Atman

Interested in Tech, Fitness, Investing, Macroeconomics, Aviation, Music. Generally curious. Views my own.

[Atman @HeyAtman]

@gordic_aleksa The (2) was a trick we used a lot in Parameter Golf, however they were really tiny models

[Atman @HeyAtman]

@woodchipdaddy @difficultyang Exactly. The ability to go from spec to impl is the key.

[Atman @HeyAtman]

@shantanugoel @VazeKshitij @shreyadoesstuff Highly reccomend Nand2Tetris and CS:APP. Must read.

[Atman @HeyAtman]

@aaryan_kakad yes this is huge but it's better to treat benchmarks as a point on a multi-dimensional axis then to treat them as scalar numbers

[Atman @HeyAtman]

@serenaa_ge thank you so much! amazing work with the bench!!

[Atman @HeyAtman]

@jeremyphoward @Zai_org I'm really awed, shocked and surprised (what a whirlwind)! It did really well in some of my testing. Profound implications.

[Atman @HeyAtman]

Who knew parsing PDF was harder than building AGI

[Atman @HeyAtman]

@NirantK Personally speaking I don’t think this is the case in the current Opus 4.8 model.

[Atman @HeyAtman]

To be fair, Opus, GPT, etc. are extremely capable models.

[Atman @HeyAtman]

Experiencing a model with a step-change in intelligence is such a powerful feeling. It's been a couple of days since we lost access to Fable, but it's hard to not judge any other models (however good) to it and feel like the bar shifted somewhere you can't unsee.

[Mike Sowden @Mikeachim]

Zero.

[gabriel @gabriel1]

first profitable ai company reinvests everything into trying to save millions of lives and to make it as generally available as possible

[Atman @HeyAtman]

@JakeABoggs would love to see scores on your bench

[Atman @HeyAtman]

People in the future will gasp at the fact that we used to manually press 'Enter' on coding agents to allow them to do stuff.

[Atman @HeyAtman]

From 10 claude-3 opus requests per day to 10s and 100s of opus requests everyday, we've come a long long way with AI use.

[Damon Chen @damengchen]

Michael email me in early 2024, and I joint Cursor

as a paying customer 🎉

[Atman @HeyAtman]

good folks at @datacurve, may we please have updated DeepSWE benchmark results? amazing work with your bench!

[Atman @HeyAtman]

@rishabhm How are you finding the Opus models for C++? Last you mentioned they were far short of your standards, has it improved since?

[Sidu Ponnappa @ponnappa]

people keep talking about Sovereign AI as an India Inc challenge, but it's actually an India Unc challenge.

#SorryNotSorry

[Atman @HeyAtman]

@vijaythirumalai capital*

[Atman @HeyAtman]

Oh god. Banning frontier models is not the panacea that you think it is. It’s a really really bad policy. Neither is the ban enforceable and nor does it appreciate the competitive aspects of today’s world.

You can’t take a risk-capita, infra and talent problem and use archaic banning to think it’ll fix it.