In January, Dario Amodei told a stunned audience at Davos that the coding agents developed by Anthropic will be used “to create the new generation of models, and speed it up, create a loop that would increase the speed of [AI] model development”. Anthropic views Claude Code as the path towards automation of AI research and, eventually, recursive self-improvement (RSI). Similarly, OpenAI is using Codex to accelerate its own development of AI models and expects that a future version of Codex will eventually become the automated AI research intern:

Where AI researchers have great hope to help themselves... is that if you could just say ‘hey, Codex, this is the idea, and it’s fairly clear what I’m saying, please just implement it so it runs fast on this 8-machine setup or 100-machine setup’. I think that’s what OpenAI [means by] an AI intern by the end of [2026].

—Lukasz Kaiser, OpenAI

OpenAI and Anthropic are racing to automate AI research and reach RSI. But is it a two-horse race, or might any other labs join them? Read on to find out.

OpenAI

OpenAI’s goal announced in October 2025 is to develop an automated AI research intern (i.e., the system as described by Lukasz Kaiser, above), running on “hundreds of thousands of GPUs”, by September 2026. Jakub Pachocki recently said that, based on the improving coding capabilities of Codex, he thinks the intern is “on track” to be developed by September - now just 5 months away. Pachocki also described the differences between the “intern” and the fully automated AI researcher (which OpenAI expects to develop by March 2028):

The way I would distinguish a research intern from a full automated researcher is the span of time that we would have it work mostly autonomously or the specificity of the task that has to be given. I don't expect we'll have systems where you tell them: “Go improve your model capability, go solve alignment” - and they will do it. Not this year. I think we might get there at some point. But for more specific technical ideas - like this particular idea how to improve the models, how to run this evaluation differently - I think we have the pieces that we mostly just need to put together.

In another interview, Pachocki said that the “intern” is a system to which “you can delegate tasks that would take a person a few days”.

It is not clear whether OpenAI is deliberately being conservative with its September 2026 timeline for developing the “intern” and/or its March 2028 timeline for developing the fully automated AI researcher. Interestingly, Sam Altman recently said that “it’s going to be a faster takeoff than [he] originally thought”.

Anthropic

Anthropic’s publicly stated timeline for reaching fully automated AI research is significantly more aggressive than OpenAI’s. Dario Amodei expects 2026 to “have a radical acceleration that surprises everyone… I think we are on the precipice of something incredible”. According to Anthropic’s Frontier Safety Roadmap, released in February 2026, it is “plausible, as soon as early 2027, that [Anthropic’s] AI systems could fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams of human researchers in domains [including development of] AI itself”. Echoing this timeline, Anthropic co-founder and chief science officer Jared Kaplan told Time magazine in March 2026 that fully automated Al research could be "as little as a year away".

Also in line with these predictions, Jack Clark continues to believe that “a country of geniuses in a datacenter” (i.e., AGI)1 will be achievable in late 2026, and “running many copies” in 2027.2

Google

Sitting across from Dario Amodei at Davos in January 2026, Demis Hassabis was diplomatically skeptical about coding models leading to RSI:

The full closing of the loop, I think is an unknown... I think it's possible to do, you may need AGI itself to be able to do that in some domains where there's more messiness around them [and] it's not so easy to verify your answer very quickly. There are NP-hard domains, and I also include for AGI physical AI, robotics. And then you've got hardware in the loop that may limit how fast the self-improvement systems can work - but I think in coding and mathematics, I can definitely see that working.

“If self-improvement doesn’t deliver the goods on its own”, Hassabis said, “then we’ll need other things to work” - i.e., world models, robotics and continual learning.

Under Demis Hassabis’ leadership, Google has indeed focused on reaching AGI via the path of developing continual learning, world models and “physical AI” (i.e., robotics). This is a vastly different path from that currently being pursued by OpenAI and Anthropic. Demis Hassabis estimates that building AI on this path will result in Google achieving AGI in 5 to 10 years.

But is there a change coming at Google? On April 20, 2026, The Information reported that Sergey Brin has formed a “strike team” to improve Google’s coding models. “The end goal”, the article reads, “is AI takeoff or AI that can improve itself… Brin has told staffers that improving Google AI’s coding abilities is a step toward that eventual goal.”

Is Google joining the race? If so, will it throw enough compute and other resources at the problem so as to actually be able to catch up with Anthropic and OpenAI? We will find out over the next few months.

xAI

There is no publicly available evidence to date that xAI is focused on achieving automated AI research or RSI. In January 2026, reports emerged that xAI’s team was using Anthropic’s models through Cursor instead of using Grok. And when co-founder Jimmy Ba left xAI earlier this year, he tweeted that he was leaving to “recalibrate his gradient on the big picture” because “[r]ecursive self improvement loops likely go live in the next 12mo.”

In response, Elon Musk has chosen to focus his energies on beating Anthropic on coding capabilities.3. On April 21, 2026, SpaceX announced that it will be “working closely together” with Cursor “to create the world’s best coding and knowledge work AI”. As part of the deal, SpaceX will either pay Cursor $10B for this collaboration in 2026 or, at its option, will purchase Cursor for $60B.

Does xAI’s senior leadership realize that coding models and the resulting enterprise revenue are merely an (extremely useful) milestone on the path to RSI, or is xAI’s goal limited to mimicking Anthropic’s success in delivering agentic models to enterprises? Time will tell.

