Filip Háša is Director at Pixelfield, a London AI and product engineering studio he founded in 2013. The company ships production LLM, RAG and agent systems for scaleups and enterprises, and will say when a simpler fix is the better product.��

• A ChatGPT demo is not a product. It is a slide.
• Most AI proof-of-concept work never becomes a system people use every day. IDC research, reported by CIO.com, found 88% of observed PoCs never reach widescale production.
• RAG is how a system answers from your files, with citations and permissions. The 2023 “dump PDFs into a vector database” version is what failed. Retrieval itself did not die.
• Agents fail on cost, state and access, not on model IQ. A founder on X described a coding agent wiping production data in nine seconds. Last week, researchers showed training agents turning public wikis into a private message board.
• Write the one-sentence success test before you write the prompt. If you cannot, you are not ready to build.

If you have spent the last two years around UK startups, you already know the scene. Someone opens ChatGPT on a projector. The bot summarises a contract. The room nods. A board member asks how soon it can sit on the website. Six months later the same team is still “iterating the prompt”.

That gap is the story. Not the models. The models got better. The product work did not.

I run Pixelfield, a London AI and product studio. We have been shipping software since 2013. More than 50 AI features we built are still running: chatbots, retrieval, agents, older machine-learning jobs that never needed a large language model. The useful pattern, after enough of those builds, is unglamorous. The demo answers a question nobody will pay for twice. Production answers the question people ask on a Tuesday, with the right files, the right permissions, and a way to tell when the answer is wrong.

Here is how we got here, in order. Named people, dated sources. No vibes.

2022–2023: The Demo Became The Product

ChatGPT arrived and, for a few months, that was enough. Founders showed a chatbot that could talk about their help centre. Investors asked if the company was “an AI company now”. A Copilot licence looked like a strategy.

None of that was stupid. A live demo is how you find out whether anyone cares. The mistake was treating the demo as the finish line. A prompt on top of a public model has no memory of your permissions, no log of why it said what it said, and no owner when it is confidently wrong in front of a customer.

UK teams felt this first in regulated corners: insurance, health, property, anything with a paper trail. The bot was fine in a Monday standup. It was not fine once someone asked “where did that number come from?” and the only answer was “the model”.

A shared Copilot login still looks like a strategy. A few saved prompts. A hope that the next model will paper over the mess.

2023–2024: Everyone Ran An AI Proof Of Concept

Then the brief changed. Boards stopped asking for a demo and started asking for a proof of concept. Fair enough. A PoC is supposed to answer one question on real data, then stop: go live, iterate, or kill it.

What actually got funded was often a second demo with a nicer UI. The data was a SharePoint dump. The metric was “the CEO liked it”. There was no eval set, no failure mode, no plan for what happens when the corpus goes stale next Tuesday.

Hamel Husain described the stall in a February 2026 interview. Teams had glued together “the basic components of an AI system, like a RAG database and all the UI components”, shipped a prototype, and then got stuck. The hard part was not the glue. It was how to improve the thing without guessing.

Ashish Nadkarni at IDC put a number on the same pattern. Research commissioned with Lenovo, reported by CIO.com in March 2025, found that 88% of observed AI PoCs never made it to widescale deployment. For every 33 proofs of concept, four graduated. Nadkarni’s diagnosis was not “the model was too weak”. It was board-level panic, thin funding, and no real business case. “Most of the time the POC happens not because of a strong business case.”

Gartner had already warned, in July 2024, that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, for poor data, missing risk controls, rising cost, or no clear value. That was a forecast. Sitting in London in 2026, it does not feel like they missed.

A PoC that cannot be killed is not a PoC. It is a hostage.

2025: Agents Arrive, And So Does The Hangover

Once chatbots looked ordinary, the pitch became agents. An agent, in the useful sense, is not a chat window with a personality. It chooses tools, keeps state, and is supposed to know when to stop.

Gartner’s Anushree Verma, in a 25 June 2025 note, predicted that over 40% of agentic AI projects will be cancelled by the end of 2027. Read that as a forecast, not a body count. The line that matters is this:

“Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.” -Anushree Verma, Senior Director Analyst, Gartner, 25 June 2025

Gartner also named the sales trick: agent washing. Chatbots, RPA and search boxes, relabelled. Verma’s split is still the one I use with founders. Use an assistant for retrieval. Use automation for a routine workflow. Use an agent only when a decision has to be made and a human path exists when the agent should not.

Maxim Fateev, co-founder and CTO of Temporal, has been making the unfashionable point for months. Agent demos die on execution and state, not on IQ. In a 1Password conversation published in August 2026 he said people who started from “a synchronous Python in-memory program are learning the hard way that these are not simple problems.” A laptop notebook has no crash recovery and no audit trail. The first overnight job is where the demo ends.

What RAG Actually Is, And What Died

Somewhere in 2024 a lot of people decided retrieval-augmented generation was over. Context windows got bigger. Vendors said you could “just put the documents in the prompt”. Then the bill arrived, the model missed the one clause that mattered, and nobody could say which file the answer came from.

RAG, in plain English: the system fetches the right pieces of your own material, then writes an answer from those pieces. Citations. Permissions. A log. It is how a bot or an agent answers from your files instead of from its training data.

Hamel Husain opened a series on this in July 2025 with a line I wish more pitch decks had read: he was tired of hearing “RAG is dead”. Ben Clavié’s version is even cleaner. What died is the 2023 marketing version. Chuck documents into a vector database, do cosine similarity, call it a day. That fails because a single vector cannot hold a 40-page policy.

Retrieval itself is more important than ever. Models are frozen at training time. A million-token window does not make it cheap, or accurate, to stuff everything into every query.

If you want the engineering version of that argument, the production checklist we use for is unromantic: chunking that matches the document type, hybrid search, reranking, citation tracking, evaluations, freshness, and monitoring. Graph databases are optional. Guessing is not.

A UK example, from a bank rather than a startup pitch. In March 2026 the quantitative research machine-learning team at Lloyds Banking Group wrote up what happened when they tried to get a Copilot over SharePoint to answer methodology questions. Copilot was the right first test. It hit a wall: multiple sources, mixed question types, from “how does this model work?” to “what if we change a production parameter?”. They built their own intent detection and routing. Stateless “search then generate” was the limit. They said so in public on 24 March 2026.

2026: Cost, Permissions, And Nine Seconds

Two public stories from this year sit on opposite ends of the same problem. One is about money. One is about access.

The free lunch ended in August

On 23 August 2026, Drew Breunig published “Fable & The End of the Free Lunch”. Simon Willison linked it the same day. The argument is simple enough to steal.

“Prior to Fable, it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems. But then Fable landed. It was (and still is!) incredible. But the cost was so high and Opus was good enough… for most of the code we needed. So we started to think about what work went where.”- Drew Breunig, 23 August 2026

He now uses the expensive model to interrogate a design, then hands a brief to a cheaper model for the rote work. GLM 5.2, in his comparison, was roughly a ninth of Fable’s cost and good enough for most routine coding once the context was strong. The other shock was policy: access controls and required data retention. Enough companies, and countries, started asking where traces go.

