Neil Traft

I found this refreshingly honest and forthright statement about AI use in a recent paper! IMO this is exactly the right way to do it. ieeexplore.ieee.org/document/115...

I’m starting a blog! My first post is on how 3rd party embedded evaluators seem totally unsuited to addressing the problems we are currently facing, and what the real problem is. stellabiderman.ai/blog/embedde...

I just had an odd thought: AI "alignment" is like the Halting Problem. Alignment at model training time is obviously not possible, in the same way that a halting decider is not possible: you can't magically know ahead of time whether or not an algorithm will "make a boo-boo".

Screw-ups should be costly. There should be a high government fine on top of it, plus paying to clean-up and re-secure the systems hacked. These are equivalent to industrial accidents and should be treated as such, imo

Astra beats NetHack: kenforthewin.github.io/blog/posts/l...

Modifying an NCA phenotype with LoRA adapters (showing there are low-dimensional modifications that can scale/skew the resulting image). Bringing together two of my favorite AI concepts, NCAs and PEFT (parameter efficient fine-tuning). 🙂

In our Nature paper, we introduce the first superhuman Stratego AI, which we built using general techniques that we developed for RL & test-time compute under imperfect information. www.nature.com/articles/s41...

I asked Claude who'd be dumb enough to invent a dumb fruit like guava and it made me this

"Despite the seemingly novel power these tech oligarchs wield in the space sector, their position today does not reflect a break from the space industry of the past but is a direct result of foundational decisions made at the origins of the US space programme." www.tandfonline.com/doi/full/10....

"According to the tenets of rational choice theory, you can answer [any question] by forecasting a dollar value and probability of every outcome. … This is ludicrous if you think about it for two seconds. And yet it’s become a standard social convention..." www.argmin.net/p/pricing-co...

Hilarious anecdote at the end of "Patterns, Predictions, and Actions": Since an income tax appeared too unpopular, King William III created a property tax based on the number of windows in one's house. This worked at first... until people started to brick up their windows. 🙈

A common issue with position papers is that they leave the reader wondering “okay, but what should I actually do”? To address this we provide open problems on a wide variety of topics throughout to illustrate our perspectives and guide future research

Happy New Year! Here's some microorganism porn. vimeo.com/1024357218 (srsly though. it's mesmerizing.)

claude code is fucking insane i know literally NOTHING about Hegel. ZERO. and it just built me a complete system of German idealism

Fascinating story: 1. Fear-mongering about peanut allergies multiplies in the late 90s. 2. AAUP advises "no peanuts til age 3", despite NO evidence. 3. This advice CAUSES a massive spike in peanut allergies! 4. They had to do a large study just to undo the damage & convince everyone it was safe!

Astonishing: the default behavior of Chrome's "Listen to this page" feature is NOT to simply read you the page, LIKE YOU WOULD EXPECT, but instead to make one of those goofy NotebookLM AI podcasts... and I had to search the Help docs to figure out how to change it. Thanks, Google.

MAORL (Multi-Agent Olof RL) arxiv.org/abs/2512.16705

In applying AI to material science, biology, etc, capitalism is trying to shed science. The point is to substitute the engineering of a machine that can generate what science has hitherto done, but without having people know things. Knowledge ultimately residing in private property is the dream.

[In considering candidates] "I actually view publication volume completely negatively... I've learned, from experience, that those with a really high volume [don't have a depth of knowledge]." —@abeirami.bsky.social #slowscience FTW

I thought this was a fascinating way to motivate model merging: to be able to *interpolate* between model classes, rather than be forced into discrete choices. (models trained on different datasets, models of different sizes, etc.) "Interpolation and alignment are two sides of the same coin."

Did you know that you could create hybrids across different neural net architectures—*without* updating the weights?? Come check out my poster at the UniReps @unireps.bsky.social workshop at 3:45pm today at #NeurIPS! (Or you can already browse the posters on the wall in 20D throughout the day.) 😃

Check out our new work on using low-rank perturbations to make evolution strategies work for billion-parameter models.

What if instead of buying software from a store, you could grow it in your garden?

What an incredible tool! I think I'll be returning to this a lot over the next month! The most interesting papers might be the ones that seem misclassified or out of place...

It's too easy to write garbage survey papers and get lots of citations. Over time, I've learned to ignore most survey papers and recognize some markers of quality.

The blog post is available: blog.arxiv.org/2025/10/31/a...

It's been 8 months since I notified @springernature.com of a clear-cut case of fake references in a book chapter which they published. They have still not taken any action as they claim to still be investigating. I will not review for Springer again until this matter is satisfactorily resolved.

UBC’s department of Computer Science invites applications for up to 2 full-time tenure-track positions. The department is particularly interested in researchers in: visualization, robotics, reinforcement learning, data management, and data mining. @cs.ubc.ca science.ubc.ca/about/careers

Translation: You have no new messages.

This year's ALICE guest speakers 🧑‍🔬 - Angel Goñi-Moreno - @angelgm.bsky.social - Alyssa Adams - @alyssa-m-adams.bsky.social - Alexander Mordvintsev - Eric Medvet - @ericmedvetts.bsky.social - Kyrre Glette - @kyrre2000.bsky.social - Stefano Nichele - @stenichele.bsky.social - Susan Stepney

"Consider another human difference: financial solvency, which can be measured and quantified, just like IQ. It is heritable in twin studies and (less so) in SNPs. But would you bet that sooner or later we are going to know 'what’s going on' with your bank account at the level of genes?" Same for IQ.

GPT now has ads. 💀 They are voluntary... for now.

Q. Who aligns the aligners? A. alignmentalignment.ai Today I’m humbled to announce an epoch-defining event: the launch of the 𝗖𝗲𝗻𝘁𝗲𝗿 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗔𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝗼𝗳 𝗔𝗜 𝗔𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝗖𝗲𝗻𝘁𝗲𝗿𝘀.

At #AutoML25, Dr. Manuela Veloso calls for 3 major directions for AI research: 1) AI must know what it doesn't know (uncertainty). 2) AI must be able to continually improve (continual learning). 3) AI systems must include humans at all times to be more robust (automation is always only partial).

No mires pabajo

Merely a day after I call out OpenAI for being so closed (bsky.app/profile/ntra...), they release an open weights GPT. Glad they're finally reading my skeets! 😏