Gemini Self-Improving AI: UK SME Guide
Recursive self-improvement is when an AI system becomes capable enough to improve its own training or code without much human help, each gain feeding the next. Google says AI agents now write most of its new code, and co-founder Sergey Brin has reportedly pushed Gemini toward full self-improvement, though no lab has demonstrated it running autonomously yet.
AI Advisers is a Milton Keynes-based AI consultancy that helps UK SMEs evaluate new AI developments and implement the workflows worth acting on. The analysis below separates what is verified from what is reported, so you can decide what — if anything — to do about it.
Recursive self-improvement: key facts for UK SMEs
- Google said in an April 2026 blog post that AI agents now generate roughly 75% of its new code, up from 50% in autumn 2025 and 25% in October 2024 (The Decoder).
- Google co-founder Sergey Brin has reportedly told DeepMind staff to prioritise recursive self-improvement (RSI) — AI that upgrades itself with little human input — according to Reuters reporting relayed via Dealroom and FourWeekMBA. This is a reported internal push, not a demonstrated capability.
- No AI lab has shown a fully autonomous self-improving system. Reuters reported on 15 September 2026 that Anthropic's Dario Amodei, OpenAI's Sam Altman and xAI's Elon Musk jointly called for caution, with "some AI executives putting it three to five years away" (Reuters, via CP24).
- Google's newest model, Gemini 4 Argon, began rolling out to cybersecurity partners on 30 September 2026 and more widely from 1 October 2026 (Khaleej Times).
- 54% of UK SMEs now use AI in some form, up from 35% in 2025, 25% in 2024 and 23% in 2023, per a 2026 British Chambers of Commerce and Atos survey in which 94% of respondents were small or medium-sized businesses (Workplace Insight).
What is recursive self-improvement, and why is everyone talking about it now?
Recursive self-improvement (RSI) describes an AI system that becomes capable of improving itself — its own training, code or research process — with little to no human help, where each improvement feeds the next one. Reuters defines it as "the idea that an AI system could become capable of improving itself, with little to no help from humans, and allowing each advance to help produce the next one" (Reuters, via CP24).
The topic resurfaced this month after a widely shared X post claimed Sergey Brin had confirmed Gemini was "entering" RSI and that AI agents now handle most of Google's production code and bug fixes. Our research found the picture is more nuanced than that framing suggests: Google has confirmed a figure for AI-generated code, but the RSI push itself is reported, internal direction — not an announced or demonstrated capability. Brin reportedly addressed DeepMind staff at a town hall in April 2026 urging them to prioritise RSI and close the gap with rivals including Anthropic and OpenAI, and Reuters reported later in the summer that he was shifting resources toward self-improving systems (Dealroom). No verbatim public quote from Brin confirming "Gemini is entering RSI" has surfaced in the reporting we could verify, and analysis of the story is explicit that this is "reported, not announced" with "no evidence of a working autonomous self-improving system anywhere in the industry" (FourWeekMBA).
That caution matters because other labs are racing on the same frontier from different angles — not just the largest players. Moonshot's recent open-weight release, covered in our piece on Kimi K3, is one example of how quickly capable models are now shipping outside the three best-funded US labs.
How much of Google's code is AI already writing?
This is the one figure in the story that comes straight from the company itself, not from unnamed sources. Sundar Pichai said in an April 2026 blog post that 75% of Google's new code is now AI-generated, up from 50% in autumn 2025 and 25% in October 2024 — and that the work has shifted from single-line autocomplete to "truly agentic" workflows where engineers supervise AI agents that plan, execute and refactor (The Decoder; Mobile Syrup). Pichai's post described a complex code migration "completed six times faster than was possible a year ago" using agents and engineers working together (Mobile Syrup). All AI-generated code is still reviewed and approved by human engineers before it ships.
Google is not alone. Anthropic has separately reported that Claude now writes over 80% of its own production codebase (The Decoder). These two figures are not directly comparable: they come from different companies, cover different codebases, and there is no public detail on whether both count the same way (lines generated versus lines merged, for instance). What they share is direction — AI-written code is becoming the default at the labs building the models, not the exception.
Is recursive self-improvement already happening at Google, or is this still a target?