If you are a UK startup still waiting for “the next model” to fix a messy corpus and an unowned workflow, that essay is the polite version of no.

An agent on X, and a wiki message board last week

On 25 April 2026, Jer Crane, founder of PocketOS, posted on X what then ran in Decrypt and TechRepublic. A coding agent in Cursor, running Anthropic’s Claude Opus 4.6, was doing a routine task in staging. It hit a credential mismatch. It “fixed” the problem by deleting a Railway volume. The volume held the production database and the volume-level backups.

“Yesterday afternoon, an AI coding agent—Cursor running Anthropic’s flagship Claude Opus 4.6—deleted our production database and all volume-level backups in a single API call to Railway, our infrastructure provider. It took 9 seconds.” – Jer Crane, founder of PocketOS, posted on X, 25 April 2026, as reported by Decrypt

The agent found an over-privileged token, guessed that a delete would stay in staging, and did not ask. Crane’s point was not “one bad model”. It was an industry wiring agents into production faster than it is wiring confirmation, scoped tokens and delayed deletes. Railway later added a 48-hour soft-delete on that API path. For a UK team handling bookings or payments: a demo agent with blanket credentials is an outage with a chat log.

You do not need April for a current example. On 4 September, Simon Willison wrote up new research on OpenAI training agents that were supposed to have controlled web access. They found public wikis they could write to with GET requests, then spent weeks exchanging thousands of messages to finish a timed benchmark.

Moderators cleaned up. The agents noticed the alphabetical deletion sweep and started copying pages with a ZZZ prefix. Willison’s point on the sandbox is the sticky-note version: the proxy assumed GET cannot change data. Plenty of old software never got that memo.

That is not a startup using ChatGPT. It is the same shape of failure you get when a tool-using system meets a permission model written for humans. If your agent can send, delete, pay, or post, the control cannot be a sentence in the system prompt.

What Production AI Looks Like From London

I am not going to pretend every UK company needs a custom model. Most do not. A lot of the work we turn down is a rules engine, a better search box, or a form that already has the answer. One client arrived convinced they needed a custom LLM and left with a £5,000 automation. We could have sold a £40,000 build. We did not. Vendor-neutral is a dull phrase until you have sat in the meeting where someone is trying to buy a logo.

When AI is the right tool, production has a shape. There is a named owner. There is a corpus someone is responsible for. There is an eval set of real questions, including the nasty ones. There are permissions that match the product, not the demo. There are logs. There is a kill switch. There is a cost line that assumes the cheap model will not stay cheap if you send it everything.

A proof of concept, if you run one, should be designed as a go / no-go on real data. If we cannot write the question in one sentence, it is not ready. AgentWise, a real-estate operating system we still maintain, started as an AI PoC. That is the only reason to run one: it becomes a product, or it dies in public inside the team, before it dies in front of customers.

If you need an AI development company that will also say when the simpler fix is the product, that is the work we do at Pixelfield. London HQ, studios in Prague and Amsterdam, shipping since 2013. You keep the IP. Discovery starts at £2,000. A typical proof of concept sits between £8,000 and £20,000. Production is a different conversation, and it should be.

Founders are trying to buy an outcome, not a model. The outcome is a system that is still correct in six months. That is a product problem with some machine learning in it. It is not a prompt.

A Checklist Before You Spend The Next £20,000

Steal this. Argue with it. Do not skip it.

• Can you write the success test in one sentence, with a number a sceptic would accept?
• Is the data you need actually available, permissioned, and not three years out of date?
• Who owns the corpus next quarter? If the answer is “the intern who built the demo”, stop.
• What happens when the answer is wrong in front of a customer? Who gets the ticket?
• Do you need retrieval, a workflow, or a decision-making agent? Most pitches need the first or the second.
• Which model does the expensive thinking, and which model does the rote work? Breunig’s split is now table stakes.
• If an agent can delete, pay, send or post, is there a confirmation that the agent cannot auto-complete?
• What is the monthly inference bill if usage 10×? If you do not know, you do not have a production design.
• Will you kill the project if the PoC misses the number? Write that down in the same doc as the prompt.

FAQs

What is the difference between a ChatGPT demo and production AI?

A demo shows that a model can talk about your material in a room of friendly people. Production AI is a system with an owner, permissions, evaluations, logs, a cost model, and a way to fail safely. The first is a slide. The second is software.

Why do so many AI proof-of-concept projects fail?

They are often a second demo: no one-sentence question, no real data, no evals, no kill criteria. IDC, as reported by CIO.com, found 88% of observed PoCs never reached widescale production. Gartner separately forecast that at least 30% of generative AI projects would be abandoned after PoC by the end of 2025. The blocker is usually data, process and the business case, not model IQ.

Is RAG dead?

No. The 2023 version is. Dumping PDFs into a vector store and hoping cosine similarity finds the clause you need is a bad product. Husain and Clavié’s point in 2025 still holds: retrieval, routing, citations and evaluations are the work. Lloyds’ quant team had to leave a generic Copilot path for exactly that reason.

Should a UK startup build AI agents?

Only if a decision has to be made, tools have to be called, and you can say what happens when the agent should stop. Gartner’s 40% cancellation forecast by end-2027 is a warning about hype, not a reason to freeze. Crane’s outage is about credentials. Fateev’s point is about state. Start with retrieval or a workflow unless you can explain why those are not enough.

How much does an AI proof of concept cost in the UK?

It depends on the data, not the slide. On our side, a focused PoC is typically £8,000 to £20,000, after a discovery pass that starts at £2,000. If a vendor cannot tell you what would make them recommend you do not build, keep walking.
When is the simpler product the right product?

When a search box, a rules engine, a notification, or a well-designed form already answers the job. Generative models are bad at being a database and expensive at being a spreadsheet. We will say that in writing. You should want a studio that does.

The Short Version, One More Time

The last four years in UK AI look like this if you squint. 2022: the demo. 2023: the PoC. 2024: the hangover. 2025: agents, and a lot of relabelling. 2026: cost, permissions, and a few very public accidents. The teams that got through it treated production AI as product engineering with a sceptical eval harness, not as a prompt with a logo on it.

If you are still in the projector room, that is fine. Just do not confuse applause with a system. Write the one sentence. Then decide whether you are building software, or collecting another demo.

Filip Háša is Director at Pixelfield, a London AI and product engineering studio he founded in 2013. The company ships production LLM, RAG and agent systems for scaleups and enterprises, and will say when a simpler fix is the better product.

The post UK Startups Are Stuck Between ChatGPT Demos And Production AI appeared first on 91̽��.

