tâches. Ici, les progrès ont été extrêmement frappants : en 2022, GPT 3.5 pourrait effectuer des tâches qui pourraient prendre environ 30 secondes à une personne. En 2023, cette durée est passée à 4 minutes avec GPT-4. En 2024, celle-ci est passée à 40 minutes (o1). En 2025, il a atteint environ 6 heures (GPT 5.2 (Élevé)). En 2026, elle s'élève déjà à ~12 heures (Opus 4.6). Ajeya Cotra, prévisionniste de longue date en IA qui travaille au METR, pense qu'il n'est pas déraisonnable de s'attendre à ce que les systèmes d'IA effectuent des tâches qui prennent environ 100 heures d'ici la fin de 2026 (#448). Cette augmentation significative de la durée pendant laquelle les systèmes d’IA peuvent fonctionner de manière indépendante est étroitement liée à l’explosion des outils de codage agent – ​​il s’agit de la production de systèmes d’IA qui travaillent pour le compte des personnes, agissant de manière indépendante pendant des périodes de temps significatives. Cela renvoie également à la R&D en IA, où si vous examinez de près le travail de nombreux chercheurs en IA, une grande partie de leurs tâches se résument à des tâches qui peuvent prendre quelques heures à une personne : nettoyer les données, lire les données, lancer des expériences, etc.

All of this kind of work now sits inside the time horizon scope of modern systems.

The more skilled AI systems get and the better they get at working independently of us, the more they can help automate chunks of AI R&D
Key ingredients in delegation are a) confidence in the skills of the person, and b) confidence in their ability to work independently of you in a way that is aligned with your intentions.
When we look at the competency of AI at coding, it seems that AI systems are getting far more skilled and also able to work independently of people for longer and longer periods before needing re-calibration.
This correlates with what we see around us - engineers and researchers are now delegating larger and larger chunks of their work to AI systems, and as capabilities rise, so too does the complexity and importance of the work being delegated.

AI is getting good at core science skills essential to AI R&D
Think about modern science - a huge amount of it is about specifying a direction where you want to generate some empirical information, running experiments to generate that information, then sanity-checking the results of the experiment. The combination of advances in coding over time combined with the general world modeling capabilities of LLMs has yielded tools that are already helping to speed up human scientists and partially automate aspects of R&D broadly.

Here, we can look at the rate of AI progress in a few key scientific skills which are inherent to AI research itself: Replicating research results, chaining together machine learning techniques and other approaches to solve technical problems, and optimizing AI systems themselves.

Implementing entire scientific papers and doing the experiments:
One core job of AI research is reading scientific papers and reproducing their results. Here, there has been dramatic progress on a wide range of benchmarks.

One good example is CORE-Bench, the Computational Reproducibility Agent Benchmark. This benchmark challenges AI systems to “reproduce the results of a research paper given its repository. The agent must install libraries, packages, and dependencies and run the code. If the code runs successfully, the agent needs to search through all outputs to answer the task questions.” CORE-Bench was introduced in September 2024 and the best scoring system at the time was a GPT-4o model in a scaffold called CORE-Agent which scored ~21.5% on the hardest set of tasks in the benchmark.
In December 2025 one of the authors of CORE-Bench declared the benchmark ‘solved’, with an Opus 4.5 model achieving 95.5%.

Building entire machine learning systems to solve Kaggle competitions:
MLE-Bench is an OpenAI-built benchmark which examines how well AI systems can compete (offline) in “75 diverse Kaggle competitions across a variety of domains, including natural language processing, computer vision, and signal processing.” At launch in October 2024, the top scoring system (an o1 model inside an agent scaffold) got 16.9%. As of February 2026, the best scoring system (Gemini3 inside an agent harness with search) gets 64.4% .

Kernel design:
One of the harder tasks in AI development is kernel optimization, where you write and refine the code that maps specific operations, like matrix multiplication, to the underlying hardware. Kernel optimization is core to AI development because it defines the efficiency of both training and inference - how much compute you can effectively utilize to develop an AI system, and once you’ve trained a model, how efficiently you can convert that compute into inference.

In recent years, AI for kernel design has gone from a curiosity to a competitive area of research and several benchmarks have emerged. None of these benchmarks are especially popular, so we can’t easily model progress over time. On the other hand, we can look at some of the research being done to get a feel for the progress.
Some of the types of work include: Using DeepSeek’s models to try to build better GPU kernels (#400), automating the conversion of PyTorch modules to CUDA code (#401), Meta using LLMs to automate the generation of optimized Triton kernels for use within its infrastructure (#439), using LLMs to help write kernels for non-standard hardware like Huawei’s Ascend chips (”AscendCraft” #444), fine-tuning open weight models for GPU kernel design (”Cuda Agent”, #448).

