OpenAI prepares for recursive self-improvement. Frontier lab boss says: "Ignore the hype."

OpenAI has committed itself to developing AI models capable of recursive self-improvement (RSI), the ability to independently improve their own capabilities.

But it also admitted: "We do not yet know how to safely get all the way to aligned, full RSI."

Yesterday, OpenAI announced it was "making strong progress" on the development of an “automated AI researcher” by March 2028.

It also shared a “detailed snapshot” of how agents are already accelerating its own AI research by writing code, running experiments, debugging research systems, and helping to improve other models

Behind the scenes, it has also formed a team dedicated to researching RSI and is hiring engineers on salaries of up to $500,000 to build “AI systems that accelerate and ultimately conduct high-quality research at OpenAI”.

None of that is RSI - which Anthropic previously said was not "inevitable".

So what is currently stopping us from birthing AI that can build itself?

In an article for Machine, Zuzanna Stamirowska, CEO of frontier AI lab Pathway, argues that the industry is overlooking fundamental limitations in today's technology.

“Ignore the hype,” she writes. “We’re not on track for self-improving AI.”

Here's her article in full.

Recursive self-improvement (RSI), the inflection point at which AI systems start to autonomously design and develop their own successor, is nothing new as a concept. But it’s been the subject of fresh interest and hype over recent months. Startups have raised hundreds of millions with the explicit goal to build such a system, and Anthropic caught attention with a June blog on its progress towards RSI, claiming AI is already accelerating the development of AI systems. 

As agents take on increasingly longer model-development workflows, from implementation to evaluation, the prospect of AI-enabled self improvement appears less speculative. The time-task horizon, which measures how long an AI agent can complete a task autonomously without human intervention, has doubled roughly every 7 months since 2019. Advocates of recursive self-improvement view this trend as the start of a continually upward curve toward a complete autonomy of self-improvement that is comparable to Moore’s Law. 

Yet these comparisons break down on closer inspection. If the AI industry really cares about reaching RSI, we can’t overlook the value of introspection. Settling for the current RSI narrative risks being complacent about what achieving RSI truly requires. 

‘Just engineering’ and the perils of misappropriating Moore’s Law

Conventional wisdom in recent discussions around RSI is that the LLM capability gains that put AI on the path to RSI are a matter of engineering and upward scale (of data, compute and parameters). Well-known AI researcher Andrej Karpathy captured this sentiment in a March tweet that went on to be quoted in recent high-profile media articles discussing RSI. Layers of iteration and gradual engineering alone, we’re told, are on track to achieving it. 

Yet making steady progress through engineering is one thing; creating AI that can independently improve itself is quite the other. 

This brings us to another, closely linked and commonly held assumption among RSI evangelists: that progress towards RSI is tracking against a Moore’s Law-style curve (in step with the exponential advances seen in computing power as more transistors are added to microchips over time, as observed by Gordon Moore). But this analogy is weak. Moore’s Law doesn’t simply entail fitting more transistors on the same chips. It has been accompanied by several generations of microchip advances serving as step changes in underlying technology that shifted the art of the possible. Those shouting from the rooftops that RSI is imminent aren’t talking about a similar step change from the transformer, the dominant architecture that all current leading models are built on. 

As mentioned, there’s no question that there have been serious capability gains made by this current generation of models. According to METR’s latest analysis, today’s most advanced frontier models achieve a time-task horizon of three hours (with a 80% success rate) working on software tasks autonomously. But the minimum requirement for AI systems to achieve full RSI-like self-improvement, in time-task horizon terms, should be closer to 40 hours with a >80% success rate. 

READ MORE: System shocks: Rogue OpenAI agents hijacked a wiki. Humans attacked the power grid

That is a real signal: RSI depends on AI systems being able to sustain long research and development cycles with minimal human oversight. But duration alone is not enough. An agent could spend days optimising parameters without producing a genuinely new insight. 

The bigger challenge is moving beyond the limits of today’s transformer-based models. Current systems are trained to predict the next token based on vast quantities of data, relying largely on knowledge acquired during training. 

To get closer to RSI, AI will need to do more than generate better outputs. It must be able to learn and adapt from experience, retain knowledge and discover new approaches independently, meaning the transformer-centric stack must be rethought entirely. 