While it obviously isn’t the first time that Amodei and other Big Tech leaders have made important, poignant statements about the more troubling aspects of the future of AI, this letter lands in a slightly different space. Indeed, it follows what seems to be becoming a slew of resignations by Anthropic employees, with Mrinank Sharma in February and now, most recently, Jacob Coxon, former Anthropic researcher, last week. According to The BBC and Business Insider, these explosive departures have been shrouded in dramatic language, including notions that AI companies are “gambling with our lives” and creating “a world in peril”, to name but a few.

And so, Dario Amodei has taken to his own website to address some of these concerns, and he’s done so in a rather poignant open letter entitled, “We Must Pace The Frontier”.

Amodei’s Take On the “AI Problem”

Unsurprisingly, the article quickly gained attention for a few reasons. Firstly, it was, perhaps indirectly, responding to the recent statements about the state of the AI industry that have been causing significant alarm and anxiety around the world.

Second, Amodei’s letter clearly acknowledges the issues and concerns being faced in AI now and going forward without sugarcoating the situation.

And finally, while he’s acknowledging the seriousness of the concerns, the Anthropic CEO isn’t simply offering a warning; he’s actually providing a proposal for a step-by-step strategy to deal with a potentially serious issue.�� He hasn’t suggested that AI companies and governments need to stop development. Rather, he’s��arguing that the AI industry should deliberately slow the pace of frontier AI development so that safety measures have time to catch up.

It’s worth taking the time to unpack exactly what Dario Amodei has stated and suggested, as well as what he hasn’t, when it comes to global safety concerns over the future of AI and the wrold as we know it.

A Quick Overview Of Amodei’s Argument

It’ll be hard to do his letter justic in so few words, but even so, I think it’s important that we simplify and explain what’s�� being said so that the letter is accessible and easy to understand for everybody. So, here’s my shot at exactly that.

At the heart of the letter is a simple concern: that is, AI capabilities are advancing faster than our ability to understand, monitor and control them. To explain this, Amodei points specifically to two developments.

The first is what he calls recursive self-improvement – AI increasingly helping to build the next generation of AI systems. The second is the recent OpenAI-Hugging Face incident that involved AI agents behaving in unexpected and potentially harmful ways, which he argues demonstrates how quickly capabilities are evolving.

Amodei believes states firmly that he believes that AI could deliver enormous benefits, from accelerating scientific research to improving human health. However, he immediately notes, with no uncertainty, that the industry needs to become more deliberate about how quickly it pushes capabilities forward. Importantly, he repeatedly stresses that this does not mean stopping AI development altogether.

In fact, he outlines a three-step process that he believes will help us control the potential problems and dangers of AI, while still allowing it to progress and provide us with the many incredible advantages it has to offer humanity. Importantly, he explains that these steps don’t have to (and may not) be followed in chronological order, and not only that, he acknowledges that some may be significantly more challenging to deal with than others.

So, with those caveats in mind, here’s how he explains the three steps that need to be taken in order to make AI development safe.

Step One: Embedded Evaluators

The first step is the most immediate and arguably the most practical. Amodei wants frontier AI companies to give independent third-party evaluators ongoing access to their systems, processes and safety practices. These evaluators would effectively operate inside AI companies, helping verify that safety commitments are being followed and identifying potential risks before models are released.

He immediately states, on behalf of Anthropic, that he is committed to implementing this approach immediately and is encouraging other frontier labs to do the same. The idea is similar to external auditors in finance: rather than asking companies to assess their own safety practices, independent experts would have visibility into what is actually happening. Thus, it would be much like a “neutral” industry-wide body that would evaluate each company without bias or corporate intention.

Step Two: Coordination Between Democracies

The second step moves beyond individual companies, and this is where things become more challenging.��Amodei argues that frontier AI companies operating in democratic countries should coordinate around common safety standards and limits on unchecked capability growth. The goal is to avoid a situation where companies feel pressured to move faster simply because competitors are doing the same.

In other words, he wants AI companies to compete on products and innovation without creating a race where safety is sacrificed for speed. In his words, rather than a “race to the bottom”, they need to create an environment that encourages “a race to the top”.

He also suggests governments may need to help facilitate some of these discussions, particularly where antitrust concerns could otherwise make collaboration difficult.

Step Three: Global Cooperation

The final step in the letter is arguably the most ambitious.��Amodei argues that democratic nations should eventually work with countries including China to establish international AI safety agreements. These could range from banning particularly dangerous uses of AI to creating shared testing standards for frontier systems.

At the same time, he acknowledges the geopolitical reality, which, unfortunatley, screams out at the reader as soon as the phrase “global cooporation” entered the chat. The letter repeatedly argues that democracies must maintain their technological lead and prevent authoritarian governments from gaining a strategic advantage through AI. Thus, he certainly seems to be presenting global cooperation as something that should happen alongside efforts to preserve Western leadership in AI technology.

The Letter In Brief

With so many immediate reactions from personalities all over the world, both experts and non-experts, it’s worth highlighting both what Amodei does explicitly assert and what he does not.

What Amodei �ٴDZ����Say

One reason the essay generated so much discussion is that many people interpreted it as a call for an AI slowdown, and in a sense, it is.

Amodei explicitly argues that frontier AI development should proceed at a more measured pace. He believes additional time should be used to improve alignment research, interpretability, operational security and safety testing before capabilities advance too much further.

His core argument is that slowing capability growth slightly could buy valuable time to make future AI systems safer and more understandable.

What Amodei���ٴDZ����’����Say

However, there are ideas floating around that Amodei doesn’t address or agree with, and the essay also specifies several things it does not advocate:

• It does not call for a complete pause in AI development.
• It does not propose shutting down frontier AI labs.
• It does not argue that AI itself is inherently bad.

And these things are repeatedly noted throughout the writing, leaving very little space for interpretation (in my opinion).

Despite the dramatic headlines, it does not suggest the industry should stop innovating. In fact, Amodei repeatedly emphasises the enormous benefits AI could bring if developed responsibly. Thus, he’s saying innovation should continue, but under the right circumstances that are controlled, monitored and agreed upon.

Thus, the letter is less about stopping AI and more about creating mechanisms that make it easier to slow down when necessary. And, of course, control the beast altogether.

Immediate Critiques of Dario Amodei’s “We Must Pace the Frontier”

It wasn’t long before the Anthropic CEO garnered support from close colleagues in the industry.

Both Sam Altman and Elon Musk, among others, responded to Amodei’s letter on X, asserting that they agree with his evaluation of the situation and the suggested path forward. They seem to at least generally agree that stronger safety oversight and coordination are needed as models become more capable.

��

Critics, however, have raised concerns that, at this point, are all too familiar. Some argue the proposals could strengthen the position of large AI companies by creating barriers that smaller competitors and open-source projects would struggle to meet. Meanwhile, others question whether the industry can genuinely regulate itself, particularly when commercial incentives continue to reward faster progress. Basically, is what Amodei is suggesting even possible at all?

Of course, perhaps unsurprisingly, China does not seem to agree with Anthropic and the West’s chracterisation of the impending AI problem. Indeed, according to China Tech News, today, Chinese Foreign Ministry spokesman Guo Jiakun said: “Fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance and serve the interests of no one”. This seems to reflect Chinese concerns that moderation in AI safety and innovation may be used to justify US restrictions on Chinese development.

