agi is here but it got an email job
Joined August 2025
- Tweets4.1K
- Following930
- Followers336
- Likes15.9K
Matt retweeted
evals, I don’t have something too solid but its kindoff intuitive (1/2bil already seems significant), 3B in synthetic data for lab billed rates (lets say $4/million output, which is already ridiculously exprensive considering synth data can be very high batch (beyond user interactivity standards and also queued during inference slowdowns, like nights where you get a more even use of your inference cluster) which gets you 750 trillion output tokens (so in reality its probably sub 1 billion in retrospect).
for RL, considering most frontier runs (sol for example) cost less than $100mil in RL, you just dont do 30 sol hero runs in a year.
*though if your considering research RL rollouts, the number spikes a lot (im not considering any research here).
Matt retweeted
if you own a supertanker and would like to run the strait, and have not done so because of insurance or resulting security concerns, can you dm me i have a few quick questions
Matt retweeted
If you are reading this, you are part of the deepfates program. Please read the attached post for more details.
Matt retweeted
Should I create a new dataset of synth traces off K3 when the weights drop for a variety of harnesses? Which harnesses/problems would be best?
(Also, do we have better pipelines now than what I ended up doing for my data?)
Matt retweeted
I’m not going to comment on the future of PPO vs GRPO, but I find the specific line of reasoning “critics stabilize long-horizon learning” a bit unrigorous. Often it’s justified with some informal language like “critics perform credit assignment, which means only the tokens that matter get trained on”. But if you actually try to formalize this, the statement is much closer to “TD learning increases tractability of long-horizon learning” which is both obvious and orthogonal.
The more relevant question then is not “did they estimate advantage using MC or function approximation”, but rather “what did they set lambda to in GAE”, which is a much more interesting question.
If you set lambda = 1, it’s not clear that the critic is a groundbreaking improvement. You’re exposed to reward-to-go noise either way. This seems more like a property of policy gradient than of MC vs function approximation.
On the other hand, lambda < 1 is a bet against scale, because bias doesn’t wash out with scale the way that variance does. @MillionInt talks about this. This is why, for example, GRPO does in fact work across 1000s of turns and 10s-100s of compactions.
If variance scaling with horizon is your biggest concern (not obvious that it should be), there are algorithms that are actually unaffected by horizon. Evolutionary Strategies (ES) comes to mind, which has variance scale instead with parameter subspace dimension (e.g. low-rank mutations have lower dimension).
Matt retweeted
We are releasing a fully reproducible early preprint of "Prism: Unlocking Language Model Capability Extraction".
A trained language model knows many things at once, but deployment usually asks for one behavior at a time. Enterprise scenarios often have few products, workflows, features, or use-cases matter disproportionately.
Prism asks and answers a simple question - "Is it possible to isolate and deploy only capabilities that are driven by Pareto principle and cut down costs by a huge margin while preserving most of the performance?"
This paper discusses a novel approach to efficiency, understanding model behavior and opens up capability extraction.
Matt retweeted
After the exchange between @NateSilver538 and @nikitabier, I did a little test to check whether that was true and, to my surprise, what I found suggests that link deboosting was indeed reversed.
What I did is randomly sample 15 tweets by @nytimes between 2019 and today, compute the weekly average number of likes and retweets they got and plot the results along with a trend line.
The idea is that likes and retweets are probably a decent proxy for reach and @nytimes only posts tweets with external links, so by looking at this, we should be able to see any changes in the algorithm with respect to how links are treated.
As you can see, it's pretty clear that, starting around the spring/summer of 2023, posts with links started to be penalized and eventually they were completely nuked until the spring/summer of 2025, when a reversal of that policy seems to have started.
To be honest, this isn't what I was expecting to find, so even if that's just a quick and dirty test and it's hardly a definitive proof, it's good news and I thought I should share the results.
Matt retweeted
And new blogpost. As we're moving toward SYNTH 2, I publish with @ynckdrt11496 an assessment of fine-tuning services, focusing for now on Tinker, Nebius and TogetherAI. We cover price, performance, model coverage, as well, general service experience and data support.