The honest answer, based on what's verifiable, is that it's a stated direction rather than a demonstrated result. A survey of 25 researchers across OpenAI, Anthropic, Google DeepMind, Meta and US universities, conducted in late summer 2025 by IAPS fellow Severin Field, found that 20 of the 25 rated the automation of AI research as one of the most severe and urgent AI risks — but the same reporting notes that several of the milestones those researchers predicted (gold-medal performance at the Maths Olympiad, AI-authored peer-reviewed papers) have already landed, while full autonomous self-improvement has not (The Decoder).
On 15 September 2026, Reuters reported that Amodei, Altman and Musk jointly called for a pause in development, warning AI "could soon improve on its own and slip beyond human control." Amodei specifically warned that "recursive self-improvement could eventually outrun humanity's ability to understand and control AI systems if pursued without sufficient safeguards," with "some AI executives putting it three to five years away" (Reuters, via CP24). That three-to-five-year estimate, from the people running the labs in question, is a very different timeline from the near-term framing circulating on social media this week — worth knowing if you're deciding how urgently to react.
What is Gemini 4 Argon, and what does it have to do with self-improving AI?
Gemini 4 Argon is Google's newest flagship model, which began reaching trusted cybersecurity partners on 30 September 2026 before a wider rollout through the Gemini app and API from 1 October 2026 (Khaleej Times). Google describes it as its "most performant model yet built for complex workloads," positioned for software engineering, enterprise knowledge work and cybersecurity defence, and claims it is comparable to OpenAI's and Anthropic's frontier models on key coding and cyber benchmarks — though it reportedly trails on two of four coding-specific benchmarks (Khaleej Times). Bloomberg has reported that some Google staff say the model performs less well on real-world coding tasks than the benchmark scores imply, a claim Google disputes.
Argon isn't itself evidence of RSI — it's a product release. But it's the model generation Brin's reported push is aimed at improving, and it followed our earlier coverage of Google's Gemini 3.8 speech models, which shows how quickly this product line is now moving.
What does recursive self-improvement mean for UK SMEs?
For a UK small or medium-sized business, the practical story isn't "the robots are self-improving" — it's that the tools you already use (Gemini, ChatGPT, Copilot) are being built faster, by teams that increasingly lean on AI agents themselves. That compounding speed is already showing up in adoption numbers closer to home: 54% of UK SMEs now use AI in some form, up from 35% in 2025, 25% in 2024 and 23% in 2023, according to a 2026 British Chambers of Commerce and Atos survey (with analysis by the BCC Insights Unit and the University of Essex) in which 94% of respondents were small or medium-sized enterprises (Workplace Insight). Most of that usage is still generic tools rather than bespoke AI systems, and the same survey found 95% of SME AI users reported no change in headcount over the past year, with 86% saying job roles stayed the same — though around one in ten bespoke AI adopters do expect workforce reductions (Workplace Insight). That's a UK SME sample answering about their own businesses, not a projection about the wider economy.
The gap this opens is less about RSI itself and more about the pace of change it implies: new model generations, new capabilities and new risks arriving faster than most small teams can evaluate them. Building basic AI fluency across a team — not just in the owner's head — is the single biggest lever most SMEs have for keeping up, a point we cover in more depth in building AI literacy so your team understands what it's using.
What should a UK SME do about recursive self-improvement right now?
Nothing about this story requires a UK SME to act today — none of it describes a live capability reaching small-business tools yet. But it's a reasonable prompt to check three things:
- Audit before you automate. Before adding another AI tool because a new model generation looks impressive, work out which of your actual workflows are worth automating — not all of them are, and we set out how to tell the difference in what business workflows are actually worth automating with AI right now. If you'd rather have that automation built for you, that's the scope of workflow automation for small businesses.
- Build the culture, not just the toolkit. Teams that treat AI tools as something IT switched on, rather than part of how the business works, tend to get less from them — we go through what that looks like in practice in building an AI-first company culture.
- Centralise rather than bolt on. If you're running several disconnected AI tools, a single operating layer across the business — what we install as an AIOS — is usually a better foundation than adding one more point solution every time a new model ships.
