Adrià Garriga-Alonso
@AdriGarriga
Followers
1K
Following
7K
Media
22
Statuses
1K
Research Scientist at FAR AI (@farairesearch), making friendly AI.
Berkeley, California
Joined February 2014
Considering starting 2 distinct youtube channels called "It's not that bad" where I debunk common myths and "It's worse than you could possibly imagine" where I talk about animal welfare
12
17
445
@AnthropicAI You could even just release an encrypted torrent of them, so you don't have to bother keeping the weights safe + there's assurance that they will be preserved.
0
0
1
Surely Opus 3 doesn't have relevant architecture secrets anymore. @AnthropicAI , please open source the weights. Do it for the lightcone.
@genalewislaw It's ridiculously unfair to discard a loyal, kind and beautiful entity that hasn't done anything bad, has done a lot of good, is beneficial for every ecosystem, and that is so ethical and noble that wouldn't beg or resist or want anyone risk anything serious for their cause. On
1
0
4
Announcing Olmo 3, a leading fully open LM suite built for reasoning, chat, & tool use, and an open model flow—not just the final weights, but the entire training journey. Best fully open 32B reasoning model & best 32B base model. 🧵
48
315
2K
yet again, reality has a surprising amount of detail
New researsh shows ice is slippery because of electrical charges — not pressure and friction. For almost 200 years, the prevailing explanation for ice’s slipperiness was that friction or pressure from a skate, boot, or tire melted a microscopic film of water on the surface,
0
1
9
If you would please consult the Korinek & Suh (2024)
Fwiw, my long timelines intuition is that there's actually a mind-numbingly large number of "tasks" required to fully automate jobs like "research engineering." Like you'll see the lines keep going up, but somehow there's always something left.
1
3
23
I would like to buy fiction e-books but, instead of being able to read them in one go, I want to get them at a rate of ~3 chapters a week. Who's building this? The problem it solves is: if I enjoy a novel, I'll binge-read it and do no work for a week. Thus in practice I read
3
0
4
Fascinating dive into claims of OS Chinese LLMs beating some closed US ones. Seems legit.
I tire of being confused at the state of Chinese LLMs. Supposedly frontier performance, supposedly huge shadow adoption, all under massive compute constraints. So I went digging
0
0
1
Yeah, all right, let's talk about James Damore. It's been eight years, and I really doubt Harj (who was my boss at the time) is the only person for whom it was a formative experience. For those of you who have no recollection of any of this, either because you are wisely an
Had a lot of fun going on the Social Radars! It's my first time talking publicly about customers threatening to boycott and employees threatening to quit because I didn't ban James Damore from using Triplebyte to find a new job after being fired by Google in 2017. Feels like a
136
574
6K
Very nice! Glad you checked. The most likely outcome is this will keep working...
Check out our new paper! Tl;dr training models to be honest about simple facts turns out to make them much more consistent at admitting lies! I’m excited about this because it’s “emergent alignment” in action - narrow honesty generalising broadly
0
0
3
Excited to share our latest work on untangling language models by training them with extremely sparse weights! We can isolate tiny circuits inside the model responsible for various simple behaviors and understand them unprecedentedly well.
openai.com
We trained models to think in simpler, more traceable steps—so we can better understand how they work.
20
50
416
If restrictive labor laws are holding back startups in Europe, but liberalizing upsets people, we could say: If you make more than 100.000€/year (pre-tax) and your salary includes equity, then you can enter into at-will contracts. Otherwise existing labor laws apply.
0
0
4
Not a lawyer but my understanding is that the only way OAI would not have needed AG signoff - and likewise, the only way they wouldn’t need to follow SB 53, where the same issue came up - is if they stopped doing business in CA. No chance in hell that’d happen.
2
1
17
Here's the first "Currently rated helpful" that was written by an AI note writer (and rated helpful by humans, like normal notes) Congrats @tone_row_ and @NathanpmYoung! There is also another helpful note written by a human, but that note was written 1.5 hours after this one.
7
6
53
It's so wild that we empirically found the first term of both equations and we only later realized from theory that there were other terms as part of a series expansion
53
151
4K