Grok3, Quantum Chips, and Google's AI Co-Scientist – Live and Learn #62

✨ Quote ✨

Questions reshape your reality. We all think what we want is answers. We don’t, actually. Answers are dead things. Questions are animating. What we want is great questions.

– Packy McCormick - (source)

🖇️ Links 🖇️

Helix Announcement by Figure. Helix is the new large language action model that Figure is putting into their humanoid robots. The model allows the robots to generalize and handle items they haven't encountered before. With this, Figure gets one step closer to robots that can be dropped into any setting and just start working on tasks given to them in natural language. With Helix, robots can now understand and execute novel tasks, autonomously. I recommend checking out their demo video for what the first iteration of this looks like.

Majorana 1 - A New Quantum Computing Chip by Microsoft. Microsoft announced their new quantum computing architecture. Their approach is working with a novel quasi-particle known as a Majorana mode. In theory, these quasi-particles have properties that make them suitable to carry out quantum computation. For now, most of their claims seem to be more marketing speak than a fundamental breakthrough in quantum computing though. What they have is a theoretical approach that is better than others in terms of scaling. But they have failed before trying to demonstrate the existence of Majorana modes. And right now they don't have a solid demonstration of even one physical Qbit, far less the millions they claim will be possible. What they have instead, is an experimental observation showing that Majorana modes are real and have the expected properties. Let's see if they build this out into something real over the next few years without running into unexpected problems along the way. For the whole topic, I found this discussion on HN quite interesting, as well as the Scott Aaronson FAQ and this Arstechnica report, all providing more details. Microsoft's CEO Satya Nadella also appeared on the Dwarkesh Patel podcast and they covered the Majorana 1 in their conversation. They also talked a lot about AI and Microsoft's strategy in RnD across different fields and its role as a hyper-scaler going into the future.

Grok3 by xAI. Elon Musks xAI released their next generation model Grok3 which seems to be better than almost everything else out there, representing another step forward for LLMs. Watch the announcement livestream where they show more of Grok3's capabilities. Somehow, breakthroughs like this stopped exciting me as much as they used to... I still think it's nice that we have another LLM performing better on benchmarks than before, but it's not as much of a WTF moment as these announcements used to be just a year back.

Muse – A Generative Gameplay AI Model by Microsoft. This announcement is Microsoft's version of Google's Genie 2 model, that I covered back in December 2024. Muse can generate low-resolution gameplay footage solely from human keyboard input. The model gets to see the first few frames of a game and the keys the user is pressing during that. Then the game turns off and the AI has to re-create the scenery while the human keeps playing the AI-generated version of the game, for up to two minutes. It is remarkable how accurate these models have become at generating interactive games and they start to exhibit an understanding of the underlying game mechanics.

AI Co-Scientist by Google. Google seems to have distilled the scientific method into a set of agentic LLMs acting together. In their own words: "The AI co-scientist system is intended to uncover new, original knowledge and to formulate demonstrably novel research hypotheses and proposals, building upon prior evidence and tailored to specific research objectives." The idea here is that Google's AI Co-Scientist can be used to accelerate research progress. Ask questions about your research and have another mind help you along by providing testable novel hypotheses and arguments. To enable this, their system uses different sub-agents for tasks like Idea Generation, Reflection, Ranking, and Idea Evolution. They built this, taking into account the self-play and recursive-self improvement ideas from Alpha Zero. This leads them to an ELO system, where the algorithm can play itself in a tournament-style setting, improving the quality of the hypotheses it generates over time as it gets more training experience, increasing its ELO. This internal ELO evaluation metric correlates with performance on the GPGA benchmark, a set of very hard scientific questions. This shows that ELO represents something useful, it is a valid training signal for reinforcement learning of these systems. The AI Co-Scientist's output quality also scales with inference compute, much like o1 or o3. The more time and compute you give the model to think and reason through an answer, the better the answer becomes. To me, this seems like a crazy breakthrough, and I am looking forward to the scientific equivalent of move 37. But it also makes me scared of the rate of progress that is yet to come as these types of systems get better and gain widespread adoption in the research community.

Using AI to decode language from the brain by Meta FAIR. Meta has built an AI that can decode what a human is typing with up to 80% accuracy from non-invasive brain scans. Right now this needs to be done in a shielded room with a giant machine and the practical implications are minor. But in the not-so-distant future, this might change drastically. During the experiments they learned something about how brains produce language though. What they found is that brains process language top-down—starting with higher-level concepts, like the meaning of a sentence, and then breaking it down step by step into phrases, words, syllables, and finally letters, before driving the motor actions necessary to type. What a wild world, where AI is literally able to read our minds...

🌌 Travel 🌌

The last two weeks I didn't travel any further with the bicycle (again). I like my days here, in the small Finca in the middle of nowhere, programming, reading, learning about shaders, and generally living a laid-back life. I created a small demo art project for the fun of it. Feel free to play around with it and create something beautiful.

🎶 Song 🎶

Boogie 99 by Luca Sestak

Youtube Music | Spotify

---

That's all for this time. I hope you found this newsletter useful, beautiful, or even both!

Have ideas for improving it? As always please let me know.

Cheers,

– Rico