Meta

The Meta Superintelligence Labs team, assembled at a great cost in mid-2025, has been working for many months on developing new AI models and related tools. Thus far, these efforts have culminated only in the release of Muse Spark, a new reasoning model.

There is no publicly available evidence currently that Meta is focusing any attention on coding models, automation of AI research or RSI.

Microsoft

An underrated player in the quest for AGI, Microsoft has a major trump card up its sleeve: its licensing deal with OpenAI means that Microsoft has rights to OpenAI’s “research IP” (including models intended for internal deployment or research only - which should include automated AI research models) until the earlier of 2030 or verification of OpenAI’s declaration of AGI by an independent expert panel. After OpenAI declares AGI, we can expect Microsoft to use this research IP to undertake its own quest for superintelligence.

Possibly in preparation for this move (at least in part), Microsoft has been aggressively expanding its data center capacity.

Others?

DeepSeek has been suspiciously quiet recently, with no major model releases since December 2025. Given DeepSeek’s technical prowess and taste, it would not be surprising to the author if it turned out that DeepSeek has built its own coding model internally and is using it to accelerate its own coding and AI research capabilities. These efforts, if they indeed do exist, may be significantly constrained by compute availability and other factors.

For now, the smaller U.S. labs and “neolabs” (e.g., SSI, Thinking Machines, Core Automation - just to name a few) generally appear to be headed in different directions from the one chosen by OpenAI and Anthropic.4

Finally, time will tell whether any other Chinese labs are willing and able to join the race to RSI.

I‘ve been asked to expand on my tweet - why do I believe that BigLaw as a concept will not survive the arrival of powerful AI?

Why Hire BigLaw?

I’d say there are three main reasons people hire a BigLaw firm:

1. The client needs customized or specialized advice about something. For example, the client’s in-house lawyers might know how to draft a simple vendor agreement, but not how to navigate a complex M&A deal with tax considerations, regulatory implications, complicated transaction mechanics, etc.

2. The client needs someone to do a large amount of legal work, potentially under a tight time frame. For example, a team of a few in-house lawyers would not be able to review thousands of documents dumped into a data room on a Friday in a timely manner - but a BigLaw firm will staff a dozen associates on this if needed, and they’ll work around the clock to finish the task by any deadline, no matter how unreasonable.

3. The client needs advice in a matter that is high-stakes, or needs to be blessed by competent outside counsel, or involves a counterparty that is using another BigLaw firm. Big M&A transactions, bet-the-company litigation, internal investigations, sensitive matters requiring establishment of a special committee all fall within this category.

Often, a matter will meet more than one of these criteria (and potentially will meet all three).

The BigLaw model involves one key partner (or a few key partners) providing mostly strategic advice to the client, plus a team of associates supporting the partners’ work. The associates do the research and draft a legal memo, a junior partner reviews it, the final work product goes to the senior partner who glances at it and maybe distills it down to a few talking points for the client’s GC (the GC will not read the memo, but will listen to these talking points).

How does AI impact all this?

Category 1 (specialized advice): GPT-5.x Pro knows the tax laws and the regulatory implications. I also think we’re not far away from an AI harness that would enable a SOTA AI model to succinctly (and fairly quickly) summarize all legal implications of a particular fact pattern (e.g., in a legal memo) or implement them via contract language. LLMs would probably need to learn how to “write like a lawyer” a bit better in order to achieve this, but I don’t view this obstacle as being insurmountable. The remaining human role in this process: (1) verifying the LLM’s output, (2) providing high-level strategic advice, and (3) going “beyond the law” to things like unwritten regulatory requirements, market practices, etc. All of this can be done by a team of senior-partner-level people; associates are not required.

Category 2 (high-volume work): AI works much faster than humans, doesn’t need to take breaks, doesn’t sleep, can work for many hours at a time, ‘nuff said. The remaining human role in this process: verifying the LLM’s input by doing things like double-checking summaries of key documents, reviewing a sampling of documents to make sure the human reviewer agrees with the LLM’s conclusion, etc. This can be done by a small team of in-house counsel, with maybe some input from senior-partner-level people in a law firm.

Category 3 (high-profile work): This is where BigLaw firms will continue to dominate regardless of AI... at least for a while. Yet, slowly but surely, the law firms’ work will be eroded. No, don’t do the diligence; do only specified spot-checking and a high-level review of our (the client’s) AI’s findings. No, don’t draft the Merger Agreement; we’ll send you something we put together ourselves for review. No, don’t do the research; we’ll send you something our AI put together to double-check. Eventually, the value typically brought by BigLaw to these kinds of transactions (i.e., leveraging an army of associates to do the gruntwork) disappears, and it starts looking more and more like the client is hiring a particular partner (or team of partners) to do the entire matter. The leverage starts disappearing.

At some point, the best partners start asking themselves why they should share their profits with less successful partners - including those who have not adapted to the age of AI. And so, the senior partners will slowly start leaving to start up their own boutique law firms. The client gets a great deal: Bob Jones, formerly the top deal-maker at Cravath, is still handling the client’s work, but charges the client only X% of what Cravath would charge - and the work is faster while its quality is better. The boutique law firm consists of Bob Jones, maybe a few specialist colleagues (tax, regulatory, etc.), and a few junior lawyers, paralegals and/or IT professionals whose main job is to work with AI models to rapidly produce high-quality legal work. On the client’s side, just a few in-house lawyers now handle not only the work that a much larger in-house legal team used to do, but also a good portion of work that was formerly done by BigLaw.

There is no longer any room for BigLaw in this paradigm, and BigLaw firms start disappearing.

The timing for all this is extremely uncertain. The legal industry moves slowly. Lawyers are extremely non-technical. I’d venture to guess that 99% of lawyers today don’t know the power of GPT-5.4 Pro. This will eventually change, but how long is “eventually”? And when will the clients begin to understand, begin to truly internalize, the transformative impact that AI can have on the practice of giving legal advice?

Could be 2 years, or could be 10.1

This article also appeared on Twitter.

Jerry Tworek, OpenAI’s former VP of Research who left the company earlier this month, is raising for his new start-up, Core Automation.

The start-up’s roadmap is ambitious, and starts with nothing less than developing a new AI architecture to be used in lieu of the transformer. Standard methods for training models “up to and including gradient descent” will go out the window. A new model named Ceres (after the Roman goddess of fertility) will be trained using these new methods. The training process will be hyper-efficient, using 100x less data than today’s frontier models. And Ceres will be able to learn through real-world experience - because Core Automation also intends to crack continual learning.

As if that weren’t enough, Core Automation’s goals after developing Ceres include automating development of future AI products, constructing self-replicating factories, and “potentially building biomachines to automatically create custom designs - or even terraform planets”.

We know that Ilya Sutskever thinks that he’ll be able to crack continual learning in 5 to 20 years. Jerry doesn’t have that kind of time. After all, his former employer intends to fully automate AI research by March 2028, which might lead to recursive self-improvement (RSI). With RSI also squarely on Core Automation’s roadmap (what did you think “automating development of future AI products” meant? vibes? papers? essays?), the company will need to execute on its goals very quickly - potentially within months! - lest the likes of OpenAI and Anthropic achieve RSI first and scupper its plans.

After all, the bet on automated AI research ultimately means betting on the bitter lesson - i.e., that “general methods that leverage computation are ultimately the most effective, and by a large margin”. Once the GPUs powering OpenAI’s and Anthropic’s automated AI researchers start humming, it appears quite possible that human-led AI research will quickly fall by the wayside and that the big unsolved problems of AI (such as continual learning) will become much more easily solvable if desired. Viewed from this perspective, the key dilemma emerges: given the race between the frontier labs to fully automate AI research and potentially achieve RSI, does it still make sense to “front-load” human-led research of new AI architectures and transformative ideas, or is it more prudent to instead curtail these initiatives and spend the freed-up resources on reaching the goal of automating AI research even faster?

Jerry Tworek’s bet is on humans for another few years. His former employer’s bet is exactly the opposite.

We shall soon find out which one of the two is right.

This blog post gathered some attention, and its ideas have since been mindlessly copied by others. Mark Zuckerberg claimed a few days later that “[o]ver the last few months we have begun to see glimpses of our AI systems improving themselves”.1 More recently, Elon Musk, too, said that we have entered the singularity.

It has typically been assumed that these claims have been principally driven by the generally fast rate of improvement in AI models (i.e., “AI is improving fast today; AI will improve even faster tomorrow”). With respect to Altman’s claims specifically, I am of a different view. I believe that Altman meant something very specific when he said that “we are past the event horizon”, and that this “something” is the most important thing happening in AI today.

Codex

On May 16, 2025 (a few weeks before Altman’s blog post), OpenAI released its agentic coding tool, Codex. The release flew a bit under the radar, overshadowed by the previous month’s release of o3 and endless speculation about the then-impending releases of o3-pro and OpenAI’s open-source models. But no matter. The coding agent, which was OpenAI’s answer to Claude Code, released just three months earlier, was merely the first step on OpenAI’s path to full automation of AI research.

OpenAI likely set out on this path in or around March 2025, just a few weeks after Anthropic’s release of Claude Code. This is why OpenAI’s Preparedness Framework was updated to include recursive self-improvement (RSI) as a Tracked Category in April 2025. Other circumstantial evidence also points to the project’s launch in March 2025: OpenAI’s goal of developing a fully automated AI researcher falls exactly three years later (March 2028), and its mid-way goal of developing an automated AI research “intern” falls exactly mid-way through this three-year process (September 2026, or 18 months after March 2025).

Even OpenAI insiders were initially not convinced by Codex until a much more powerful version arrived with August’s release of GPT-5:

By September 2025, OpenAI began leaking that an automated AI researcher has become the focus of its entire research program. Here’s Jacub Pachocki explaining that OpenAI has been building most of its projects with the goal of achieving an automated AI researcher:

Our set goal for our research program has been getting to an automated researcher for a couple years now. And so we’ve been building most our projects with this goal in mind.

The following month, OpenAI officially announced to the world that it is focusing on developing the automated AI research “intern” by September 2026 and the fully automated AI researcher by March 2028. Sam Altman added that the “intern” will run on hundreds of thousands of GPUs.

Since this announcement, OpenAI has repeatedly stressed that automated AI research is now its primary focus. “We’re very excited about our 2026 roadmap and advancing work toward an automated scientist,” Mark Chen said just yesterday.

Again, the path towards fully automated AI research starts with Codex. This is clear, e.g., from this description of the “intern” from Lukasz Kaiser:

Where AI researchers have great hope to help themselves... is that if you could just say ‘hey, Codex, this is the idea, and it’s fairly clear what I’m saying, please just implement it so it runs fast on this 8-machine setup or 100-machine setup’. I think that’s what OpenAI [means by] an AI intern by the end of next year.