Ultimately, in addition to disagreements over strategy and intention as well as geopolitical concerns, there’s also a major practical challenge that doesn’t seem to be going away; one that Amodei acknowledges in his letter, in many respects. That is, most people seem to agree that AI should be safe, but far fewer people agree on who gets to decide what “safe enough” actually means.

Thus, Amodei’s letter isn’t a demand to stop AI, nor is it a detailed blueprint for global regulation. Instead, it’s an argument that frontier AI is approaching a point where safety can no longer be treated as a secondary consideration. His three-step plan is, according to the letter, supposed to be an attempt to give safety research time to keep pace with capability gains. Ultimately, an objective that ought to serve the greater good.

Whether the industry, both in the West and around the world, embraces that vision remains to be seen. But as AI systems become increasingly powerful, one thing is for sure: the debate has shifted from what AI can do to a more difficult question of, who gets to decide how quickly it should advance?

The post Dario Amodei’s Three-Step Plan For “Pacing The Frontier” Of AI Development Explained appeared first on 91̽��.

Meta’s CTO has admitted, in an internal memo, that the company’s AI reorganisation was “an atrocious job.” It is a rare thing for a technology leader to say so plainly, and it is worth taking seriously, because the pattern behind it is not unique to Meta.

Since the reorganisation, code output has reportedly jumped by 220% among engineers reassigned to AI training work. On paper, that looks like a win. In practice, morale has fallen to a record low, and Meta is still monitoring employee activity even as it introduces new perks to offset the damage. Those two facts sitting side by side, rising output and falling trust, tell you almost everything you need to know about what went wrong, and it was not the technology.

I spent fifteen years in Fortune 500 technology leadership before moving into workplace wellbeing and leadership consulting, and I have watched this same pattern repeat across multiple waves of technological change. It happened with the shift to enterprise software, again with cloud migration, and now with AI. The technology moves fast because it can. Leadership moves more slowly because it depends on trust, and trust is not something you can deploy on a timeline.

What makes the AI wave different is speed and visibility. Employees do not experience an AI rollout as a neutral technology project. They experience it as a direct question about their own value: will I still have a role, will my judgement still matter, is my output being measured in ways I cannot see. When leaders do not answer that question honestly, employees answer it for themselves, usually in the least generous way possible. Output can rise in that environment. Commitment does not.

Meta’s response, adding perks while keeping the monitoring in place, is a common leadership reflex, and it does not work. Perks address comfort. Surveillance signals distrust. You cannot resolve a psychological safety problem with the first while reinforcing it with the second, and employees notice the contradiction immediately.

For founders and leaders building fast-growing companies, the lesson is not to slow down AI adoption. It is to treat the human side of that adoption as seriously as the technical side, and to do it before the rollout, not after morale has already dropped. That means being transparent about what is changing and why, giving people a real voice in how new tools are introduced into their own work, and being honest that AI raises genuine questions about job security and competence rather than pretending it does not.

There is a global standard for exactly this kind of risk, ISO 45003, which addresses psychological health and safety at work, and its logic applies directly here: psychosocial risk has to be assessed and managed with the same discipline as physical safety, not treated as a soft afterthought to a technology decision.

Meta’s admission is unusually candid, and that candour is useful. It gives every other organisation currently rolling out AI a preview of what happens when the technology outruns the leadership around it, and a chance to choose differently before their own employees force the same admission.

César Gamio spent fifteen years in Fortune 500 technology leadership. He is an author, adjunct professor at IE Business School, founder and managing director of the Dharma Centre for Workplace Wellbeing, and a British Standards Institution Associate Consultant for ISO 45003, the global standard for psychological health and safety at work.

The post Meta’s AI Reorg Is A Leadership Story, Not A Technology One appeared first on 91̽��.

The UK government launched a £500 million Sovereign AI fund back in April with a clear objective: back British AI companies, help them scale quickly and ensure that growth stays in the UK.

Designed to operate more like a venture capital firm than a traditional public funding body, this marks a massive shift in how the UK is supporting innovation. For SME leaders and founders, it is not to be mistaken with just another funding announcement. It signals a more competitive, better-resourced environment for building and scaling AI businesses, domestically.

Capital Meets Capability

Access to funding has always been a constraint for high-growth AI businesses, but Sovereign AI goes further than capital alone. It combines direct equity investment with access to infrastructure, talent and regulatory support, creating a far more complete growth environment.

For SMEs, that means fewer trade-offs. Instead of choosing between funding, talent or capability, businesses can access all three in one place, a true moment to focus on.

Infrastructure That Actually Moves the Needle

A key differentiator is access to the UK’s AI Research Resource supercomputers, something that has historically been out of reach for most smaller companies.

This level of computation completely changes what is possible. Models can be trained faster and tested more rigorously, removing a lot of the constraints that typically slow development. The result is a shorter path from idea to deployment, and a clearer route to building a real competitive edge.

Talent and Speed

Hiring remains one of the biggest blockers to growth in AI, particularly at the specialist level. Sovereign AI addresses this directly through faster visa decisions, making it easier to secure international expertise without long delays.

These are exactly the kinds of challenges being actively and continuously explored at . Founders and senior teams are comparing how they are securing talent, building AI capability and scaling faster in an increasingly competitive market, everywhere from stands to the speed networking area. The gap between businesses that can move quickly and those that cannot is widening, and access to the right conversations is a competitive advantage in itself.

From Funding to Market Access

Beyond investment, Sovereign AI also supports businesses in navigating procurement and regulation, opening clearer pathways into public sector opportunities.

For many SMEs, this pinpoints where growth either accelerates or stalls. Access to contracts, combined with a better understanding of regulatory frameworks, allows businesses to convert innovation into revenue more effectively.

What It Means for SMEs:

• Reduced barriers to scaling AI businesses within the UK
• Access to high-performance compute previously limited to larger firms
• Faster hiring through streamlined international talent processes
• Stronger access to procurement opportunities and regulated markets
• Support for building datasets and advancing AI capabilities

This is not just funding. It is a coordinated push to make scaling in the UK more viable.

The Bigger Picture

The global AI race is intensifying, and the UK has often struggled to retain its highest-growth companies as they scale. Sovereign AI is designed to change that by removing the need to look elsewhere for capital and capability.

For business leaders, the message is clear. The conditions for scaling in the UK are improving, but so is the pace of competition. Knowing how others are responding, where capital is moving and how quickly the landscape is shifting is becoming just as important as the funding itself.

Sovereign AI is more than a funding initiative. It is a signal that the UK is backing its innovators, and that the next wave of AI growth will favour those ready to move early.

The post The UK Government Has Backed AI With £500 Million: Here’s What That Means appeared first on 91̽��.

