hci @StanfordNLP, writing https://xitter.cf/t.co/M7hdeAExFk previously: interp @TransluceAI, hotel manager, 2x IMO gold
sf
Joined November 2020
- Tweets343
- Following499
- Followers1.8K
- Likes5.2K
vincent! retweeted
It feels worth saying this at a top level too: to the extent that the positive update causing everyone to freak out is about MLP neurons being useful, a paper by @TransluceAI came out a year ago which covers similar ideas.
This is the Transluce project I've been most excited for, because of
1) piloting a setup where multiple labs collaborate on transparent 3rd-party audits of models
2) the general approach of large-scale user simulation + validation on human data
Please check it out!
ok so it's time to give up my other activities and become a supply chain blogger 🤦
Did you know the typical apple in a grocery store is up to 12 months old? That farmed fish make up half of all seafood? This deep dive by @vvhuang_ maps the supply chains of four key industries that make our modern world: fossil fuels, metals, minerals, and agriculture mindslice.substack.com/p/eve…
very cool results from @alephneuro 🔊🧠
this is the most inevitable team i've ever met. they find creative and efficient solutions to every technical problem they encounter, whether it's hardware or physics or ML. consider reaching out if you're interested in brains and imaging!
vincent! retweeted
Congrats to BART rider and Oakland legend Alysa Liu on winning a gold medal at the Olympics and making the Bay Area proud!
vincent! retweeted
New blog post on introspection for interpretability, and why I think training models to self-explain is a promising frontier for interpretability research:
vincent! retweeted
Continual learning from natural language is data-hungry. Can we make it sample-efficient? SIEVE distills natural language context (instructions, feedback, rules, etc.) into model weights using as few as 3 examples only of queries—outperforming prior methods and even in-context learning baselines. (1/n)
there’s a whole branch of psychology (Self-Determination Theory) which thoroughly breaks down "Agency" + where it comes from + why people fail to exhibit it
(and notably, SDT has survived the replication crisis!)
here’s a good overview which summarizes what SDT says about: intrinsic vs extrinsic motivation, locus of control, social connectedness
selfdeterminationtheory.org/…
love when i’m in a 3-person group chat with 2 friends who are dating and then they start hashing out their relationship in front of me using a fictional character i wrote
the Let It Go sequence from Frozen is incredibly profound and you probably never realized if you haven't gone back to rewatch it as an adult
We trained a decoder to read the internal activations of an LLM and answer questions about what the model will think about or do next.
We find that this decoder can understand LLM behaviors, even when the model itself is confused! (for instance, if the model has been jailbroken)
Very grateful to teammates @damichoi95 @_ddjohnson @cogconfluence @JacobSteinhardt for all their help, and to @aryaman2020 @ZhengxuanZenWu for many relevant office chats!
Check out our paper at arxiv.org/abs/2512.15712 for more details on the training setup and other experiments.