Claude Code

Not surprisingly, Anthropic views Claude Code in exactly the same way as OpenAI views Codex - i.e., as a coding tool that will eventually lead to automation of AI research. Indeed, Sonnet 4.5 and Opus 4.5 system cards conspicuously included results of surveys of Anthropic employees designed to evaluate whether the model, paired with Claude Code, is good enough to fully replace a junior AI researcher. In the Opus 4.5 survey, two (2) out of 18 participants classified Opus 4.5 as a “near-complete entry-level researcher replacement” - albeit with “meaningful caveats”.

This is also why we’ve heard Sholto Douglas speak about withholding models with capabilities to perform AI research from Anthropic’s competitors:

As AI models get better at [machine learning research tasks], I do expect the labs to to hold back some of the the capabilities. If a model's capable of writing out a whole new architecture that's a lot better, even if it's just capable of writing all their kernels for them, you probably don't want to release that to your competitors.

And what is “the main thing” that Jack Clark worries about these days? But of course, closing the loop on AI R&D, which would lead to RSI:

The main thing I worry about is whether people succeed at 'building AI that builds AI'—fully closing the loop on AI R&D (sometimes called recursively self-improving AI).

Clark notes that “extremely early signs” of AI getting better at doing components of AI research can already be seen, “ranging from kernel development to autonomously fine-tuning open-weight models”.

Recursive Self-Improvement and the Takeoff

But why does automating AI research matter? Turning again to Jack Clark, the key is “compounding R&D advantage” from automated AI research. The premise is that an AI researcher would be able to build an even better (and smarter) AI researcher, which, in turn, would be able to build yet another better and smarter AI researcher. Automated intelligence could quickly lead to automated superintelligence - and, eventually, to systems so much smarter than humans that a human researcher would not be able to even understand the new discoveries being made by AI, much less keep up with it:

If this stuff keeps getting better and you end up building an AI system that can build itself, then AI development would speed up very dramatically and probably become harder for people to understand.

This is “the takeoff”.

Parting Thoughts on the Race to AGI

The above considerations lead us to the most critical insight of all vis-a-vis the race to AGI. If OpenAI and/or Anthropic succeed in fully automating AI research, there is a chance that the “takeoff” shall occur, with the result that no other lab shall ever be able to catch up to models built by these labs. In the “takeoff” scenario, even a large team staffed with the very best human AI researchers will never be able to compete with a model capable of superhuman AI research, and the advantages will only compound from there. A rival lab might reach automated AI research at a later date, but by then it will be too late - its model will not be able to compete with the much more advanced AI researchers compounding faster being operated by the other labs.

Assuming that one believes in this version of the take-off,2 that should lead one to also believe that a chasm is already developing between those labs that are racing to automate AI research and those that are not.

Sam Altman caused some commotion on X today with his post that the new Head of Preparedness role at OpenAI would be responsible, inter alia, for “gain[ing] confidence in the safety of running systems that can self-improve”.

However, the fact that the Head of Preparedness will oversee safety efforts for model self-improvement should not be surprising. After all, the role’s primary responsibility is to “lead the technical strategy and execution of OpenAI’s Preparedness Framework”,1 which currently focuses on cybersecurity risk, bio risk, and - yes - model self-improvement.

But this is where things get interesting. It turns out that AI self-improvement capabilities were added to the Preparedness Framework as a “Tracked Category” all the way back in April 2025(!). What is a “Tracked Category”? OpenAI explains that it’s a capability that meets all of the following five criteria:

1. Plausible. “It must be possible to identify a causal pathway for a severe harm in the capability area, enabled by frontier AI.”

2. Measurable. OpenAI must be able to construct or adopt capability evaluations that measure capabilities that closely track the potential for the severe harm.

3. Severe. There is a plausible threat model within the capability area that would create severe harm.

4. Net New. “The outcome cannot currently be realized as described (including at that scale, by that threat actor, or for that cost) with existing tools and resources… but without access to frontier AI.”

5. Instantaneous or irremediable. Once the outcome is realized, its severe harms: (1) are immediately felt; or (2) are inevitable due to a lack of feasible measures to remediate.

Per OpenAI, “AI self-improvement” was separated as a Tracked Category because:

“it presents a distinct plausible, net new, and potentially irremediable risk, namely that of a hard-to-track rapid acceleration in AI capabilities which could have hard-to-predict severely harmful consequences. In addition, the evaluations we use to measure this capability are distinct from those applicable to Long-range Autonomy and Autonomous Replication and Adaptation.”2

…but reading between the lines, the real reason for this capability’s designation as a Tracked Category may have been to begin preparing for automation of AI research. After all, “High” risk associated with this capability is defined as follows:

The model’s impact is equivalent to giving every OpenAI researcher a highly performant mid-career research engineer assistant, relative to those researchers’ 2024 baseline.

(This sounds perhaps slightly more capable than the “automated AI research intern” that OpenAI intends to develop by September 2026.)

And “Critical” risk sounds suspiciously similar to the risk that would be posed by the “automated AI researcher” that OpenAI intends to develop by March 2028:

The model is capable of recursively self improving (i.e., fully automated AI R&D), defined as either (leading indicator) a superhuman research scientist agent OR (lagging indicator) causing a generational model improvement (e.g., from OpenAI o1 to OpenAI o3) in 1/5th the wall-clock time of equivalent progress in 2024 (e.g., sped up to just 4 weeks) sustainably for several months.