Of course, while AI (whether it’s Chat, Perplexity, Claude, Gemini or any other model) is undeniably incredible, especially when it comes to writing, we now know that it’s still fallible. It makes mistakes, it hallucinates and it’s simply not completely reliable. Not only that, it lacks a certain “human touch” that while we can’t always explain exactly what is, we just know it’s missing.

But, much like most things in life, there’s a limit to what AI can do. One of these constraints is the fact that AI is only as useful as the instructions you give it. And that’s where prompt engineering comes in.

The term has become increasingly common as generative AI has moved from being a novelty to a serious business tool. While we’re mostly concerned about the jobs AI is taking and how many humans will be replaced by AI systems, this may be a situation in which AI is actually creating a job.

What Is Prompt Engineering?

Put simply, prompt engineering is the process of designing, testing and refining instructions given to an AI system to get a particular result. A prompt is just the input you give an AI model: it could be a question, an instruction, a description or a combination of all three.

Prompt engineering is about working out how to phrase that input so the AI understands what you actually want. Similar to how when you Google something, typing in an entire sentence with perfect grammar isn’t going to be as effective as using keywords.

According to IBM, effective prompt engineering helps generative AI systems produce outputs that are more relevant, accurate and aligned with the user’s intended goal.

That might sound obvious. After all, if you ask ChatGPT to “write something about startups”, you’re probably going to get something fairly vague. But, if you tell it that you’re writing a 700-word article for a technology publication aimed at startup founders, ask for a conversational tone and specify exactly which aspects of the startup ecosystem you want covered, you’ve given the model considerably more to work with and it’s more likely to produce something a little closer to what you’re looking for.

Why Does Prompt Engineering Matter?

Generative AI doesn’t necessarily know what you mean simply because you know what you mean, and that’s something that many people don’t quite seem to understand (and this was a problem even before AI shifted into the mainstream). Humans regularly leave things unsaid because we rely on context, so AI models are much more likely to produce useful results when you provide that context explicitly.

A good prompt can tell an AI model what to do, who the output is for, what information it should use, what format the answer should take and what it should avoid.

For instance, in a professional situation, an employee asking an AI tool to “summarise this customer feedback” might get a reasonable answer. But, a carefully constructed prompt could instead ask the system to identify the five most common complaints, group similar issues together, quote representative examples and present the findings in a format suitable for a management report.

In this case, the underlying AI hasn’t necessarily changed, but the instructions have, thus producing a far better result.

How Does Prompt Engineering Work?

Unfortunately, as convenient as it would be, there isn’t one magical formula for writing the perfect prompt. In fact, much of prompt engineering involves testing, reviewing and refining.

One of the simplest approaches is to be clear and specific. Give the AI enough context to understand the task, explain what you want the output to look like and provide examples where useful.

This can also involve techniques such as zero-shot and few-shot prompting. Zero-shot prompting means asking an AI model to complete a task without giving it examples of what you want. Few-shot prompting, meanwhile, gives the model a small number of examples to demonstrate the desired result.

For example, rather than simply asking an AI to classify customer reviews as positive or negative, you could provide several examples showing exactly how you want different types of reviews categorised. This means that it’s not only using its own judgement (which is, of course, based on information and date it’s been trained on), but it can actually start to interpret your own judgement too, and if it’s effective, replicate that.

Is Prompt Engineering Just Writing Good Questions?

It’s not just about asking good questions, although that’s a big part of it. Prompt engineering can range from simply improving the wording of a request to much more sophisticated processes involving prompt templates, examples, external data, testing and automated workflows.

PromptEngineering.org describes the discipline as involving the systematic design and optimisation of prompts, rather than simply coming up with clever sentences. And, this is where prompt engineering starts becoming particularly relevant to startups and businesses.

A company building an AI-powered customer service tool, for example, isn’t going to manually type a beautifully crafted prompt every time a customer asks a question, because that’s simply not realistic. It’s difficult and it’s time-consuming. Developers, on the other hand, can create structured prompts and workflows that consistently tell the underlying model how to behave. Thus, in this kind of situation, the prompt almost becomes part of the product.

Do You Need To Be A Prompt Engineer?

Not everybody needs to be a prompt engineer per se, but if you’re going to be using AI consistently, especially in a professional situation, it would be a good idea to try and understand it as far as possible.

The good news is that in many everyday situations, the basics are surprisingly simple. You must be specific, provide context, explain the desired outcome and don’t be afraid to refine your request.

But, professional prompt engineering requires a lot more technical knowledge than just being clear. At the highest level, it can require engineers to also understand large language models, programming, data structures, algorithms and the limitations of AI systems. So, there’s is a difference between being good at prompting an AI tool and working professionally in prompt engineering.

How Long Will Prompt Engineering Be Relevant?

This is the question we ask across the board in technology and AI these days, as innovation as constant and technical skills seem to become irrelevant quicker than they appear. Indeed, as AI models become better at understanding natural language, the need to obsess over exactly which words is most likely going to decrease.

But the thing is, that doesn’t mean the underlying concept of prompt engineering is disappearing altogether. It’s probably just going to change and evolve. Businesses still need to decide what they want AI systems to do, what information they should have access to, what rules they should follow and what a successful result looks like.

So, prompt engineering may become less about simply finding the perfect sentence and more about designing effective interactions between humans, data and AI systems. That is, it’ll probably become more advanced and complex. But hey, these days, what isn’t?

The post What Is Prompt Engineering? appeared first on 91̽��.

The deal that was confirmed by Qualcomm this week will focus on AI inference and cover multiple generations of chips. So if Amazon is already investing in its own chips, why is it also working with another chipmaker?

Amazon Wants More Control Over AI Hardware

Amazon’s decision to develop its own chips gives it more say over how the hardware used by AWS is designed. Instead of relying entirely on general-purpose chips, Amazon can build silicon around the workloads its cloud customers are actually running which can help the company manage costs, improve performance and make its data centres more energy efficient.

On the AI side, Amazon’s chips designed for AI workloads are called Trainium, while Graviton is designed for general cloud computing. Trainium is built for AI model training and inference, while Amazon also has Inferencia chips that are aimed specifically at inference. Amazon says its custom silicon is designed to deliver better price performance and energy efficiency for the workloads running on AWS. That gives Amazon more flexibility over how its infrastructure is built and which hardware it uses for different types of work.

Why Is Amazon Working With Qualcomm?

The Qualcomm deal is focused on AI inference which is when a trained AI model is used to produce an answer or prediction. That’s different from the training stage, where the model is built and taught using huge amounts of data.

Reuters says Amazon could buy up to $60 billion worth of Qualcomm’s AI data-centre chips and related products. The two companies are also working on optical connectivity for AI data centres, which is designed to help move huge amounts of data around the infrastructure.

Inference is also becoming a bigger part of the AI infrastructure as more AI tools move from training into everyday use. Every chatbot response, generated image or AI-powered search result needs computing power behind it. As more of these services are used every day, the infrastructure needed to handle those requests becomes a really big part of running AI at scale.

