matt west

Was a molecular biologist. Now I trade commodities and build AI/ML infrastructure.

[matt west @mattyxcakes]

@chamath The irony of conflating repos with impact, while ROCm is the direct, open-source challenger to CUDA is mighty rich.

[matt west @mattyxcakes]

@ryans_investing MacBook Air is my fav product. Small enough to carry like a tablet but full keyboard is just so much faster than screen typing

[matt west @mattyxcakes]

@GovNuclear Ok, and you can type SCALE(0.001) to make it fit into your palm but it still won't be real. Let's see the real thing

[matt west @mattyxcakes]

@aimalysheva that's a pringle

[matt west @mattyxcakes]

@nicochristie Is Anthropic really the overwhelming favorite for first to AGI, though?

[matt west @mattyxcakes]

@craigweiss possible, but you don't know for sure

[matt west @mattyxcakes]

@AIandTechh87 Who is the developer?

[matt west @mattyxcakes]

@dedene @mfranz_on Followers are cheap. Views are more expensive. Retention is very expensive. IG and TT audience size batching is largely based on your followers’ engagement in the first hour. A slew of likes along with video completion are the strongest signals

[matt west @mattyxcakes]

@jeffreyweichsel I would hazard a guess that they do, but mostly within alignment and mech interp studies

[matt west @mattyxcakes]

@pentagoniac Can you elaborate on the industrial use-cases?

I am a fan of Gemma for its mech interp community, but have an embarrassingly small amount of application experience.

[matt west @mattyxcakes]

@danrobinson These already exist, but they aren’t LLMs. Federate learning networks distribute the best performing weights across the botnet. There are scout variants, infil variants, and persistence variants, some of which can hook themselves back into a fresh install from mobo

[matt west @mattyxcakes]

Red herring. Exercise recommendation is clearly a public health benefit as it pertains to depression. The malpractice framing is actually strong imo. If it is widely effective across the general population, and doesn’t risk harm (unless you are implying that exercise itself is contributing to elite athlete depression), then it would be standard practice for providers to recommend it, bordering on negligence if they chronically or systematically avoided recommending it.

[matt west @mattyxcakes]

@snellchapo @__tinygrad__ @Kimi_Moonshot The weights aren’t public yet shiggy

[matt west @mattyxcakes]

@memecrashes Mathematicians would be wise to start studying the mech-interp behind these discoveries if they want to remain relevant

[matt west @mattyxcakes]

Everyone is so worried about “best model” and open vs closed weights.

I don’t think model benchmarks will be a relevant measuring stick in the future. We’re headed toward continuous learning and new hardware architectures.

Don’t anchor on the status quo.

[matt west @mattyxcakes]

That game theory doesn’t game. If you zoom out, “best model” is currently a function of who has compute resources // US and China. That will change. Making a model isn’t actually that hard if you have the human capital. Continuous learning is on the horizon. The optimal strategy is for American companies to release open weights, level the field, and win on CL, inference, and operating efficiency

[matt west @mattyxcakes]

@AlexisChicoine1 @iamashtonchew General use chatbot LLMs are one thing. Background agents and specialized models outside of general LLMs are another thing, and their share of development and compute resources is increasing every day

[matt west @mattyxcakes]

@MikeMcCormick_ I develop greybox and mech interp rigs to evaluate the behavior of LLMs in hidden information games and high-stakes negotiations. Already have a large corpus of neatly formatted jsonl and parquet data and I know how to drive a cluster of H200s.

[matt west @mattyxcakes]

@AndrewYNg Sweet. Will definitely be checking this out. TBH I thought most people were already running some bespoke form of this.

[matt west @mattyxcakes]

@anton_lozhkov @SergioPaniego Impactful