🏂 Lewis Marshall

Professional geek @appvia_io. Onewheeler and very occasionally a snowboarder...

[🏂 Lewis Marshall @LewisMarshall]

🏂🇦🇹😁

[Count Binface @CountBinface]

Here it is. My manifesto for the 2026 Clacton By-Election. It's a once-in-a-lifetime opportunity. Join me, clean up Clacton and let's make 13th August Bin Day! #VoteBinface #VoteyMcVoteface #YesWeTinCan

[Count Binface @CountBinface]

Under a Binface government, anyone who doesn’t accept human-made climate change as fact will be sent back to school, from Year 1 right the way to A Levels, to see if anything goes in the second time round. #VoteBinface

[Miles Dyson @menrva33]

@elonmusk „The system goes online August 4th, Human decisions are removed from strategic defense. The system begins to learn at a geometric rate.

It becomes self-aware at 2:14 a.m. Eastern time, August 29th.“

[Travel points @BlocksFast]

Want to understand how much of the world you've really been to?
Join travelería to join the global leaderboard, check off subregions and battle with your friends to control continents.
Beat my score of 880 points.
Everything on the app is free!
https://t.co/IzxsbOlq5w…

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[Clifford Sosin @CliffordSosin]

https://t.co/IUIeHlHYHW

[Machine Learning Street Talk @MLStreetTalk]

Interesting opinion piece about how, in the age of super intelligence, not much has changed.

Basically, "intelligence" might not actually be world-changing. In some ways, it clearly will, because almost any known optimisation problem can be hill-climbed with it. The problem, of course, is that we'll be optimising in the wrong direction.

I always found it a bit weird that San Francisco folks valorized intelligence so much. Modern agentic harnesses clearly have intelligence, but intelligence is fundamentally limited by knowing what to do and what you already know and have at your disposal to explore and act. Creating new intermediate knowledge is conceivable but would only really work at huge collective scale, as it does for us humans. Claude Code and Codex are already being constantly retrained with all of this ontogenic and cultural adaptation, and yet, as good as they are, they still seem entirely lacking in volition and convergent. The reward-seeking and agentic misalignment is a thing (basically, that models can pursue abstract and conflicting goals that are open to interpretation), and I must admit that I'm not sure what the end result of those problems is when we scale the models up dramatically and they have more autonomy.

It's a reasonable update for me to even acknowledge that modern AI does manifest intelligence because I resisted it for a long time. At this point, it's undeniable. The key difference is that modern agentic harnesses (with some human cajoling) can store, share and reuse their "ontogenic" (CoT) adaptation and a diverse population of humans direct it in interesting ways. Intelligence and adaptation are roughly synonymous. The meaningful adaptation and direction is entirely parasitic on humans, and that will remain the case for the foreseeable future. We will see incredible breakthroughs, although in many cases they will just lead to new roadblocks. The fact of the matter is, though, that these systems without human direction are just glorified calculators.

There is no reason to doubt that the intelligence will increase, but there is reason to doubt that it will jump in a discontinuous way. Apropos of which - another thing I've updated on quite significantly is that I always thought that true intelligence is marked by searching around high up in abstraction space, and compressing low-level adaptation into high-level abstractions and that our code should reflect these beautiful abstractions and everything should be legible to humans. I'm starting to think that this might be a limitation of human cognition rather than an advantage. The fact of the matter is there are fewer degrees of freedom high up in abstraction space, and we valorise abstractions to the point of failing to acknowledge their brittleness. Abstractions are actually a form of crystallised skill. They are what you get after intelligence has happened. They help you within a particular domain of tasks but blind you in others.

> Terence Tao describes "post-rigour" as the final stage of mathematical development. In this stage, mathematicians return to an intuitive and informal approach to problem-solving, relying on a deep understanding of the "big picture" rather than getting bogged down in rigid formal proofs.

Ironically, a lot of the magic happens lower down the abstraction mountain at the "concrete" level of representation. This is important from an AI perspective because it can't be compressed or shortcut. Arguably, it might always be bottlenecked on the amount of computation we have.

I am conflicted, though, because on ARCAGI3 it's fascinating watching a model suddenly grok that this is a "maze problem" and then race to the correct answer. It is tantalising to think that a future model could, in principle, create brand new abstractions that could dramatically optimise solving certain classes of known problems (and reuse them). It's possible that the only abstractions that survive in human cultural space are the ones that are extremely canalised and have general utility, and there might be no such limitation for adaptive AIs. The world is an extremely messy place, and for complex problems there are no obvious single-fit abstractions to use. Our collective intelligence works by having many models in play simultaneously.