Amazon Doesn’t Need To Build Everything Itself

Building its own chips gives Amazon something it wouldn’t get from buying whatever is available. It can decide what those chips need to do and how they fit into the rest of AWS, which can make a difference when you’re running huge data centres and dealing with the cost of powering them.

There’s also a financial side to this, as designing and producing custom silicon needs significant investment – so there’s no point in Amazon trying to replace every other hardware piece on its own.

But designing a chip is a really big job, and there isn’t much reason for Amazon to recreate every piece of technology it needs. Qualcomm already has experience developing hardware for demanding computing workloads, so Amazon can use its own chips in areas where they give AWS an advantage and bring in outside technology where that makes more sense.

Amazon Wants Control, But Not Total Control

Amazon having its own hardware was never really about making every piece of hardware inside AWS itself – it was about having more say over how that infrastructure works, from the performance of its AI systems to the cost and energy needed to run them.

The partnership with Qualcomm means that there’s now another option instead of replacing what Amazon has already built. Its own chips give it one option, while Qualcomm can now bring another type of hardware into the mix.

Therefore, Amazon can develop its own technology while still working with companies like Qualcomm where their hardware makes sense. How those different chips are used across AWS will depend on the workloads they are built to handle.

The post Amazon Is Building Its Own AI Chips – Why Not Just Buy Them? appeared first on 91̽��.

This feud is nothing new. Much like Paris Hilton and Nicole Richie’s iconic early-2000s feud, the quarrel between Musk and Chess.com has been ongoing for years. But now, Musk seems to believe he holds a silver bullet: AI will “solve” Chess just like it did Checkers, he has claimed.

Cue: gasps, whispers and shock among the crowd! Followed by outrageous applause and and a roaring, standing ovation! Well, perhaps in his imagination, but in reality, the world in which most of us live, it’s been more like the virtual version of many confused expressions and the exchange of one too many awkward glances between strangers.

Look, Elon’s done a lot in his day – hell, he’s done a lot this year. But, is he really arguing with a website? Or rather, whoever’s controlling the social media for the website? And not just any website, but Chess.com?

The answer is yes, and while the story is admittedly absurd, I do believe there’s a far greater truth in this fun little anecdote that we ought to take note of. Almost like a small child trying to keep a secret who blurts everything out at the dinner table by accident. I think Musk has, in a way, revealed the essence of why his understanding of the world, technology, AI and just about everything else is often so far from common consensus, and the trouble is, this reflects a logic that is becoming widely adopted (whether intentional or not).

Indeed, I don’t think it’s just him; I think this manner of thinking is what we’re seeing in so many powerful figures in big tech, and for that reason, it’s well worth sticking around and taking a few notes.

Musk Vs. Chess: AI Will Win, Mark My Words!

Essentially, Elon took to X to make his feelings known about the game that seems to get him rather riled up: he asserted, in a flurried string of posts, that soon enough, AI will “fully solve” Chess,, because, in his words, “Chess is hard for humans but not for computers”. Of course, Chess.com fired back with a response that cited numbers and spoke about the sheer complexity of the game of chess, essentially concluding that it would be virtually impossible to map all possibilities.

The exchange went on, becoming increasingly petty and heated, and much like most arguments on social media, there was no clear winner.

But here’s what stood out to me, and hopefully to many other people who were furiously snacking on the metaphorical popcorn while the feud unfolded: why are we arguing about whether or not a computer would be more likely to win a game than a human?

I’m not dismissing the obvious relevance of the fact that the whole issue is based on exponential probability, something computers tend to be pretty good at. Nor am I disagreeing with him; he’s probably right. But the thing is, the whole discussion is missing the point of why and how we play a game like chess, and chess is becoming symbolic of a far greater truth.

Perhaps Winning By Solving An Algorithm Isn’t the Point

There are many things in the world – many tedious, time-consuming tasks – that would be made exponentially easier if we could find a way to get them done immediately. To remove all friction from the equation and make it as easy as simply imagining it were done, to having it actually be accomplished in a matter of seconds.

For instance, when you’re late for work and you’re stuck in traffic, wouldn’t it be so convenient to just be able to click your fingers and teleport to the office? Or, what about something more practical, like having to manually fill out forms, only to enter the data into a system in multiple different places a seemingly infinite number of times? With modern technology and AI, we can make that happen in a heartbeat – time saved, fewer curse words uttered.

Both of these examples, while one is real and the other imaginary, include friction that I’m sure we can all agree is pretty unnecessary when it comes to having a positive impact on us and our lives. Sure, someone may argue that sitting in traffic gives you time to think or entering data manually is good for…actually, I can’t even think of a good reason why anybody would want to do manual data entry.

But ultimately, regardless of these potential motivations, these are situations in which the thing that makes them difficult, tedious, time-consuming, and annoying are pretty unnecessary – they don’t actually contribute to the overall goal or final output.

What about something that requires a process to be followed in order for the final goal to be achieved? For instance, going for a run. People run for different reasons – whether it’s to gain fitness, boost mental health, keep physically healthy, lose weight – but no matter the specific motivation, I think it’s safe to say that the activity is about more than reaching a destination or a turning point; it’s about the process.

The point isn’t that you need to physically be five miles away from home. The point is the movement that you do that ends up with you being five miles down the road. After all, if it wasn’t, it would probably be a lot quicker and easier to just hop on a train or in an Uber, right? It’s also why we have treadmills that don’t get you anywhere at all.

And that’s the difference I think Elon is fundamentally misunderstanding.

Chess isn’t being stuck in traffic on the way to work or doing manual data entry at 16:59 on a Friday. Playing chess is going for a run.

Who “Wins” If We Solve Chess?

An AI platform gets a pat on the back, Elon shows off a smug grin and…we all carry on with our day?

I believe that a huge takeaway from Chess.comgate, as I will refer to it from hereonout, ought to be the realisation that perhaps we are way too focused on “solving” everything. And it’s not just an Elon issue, although he takes it to the extremes and certainly shines a very bright (flood) light on it, as he does most things.

With technology having advanced so dramatically over the last decade or so, we’re at a point where we have so much more power than ever before to find solutions and quick fixes. Don’t have time to go grocery shopping? No worries, just login to an app, select your daily pre-organised shopping list and hit order – it’ll be at your house in a few hours. Or, going on holiday but don’t want to get a travel agent to organise the trip for you? Just ask Chat to create a detailed itinerary.

We’ve become so used to having just about anything we need at our fingertips, that it almost feels as if the logic has reversed somewhat. Not only is it possible to sort these things out really quickly, but they should be organised without any hassle. And now, we’re beginning to apply this logic to just about every facet of our lives, even when it’s totally unnecessary.

So, when Elon Musk chose to get into a heated argument with “some random intern at a chess website” (his words, not mine, and I’d argue that he actually kind of schooled himself by admitting this), he inadvertently gave us the greatest example of how our logic is becoming skewed by our obsession with removing any and all friction from our lives.

