@GovBobFerguson Shame on you for what you are doing to this state and the pain you are bringing on those of us who will remain.
English
Will Dabney
165 posts
Will Dabney
@wwdabney
Research scientist at DeepMind. On the critical path to AGI. Also, a persistent optimist.
Seattle, WA Katılım Mayıs 2015
77 Takip Edilen1.5K Takipçiler
It's time to change our state's upside-down tax system with a Millionaires' Tax, one that puts a significant percentage of the revenue generated back into the pockets of Washingtonians to make life more affordable.
nytimes.com/2026/02/23/us/…
English
@weheartseattle Buying a book you disagree with does far more to encourage the author and publisher than burning it does to discourage them. This is not a well thought out idea, not unlike and for similar reasons as the policy you dislike.
English
Call to Action: Gregg Colburn is an associate professor of real estate in the University of Washington’s College of Built Environments. In his research, Gregg studies housing policy, housing markets, housing affordability, and homelessness. Gregg is the author of Homelessness is a Housing Problem that Washington elected leaders have turned to as their solutions manual. This book has misguided policy makers nation wide and killed our loved ones. Buy this book and post pictures of you burning it.
English
@VijayInWA @GovBobFerguson Started from the bottom now we’re here :(
English
Under @GovBobFerguson's income tax proposal Seattle residents would now have the top income tax rate in the whole country, eclipsing the eye-watering tax rates of New York city and California.
The incentive for entrepreneurs to move to our state will be completely destroyed.
English
Will Dabney retweetledi
Thrilled to share our new #NeurIPS2025 paper done at @GoogleDeepMind, Plasticity as the Mirror of Empowerment
We prove every agent faces a trade-off between its capacity to adapt (plasticity) and its capacity to steer (empowerment)
Paper: david-abel.github.io/plasticity.pdf
🧵🧵🧵👇
English
@RockChartrand @iBringBalance It’s like if a really great carpenter cargo-culted an airplane and it turned out to be a pretty amazing glider.
So we launch these things with more and more force, and they glide and some people say they are airplanes. They aren’t but we can still learn a lot about aerodynamics.
English
@iBringBalance So how does it work? Do you just install an app and then magic happens?
English
These kinds of takes can only come from people who don't understand how the technology actually works 👇
Rock Chartrand🤑@RockChartrand
AI doesn’t have a brain, and it never will. It doesn’t think, understand, or judge, it calculates. You can stuff every neural network into a metal shell and call it a “robot,” but all you’ve done is build a very expensive parrot with power tools. A plumber has to improvise, sense pressure, recognize smells, make moral calls about safety, and adapt to a thousand unpredictable details. AI can’t even comprehend what a leak is, let alone why it matters. It has no consciousness, no goals, no curiosity. It executes instructions, and calling that a brain is like calling a calculator a mathematician.
English
@NandoDF Both? If I have a way of evaluating independently of this data, then why not try both. If not, then I want to hold out some data anyway and again, why not both. Any argument pales in the light of reality.
English
If you have 10K data instances, would you:
1. SFT an LLM with 10K data, or
2. Learn a reward with 5K, and RL the LLM on the remaining 5K with the learned reward
3. Other (explain)?
English
@tw_killian If I say it publicly it will invariably end up in the training data, then I can’t use it to evaluate frequency bias.
I multiplex my LLM usage pretty heavily and am happy to add this in the mix! I sort of want to find a way to do that automatically, ensemble of LLMs approach.
English
@wwdabney I’d love to see what that prompt is!!
I’m thrilled that you’re enjoying the model!
English
#K2Think (🏔️💭) is now live.
We're proud of this model that punches well above its weights, developed primarily for mathematical reasoning but has shown itself to be quite versatile.
As a fully deployed reasoning system at k2think.ai you can test it for yourself!
MBZUAI@mbzuai
Introducing K2 Think - a breakthrough in advanced AI reasoning. Developed by MBZUAI’s Institute of Foundation Models and @G42ai, K2 Think delivers frontier reasoning performance at a fraction of the size of today’s largest systems. Smaller. Smarter. Open to the world. Available now: K2Think.Ai/K2Think #K2Think #AI #OpenSource #MBZUAI #G42 #Innovation
English
@tdietterich There are scenarios where this would be acceptable (esp for “data”). A bigger question imo is: what standard should arxiv require of submissions? Then we ask: Is formal writing included in that standard?
Give us the abstract principle, then let’s ask about the specific instance.
English
We suspect this is an LLM writing style. Traditional scientific papers are written in complete sentences. Is this an emerging change in acceptable writing style? Or should writing like this be rejected from arXiv as not conforming to scientific standards?
English
Question for ML researchers about writing style? At @arxiv, we are seeing many submissions written with either bullets or bold-faced paragraph lead in. Often as incomplete sentences. For example:
English
Just signed a book deal for The RLHF Book, excited to make improvements to it this fall and get physical copies in your hands soon :)
(rlhfbook dot com)
English
Back at it! I'm happy to share that I joined DeepMind in London full-time, and I will be working on automated tools for neuroscientific discovery.
Disclaimer: I am not a go player, I just love this painting.
English
@tejasdkulkarni @Copilot It is pretty amazing what you all have accomplished. Thinking back to early musings that you talked about on our walks after work. The vision for the future is even more impressive.
Any thoughts on meshes vs splatting vs some sort of hybrid?
English
Amazing setting for @RL_Conference in Edmonton. The weather feels perfect and the bus pass included in registration 🤌
English
@tw_killian @llm360 @allen_ai I don’t know how to decide objectively which country is more free or supportive of open scientific inquiry. Also, a coordinated effort might tend towards the union of the members’ biases instead of their intersection.
I’d prefer having a variety of open models w/ diff biases.
English
Nathan is hitting many of the right notes. I don’t fully buy the need for an “American” open model, but a wholly free + open scientific infrastructure/industry is needed. Thankfully we’ve a handful of well aligned orgs in place such as @llm360 + @allen_ai already.
Nathan Lambert@natolambert
America needs to take open models more seriously. This summer the early lead in open model adoption of the US via Llama has been overtaken by Chinese models. With The American Truly Open Models (ATOM) Project we're looking to build support and express the urgency of this issue.
English
Will Dabney retweetledi
Official results are in - Gemini achieved gold-medal level in the International Mathematical Olympiad! 🏆 An advanced version was able to solve 5 out of 6 problems. Incredible progress - huge congrats to @lmthang and the team! deepmind.google/discover/blog/…
English
@marcgbellemare Depends on your definition of question answering? We can ask questions (sourcing from theory of computation or philosophy depending on preference) which cannot be answered.
Probably, even for the type you mean, there are questions that can only be answered probabilistically.
English
A can-of-worms type question to all of you, from an RL researcher turned NLP startup founder: assuming a benchmark with no annotator error, is 100% accuracy on question answering possible?
English
@johanobandoc Repeatedly smashing “submit” for the one on screen but will definitely clear cache once this one goes through.
English
@wwdabney It’s working on my end! Maybe try clearing your browser cache. 😊
English
OpenReview seems to be essentially down? I guess I’m not the only one submitting last minute NeurIPS reviews.
English