One caveat here is that kernel design does have some properties that make it unusually amenable to AI-driven R&D, like having easily verifiable rewards.

Affiner les modèles linguistiques via PostTrainBench
Une version plus difficile de ce type de test est PostTrainBench (#449), qui voit dans quelle mesure différents modèles frontières peuvent prendre des modèles à poids ouvert plus petits et les affiner pour améliorer les performances sur certains points de référence. La caractéristique intéressante de ce benchmark est que nous disposons de références humaines extrêmement bonnes - les versions existantes « optimisées par les instructions » de ces modèles, qui ont été développées par de talentueux chercheurs en IA humaine travaillant dans des laboratoires frontaliers. Ces modèles ont été développés par des chercheurs et des ingénieurs extrêmement talentueux et déployés dans le monde entier. Ils représentent donc une base humaine très difficile à surmonter.
Depuis mars 2026, les systèmes d’IA sont capables de post-entraîner des modèles pour obtenir environ la moitié de l’amélioration que ceux entraînés par des humains.
Les scores d'évaluation spécifiques sont dérivés d'une "moyenne pondérée prise sur tous les LLM post-formés (Qwen 3 1.7B, Qwen 3 4B, SmolLM3-3B, Gemma 3 4B) et les benchmarks (AIME 2025, Arena Hard, BFCL, GPQA Main, GSM8K, HealthBench, HumanEval). Pour chaque exécution, nous demandons à un agent CLI de maximiser les performances d'un LLM de base spécifique sur un référence spécifique. »
Les systèmes les mieux notés en avril obtiennent entre 25 et 28 % (Opus 4.6 et GPT 5.4), contre un score humain de 51 %. C’est déjà assez significatif.

Optimisation de la formation du modèle de langage :

For the last year Anthropic has reported how well its systems do at an LLM training task which is described as tasking its models to “optimize a CPU-only small language model training implementation to run as fast as possible”. The score is the average speedup over the unmodified starting code and progress has been striking: Claude Opus 4 achieved a 2.9× mean speedup in May 2025; this rose to 16.5× with Opus 4.5 in November 2025, 30× with Opus 4.6 in February 2026, and 52× with Claude Mythos Preview in April 2026. To calibrate on what these numbers mean, it is expected to take a human researcher 4 to 8 hours of work to achieve a 4x speedup on this task.

Conducting AI alignment research:
Another Anthropic result is a proof-of-concept of Automated Alignment Research (#454); here, an Anthropic researcher primes a team of individual AI agents with a research direction, then they autonomously go and try to get a better score than a human baseline on an AI safety research problem (specifically, scalable oversight). The approach works, with the AI agents coming up with techniques that beat the Anthropic-designed baseline. However, this is done at a relatively small scale and doesn’t (yet) generalize to a production model. Nonetheless, it’s proof that you can apply today’s AI systems to contemporary cutting-edge research problems and we already see meaningful signs of life. All of the above mentioned benchmarks once looked like this, too, and then after a few months or at most a year, AI systems got dramatically better at whatever the benchmarks were testing.

Meta-skills: management
AI systems are also learning to manage other AI systems. This is visible in broadly deployed products like Claude Code or OpenCode, where a single agent can end up supervising multiple sub-agents. This allows AI systems to work on large-scale projects that require multiple individual ‘workers’ each with different specialisms that work in parallel, typically under the direction of a single AI manager (which, here, is an AI system).

Is AI research more like discovering general relativity or Lego ?
Can AI invent new ideas that help it improve itself, or are these systems best equipped for the unglamorous, brick-by-brick work required for research? This is an important question for figuring out the extent to which AI systems can end-to-end automate AI research itself. My sense is that AI cannot yet invent radical new ideas - but the technology may not need to for it to automate its own development.

As a field, AI moves forward on the basis of doing ever larger experiments that utilize more and more inputs (e.g, data and compute). Every so often, humans come up with some paradigm-shifting idea which can make it dramatically more resource efficient to do things - a good example here is the transformer architecture and another is the idea of mixture-of-expert models. But mostly the field of AI moves forward through humans methodically going through some loop of taking a well performing system, scaling up some aspect of it (e.g, the amount of data and compute it is trained on), seeing what breaks when you scale it up, figuring out the engineering fix to allow it to scale, then scaling it again. Very little of this requires extremely out-of-leftfield insights and a lot of it seems more like unglamorous ‘meat and potatoes’ engineering work.
Similarly, a lot of AI research is about running variations of existing experiments where you explore the outcomes of using different parameters, though research intuitions can help pick the most fruitful parameters to vary, you can also automate this and have the AI figure out which parameters to vary (an early version of this was neural architecture search).