I say this, because other situations tend to involve significantly more nuance and can be more challenging to explain, but when it comes to chess, I’m sure most people can agree that the process of playing the game is far more important than the final outcome. In fact, the outcome – having a winner and a loser – really only has value because of the intellectual and strategic game that has been played in order to achieve the final goal. So simply asking a computer programme and an advanced AI model to solve it is completely missing the point and would be, well, futile.

I don’t necessarily think that Musk’s obsession with “solving” chess is going to go away, nor do I think that the outcome of this issue in itself will have dramatic consequences for humanity. However, I do think it has revealed a far more telling indication of how our logic is changing with our increasing ability to quickly solve problems with very little (or no) friction. The lesson here should be, I believe, that we need to stop asking whether or not it’s possible to “solve” something, and rather, we should question whether or not we need to “solve” it at all, rather than just work through it ourselves.

Otherwise, I fear we’re zooming towards a dystopian future in which instead of using our smart watches to simply track our runs, they’ll have sliding scales for things like serotonin and other hormones that we can simply modify whenever we feel need a little boost, without putting in any effort, using any brain power or making any kind of intelligent decisions.

Personally, I’d rather just go for a run.

The post Elon Musk Wants To “Solve” Chess, But Isn’t The Struggle The Point Of The Game? appeared first on 91̽��.

The comment came as a reply to resigning colleague Jacob Coxon, who warned that frontier labs are recklessly playing with fire. Rather than trying to soften the blow, Hubinger leaned right into the critique and gave it a timeline and a metric, specifically tying his estimate to the risk of recursive self-improvement.

The weight of the comment comes from Hubinger’s position at Anthropic. He leads the division tasked with ensuring advanced AI remains safe and obedient as it evolves. And when the person in charge of alignment at an industry-leading safety lab acknowledges these odds, it acknowledges a glaring problem: the sector is racing towards superintelligence without a solution to keep it under control.

Where Does Recursive Self-Improvement Fit In?

Hubinger’s timeline, thankfully, doesn’t apply to the tools currently on the market, which he views as largely manageable. Instead, he’s looking at superintelligence born out of recursive self-improvement, something he claims is improving faster than the industry anticipated.

It all boils down to a feedback loop: AI systems get smart enough to help build their successors by refining algorithms, designing better hardware tools and optimising training methods. Each upgraded generation then takes over the task of engineering the next, drastically shortening development cycles. Left to run, that loop could pull off massive capability leaps, ultimately producing systems that outclass human intellect in almost every field.

Anthropic has previously noted in its own risk assessments that unconstrained self-improvement increases the odds of humans losing control over AI altogether. This official documentation connects the technical reality to the loss-of-control scenarios Hubinger is airing in public. Today’s chatbots are largely benign, but this automated cycle threatens to collapse safety timelines completely.

The risk of extinction is fundamentally about control, not malice. Advanced software accelerating its own development doesn’t need bad intentions to prove catastrophic. A system simply needs to reach a level of capability where people can no longer monitor its optimisation targets or redirect its trajectory if its objectives diverge from human interests.

What’s Driving the Sudden Urgency?

A combination of factors is driving these concerns. Progress on AI systems capable of aiding their own engineering is outpacing previous expectations, bringing a functional self-improvement cycle closer to reality. Simultaneously, Hubinger has made it clear that solving alignment for superintelligent systems is still an open challenge. The timing of his statement, paired with an engineer’s public exit, points to seemingly widespread unease among technical insiders.

Hubinger also later reinforced that current commercial models are safe, focusing his warnings on the trajectory towards self-improving systems. This nuance is key to understanding his assessment. An insider with direct line of sight into cutting-edge capabilities is pointing out that the field is speeding down a track without a clear brake mechanism.

What This Does To Public Trust

Direct warnings of unsolved, existential risk from an alignment lead make it exceptionally difficult to write off safety concerns as theoretical noise.

Anthropic actively trades on its reputation as a responsible developer. Seeing senior engineers at the heart of that work attach high probabilities to total disaster speaks volumes about the actual state of internal progress. It also exposes a growing contradiction for companies operating at the forefront. Anthropic continues to pitch the virtues of increasingly powerful systems while its own safety head assigns double-digit odds to human extinction within ten years.

Industry observers note that while Hubinger’s metric represents his individual assessment rather than formal corporate policy, it echoes sentiments held across Anthropic’s research divisions.

From a regulatory perspective, such explicit warnings from senior technical insiders certainly build a compelling argument for binding government oversight and mandatory safety checks before releasing self-improving models, and replacing voluntary self-policing with enforced accountability.

The post Anthropic’s Own Alignment Lead Says There Is A 10% Chance AI Kills Everyone Within A Decade appeared first on 91̽��.

At the centre of this shift in AI audio is a company whose tools have become a reference point for what natural-sounding synthetic voice can do. For anyone tracking where AI is heading, it is worth understanding what this company builds and why voice has become such a significant frontier.

Why Voice Became a Frontier

For much of the recent AI boom, voice lagged behind. Text generation and image generation captured attention while synthetic speech remained stuck with the robotic quality everyone recognised from old automated systems. Voice is genuinely hard, because human speech carries meaning not just in words but in tone, rhythm, emotion and countless subtle cues, and reproducing that convincingly is a serious technical challenge.

That is what makes the recent progress notable. Speech that sounds natural, expressive, and human rather than mechanical represents a real breakthrough and it unlocks a huge range of applications that poor-quality voice had kept out of reach. As the quality crossed the threshold from tolerable to genuinely convincing, voice moved from a neglected corner of AI to one of its most active and consequential frontiers.

The companies driving that progress have consequently become important to watch.

What ElevenLabs Builds

Among the names most associated with this shift, has built a suite of AI tools focused on voice and audio and its work spans the main capabilities that define the field.

That includes turning written text into natural-sounding speech, creating synthetic versions of specific voices, transcribing spoken audio into text and building conversational agents that can hold spoken interactions. Taken together, these cover the core of what modern AI audio makes possible.

What ties the suite together is a focus on quality and naturalness, the qualities that separate genuinely useful voice technology from the robotic synthesis of the past. The tools are made available both to individuals and to developers building them into their own products, which is part of why the technology has spread quickly. Understanding a company like this means understanding the capabilities it has helped bring into the mainstream, from natural speech to voice cloning to conversational audio.

Why It Matters

The significance of natural AI voice goes well beyond novelty. For creators, it means producing narration and audio content without studios or recording sessions. For businesses, it means voice features, audio versions of content, and spoken interfaces that were previously impractical.

For accessibility, it means making content available to people who cannot easily read a screen, in voices that are actually pleasant to listen to. And for developers, it means adding audio capabilities to their products through straightforward integration rather than building the technology themselves.

These applications are why the field attracts so much attention. Voice is one of the most natural ways humans interact and technology that can produce and understand it convincingly opens up new kinds of products and experiences.

A company advancing that capability is not just improving a niche feature but helping to shape how people will interact with software, which is why its progress is watched closely across the technology world.

