New blog post about a side project I've been tipping away at. I've build an app to explore open-ended idea discovery using LLMs in a multi-agent evolutionary algorithm. I've also open source the code so you can play with yourself

Might be a lack of imagination on my part but my default expectation is for the $6.5bn acquisition on Jony Ive's io to result in OpenAI wearable pendant like the so many already out there (e.g. Limitless below) Hopefully the was something more exciting in the deal.

It's an exciting time for scientists everywhere 🧪

Work like AlphaEvolve is what for me has always been the clear distinction between Google DeepMind and OpenAI. 🧵

Delighted to see our Wave device receive breakthrough designation from the FDA #neosky #medsky www.irishtimes.com/business/202...

Our seizure detection paper is finally out! Published in npj Digital Medicine. Lots of work went into this one co-led with @johnotoole.bsky.social. Blog: www.cergenx.com/blog/scaling... Paper: www.nature.com/articles/s41... #Neosky

Want a job in robotics in New York? faunarobotics.com

Interesting paper showing how LLMs change their representational geometry in-context to match a task structure: arxiv.org/abs/2501.00070 Here, the model’s latent representations show a grid structure matching the task. #MLSKy #NeuroAI

I think we are only scratching the surface of what current LLMs can do. We're fumbling in the dark with prompts. We might just be starting to find the light. New blog post open.substack.com/pub/selfsupe...

So glad that Quanta Magazine exists. Absolutely love their top breakthroughs of the year series and it's really one of the few science magazines where you can learn things deeper that very surface level. They actually try to get to meat of it. youtu.be/UhG56kltfP4?...

"Ants solving a geometric problem and it's amazing." This is a great example of 'emergent phenomenon': - None of the ants understand the problem they're solving. - None of them can see the whole shape. - A series of small decisions or rules add up to something with a new layer of complexity.

Might be of interest to folks in 🧪 and ⚛️. I think there are some strong parallelism between what's happened in the last few years of AI with what happened in last few centuries of physics.

First post on substack on something I've been thinking about for a while. Check it out here open.substack.com/pub/selfsupe...

Goalposts moving at lightning speed. Is _anyone_ surprised?

Interesting that there's a bit of diversity developing in these model offerings. We had o1 the test time compute stuff, and Gemini are going all in on the "native" integration of various modalities in and out, with a focus on agents. blog.google/products/gem...

@quantamagazine.bsky.social really is always such a joy to read. Excellent write up of the surface code and Google's result with their new quantum computer chip Willow 🧪⚛️ www.quantamagazine.org/quantum-comp...

Just adding to 🧪 🙂

First baby steps in #AI4Science for Neonatal EEG! We built the first neonatal seizure detection AI at expert level. It reveals new insights into seizure progression in babies & differences between localized vs generalized seizures. #AES2024 #Neonatal #EEG #AI

If an #H5N1 pandemic starts tomorrow or in three months, there will be little mystery as to how it happened. The conditions are all there. They have been for a while. So in some ways the more interesting question to me at the moment is: Why aren’t we in a pandemic yet? Story here, 🧵 to come: 🧪#IDSky

Individual grants of up to $1m from AI for Math Fund by Renaissance Philanthropy. Very cool 🧪 renaissancephilanthropy.org/initiatives/...

Text-to-video game getting pretty good 😲 goo.gle/4g3Oz5d

🧪⚛️

Finally catching up on this Test Time Training paper. Thoughts - very nice results - interesting that TTT helps smaller models more (sign that scale is disproportionately useful for more memorization?) - What a pain, I really hope we don't have to do this 😅 arxiv.org/abs/2411.07279

New study: AI _increases_ burnout in Radiologists Highlighted difference between screening and clinical settings. AI can actually increase workload in clinical settings when an abnormality is found due to time spent in differential diagnosis. jamanetwork.com/journals/jam...

Pretty alarming discovery. They found you can totally destroy a model by pruning a single weight. Also have some nice results that use this to improve quantization. My main concern though is what this tells us about how brittle these models can be. arxiv.org/html/2411.07...

Health AI Developer Foundations (HAI-DEF) is a new suite of open weight models to help developers more easily build AI models for healthcare applications. The initial launch is focused on imaging applications in radiology, dermatology and pathology. research.google/blog/helping...

The hardest thing about building Artificial Intelligence is that people keep redefining intelligence as the thing that AI models can't do yet

Interesting work from Dr Thomas Helfer on using AI to improve fidelity of Black Hole simulations. We overlapped in TPPC group in King's College London during our PhDs. An huge amount PhDs from there are now working with AI for various applications in science and industry. arxiv.org/abs/2411.02453

Catching up on this. So glad there's still people at the top of AI like Demis Hassabis who are making sure its application to science is given the air time it deserves and even pushes the latest LLMSys leaderboard out of the news for 24 hrs. www.youtube.com/watch?v=nQKm...