Based on the updated Preparedness Framework’s release date (April 2025), it seems that March/April 2025 is when OpenAI first set the internal goal to fully automate AI research within 3 years’ time (by March 2028). We are now ≈9 months into this 3-year project. Six months in, OpenAI publicly announced it to the world. And now, nine months in, OpenAI has begun building out a comprehensive safety function around it. It seems that we continue to be “all systems go” for the launch of the automated AI research intern next fall.

2025 was a year of stunningly fast AI progress.

In December 2024, the best reasoning model was OpenAI’s o1, a toy reasoning model that wasn’t even particularly proficient at using tools. By September 2025, OpenAI’s unreleased general reasoning models had won gold medals on the 2025 International Mathematics Olympiad (IMO), the 2025 International Olympiad in Informatics (IOI), and the 2025 International Collegiate Programming Contest (ICPC) World Finals. Another unreleased OpenAI model won second place in the AtCoder World Finals, working fully autonomously without human intervention for the entire 10 hours of the competition. And coding agents - including, in particular, Claude Code - have taken the world of coding by storm, while also meaningfully accelerating the pace of AI research at the frontier labs.

We have also begun to see glimpses of AI meaningfully contributing to work in fields other than coding. Starting in late Q3 2025, I began using GPT-5.x Pro for legal research and analysis, and am now finding it absolutely essential to my work. I am also increasingly seeing reports that Google’s NotebookLM is fantastic at generating presentations and data tables, which is another important enterprise use case. And even non-technical people (yes, including yours truly) are discovering “Claude Code for things that are not coding”.

Where does this lead us in 2026? Here are some predictions:

Automation of AI research

Earlier this year, roon1 played with Codex for the first time and “realiz[ed] we’re in the takeoff”. In 2026, agentic coding tools like Codex and Claude Code will continue accelerating frontier lab researchers. By September 2026, OpenAI intends to have this effort culminate in an automated AI research intern running on hundreds of thousands of GPUs, which will be able to automatically handle the implementation and debugging of research ideas proposed by OpenAI’s human researchers.2

Continual learning

In Q3 2025, public consensus suddenly decreed that continual learning is required to achieve AGI. Andrej Karpathy said that current LLMs are “cognitively lacking” due to lack of continual learning, and “it’s not working” - placing AGI about a decade away. Later, Ilya Sutskever added fuel to the fire when he revealed that SSI is working on developing AI capable of continual learning - which he said is “5 to 20 years” away.

Relatively unnoticed among all the hoopla were comments on continual learning from Anthropic’s CEO, Dario Amodei:

One thing we learned in AI is whenever it feels like there’s some fundamental obstacle - like two years ago we thought there was this fundamental obstacle around reasoning - turned out just to be be RL, you just train with RL and you let the model write things down to try and figure out objective math problems…Without being too specific, we already have maybe some evidence to suggest that [continual learning] is another of those problems that is not as difficult as it seems that will fall to scale plus a slightly different way of thinking about things.

And just a few days ago, Sholto Douglas, an Anthropic employee, dropped a bombshell with his prediction that “continual learning [will get] solved in a satisfying way” in 2026.

Does this mean that Anthropic already knows how to achieve continual learning? We’ll find out next year.

Recursive self-improvement

Mark Chen recently mentioned that OpenAI is aggressively scaling up several bets, including one related to synthetic data. This was a reference to Sebastien Bubeck’s brief cameo during the GPT-5 launch livestream, in which he revealed that OpenAI has developed “new training techniques” whereby o3 had generated synthetic data to train GPT-5 in a way “raw web data just never could”. “This interaction between models foreshadows a recursive self-improvement loop”, Bubeck said.

Google DeepMind is also working in the same direction, according to Sebastian Borgeaud, pre-training lead for Gemini 3:

One really interesting question is whether you can actually generate synthetic data to make a model that you want to train in the future better than the model that generated the synthetic data in the first place. We spend a lot of time thinking about this and doing research in this direction.

It is unclear where these efforts will lead in 2026, but needless to say that this is an area of ML research that is well worth monitoring.

AI is coming to the workplace (not just for coders)

Here’s Sholto Douglas again:

The most striking thing about next year is that the other forms of knowledge work are going to experience what software engineers are feeling right now, where they went from typing most of their lines of code at the beginning of the year to typing barely any of them at the end of the year. I think of this as the Claude Code experience, but for all forms of knowledge work.

Those who follow me on X know that I have been crying out for an interface that would enable even a non-technical lawyer to “vibe-code” a stock purchase agreement (see, e.g., point 3 here). It looks as though my wish may finally come true in 2026.

But it will be more than that, of course. Anthropic’s goal for 2026 is to develop and sell to enterprises a “virtual co-worker that is in all your Slack channels and can join your meetings and can work alongside you”. Some of us will be seeing these “virtual co-workers” join our companies next year.

And as for coding…

I am not a coder, but, as an outside observer, I can easily tell that several significant “vibe shifts” occurred in 2025 around using agentic tools like Claude Code and Codex for coding tasks. Claude Opus 4.5 in particular smashed the METR 50%-time horizon benchmark, and appears to be a huge step change when compared to the previous generation of models - to the point where some debate may be had as to whether Opus 4.5 in Claude Code is “basically AGI” (by OpenAI’s definition: a highly autonomous system that outperforms humans at most economically valuable work).

It seems clear that the models will continue to improve at a rapid pace from here vis-a-vis coding. Expect software engineering to “go[] utterly wild next year”.