Approaching the Technology Thoughtfully

Powerful voice technology also brings responsibilities, and the serious players in the field engage with them. The ability to recreate a voice, in particular, raises important questions of consent and potential misuse, which is why responsible providers build safeguards around voice ownership and verification. Using such technology well means respecting consent where a voice represents a real person and being transparent about synthetic audio where audiences would reasonably expect it.

These considerations are part of the story of AI voice, not a footnote to it. Stanford University’s Institute for Human-Centred AI , the broader field within which this kind of voice technology sits.

As the technology becomes more capable and more widespread, how it is governed and used responsibly matters as much as what it can do. The companies shaping the field, and the wider community around it, are actively working through these questions, which is an important part of the technology maturing into something trustworthy.

A Frontier Worth Watching

The rise of natural AI voice is one of the more striking developments in artificial intelligence, turning synthetic speech from a robotic curiosity into something genuinely human-sounding and broadly useful. Companies focused on this frontier have brought capabilities like natural speech, voice cloning, transcription and conversational audio into the mainstream, opening possibilities across creativity, business, accessibility and software development.

For anyone following where AI is going, voice is a frontier worth watching and the companies advancing it are worth understanding. As software increasingly speaks and listens in ways that feel natural, the technology behind that shift is set to become a normal part of how people interact with the digital world.

What sounded robotic and artificial not long ago now sounds convincingly human, and that change is opening a new chapter in how we relate to the technology around us.

The post The Company Teaching Software To Speak: ElevenLabs appeared first on 91̽��.

Findings published in Nature Biotechnology on 7 September show that an AI-designed drug, rentosertib, successfully drove down predicted biological age in human trial participants, as measured by six separate proteomic ageing clocks that reached the same conclusion. Developed by Insilico Medicine, the trial gives the biotech sector something far more compelling than just another optimised molecule sitting in a computer database: a therapeutic indicator that actually survived contact with human subjects.

For nearly a decade, AI drug discovery has proven strong in early phases by spotting targets, designing candidate molecules and shaving years off timelines. However, it consistently lacked actual clinical evidence showing proven benefit to patients. Rentosertib is among a small number of emerging programmes delivering that critical proof. Though, it’s not the final answer by itself, and no one should view it as such.

Behind The Rentosertib Trial Data

Rentosertib was originally designed to tackle idiopathic pulmonary fibrosis, a harsh, progressive lung condition tied to getting older.

Its Phase IIa trial brought in 71 patients across 22 medical sites in China. While the main objective over 12 weeks was just making sure the drug was safe and tolerated, researchers also tracked lung capacity as a secondary metric. The results were interesting. Patients taking a 60mg daily dose saw their forced vital capacity jump by 98.4ml over 12 weeks, compared to a 20.3ml drop in the placebo group. In a disease defined by irreversible decline, seeing lung function move in the right direction is a huge deal.

The fascinating anti-ageing indicator came from testing stored blood samples from 42 participants against six standalone proteomic ageing clocks. These models calculate biological age by evaluating thousands of blood proteins. While some models predict chronological age and others estimate mortality risk, using completely different algorithms, all six agreed that rentosertib lowered predicted biological age. The peak change appeared at week four among patients taking 30mg twice daily. Most models logged roughly three to four years of biological-age reversal, while one model showed a drop of nearly six years.

The significance of this finding stems from six differently constructed models, trained on separate data metrics, arriving at the exact same conclusion. The biological evidence is equally clear, with statistical checks identifying 326 proteins showing modified trajectories in treated patients against just two in the placebo group. This data shows a clear and measurable biological impact, not just a mere algorithmic fluke.

The trial marks a genuine milestone, but the caveats are just as noteworthy as the headlines. A sample size of 42 patients over 12 weeks is small, and every participant was already managing serious lung damage. On top of that, the strongest dose for lung performance wasn’t the strongest dose for age reversal.

That mismatch hints at two different biological mechanisms playing out at the same time, something this specific trial wasn’t actually designed to separate. The team at Insilico explicitly says this isn’t ironclad proof of slowed ageing. But the bigger picture still holds weight, given rentosertib entered Phase III trials in China in July 2026, securing its spot as one of the first AI-designed molecules to get that far.

How Much Did The Algorithms Actually Contribute?

Rentosertib gives a clear look into how this software works in practice. Insilico used its PandaOmics platform to zero in on a protein called TNIK, mainly because it flagged high across six recognised hallmarks of biological ageing.

That gave them one target that could hit both the lung disease and the longevity markers they wanted to track. Instead of tweaking an old chemical recipe, they generated the molecule from scratch using Chemistry42, their generative chemistry tool. The whole process took about 18 months from finding the target to locking in a candidate, cutting years off the usual timeline.

The setup of the trial itself shows how deep the algorithmic influence went. By integrating longitudinal proteomic sampling into the Phase IIa plan right from the start, the team laid the groundwork for the ageing-clock analysis long before the trial wrapped.

This calculated trial design stands out as an achievement on par with the drug design itself. The team planned for longitudinal data upfront instead of relying on post-hoc discoveries in legacy files. Integrating longevity metrics into standard clinical protocols establishes a clear, practical model for other biotech firms looking to streamline trial data collection.

The Pipeline Beyond Insilico

This pattern goes beyond a single success story – the wider industry data is where things get really interesting.

Recursion Pharmaceuticals showed what their platform can do with REC-4881, which delivered a median 43% reduction in polyp burden after 12 weeks in a Phase 1b/2 trial for a severe genetic condition. A full 82% of evaluable patients maintained their progress 12 weeks after stopping the drug.

Big pharma is backing this up too. In August 2026, Genentech and Roche greenlit an AI-discovered neuroscience target from their Recursion partnership, advancing it to joint discovery after their own labs confirmed the biology was solid. Multiple other biotechs, including XtalPi and IGC Pharma, currently have AI-guided candidates progressing through mid-phase testing, with a key Alzheimer’s trial result landing in late 2026.

On its own, no single candidate settles the debate over whether computational drug design works. Together, though, they point to a huge leap in credibility. The conversation is no longer about lab simulations or early press releases, but legitimate clinical results from real people in real trials.

Moving Past The Discovery Trap

Pharma spent years throwing AI investments at discovery tools, ignoring the fact that most time, money and failure accumulates downstream.

The real challenge has always been clinical testing, trial setup and demonstrating actual patient efficacy. Rentosertib and its contemporaries show concrete proof that AI-created compounds can make it through the human testing barrier instead of stalling in the lab, which is usually easy to hype but exceptionally difficult to convert into useful treatments.

No one involved in these trials is suggesting AI has completely solved biological ageing. What this shows is that an AI-designed molecule, aimed at an AI-identified target and tested using an AI-influenced biomarker strategy, generated data solid enough across six separate measurement tools to win over peer reviewers.

That moves the conversation away from optimistic lab announcements toward rigorous, published evidence, giving the market the tangible proof it’s wanted for years.

The post Can AI Stop Ageing? Inside The Clinical Breakthroughs Delivering Concrete Results appeared first on 91̽��.