Tem @TheMaximu5

Lived in the steppes as a nomad. Chess enthusiast, Information Economist, Ed-Tech, History lover. Seeking to establish better and new equilibria. #IndustrialAI instagram.com/dontknowmadtem/ Japan, Aichi Joined July 2015

Tweets

282
Followers

104
Following

2K
Likes

3K

Tem @TheMaximu5

a day ago

TIL @RichardSSutton is a chess player and just like me greatly enjoys the way AlphaZero plays in styles reminiscent of Tal -> sacrificing for the sake of sacrificing for long-term advantages be it positional or even tactical but of tendency to persist for many moves later on.

0 0 0 6 0

Taelin @VictorTaelin

3 weeks ago

that feeling of dread and despair when the AI is clearly unable to grasp what you're asking, and you realize you'll need to code manually as did the old Mayans

69 90 1K 51K 85

Tom Dörr @tom_doerr

4 weeks ago

What it feels like to merge various DSPy pipelines into one program to run end-to-end optimization

3 4 23 6K 6

Download Gif

Jack Morris @jxmnop

4 weeks ago

my grandmother’s lifetime - color film - talking movies - television - cars - airplanes - plastics - penicillin - transistors - computers - the Internet - the Concorde my lifetime - gpt-1 - gpt-2 - gpt-3 - gpt-4 - gpt-5 - the iPad - ozempic

117 539 14K 288K 659

dr. nick saunders ⏾ @nksaunders

3 years ago

I used chatgpt for research for the first time today, asking it to convert python code to fortran 90 so I can run some custom mesa models. I’ve created a reverse grad student

5 7 251 21K 13

Tem @TheMaximu5

a month ago

Still cant tell if its just a skill issue with using gpt-5-high for codex or that claude is still the superior terminal agentic coder

0 0 0 39 0

Simo Ryu @cloneofsimo

a month ago

So we went from "LLM is memorizing dataset" to "LLM is not reasoning" to "LLM cannot do long / complex math proving" to "Math that LLM is doing is not REAL math. LLM can't do REAL math" Where do we go from now?

Edward Frenkel @edfrenkel

a month ago

252 235 2K 1.0M 602

154 83 1K 232K 215

Ethan Mollick @emollick

a month ago

At this point, Google Gemini has the worst training opt-out for consumers of any of the big players. Claude doesn't train on your data at all, OpenAI and Grok let you turn off training at will, but you can only turn off training for Gemini if you also disable your chat history.

83 99 1K 121K 268

Jeremy Howard @jeremyphoward

2 months ago

@jachiam0 I think the problem is that a lot of research activity and marketing has been over-focused on math, which isn't as significant in practice as a lot of folks seem to assume.

1 1 11 649 1

Harrison Kinsley @Sentdex

2 months ago

Quite the chart

65 405 3K 350K 863

Download Image

🦋/acc 🌲 @argyros_selini

2 months ago

What xAI researchers see when they run “nvidia-smi”

Elon Musk @elonmusk

2 months ago

What xAI researchers see when they run “nvidia-smi” https://t.co/PFWGF5GYHd

7K 11K 74K 15.1M 8K

7 30 706 32K 42

Download Gif

Alexander Wei @alexwei_

2 months ago

On IMO P6 (without going into too much detail about our setup), the model "knew" it didn't have a correct solution. The model knowing when it didn't know was one of the early signs of life that made us excited about the underlying research direction!

Daniel Litt @littmath

2 months ago

14 36 441 225K 102

80 164 2K 280K 304

Omar Khattab @lateinteraction

2 months ago

It's clear that the next big thing after the shift from RLHF to "RLVR"* is scaling reward models ("verifiers") for concrete capabilities, not just average human preferences. This actually kinda looks very similar to RLHF. The main difference is that the verifiers here: [A] Are…

7 23 224 24K 204

Alexander Wei @alexwei_

2 months ago

8/N Btw, we are releasing GPT-5 soon, and we’re excited for you to try it. But just to be clear: the IMO gold LLM is an experimental research model. We don’t plan to release anything with this level of math capability for several months.

43 199 2K 519K 233

Zhiqing Sun @EdwardSun0909

2 months ago

Another tip: it generates a real pptx file. So you can download the artifact, open it in microsoft powerpoint app, and apply the design you want to all of them!

Isa Fulford @isafulf

2 months ago

Another tip: it generates a real pptx file. So you can download the artifact, open it in microsoft powerpoint app, and apply the design you want to all of them!

43 34 516 54K 114

1 4 48 10K 10

Isa Fulford @isafulf

2 months ago

tip for chatgpt agent slides: first ask it to do the research only, then ask it to make the slides!

43 34 516 54K 114

CHOI @arrakis_ai

2 months ago

OpenAI’s agent joined the AtCoder Finals — 10 hours, fully autonomous. It wrote code, built strategies, and improved itself — all without human help. Among the human contestants was Codeforces #76 ‘yosupo’… The agent quietly outperformed him. The day a top-tier human fell to AI…

CHOI @arrakis_ai

2 months ago

25 156 931 246K 425

Download Image

8 30 235 32K 76

Download Image

near @nearcyan

2 months ago

unlike other AGI labs, you can't block googler's crawlers or you won't be listed on google

34 112 2K 101K 339

Download Image

Han @HanchungLee

3 months ago

xjdr @_xjdr

3 months ago

https://t.co/Grjb90yQeJ

59 25 908 97K 109

3 4 64 19K 12

Download Image

Richard McElreath 🐈‍⬛ @rlmcelreath

3 months ago

Kinds of science True & Replicable: photoelectric effect False & non-Replicable: many priming results F & Replicable: bad analyses like confounded country-level regressions T & non-Replicable: badly documented experiments, unique observations I think there's a lot of F&R