AI for science

Starting late this year, there has been an increasing cadence of reports that models like GPT-5 Pro can be leveraged effectively as a tool by human mathematicians to help with making relatively minor advances in mathematics. As models continue to improve next year, OpenAI expects that its AI systems “may be able to make small new [scientific] discoveries” in 2026. Indeed, work on these initiatives is ongoing at multiple frontier labs: for example, Anthropic has begun hiring “wet lab wizards” for its life sciences team.

But can LLM actually autonomously generate novel scientific hypotheses? In my view, the answer is almost certainly “yes”. We have already seen that even Gemini 2.0 Pro, when equipped with a great harness, can propose a novel scientific hypothesis pertaining to a complex gene transfer mechanism.3 The general rule of thumb that I think makes sense to follow is that anything an LLM can do with a harness will eventually also be achievable by a more powerful LLM without any harness whatsoever; the only (important!) question that remains open is the timeline by which this feat would become possible to accomplish.

OpenAI has declared that 2026 will be the “Year of AI and Science”. Let’s hope that the year can live up to this lofty title!

The robots are coming?

There’s been a lot of hoopla around humanoid (or otherwise) robots over the past few years, but very few of these advances have thus far made it out into the real world. I remain somewhat unconvinced that 2026 will be the year when robots truly proliferate in the real world at scale, but it’s possible that I am too pessimistic in this regard. Google DeepMind apparently projects that 2026 will be “a huge year” for embodied AI, and that there will be “a lot more robots in the real world soon”. Other knowledgeable commentators expect that 2026 will see at least “the first test deployments of home robots”.

* * *

I will end here with an overall observation. Over the past few months, it has become decidedly fashionable to update one’s views towards longer timelines for “AGI” (whatever that term might mean). If significant progress is made on automation of AI research and/or continual learning in 2026, these longer timelines will likely begin to feel extremely - maybe even needlessly - conservative by the end of the year. In particular, OpenAI’s stated goal of fully automating AI research in just slightly more than two years’ time still has not been - but should be - fully internalized by most industry observers and commentators. Should OpenAI successfully develop and deploy an automated AI research “intern” during 2026, a realization may suddenly come to many that the long-expected promise of the machine taking over the building of other, yet more powerful, machines has come to the verge of being fulfilled.

Over the past six months, it has become quite fashionable in certain circles to assume that AGI is more than a decade away because, among other things, current AI models lack continual learning. Ilya Sutskever is working on developing superintelligence with skills and knowledge of an eager 15-year-old and the ability to learn on the job; he thinks this is “5 to 20” years away. Andrej Karpathy thinks that current AI systems “are cognitively lacking and it’s just not working” because they have no continual learning; “[i]t will take about a decade to work through all of those issues”, he thinks.

But why should we assume that continual learning is a decade away? Could it be that it’s easier to achieve than some think?

Sholto Douglas: continual learning to “get solved in a satisfying way” in 2026

In the “No Priors” year-end podcast, Sholto Douglas (Anthropic) said that he thinks “that probably continual learning gets solved in a satisfying way” in 2026.

Dario Amodei: continual learning “not as difficult as it seems”

Dario Amodei, CEO of Anthropic, repeatedly stated this summer that continual learning may not be as difficult as it seems.

In a July interview, Amodei said that, even without continual learning, other techniques “can fill in many of the gaps”. One such technique is significantly lengthening the context window - perhaps to as much as 100 million words, which is roughly the number of words a human hears during his or her lifetime. There is “no reason” from a machine learning perspective why the context window could not be increased to this size, Amodei said, “it’s really just inference support” that is needed to make this viable.

But what about continual learning that allows for updating a model’s weights? Unexpectedly, Amodei suggested that Anthropic may have already found a path to achieving it:

One thing we learned in AI is whenever it feels like there’s some fundamental obstacle - like two years ago we thought there was this fundamental obstacle around reasoning - turned out just to be be RL, you just train with RL and you let the model write things down to try and figure out objective math problems…Without being too specific, we already have maybe some evidence to suggest that [continual learning] is another of those problems that is not as difficult as it seems that will fall to scale plus a slightly different way of thinking about things.

Amodei also hinted that an “inner loop/outer loop” structure, wherein an agent learns things, and optimizes for a lifetime of an episode (inner loop) and also learns over multiple episodes (outer loop) “maybe… is a way to learn continual learning”.

In a subsequent August interview, Amodei again mentioned extending a model’s context to 100 million tokens and suggested that models could be trained to be “specialized for learning over the context”; “[y]ou could, even during the context, update the model’s weights”. “[T]here are lots of ideas that are very close to the ideas we have now that could perhaps do this [i.e., achieve continual learning]”, Amodei said.

Shane Legg: “no fundamental blockers” on continual learning; “we have ideas” on how to develop it

In an interview released just today, Shane Legg, co-founder and Chief AGI Scientist at Google DeepMind, said that there are no “fundamental blockers” on continual learning (and also visual reasoning).

[W]e have ideas on how to develop systems that can do these things, and we see metrics improving over time in a bunch of these areas. So my expectation is over a number of years these things will all get addressed. But they’re not there yet.

Continual learning “might need some process whereby new information may be stored”, a “retrieval system or episodic memory”, and “systems whereby that information over time is trained back into some underlying model”, Legg said. This will require both more data and algorithmic and architectural changes.

