samsonfraser

CA, 23, swe/math/ai research, currently doing bsc in cs.

[samsonfraser @samsonfraser_]

@minilek @deni0107 any insight on how the result was achieved? ns took 10 000 concurrent agents. and how the internal model performs? and after interesting math problems, will the goal be novel research and experiments in physical labs driven by claude?

[samsonfraser @samsonfraser_]

@apples_jimmy @__alpoge__ @SebastienBubeck ga agreement on the first to build agi? just pinky promise to not capture the value of every company and employee in the entire world ok?

[samsonfraser @samsonfraser_]

@__alpoge__ they do not care at all. wouldn't it be best if you or ethan adopted the dark forest approach if you wanted to make frontier solves or contributions to math? as the traditional approach you can just get front ran.

[samsonfraser @samsonfraser_]

@timoreilly Tell it to make a /skill from the book “on writing well”. And then when you need you can use that skill and writing will be a lot better

[samsonfraser @samsonfraser_]

The key idea: each iterate is a weighted average of past points. If those points are spread out, the next step must be short (parallel-axis theorem). Long steps spread them out. So the algorithm can't take full-length steps forever, and tracking that trade-off exactly gives the constant 1.

[samsonfraser @samsonfraser_]

I halved the constant in the Frank-Wolfe Rate via a Variance Argument. Github has the preprint, proof and I also added a plain language summary so its easier to understand.

https://t.co/6Dog88iiCE

[samsonfraser @samsonfraser_]

New bound: LD²(N + H_{N−1}/3)/(N(N+1)). It's exact at N=1 and tends to LD²/N, half the classical constant. To guarantee 1% accuracy you need 101 steps, not 198.

[samsonfraser @samsonfraser_]

@SemiAnalysis_ telling me he's just blatantly lieing? damn.

[samsonfraser @samsonfraser_]

@ericzelikman is there more standard benchmarks for the technical capabilities of the model compared to others for things like math/coding?

[samsonfraser @samsonfraser_]

@__alpoge__ @jdlichtman yo devs is this patched yet?

[samsonfraser @samsonfraser_]

@TOEwithCurt Should have a convo and not just emotionless scripted questions

[samsonfraser @samsonfraser_]

lots of times i would just set off runs and go to sleep for python data runs.

the setup, if you want to copy it:
1. cowork with your project folder linked, so it reads and writes real files
2. claude in chrome for anything behind a website
3. computer access so it runs the code
4. one project holding every draft and log then give it a goal, not steps

[samsonfraser @samsonfraser_]

I did novel scientific discovery and research with claude cowork. what made cowork click for me: link a real folder on my pc, let it drive chrome for the data (MAST, ExoFOP), and give it computer access to run the code. then just talk to it like a coworker or hand it /goal. thats how i got 3 tess planet candidates, an rv candidate and built a ml pipeline paper for planet discovery. @bcherny @_catwu

[samsonfraser @samsonfraser_]

@Polymarket A month ago I found a planet candidate with claude opus 5.5. I've been building ml pipelines and finding more results. then pavel launched on twitter lol. didnt sleep and released everything today. hope people check it out.

[samsonfraser @samsonfraser_]

Claude Fable 5.1 and Opus 5.5 helped me discover a strong rv planet candidate: https://t.co/6Lgk3weu6p

built an ml pipeline for ruling out false positives and testing against planets with known fates, creating a benchmark for rv candidates: https://t.co/iqxUlloOXj

and 3 strong planet candidates from TESS data.https://t.co/IKodym2HTI

I also compared the strength of my findings against validated candidates like HD22496b: the first ESPRESSO standalone planet discovery and Analysis of the public HARPS/ESO spectroscopic archive, K. R. Sreenivas.

[samsonfraser @samsonfraser_]

this is my very first prompt for my planet candidate research. i did a sciency prompt first so it would search and do lots of work, then the rest of the time i just talk casually like boris. and use claude cowork, with claude chrome extension and also a folder for claude to dump things onto the pc. full pc access was best.