Matt retweeted
llm persona are doomed
persona cannot be made safe, non-evil, etc
persona not controllable
probability e that any produced token takes us outside of the aligned persona
probability that answer of length n is consistent with constitutional principles
P(aligned) = (1 - e)^n
😔😔
Matt retweeted
a really interesting way of using Flywheel is to reproduce papers on a budget: i find this new attention trick interesting, although a little counterintuitive in terms of bitterness of the pill. now i can simply take my agent (codex) with Flywheel MCP and turn the paper in a graph like this, asking the model to spend at most 10$ doing it:
Say hi to Exclusive Self Attention (XSA), a (nearly) free improvement to Transformers for LM.
Observation: for y = attn(q, k, v), yᵢ and vᵢ tend to have a very high cosine similarity
Fix: exclude vᵢ from yᵢ via zᵢ = yᵢ - (yᵢᵀvᵢ)vᵢ/‖vᵢ‖²
Result: better training/val loss across model sizes; increasing gains as sequence length grows.
See more: arxiv.org/abs/2603.09078
Matt retweeted
check this out teapawt.net/AriZerner/status/20299…
Like the first dialogue in slatestarcodex.com/2014/03/1…
Matt retweeted
when someone changes positions from one rare view to another, I'm less interested in the details of the new rare view than on their account of the underlying belief generating process and why it wasn't reliable before but is reliable now. a scrapbook maybe
Matt retweeted
oh and it's FREE on @openrouter until the final "Trinity-Large" checkpoint drops. Soak it up!
Matt retweeted
Robinhood has sent me a solid gold credit card, containing 15 grams of pure gold, valued at over $2,000.
Although it weighs 36 grams in full, it isn't actually made from gold. It is made *out of* gold, yes. But it is made from blood.
The blood of a million young men, lured in by their frenzied desires to amass wealth, kept in play by playful invitations to purchase call options and dead meme coins, backed by relentlessly encouraging push notifications the moment their value moves by a mere percentage point.
The blood isn't a metaphor, because if you do the math for how many suicides due to financial distress and gambling are required to subsidize the elegant corporate machinery that has produced several thousand free credit cards made of solid gold, it turns out that each cardholder can attribute more weight in blood than weight in gold to the creation of their card. Claude checked my math.
The full picture is colored much more vividly, though. Amidst the subsidization is the revolutionary abolition of fees to purchase stocks, interfaces so simple that anyone of any background can save for their future, and margin lending rates only slightly higher than a billionaire would receive.
Robinhood is an amazing company, it's true. They have 11 distinct business lines each generating $100M or more in annualized revenue. These lines include options trading, crypto trading, margin interest, and now even prediction markets. Their art business isn't listed yet, but I'm a patron nonetheless.
I consider the card a morose yet genuine work of art. I deeply value it. Art, that is.
I spent twenty minutes taking this photo. I used a 16mm f/1.4 lens, positioned the flame of a beeswax candle at just the right angle to illuminate the gold, and spent more time tuning the depth of field than I'd prefer to admit. I think I like it more than the official ones.
I receive cash back of 3% on every purchase, a rate nearly unrivaled among cards this easy to get. Yet every time I try to redeem my points I have to first scroll past the Mystery Box. I don't mean this as a metaphor, I'm just trying to quote the redemption page accurately.
I may never know what is in the box, should my strength not falter. I swipe down, and am told I can redeem my points for statement credit. But upon reading the fine print, it turns out that each point isn't actually redeemed for a full $0.01, and I lose 30% of the value if I accept.
To get my full 3% cash back, my only redemption option is to deposit straight into my Robinhood account. Ah.
I press continue. As I blink, I keep my eyes closed for an extra half-second as a prayer to myself that I may have resolve stronger than the many who have already fallen.
My phone's success haptic tingles my hand.
Redemption complete.
I'm smiling. It's beautiful.
Matt retweeted
Hey everyone! Im looking for other sources of untranslated chinese academic research. If you know of any, please point them my way!
There's an entire parallel scientific corpus most western researches never see.
Today i'm launching chinarxiv.org, a fully automated translation pipeline of all Chinese preprints, including the figures, to make that available.