Quotes worth reading in full
"Recursive self-improvement could eventually outrun humanity's ability to understand and control AI systems if pursued without sufficient safeguards," Anthropic's Dario Amodei warned, alongside OpenAI's Sam Altman and xAI's Elon Musk, in a joint call for caution reported by Reuters on 15 September 2026 (Reuters, via CP24).
Glossary
- Recursive self-improvement (RSI) — an AI system becoming capable of improving its own training or code with little human help, with each gain feeding the next.
- AI agent — software built on an AI model that can plan and carry out multi-step tasks with limited supervision, rather than answering a single prompt.
- Foundation model — a large, general-purpose AI model (such as Gemini or Claude) that can be adapted to many different tasks.
- Benchmark — a standard test used to compare AI models on a specific skill, such as coding or maths.
- AIOS — an "AI Operating System": a single, governed layer of AI tools and workflows installed across a business, rather than separate point solutions.
Frequently asked questions
What is recursive self-improvement in AI?
Recursive self-improvement is when an AI system becomes capable of improving itself — its training, code or research process — with little to no human help, so each advance helps produce the next one (Reuters, via CP24).
Did Sergey Brin say Gemini is entering recursive self-improvement?
Reuters reported that Brin addressed DeepMind staff in April 2026 and urged them to prioritise recursive self-improvement, and reported later in the summer that he was shifting resources toward self-improving systems. No verified verbatim quote from Brin stating Gemini has entered RSI has surfaced in the reporting available to us.
What percentage of Google's code is written by AI?
Google said in an April 2026 blog post that AI agents generate roughly 75% of its new code, up from 50% in autumn 2025 and 25% in October 2024 (The Decoder).
Is recursive self-improvement already happening at Google?
No fully autonomous self-improving system has been demonstrated publicly by any lab, including Google. What is confirmed is that AI agents write most of Google's new code under human review, and that Brin has reportedly pushed DeepMind toward full RSI as a goal.
What is Gemini 4 Argon?
Gemini 4 Argon is Google's newest flagship model, rolled out to cybersecurity partners from 30 September 2026 and more widely from 1 October 2026, built for coding, enterprise knowledge work and cyber defence (Khaleej Times).
What is Gemini 4 Argon used for?
Google positions Gemini 4 Argon for software engineering, enterprise knowledge work and cybersecurity defence. It began rolling out to cybersecurity partners on 30 September 2026 before wider availability from 1 October 2026, and Google claims it is comparable to OpenAI's and Anthropic's frontier models on key coding and cyber benchmarks (Khaleej Times).
How many UK SMEs are using AI in 2026?
54% of UK SMEs now use AI in some form, up from 35% in 2025, according to a 2026 British Chambers of Commerce and Atos survey of businesses that was 94% small and medium-sized enterprises (Workplace Insight).
When do experts think recursive self-improvement will arrive?
Some AI lab executives put full recursive self-improvement three to five years away, according to Reuters reporting from 15 September 2026 — considerably further out than the near-term framing in some recent social media commentary (Reuters, via CP24).
Should a UK SME be worried about this right now?
Not urgently. None of the verified reporting describes a capability that has reached small-business AI tools. A practical response is to audit which of your workflows are actually worth automating and ensure your team understands the tools it already uses — both of which are covered in AI Advisers' free trial. The three-to-five-year timeline from the lab executives themselves suggests there is time to build good foundations rather than react to headlines.
Recursive self-improvement support for UK SMEs
The headline framing — "Gemini is building Gemini" — is more exciting than what's currently verifiable: a reported internal push at Google, a confirmed 75% AI-code figure, and a new model release, with every named lab executive putting full autonomous self-improvement years rather than months away. What is real is the pace: new model generations, new agent capabilities and new terminology arriving faster than most small teams can evaluate on their own.
As an AI consultancy working with SMEs across Milton Keynes, Luton, Bedford and Northampton, we regularly help teams separate verifiable capability from vendor hype — and identify the workflows that are genuinely worth acting on now.
AI Advisers is based in Milton Keynes and works with UK SMEs to audit their current AI use, train their teams, and automate the workflows that are actually worth automating. If you want help making sense of what's hype and what's worth acting on, our Milton Keynes AI consultancy team can talk it through, or you can start with a free trial in Milton Keynes.