Ben Walker

Founder & CEO, Continuous Intelligence. Building models that learn from the world as it unfolds in continuous time. Email: [email protected]

[Ben Walker @benjaminwalker]

When I passed my DPhil viva, I wrote:

“This thesis lays the foundation for making continuous-time machine learning practical. Now it is time to make these models a reality.”

Today, that work enters its next phase with the launch of Continuous Intelligence Limited.

[Continuous Intelligence @Continu_Intel]

The real world does not arrive as tokens.

It arrives as irregular, asynchronous and multiscale streams of measurements, events and signals.

Continuous Intelligence is building stream-native AI for real-world time-series data.

[Ben Walker @benjaminwalker]

Not sure people are ready for what we’ve been cooking at @Continu_Intel

[Ben Walker @benjaminwalker]

We have come a long way in two years

[Ben Walker @benjaminwalker]

Here is one that 4o really struggles with but o1 handles easily:

A, B, and C are racing. A starts first, B second, C last. There are 4 overtakes. C never passes B and overtakes A only once. The final order is B, C, A. What was the order of overtakes?

[Ben Walker @benjaminwalker]

Maybe it is time to make these models a reality 👀

[Ben Walker @benjaminwalker]

Delighted to share that I have passed my DPhil viva for my thesis, Advances in Neural Controlled Differential Equations.

This thesis lays the foundation for making continuous-time machine learning practical. Now it is time to make these models a reality.

Thank you to Terry Lyons for the unending guidance, support, and encouragement throughout my DPhil.

Many thanks to my examiners, Marc Deisenroth and Stephen Roberts, for the spirited discussion and invaluable advice on my work.

[Ben Walker @benjaminwalker]

Want to know more? The paper, code, and dataset are now all available publicly.

Paper:
https://t.co/JO67eY4UHY
Code:
https://t.co/PVSjb1CC1O
Dataset:
https://t.co/P8U6wJEeIT

5/5

[Ben Walker @benjaminwalker]

Our new paper Chess-World-Model has been accepted to NeurIPS 2026 Evaluations & Datasets!

The most interesting result?

Models that perform almost perfectly on real games fail badly when tested on games generated from random legal moves.

Same rules. Same legal moves. Wildly different performance.

🧵 1/5

[Ben Walker @benjaminwalker]

Even after matching the random games to the human-game length distribution, the gap remains.

This shows that near-perfect held-out performance does not necessarily mean a model has learned the underlying state-transition rules.

4/5

[Continuous Intelligence @Continu_Intel]

Six papers at NeurIPS 2026 from a team of four: that's 1.5 papers per person!

Among them is our inaugural Continuous Intelligence paper, Time Flies When You Parallelise.

Preprint coming soon. In the meantime, here is a brief intro to are all six papers!

[Ben Walker @benjaminwalker]

Really enjoyed speaking at the SPRIN-D AI and Robotics conference about what we are building at @Continu_Intel !

Thank you to @IlirAliu_ for the great video!

[Ilir Aliu @IlirAliu_]

@Continu_Intel

Most AI assumes nicely ordered tokens or regularly sampled data.

Reality isn't like that. @benjaminwalker and team are building models for irregular, asynchronous streams of events and measurements unfolding continuously through time.

[Continuous Intelligence @Continu_Intel]

Two months into Continuous Intelligence, we shared our progress building foundation models for time-series data at SPRIND’s conference in Berlin!

Exciting to see so much frontier AI innovation beyond LLMs across the UK and Europe. Thank you SPRIND for bringing us together!

[Continuous Intelligence @Continu_Intel]

Continuous Intelligence has been awarded up to €3 million through SPRIND’s Next Frontier AI Challenge.

We are one of ten teams selected from 500+ applications to enter Stage 1.

Read more
https://t.co/zExphbAUxL

We’re hiring in London
https://t.co/hV1psdJwkd

[Ben Walker @benjaminwalker]

@prz_chojecki @Continu_Intel Thank you! I saw your email and will get back to you soon. It would be great to chat.

[Ben Walker @benjaminwalker]

Earlier this week, I launched @Continu_Intel. Today, I can share what comes next.

We are one of ten teams entering Stage 1 of SPRIND’s €125m Next Frontier AI Challenge.

We are turning continuous-time ML research into models that learn from a continuously evolving world.

[Continuous Intelligence @Continu_Intel]

We are delighted to be one of ten teams entering Stage 1 of SPRIND’s €125m Next Frontier AI Challenge.

We are building AI that learns directly from irregular, asynchronous and multiscale streams of real-world data.

[Jano Costard @jncstrd]

True sovereignty isn't copied. It's built. Here are the 10 teams we fund with €125M to create the next AI paradigm in Europe! Learn more about the teams and the @SPRIND Next Frontier AI Challenge 👇
1/

[Ben Walker @benjaminwalker]

Unless I’ve missed something, there are still no technical details on how they make the approach subquadratic. Anyone know how they choose which previous tokens a query should attend to without first looking at all previous tokens, or is the subquadratic claim just marketing?

[Subquadratic @subquadratic]

The transformer architecture used for ChatGPT, Gemini, and Claude has defined the last decade of AI. It also introduced a fundamental constraint: compute scales quadratically as context grows.

Longer inputs, exponentially higher costs and accuracy that degrades well before the context window limit.

SubQ changes that.

It's the first LLM that breaks the quadratic scaling constraint delivering longer context, higher accuracy, and lower cost at the same time without tradeoffs.

Read more here.
https://t.co/QZmk07GZrQ

[Ben Walker @benjaminwalker]

Using the new GPT-Image-2 to help me express what it feels like to watch a one-off specific instruction survive Codex compaction, and then get passed down forever as legend through each successive compaction

[Ben Walker @benjaminwalker]

Looks like codex and chatgpt are now down and this is the first time I have seen a foreign language in a response, is this somehow linked?

[Ben Walker @benjaminwalker]

Codex felt it could only express how totally declarative the dataclass definition should be in Russian