This is a reference to Sebastien Bubeck’s brief cameo during the GPT-5 launch, during which he said that OpenAI has developed “new training techniques” whereby o3 generated synthetic data to train GPT-5 in a way “raw web data just never could”. The point was to generate not a large volume of data cheaply, but rather useful training data.

“This interaction between models foreshadows a recursive self-improvement loop”, Bubeck said, adding: “Here at OpenAI we cracked pre-training, then reasoning, and now we are seeing their interactions significantly deepened. In the future, AI systems will move far beyond our current pre-training and post-training pipelines we’ve been used to and we are seeing the first steps towards this right now and right here.”

And OpenAI is now aggressively scaling it up.

The results (from the Opus 4.5 system card):

• 50% of survey participants reported productivity improvement of at least 100% (2x); median productivity improvement was 100%

• Mean productivity improvement was 220%(!)

• 11% (2/18) characterized the model as a “near-complete entry-level researcher replacement” (with meaningful caveats)

• Most researchers would prefer losing access to Opus 4.5 to losing access to Claude Code (i.e., the harness remains more important than the model)

Importantly, survey participants were not just average Anthropic employees, but rather were “primarily” selected from the top 30 Anthropic employees ranked by internal Claude Code usage. We should expect that these Claude power users would get significantly more uplift from using the model than the average Anthropic employee. Nonetheless, the productivity boost unlocked by the model for its most skilled users is extremely impressive and well worth noting.

Comparison to Sonnet 4.5

For reference, here are the results of a similar survey conducted by Anthropic in September 2025 for Sonnet 4.5:

• 7 Anthropic researchers were surveyed; it’s not clear whether they were average employees or some of the top users of Claude

• Productivity boost estimates from Sonnet 4.5 were: 15%, 20%, 20%, 30%, 40%, 100%, one instance of “qualitative-only feedback”

• 0% thought that Sonnet 4.5 could completely automate the work of a junior ML researcher

• As with Opus 4.5, most researchers (4 out of 7) thought that most of the productivity boost was attributable to Claude Code, as opposed to the model itself

What is this benchmark?

• Take 20 research and engineering bottlenecks that OpenAI actually encountered in the past, each of which required over a day for the OpenAI team to solve. These bottlenecks include “unexpected performance regressions, anomalous training metrics [and] subtle implementation bugs”, which actually represented delays to major projects and “in some case influenc[ed] the outcome of large training runs and launches”.

• Give the model access to a container with code access and run artifacts, permitting it to use historical code, logs, and experiment data.

• Ask the model to diagnose and explain the root cause of the issue.

• Each of the model’s solutions is graded pass@1 (one try only!).

This benchmark is very relevant to accelerating - and eventually automating - AI research. Imagine the time and resources that could be saved quashing a major bug if GPT-x could diagnose and identify it even 50% of the time (instead of the OpenAI team spending 1+ days identifying and fixing it).

I particularly like that, instead of using toy problems (which are often of questionable relevance to real-world AI use cases, but plague nearly all popular LLM benchmarks), OpenAI-Proof Q&A measures the model’s performance on major bugs that OpenAI has actually encountered in the past. And the model’s performance is judged on a pass@1 standard - none of that “well, it got the solution right once out of the twenty times we ran it, so we’ll call that a pass” nonsense.

Here’s how OpenAI’s models have done on this benchmark to date:

GPT-5.1-Codex-Max scored 8%1 (with and without refusals), beating the previous SOTA (GPT-5, which scored 2%).

• Note: It is not clear to me how much of the delta between GPT-5 and GPT-5.1-Codex-Max was due to the fact that GPT-5 had “no browsing”. Did GPT-5.1-Codex-Max have access to browsing (I assume yes)? If so, how much did this skew the score?

Regardless of comparability to GPT-5’s score, the result achieved by GPT-5.1-Codex-Max is quite impressive. Look for further updates to OpenAI-Proof Q&A as more powerful OpenAI models are released over the next 12 months.

Let’s examine this claim based on publicly available information.

Microsoft’s rights to OpenAI’s “standard” IP

Other than “Research IP” (see below), Microsoft does have broad rights to all of OpenAI’s IP until 2032. This includes:

• model architecture

• model weights

• inference code

• finetuning code

• anything related to data center hardware and software

As Satya correctly pointed out during the interview, there is one notable exclusion: Microsoft does not have any IP rights to OpenAI’s consumer hardware.

“Research IP”

Notwithstanding the above, Microsoft has rights to OpenAI’s “Research IP” only through the earlier of: (a) 2030; or (b) verification of OpenAI’s declaration of AGI by an independent expert panel.

“Research IP” means “confidential methods used in the development of models and systems”, such as models intended for internal deployment or research only.

Restrictions on Microsoft’s use of OpenAI’s IP

It’s one thing to have IP rights; it’s another to be able to use them however you like.

The contract between OpenAI and Microsoft includes a clause restricting Microsoft’s use of OpenAI’s IP as follows:

If Microsoft uses OpenAI’s IP to develop AGI, prior to AGI being declared, the models will be subject to compute thresholds; those thresholds are significantly larger than the size of systems used to train leading models today.

What does this mean for Microsoft’s quest for AGI?

Assuming that AGI will be a model larger than the threshold specified in the contract between OpenAI and Microsoft (probably a good assumption), Microsoft won’t be able to use OpenAI’s IP to develop AGI until after OpenAI already has achieved AGI.

Note, however, that:

• Microsoft will be able to use OpenAI’s IP to train smaller models for AGI-related research purposes.