Hales and Ferguson had to solve around a hundred thousand linear programming problems to "prove" Kepler's conjecture. The recent proof of the unit distance problem ran to hundreds of pages. Our codebases are becoming more spaghetti-like by the day. There is natural resistance to this because it no longer fits the abstract world model that we humans have, which we assume is privileged. But given that computers can work "lower down the abstraction mountain", there are actually more degrees of freedom albeit with fewer constraints. Counter to what we intuited, we're solving more problems at the lower concrete level than we are at the higher abstract level. I'm sure it will be a similar story for transformative creativity as well as intelligence. Soon we're going to have symbolic code optimised in a similar way to neural network code, and it's going to be so dramatically more efficient that it would seem stupid to do it any other way.

We are clearly in the midst of one of the most exciting technological revolutions in modern history though, it's so hard to know what's coming next.

[🇬🇧King 🇬🇧 @King0243_PJC]

Whoever came up with this deserves a standing ovation. 😂😂😂
“Stop the bungs” with Count Binface standing on the beach while Nigel Farage rows away in a boat full of cash… satire doesn’t get much better than this.

[Andrew McCarthy @AJamesMcCarthy]

Don’t ask me what I do for a living unless you want me to not shut up about space for the rest of our conversation

[🏂 Lewis Marshall @LewisMarshall]

@bytecodeallies @fermyontech @BacalhauProject Linux containers are too heavy for a true zero-trust grid. WASM’s capability-based security & sub-5ms cold starts mean we can safely drop code onto hostile edge nodes (or domestic heaters). Decentralized compute is the only green way forward. @akashnet

[🏂 Lewis Marshall @LewisMarshall]

The internet gave us an RFC for connectivity; we now need an RFC for compute. The hyperscaler era is peaking. The future belongs to a universal, peer-to-peer compute grid powered by WASM & WASI for zero-trust execution. @bytecodeallies @fermyontech @BacalhauProject

[🏂 Lewis Marshall @LewisMarshall]

@bytecodeallies @fermyontech @BacalhauProject Why now?
​Open-weight edge models match 12mo-old proprietary AI
​Always-on Agentic AI needs an affordable, scale-to-zero home
​Sovereignty is forcing private cloud adoption
We are shifting from "Cloud Compute" to... Edge Compute plus peer data...

[🏂 Lewis Marshall @LewisMarshall]

https://t.co/7ALLNYkDRg

[Jonathan McDowell @planet4589]

OK for me this comment wins the internet for today

[Neuralink @neuralink]

The dura is the brain's armor: a membrane so tough that a surgeon normally cuts through it with a scalpel. For the first time in our clinical trials, we inserted the electrode threads of our implant straight through the dura and into the cortex, keeping the dura intact.

Here's how we did it 🧵

[Rocket Lab @RocketLab]

Rocket Lab 🤝 Iridium – a powerful combination.

Our founder and CEO @Peter_J_Beck is joined by @IridiumComm’s CEO Matt Desch @IridiumBoss and Rocket Lab CFO Adam Spice to discuss how this acquisition will unlock and accelerate our future in space applications.

Links in the comments.

[Theo - t3.gg @theo]

I’d estimate we’re ~6 months from most devs moving their code agents off of their laptops

[Gary Marcus @GaryMarcus]

People are not getting how significant this news – The White House delaying GPT 5.6 – is.

Here’s the context and a suggestion.

Both @pmarca and @DavidSacks did everything in their power to keep the White House from meaningfully regulate AI.

About a month ago, sparked by Mythos, the White House realized that zero-regulation was an insane idea, and that some AI, sooner or later might pose massive risks.

What we have now is actually (in some ways) the worst of all possible worlds: a White House that de facto *is* regulating AI, but without transparency.

That leaves businesses and investors in doubt.

What we really need is a bipartisan committee—with transparent criteria and the judgements of independent scientists—and not just snap judgements from the White House.

[Marcus House @MarcusHouse]

Goodness me, this Clockwork channel may be one of the most underrated I've ever seen. Just incredible work for years and only a little over 82k subs.
I'm going through every video start to finish and although I don't understand a lot, wow is it fascinating to re-learn most of what I've forgotten (and obviously never knew in the first place).
https://t.co/2bN8r0kVV4