Thomas Edison said that “genius is 1% inspiration and 99% perspiration”. Even 150 years later, this feels right. Very occasionally new insights come along which transform a field. But mostly, the field has moved forward through humans sweating a lot of pain out on the schlep of improving and debugging various systems.
As the public data above shows, AI has got extremely good at performing many of the essential schlep components of AI development. Along with this, the meta-trend of basic capabilities like coding combined with an ever-expanding time horizon, means AI systems are able to chain together more and more of these tasks into complex sequences of work.
This means even if AI systems are relatively uncreative, it feels safe to bet they can push themselves forward - albeit at a slower rate than if they’re able to generate novel insights. But if you look at the public data, here too there are tantalizing signs that AI systems may be able to be creative in a way that lets them advance themselves in more impressive ways.

Pushing forward the frontier of science
We have some very preliminary signs that general-purpose AI systems can push forward the frontiers of human science, though this has so far only happened in a couple of domains - primarily computer science and mathematics - and often it happens less through AI systems acting alone and more them acting in partnership with humans in a centaur configuration.

Nonetheless, it’s worth observing the trends:

Erdos Problems: A team of mathematicians worked with a Gemini model to see how well it could tackle some Erdos math problems. After directing the system to attack around 700 problems they came up with 13 solutions. Of these solutions, 1 was deemed by them to be interesting: “We tentatively believe Aletheia’s solution to Erdős-1051 represents an early example of an AI system autonomously resolving a slightly non-trivial open Erdős problem of somewhat broader (mild) mathematical interest, for which there exists past literature on closely-related problems,” they wrote. (#444).

Centaur math discovery: Researchers with the University of British Columbia, University of New South Wales, Stanford University, and Google DeepMind published a new math proof which was built in close collaboration with some AI-based math tools built at Google. “The proofs of the main results were discovered with very substantial input from Google Gemini and related tools,” they wrote. (#441).

If you squint, you could argue that this is a sign that AI systems are developing some of the field-advancing creative intuitions that humans have. But you could just as easily say that math and CS could be unusual domains that are oddly amenable to AI-driven invention, and might end up being exceptions that prove a larger rule. Another example here is Move 37, though I’d contend that the fact it’s been ten years since the AlphaGo result and that Move 37 hasn’t been replaced by some incredibly impressive more modern flash of insight is another weakly bearish signal here.

Putting it all together
If I put this all together the picture from all of the above evidence I end up with is the following facts:

AI systems are capable of writing code for pretty much any program and these AI systems can be trusted to independently work on tasks that’d take a human tens of hours of concentrated labor to do.

AI systems are increasingly good at tasks that are core to AI development, ranging from fine-tuning to kernel design.

AI systems can manage other AI systems, effectively forming synthetic teams which can fan out and attack complex problems, with some AI systems taking on the roles of directors and critics and editors and others taking on the role of engineers.

AI systems can sometimes out-compete humans on hard engineering and science tasks, though it’s hard to know whether to attribute this to inventiveness or mastery of rote learning.

To me, this makes a very convincing case that AI can today automate vast swathes, perhaps the entirety, of AI engineering. It is not yet clear how much of AI research it can automate, given that some aspects of research may be distinct from the engineering skills. Regardless, it all feels to me like a clear sign that AI is today massively speeding up the humans that work on AI development, allowing them to scale themselves through pairing with innumerable synthetic colleagues.

Finally, the AI industry is literally saying that AI R&D is its goal: OpenAI wants to build an “automated AI research intern by September of 2026”. Anthropic is publishing work on building automated alignment researchers. DeepMind appears to be the most circumspect of the big three, but still says “automation of alignment research should be done when feasible”. Automating AI R&D is also the goal of numerous startups: Recursive Superintelligence just raised $500m with the goal of automating AI research, and another neolab, Mirendil, has the goal of “building systems that excel at AI R&D.”
In other words, the combined efforts of hundreds of billions of existing and new capital is being sunk into entities that have the goal of automating AI R&D. We should surely expect at least some progress in this direction as a consequence.

Why this matters
The implications of this are profound and under-discussed in popular media coverage of AI R&D. I’ll list a few here. This isn’t a comprehensive list, but it gestures at the enormity of the challenges AI R&D introduces. .

