Alexander Kardos-Nyheim

In my personal capacity, I actively invest in the unsolved science underneath artificial intelligence: the five problems the frontier is blocked on, and the researchers working on them.

In many parts of the field I believe we moved from exploration to exploitation too quickly. That is one reason the technology has not delivered the economic uplift expected of it in the complex parts of the economy, and one reason the market is now correcting: the capability was built out before the science underneath it could carry the weight. Capability has become abundant. The properties these five areas supply have not, and they are what the complex parts of the economy are waiting on.

Founder & CEO of Safe Sign Technologies, a frontier AI research company acquired by Thomson Reuters in 2024 before it had revenue, in the first pre-revenue acquisition in the Nasdaq-listed tech giant’s 174-year history and one of that year's more significant European AI deals. I co-lead AI research at Thomson Reuters, where the team behind Thomson released the company's frontier model in August, and I also lead the model’s adoption by some of the most sophisticated organisations in the world. I read Law at Trinity College, Cambridge, and trained and qualified as a lawyer with Allen & Overy LLP, London.

LinkedIn

Five bottlenecks

Each of the five below is an unresolved scientific problem. Each constrains what can be built above it, and each currently receives a small fraction of the capital going into the systems that depend on it. If the market is now asking what its spending has bought, the answer is under the bonnet: the vulnerabilities showing up in today’s conditions were built in at this level, and this is where they can be understood and mitigated.

Those conditions, as of November 2026: the four largest buyers of compute will spend $700bn or more on it this year, up about three quarters on 2025 and close to the whole of their operating cash flow, with more than $1tn projected for 2027. On 5 November the Nasdaq fell 5.5 per cent in a day and entered correction, with the semiconductor index down 12 per cent. Whether the revenue arrives to earn a return on that spending is the open question. The five problems below are where I would look for the answer. Each one is either adding to the cost of the buildout or standing between it and the revenue.

  1. Plasticity

    Models cannot keep learning.

    A model is frozen at the end of training. To absorb anything new it has to be rebuilt from scratch, at a cost that is now growing faster than the capability it buys. Every system that has to adapt after deployment inherits this problem, and most of them hide it behind retrieval. It is also why the buildout is a recurring cost and not a one-off: a frozen model has a useful life set by how fast the world moves, and the only way to extend it is to rebuild. The cost today is that capability is bought, then rebuilt, and never kept, and that is a large part of the bill the market is now questioning. If the problem is solved, the rebuild becomes maintenance, and every model and every chip already paid for lasts longer.

    In Nature, Dohare, Mahmood and Sutton found that standard deep learning in continual settings degrades until networks learn no better than a shallow network, and that methods based on gradient descent are not enough. Epoch AI puts a single frontier training run above one billion dollars by 2027.

  2. Calibration

    Models cannot be verified or controlled well enough to be trusted.

    I think this is why artificial intelligence has barely entered medicine, aviation, energy, defence and law, which is where most of the economic value sits. Appetite is not the obstacle and nor is regulation. The obstacle is that nobody can yet produce evidence about a learned system strong enough to carry the weight being put on it. That is also where the revenue the buildout needs is sitting. The cost today is that the sectors holding most of the value buy AI as assistance, beside the decision and at assistance prices, and the return the market is now asking about stays out of reach. If the problem is solved, a learned system can be put inside the decision, and the revenue in those sectors follows.

    Verifying properties of even simple networks is NP-complete, and six years of open competition have not moved the ceiling far: the largest network anyone has entered has 138 million parameters, orders of magnitude below anything deployed. The UK AI Security Institute states that it is not yet possible to build full safety cases that scale. RAND calculates that showing an autonomous vehicle merely matches human fatality rates would take 275 million failure-free miles.

  3. Embodied substrate

    The training data for physical interaction does not exist.

    Vision and language were handed their corpus by the internet. Touch, force and contact were not, and will not acquire one by accident. This is the constraint I think is most widely misread: the models are being funded and the instrument that would produce their data is not, at a ratio of roughly seventeen to one. The cost today is capital committed to robots that cannot yet feel, priced as if the data to teach them existed; in a correction those are the positions that reprice first. If the instruments and the collection are funded, the result is a corpus no laboratory can download, and whoever holds it owns a layer the robot companies will have to buy.

    World-model laboratories took a record $3.2bn in a year, against about $191m for every company generating the data they would learn from. The capital committed to humanoid robots rests on the layer the $191m is building. Where I have looked hardest, the limiting factor is instrumentation, not architecture.

  4. Mechanism

    Systems represent correlation, not cause.

    A system that has learned what tends to accompany what cannot be relied on to act in a situation it has not seen. Causal representation and world models are the attempt at an alternative, and both are early. The fashionable half of this problem, world models that learn to simulate what happens next, is well funded. However, the other half of the field is focused on what can be identified from data, and this is close to empty: the theory and the methods that say when a cause can be recovered from the records people actually hold, what assumptions that recovery rests on, and how to tell a causal claim from a statistical one. That half decides whether a system can be trusted to act in medicine, in an industrial process, or anywhere else somebody intervenes. The revenue case for the buildout now rests on agents that act as well as answer, and acting requires knowing what an action will do. The cost today is that agents built on correlation fail quietly once they intervene, which is where deployments stall between pilot and production, and where the revenue the buildout counts on is being lost. If the problem is solved, a system can act as well as answer, and that is the difference between a tool that gets used and one that gets paid for.