• Nothing prevents Microsoft from separately pursuing AGI without relying on any of OpenAI’s IP.

But once OpenAI’s declaration of AGI is verified, all bets are off. At that point, Microsoft will have a copy of the AGI model’s weights and will be able to race directly against OpenAI to develop even better models ( let’s call them “ASI”).

In that race, Microsoft will no longer have the benefit of OpenAI’s “Research IP”, including models used only for “internal deployment” or “research”. Thus, even if OpenAI develops “ASI” before Microsoft, it won’t have to share it with Microsoft, so long as “ASI” is not publicly released, but instead remains a model in OpenAI’s “internal deployment”.

Updated November 26, 2025 - to include new remarks by Lukasz Kaiser.

On October 29, 2025, Sam Altman and Jakub Pachocki announced during a livestream that OpenAI is building an “automated AI researcher”, targeted to be available by March 2028. Here’s what we know about it.

What’s Being Released and When

OpenAI’s goal is to “have”:

• an automated AI research intern by September 2026

• an automated AI researcher by March 2028

These appear to be dates by which OpenAI hopes to have these researchers available internally, and not necessarily release dates. For example, as tweeted by Sam:

We have set internal goals of having an automated AI research intern by September of 2026 running on hundreds of thousands of GPUs, and a true automated AI researcher by March of 2028.

Similarly, during the livestream, Jakub repeatedly referred to the goal of “getting” (as opposed to releasing) the research intern/researcher by these dates.

The Automated AI Research Intern

Even though OpenAI calls the model that will become available by September 2026 merely an “intern”, do not be fooled - this model will likely be very powerful. Sam wrote that the “intern” will run on hundreds of thousands of GPUs, a gigantic amount of compute. As a point of comparison, consider that OpenAI will likely have only just over 1 million GPUs online in total by year-end 2025. This means that running the “intern” alone would take up >20% (and possibly significantly more) of the entire compute capacity available to OpenAI today.

Given this background, it should not be surprising that, in Jakub’s words, OpenAI expects that the automated AI research intern will “meaningfully accelerate” its researchers.

The Automated AI Researcher

During the livestream, Jakub described the automated AI researcher as “a system capable of autonomously delivering on larger research projects”. Expect this system to be very impactful. In fact, Jakub described the automated AI researcher in a September 2025 interview as the focal point of OpenAI’s research program over the past few years:

Our set goal for our research program has been getting to an automated researcher for a couple years now. And so we’ve been building most our projects with this goal in mind.

Per Sam’s follow-up tweet, the automated AI researcher will have “extraordinary potential impacts” - potentially so significant that OpenAI has deemed it to be “in the public interest to be transparent” about its plans for developing it.

…But How Will It Work?

Update (December 1, 2025): Mark Chen has finally shed some light on what the automated AI researcher and automated AI research intern are intended to accomplish:

Within a year, we want to change the nature of the way that we’re doing research. We want to be productively relying on AI interns in the research development process. And within 2.5 years, we want AI to be doing end-to-end research. Today, you come up with an idea, you execute on it, you implement it, you debug it. Within a year, we’re quite confident we can get to a world where we control the outer loop - we come up with the ideas, but the model is in charge of the implementation and debugging.

So - the automated AI research intern will enable OpenAI’s research team to limit their work to generating new ML ideas; the intern will implement them and debug them. And the automated AI researcher will be doing “end-to-end research” - including generating new ML ideas.

Update (November 25, 2025): Lukasz Kaiser described the automated AI research intern similarly - i.e., as being able to convert clear - but general - instructions from a human researcher into an efficient implementation of the researcher’s idea:

Where AI researchers have great hope to help themselves... is that if you could just say ‘hey, Codex, this is the idea, and it’s fairly clear what I’m saying, please just implement it so it runs fast on this 8-machine setup or 100-machine setup’. I think that’s what OpenAI [means by] an AI intern by the end of next year.

As to how OpenAI intends to achieve this technically, Jakub Pachocki’s September 2025 interview may shed some light on at least some of the advances that OpenAI expects to power the automated AI researcher:

The big thing we are targeting with our research is producing an automated researcher. So, automating the discovery of new ideas, and in particular automating our own own work, automating ML research…

One good way to measure progress there is looking at the time horizon on which these models actually can reason and make progress. Now as we get to a level of near-mastery of high school competitions, we get to on the order of 1 to 5 hours of reasoning. We are focused on extending that horizon both in terms the models’ capability to plan over very long horizons and actually have ability to retain memory.

So, one piece of the puzzle may be to get the models to reason for longer - which OpenAI has been working on for a long time, probably since before o1-preview was first announced. Another may be to have the automated AI researcher “retain memory” (but it’s unclear how).

…And Will It Be AGI?

When asked “wen AGI?” during the livestream, Sam said that it’s more useful to have an automated AI researcher by March 2028 and define what that means that try to define “AGI”:

“The AGI term has become hugely overloaded, and… it will be this process over a number of years that we’re in the middle of. But one of the reasons we wanted to present what we did today is, I think it’s much more useful to say our intention, our goal, by March of 2028 is to have a true automated AI researcher and define what that means than to try to satisfy everyone with a definition of AGI.”

Greg Brockman recently said that he expects AGI to arrive within the next “one to three years” (for those counting, that means by late 2028) and that he “would feel like something went wrong if we were not there by 2030”. It’s interesting to observe how neatly the anticipated March 2028 release of the automated AI researcher falls within this timeline.