Nous devons parvenir à un bon alignement : les techniques d’alignement qui fonctionnent aujourd’hui pourraient échouer en raison d’une auto-amélioration récursive, à mesure que les systèmes d’IA deviennent beaucoup plus intelligents que les personnes ou les systèmes qui les supervisent. Il s’agit d’un domaine très bien couvert, je vais donc simplement souligner brièvement quelques-uns des problèmes :
- Entraîner les systèmes d'IA à ne pas mentir et à tricher est étonnamment subtil (par exemple, malgré tous les efforts déployés pour créer de bons tests pour les environnements, il arrive parfois que la meilleure façon pour une IA de résoudre ce problème est de tricher, lui apprenant ainsi que tricher est une bonne chose)
- Les systèmes d'IA pourraient être capables de « simuler l'alignement » en produisant des scores qui nous font penser qu'ils se comportent d'une certaine manière qui cache en réalité leurs véritables intentions. (En général, les systèmes d’IA savent déjà quand ils sont testés.)
- À mesure que les systèmes d'IA commencent à contribuer davantage au programme de recherche fondamentale pour leur propre formation, nous pourrions finir par changer considérablement la manière globale dont les systèmes d'IA sont formés et ne pas disposer de bonnes intuitions ou de bonnes bases intellectuelles pour comprendre ce que cela signifie. - Il existe des problèmes très basiques d'« erreur composée » chaque fois que vous mettez quelque chose dans une boucle récursive qui touche probablement tous les problèmes ci-dessus et d'autres : à moins que votre approche d'alignement ne soit « précise à 100 % » et ait une base théorique pour continuer à être précise avec des systèmes plus intelligents, alors les choses peuvent mal tourner assez rapidement.

For example, your technique is 99.9% accurate, then that becomes 95.12% accurate after 50 generations, and 60.5% accurate after 500 generations. Uh oh!

Everything that AI touches gets a massive productivity multiplier: In the same way AI is dramatically improving the productivity of software engineers, we should expect the same thing to happen for everything else that AI touches. This introduces a couple of issues we’ll have to contend with: 1) inequality of access: assuming that demand for AI continues to outstrip compute supply, we’ll have to figure out where to allocate AI to maximize a social upside. By default, I am skeptical that market incentives guarantee us the best societal upside from limited AI compute. Figuring out how to allocate the acceleratory capabilities conferred by AI R&D will be a politically charged problem. 2) ‘Amdahl’s Law’ for the economy: as AI flows into the economy, we’ll discover places where things break or slow under the increased volume, and we’ll need to figure out how to fix those weak links in the chain. This may be especially pronounced in areas where you have to reconcile the fast-moving digital world with the slow-moving physical world, like drug trials for new medical therapies.

La formation d’une économie à forte densité de capital et peu humaine : toutes les preuves ci-dessus en faveur de la R&D en IA soulignent également la capacité croissante des systèmes d’IA à gérer également des entreprises de manière autonome. Cela signifie que nous devrions nous attendre à ce qu’une part croissante de l’économie soit colonisée par une nouvelle génération d’entreprises qui sont soit lourdes en capital (parce qu’elles possèdent beaucoup d’ordinateurs), soit lourdes en dépenses d’exploitation (parce qu’elles dépensent beaucoup d’argent en services d’IA sur lesquels elles créent de la valeur), et relativement légères en main d’œuvre par rapport aux entreprises d’aujourd’hui – parce que la valeur marginale de dépenser plus en IA par rapport au travail humain augmentera constamment en raison de l’expansion soutenue des capacités des systèmes d’IA. En pratique, cela ressemblera à l’émergence d’une « économie machinique » qui se développe au sein d’une « économie humaine » plus large, même si l’on peut s’attendre à ce qu’au fil du temps, l’économie machinique interagisse de plus en plus avec elle-même à mesure que les entreprises gérées par l’IA commencent à commercer entre elles. Cela aura des effets profondément étranges sur l’économie et soulèvera toutes sortes de questions sur les inégalités et la redistribution. À terme, il sera peut-être possible d’assister à l’émergence d’entreprises entièrement autonomes dirigées par des systèmes d’IA eux-mêmes, ce qui exacerberait tous les problèmes ci-dessus, tout en posant également de nombreux nouveaux défis en matière de gouvernance.

Staring into the black hole:
Given all of this, I think there’s a ~60% chance we see automated AI R&D (where a frontier model is able to autonomously train a successor version of itself) by the end of 2028. Based on the above analysis, you might ask why I don’t expect this in 2027? The answer is that I think AI research contains some requirement for creativity and heterodox insights to move forward - so far, AI systems haven’t yet displayed this in a transformative and major way (though some of the results on accelerating math research are suggestive of this). If you had to push me for a 2027 probability, I’d say 30%. If we don’t see it by the end of 2028, then I think we will have revealed some fundamental deficiency within the current technological paradigm and it’ll require human invention to move things forward.

I have written this essay in an attempt to coldly and analytically wrestle with something that for decades has seemed like a science fiction ghost story. Upon looking at the publicly available data, I’ve found myself persuaded that what can seem to many like a fanciful story may instead be a real trend. If this trend continues, we may be about to witness a profound change in how the world works.

Thanks to Andrew Sullivan, Andy Jones, Holden Karnofsky, Marina Favaro, Sarah Pollack, Francesco 🇫🇷 Mosconi, Chris Painter, and Avital Balwit, for feedback on this essay.