    World-model laboratories have raised billions. Pure-play causal inference companies have raised roughly one European seed round between them.

  5. Cycle time

    Discovery is not yet a closed loop.

    Machines can propose and machines can synthesise. Nothing can yet decide reliably which experiment is worth running next, which is the step that would compress the time between a hypothesis and a result. Whoever closes that loop gets faster at everything else. The cost today is that discovery companies are priced on what they can propose, while the revenue arrives only when a proposal becomes a made, characterised and trusted result, and that is the part of the loop nobody has closed. If the loop is closed, the time from hypothesis to result shortens in every field that depends on experiment, and the instrument and the record that close it belong to whoever built them.

    Reviewing 2.2 million computationally predicted materials, Cheetham and Seshadri found scant evidence for compounds meeting the trifecta of novelty, credibility and utility. The leading autonomous-synthesis paper was formally corrected in Nature.

What I back

  • Companies whose central claim is scientific, and testable.
  • Positions that survive a platform replicating or bundling the capability underneath them.
  • Work at the point where the science is still mispriced, which is usually earlier than is comfortable.

What I do not back

  • Legal, tax, accounting and news technology, in full. My role prohibits me from investing in these areas: legal research, drafting, practice and matter management, and legal AI; corporate legal operations, e-billing and transaction intelligence; tax preparation, research, compliance and automation; audit, confirmation and accounting software; e-invoicing and indirect-tax compliance; risk, fraud, identity and regulatory compliance; global trade and customs; legal, tax and investigative systems for government; and news, financial information, data and events.
  • Applications built on a model somebody else owns, where the moat is a prompt, a workflow or a data partnership.
  • Anything whose advantage disappears the moment the underlying model improves.
  • Companies where the science is settled and the difficulty left is execution. That is a real business; it is not mine to invest in.
  • Companies priced on a demonstration, not a result.
  • Anything I cannot have assessed properly by someone the field would believe.

Four reasons to get in touch

You are building in one of the five

Send me the claim and whatever you have run to test it. A paper is fine, no slides required. If there was one result that made you start the company, send me that one.

Write to me about a company

You work on one of these problems and are not building anything

If you think I have any of this wrong, tell me. I am particularly interested in results I am relying on that do less work than their authors claim. I reply to all of these.

Write to me about the science

You are an institution

Nothing is being offered or marketed on this page, and nothing here is an invitation or inducement to engage in investment activity.

Write to me privately

Writing, speaking and press

I write and speak about why capital misprices science, and about what technological sovereignty requires beyond chips and data centres. Deadlines welcome. Engagements need to go through Thomson Reuters for final approval.

Write to me about a commission

Biography

Alexander Kardos-Nyheim

I read Law at Trinity College, Cambridge, and practised briefly before founding Safe Sign Technologies, an AI research company. We spent three years building models for a domain where being confidently wrong is not an acceptable failure mode, and most of that time raising from investors who could not read what we were doing. The money that did arrive was largely North American.

Thomson Reuters acquired the company in 2024, before it had revenue. I co-lead AI research there now, working on making large language models reliable enough for high-stakes use, and helped establish the Thomson Reuters and Imperial College Frontier AI Research Lab. In August the company released Thomson, its own frontier model, reaching that level by continual learning on open weights instead of pre-training from scratch. The economics matter: around $40m for the programme as a whole, a final training run under $450,000 on no more than 368 GPUs, fewer than three dozen engineers and scientists, and three months from first experiments to a model averaging 78.5 per cent across professional benchmarks against 79.5 per cent for the strongest proprietary system tested. It measures itself on whether every claim it makes is carried by the source it cites. Continual learning and verification are the first and second bottlenecks above.

I am twenty-six. The experience that shaped how I invest is the ordinary one of running a laboratory doing unglamorous work that turned out to be right, in front of people whose job was to judge it and who had no way to. Outside all of this I edit a journal and run a planning-reform organisation, both below.

Elsewhere

The Cambridge Journal of Law, Politics, and Art

I founded the journal at Cambridge and remain its Editor-in-Chief. It publishes scholars and practitioners together, in print and online, across law, politics and art; it is affiliated with Trinity College and is read in over 140 countries. Four issues so far; the third, The Human Agenda, on how human suffering has shaped law, politics and art. Partners have included Mishcon de Reya, Christie’s Education and Phillips.

cjlpa.org

Urban Development Reform Organisation

I founded UDRO and am its Chief Executive. It is a team of a few dozen legal, planning and architectural experts who investigate and intervene in London planning decisions where communities, heritage and cultural sites stand to lose, and who take their findings to the people who decide. Recent work includes the intervention against the proposed Chinese Embassy at Royal Mint Court, the Smithfield and Billingsgate markets, the Prince Charles Cinema and Brick Lane.

udro.org

Writing and coverage

Written by me

Talks

Conversations

Thomson Reuters

Coverage

A fuller list of essays, interviews and appearances is available on request.