MTS at Fleet, Early Eng @dydx, built GoldLink & Quickscope Beat Promised Consort Radahn w/o summons, Heolstor w/o friends & Lost Lace
San Francisco
Joined June 2021
- Tweets513
- Following175
- Followers610
- Likes730
Sam Leisure retweeted
on the journey of side quests: Fable and I found the first non-trivial packing of 308 squares!
ALT 308 unit squares packed in a square of side 17.9993, which is slightly smaller than the trivial 18×18 grid packing. Two diagonal bands of tilted squares run from the top-left and bottom-right edges, each in multicolour, with small gaps around them.
Sam Leisure retweeted
We have seen a growing and concerning trend of startups supplying data to China. Today we are taking a public stance against this practice
At fleet, we believe this represents one of the largest national security threats today and must be stopped. We must win the AI race for America and shape AI with the same democratic values that have shaped our nation
Our full statement here: fleetai.com/supporting-ameri…
The AI arms race isn’t just being fought over Nvidia chips. Chinese labs are matching OpenAI and Anthropic by purchasing the exact same data from US vendors like @mercor and Surge AI (who also work with the US federal govt) forbes.com/sites/annatong/20…
Sam Leisure retweeted
it turns out the interesting part of synthetic data might not be the data
increasingly in post training, simulation will matter more than distillation.
the quality of the training data now depends on how well your simulated world matches the real one.
distillation gives the model answers to copy from the teacher. simulation models a world for it to mess around in (like a fake user asking for changes, an unexpected event that triggers a chain of effects) and it learns from what happens like a real human would do
the former will become largely an inference problem at scale. the latter is a product + modeling problem that still feels very much unsolved.
i'm excited that we're rediscovering a very old problem in simulation modeling that math and operations research have been working on forever.
We're at a fork in the road where we can either work on building meaningful oversight or ignore the warnings of the smartest people on Earth
The model releases are inc. console-like. The PS5 was only a lil better than PS4 unless you look at the SSD move where you no longer have meaningful wait times.
Think this only accelerates with each new model release being a sub-linear improvement on all but niche categories.
RL shops doubling down on human data is building toward better models while simultaneously betting on the models never getting any better
Fascinating to me what the models still suck at
Writing high quality code, ripping designs (with guidance), understanding interpersonal dynamics and prioritization from slack or email: phenomenal
Writing a single coherent linear ticket: impossible
Most underrated feature of our new AI world is a hyper-scrutiny of consumer exp. & general skepticism that effort went into anything.
Slop has always been around (look at this website lol) but for the first time ever, you could spend longer consuming slop than it took to make
I thought transitioning to AI would feel like a completely different world but everyone I've met either worked in crypto before or has a deep appreciation for flashbots 😭
Sam Leisure retweeted
We have raised a $75M Series B from @IndexVentures and @usv, with continued support from @benchmark!
Memecoin terminals largely aren’t interesting as trading products anymore.
They’re interesting as training environments.
The data they captured isn’t just transactions. it’s how people behave in adversarial markets.
That’s a different asset:
