Someday I'm going to be in this picture and I'm not going to like it.

Remarkable story -- but the human interventional evidence is based on methylation clocks.

This work is simple, clear, compelling, but embellished. I first found it shocking, the notion that a model can natively understand neural network weights and productively edit them! That's really not what's happening, & many here are annoyed. BUT, it's still notable & worth understanding. I've done a bit of analysis & auditing, here are some thoughts: 1/ This paper does *not* show an agent can edit its weights ala RSI. It does *not* show a model can edit in new capabilities. It DOES show that a frontier model can remove "bugs" (reasoning loops), using the same ablation technique researchers have used to remove safety refusals (Exp 8 - Qwen). So yes, renewed concern about frontier LLMs getting access to their own weights & editing safety/CBRNE/alignment features. 2/ This paper also does *not* show that an agent can "natively" read, deeply understand, and fix model weights. What is happening is that the model can cleverly deduce the same things that mech-inturp researchers do and fix them in mostly ham-fisted ways. Still cool, but not as groundbreaking. Evidence in thread below...

When am I gonna learn to stop betting against llm's? Not yesterday.

After four adventurous years, happy to share our work. We asked: How much of RNA processing is written in the sequence, and how much depends on where it sits in the genome? We built IntronTRIP to find out. Both matter, but not quite the way we thought.
biorxiv.org/content/10.64898…

Which gene embedding should you use for your model: one learned from single-cell data, protein sequences, or a knowledge graph? Finding out usually means hours of downloading, ID mapping, and format wrangling. We built Gene Embedding Hub to skip that step. 🧬 67 human gene embedding sets across 10+ modalities 📐Aligned to up to 20,097 protein-coding genes with Ensembl IDs 🤗 Ready to load from Hugging Face: huggingface.co/spaces/BoevaL… Congrats on this amazing work Daniel Tabares Lopez and Michael Bohl!👏 🚀 Check out the preprint: biorxiv.org/content/10.64898…

We used coding agents to automatically build baselines for 36 "AI for science" deep learning methods from Nature, Science, and Cell-family publications. These baselines recovered a median of ~95% of the score of the published method, and beat them in 33% of cases.

I’m thrilled to share our new preprint, in which we describe what I think may be the best way to selectively target aneuploid cancers while sparing normal euploid tissue.

CHIP (Computational Health Informatics Program) is awesome, and we are hiring! chip.org/employment/faculty-…

Gilded-age wealth inequality, but for cognition. Bottom to top, ish: - Unresolved drug addiction + mental illness - Gambling and tiktok addiction - Social media addiction <--- YOU ARE HERE - Wealthy bubbles protected from social media ... ... - Industrial-grade AI scammers

Claim: TWAS effects are often wrong-direction due to a combination of LD and incorrect tissue context. Could that explain some of the near-symmetry when flipping the y axis of Fig. 2C from biorxiv.org/content/10.1101/… ?

How well does TWAS estimate a gene’s direction of effect on a trait? We think of this as an important stress-test for the accuracy of TWAS. In a new paper, we find that TWAS gets the sign wrong around 20-30% of the time! nature.com/articles/s41588-0…

Getting the direction of a gene’s effect on trait right is VERY important for human genetics and drug discovery. This paper clearly shows how even a widely used method like TWAS can get it wrong.

Michael Levin kills me. Because there's an interesting observation here (that maybe you can bias cortical cultures to grow in a certain direction), confounded by the lack of an effective control (checking to see if you can bias cortical cultures to grow in any other direction or if you've just made them always go left no matter what), completely muddled by some nonsense theory (something about "training cells out of disease states"). All of his papers are like this. An interesting, usually unreproducible observation; zero interest in effective controls or interrogating whether it's a robust phenomenon; a lot of handwaving on top.

Okay, so here's the thing. If you didn't read Bad Blood or the early WSJ reporting about Holmes (which I did, obsessively, because I am obsessed with scammers) you had a completely different experience of this movie. You *cannot* have engaged with that media, and then

This is what I think many have failed to grasp. T2, Epic, Roblox’s competition is going to be OpenAI, xAI, etc. The big tech companies are already making interactive world models that will compete with game experiences. This is why Mark Zuckerberg moved his bets away from XR and into AI. That is the battlefield where cutting edge tech and the network effect will merge. Games, AI, and tech are really part of the same network effect. Metcalfe’s law states that the value of a network is proportional to the square of the number of its users. In other words, the platforms with the most creators and players generate the most value. When Wada-san predicted in 2004 that Networks will transform the game industry, he meant it from a technical perspective. But he knew, and I know because I spent a decade with him, that ultimately it was about the Network of players. And so we return to the start. Games are no longer a content business. Square Enix knows it too, but the sun set on their opportunity to transition to a Network business. There were a few million players of Final Fantasy 16. There have been 14.45bn lifetime plays of Grow a Garden so far, with 21m+ peak CONCURRENTS (meaning lifetime players substantially higher). Microsoft sees this too, which is why Everything (and thus Nothing) is an Xbox. The future is unfolding before us, and given that we’re all free to spend our time as we wish (and are apparently spending it watching TikTok and playing Grow a Garden) Metcalfe’s law is the future the majority of the audience have already decided with their wallets and time that they want for games. Apparently that's growing gardens on Roblox.

Enter AI. There are many misunderstandings here. But we should not misconstrue the Microsoft layoffs with them replacing studios with AI. What happened with Microsoft was clear: AI is a one-in-a-generation change in the entire digital order. MS must either get on board, or perish. Microsoft lost mobile. Microsoft thrived in cloud because it threw the kitchen sink at Azure. Microsoft cannot lose at AI. I don't mean making game art. I mean in that it will replace many methods of communication and impact all facets of human life. Based on Xbox profitability over the last 20 years, it is guaranteed that a dollar spent by Satya on Xbox would either be burned or gained a few cents. A dollar spent on AI will either net 100x back in the future, or all is lost. Satya made his choice, and it was the rational one.

"[T]wo estimates exceeded the theoretical maximum ... of 100% ... If vaccination reduced dementia [this much], a decline might be expected ... following ... vaccination programmes; however, no corresponding decline was observed in national data from the UK, Australia, or Canada."

If you've noticed a distinct dialect emerging in frontier AI models and struggled to understand what they're talking about, you're not crazy! We turned to psycholinguistics to investigate what makes the writing style so weird, and why this might be happening 🚨New preprint! 🧵