Progress towards RSI is likely to depend not just on scaling today’s models, but on developing architectures that can continuously improve their own capabilities.  Until we solve the challenges of continual learning, memory and adaptive reasoning, genuine RSI will remain more of an engineering help than a breakthrough on its own.

Facing up to the crisis

The implications of sticking by the transformer architecture go further still. Doing so means ignoring an unsustainable reality of spiralling energy demands and costs that’s getting worse by the day. The transformer leaves the industry running on a context treadmill, compensating for weak built-in memory and limited continuous learning by using more inference-time computation. This has come in the form of Chain of Thought (CoT) reasoning, which sees models consume more and more ‘thinking’ tokens to generate more thought through responses. As a consequence, energy consumption per prompt has tracked upwards with each new model launch. 

This doesn’t just make for a growing risk of an AI-driven energy supply crunch; it’s also starting to hit enterprise bottom lines. In recent months, we’ve seen the shift to agentic workflows unleash an exponential surge in token consumption that far outpaces any downward pressure on token prices.

READ MORE: OpenAI Astra hits “critical” security threshold amid fears its reasoning will soon be “opaque”

We’ve seen Uber disclose that it blew through its annual 2026 AI coding-tool budget in a matter of four months after adopting Claude Code. In an ultimate irony,  this token furnace is seeing per token compute exceed the cost of the human labour it was meant to replace in some cases. Far from enabling a smooth path to RSI, there are more and more signs that the current trajectory is economically fragile. 

And we need to be clear-eyed about the fact that transformer shortcomings are playing a factor in the chaos. Distinguished AI experts NYU Computer Science Chair, Martin Farach-Colton, and NYU Tandon Professor of Computer Science, Julian Togelius, have both argued publicly that current AI models have limitations so fundamental that they’re unlikely to be solved by brute-force scaling alone - namely brittle reasoning, and the inability to reliably turn short-term context into durable long-term memory.

Putting model capability gains on a Moore’s Law-like trajectory is one pressing issue in AI. Controlling the unsustainable economics born out of runaway growth in CoT reasoning and agentic workflows is another. Both are going to require a step change in the underlying architecture of models.

What changes are needed to bring us closer to RSI? 

The silver lining to all of this is that a growing number of AI labs are facing these problems head-on by thinking beyond the transformer in favour of new architectures. Pushing the RSI frontier requires models that continually learn and remember from even longer horizons. Last year, we unveiled the Dragon Hatchling (BDH) architecture. BDH models break from the transformer by containing a living state that updates with every interaction. Taking cues from how learning and memory emerge in the brain, BDH models see artificial neurons strengthen and weaken over time as experience is generated over a model’s deployment in an enterprise. This is a significant shift away from transformer models, which are limited to drawing from static, pre-trained weights and therefore can’t adapt or learn to individual deployments.

Crucially, evolving memory allows these systems to build on previous learnings rather than repeatedly solving the same problems from scratch. Instead of discarding context once a session ends, BDH can accumulate expertise over time, moving the model closer to the kind of continual learning required for recursive-self improvement. 

BDH also makes a crucial break from the constraints of language. While transformer models reduce every problem to text, making them inefficient at deeper reasoning tasks that operate across multiple variables, BDH uses a latent reasoning space that enables learning, adaptation and problem solving beyond language. The language based capabilities transformer based models have pioneered are preserved, yet reasoning can be done in a richer, internal state that arrives at solutions more efficiently with lower compute. 

Together, evolving memory and latent reasoning point toward a future where AI systems become more capable and economical. Rather than relying on growing models and ever-greater amounts of compute, BDH enables intelligence to compound through experience and more efficient, economical reasoning, remembering past trials without using more tokens. The result is delivering significantly more intelligence per dollar, making AI systems cheaper to run, more autonomous and easier to scale. 

RSI – time to get real

It’s a positive that RSI is entering the public consciousness. The concept of self-improving AI has stark implications for society and is very real. But figures talking about RSI the loudest have a duty to inject realism into the discourse. Moore’s Law-like exponential acceleration in AI model capability gains requires a step change in the architecture that underpins models. The longer that debate is put off, the further away from RSI we will be.

Follow Machine on LinkedIn