Post

P(aul) Frazee
@pfrazee.com

Slow down frontier model development. Bring open weights to current frontier. Build out local inference providers, stabilize hardware prices, focus on cost and token throughput for a bit. Get our software updated to handle what's being thrown at us. I would not mind spending a beat on distribution.

15 reposts291 likes

Matty B
@mattrude.org

Smaller and locally runnable models are also easier to improve on in a smaller GPU cluster setting (AFAIK), so that feels like a good goal for the dev community in general, or at least for the GPU/compute poor. Which I think is most people on earth.

{:ok, "Joel Jucá"}
@joeljuca.com

Wow! Whatever you’re doing, lemme know. If it’s open source, even better. I’d love to contribute. AI for the good of society is smt I def wanna be involved with.

Matthew James Ⓥ
@smokeyjmirror.bsky.social

like this path. most often, if not always, tension holds balance ♎ it is tough to match the capital of frontier builders. still, we're seeing brave researchers and engineers fight it off. Don't see pause as a viable route. We are capable of being thoughtful at times, so if that equates to slower 👍

A Horse of Course
@charliehorsephotos.bsky.social

Do you receive financial or growth forecasts and wonder why they think exponential QoQ growth is possible? This will be a rug-pull. Hype the stock, promise deals to one another for future hardware purchases to muddy the actual profits and hope the bubble doesn't pop until after milking govt funds.

Friendly Skeleton
@jonathan.eylerwerve.com

"slow down development" will be aimed at open models, not the big money duopoly. Open = unknown is an easy sell. The model as a service cos will welcome certification and monitoring requirements that only they can meet. Capitalism hates competition.

Friendly Skeleton
@jonathan.eylerwerve.com

The solution to Houthis making rocket guidance is an output-cop agency that screens everyone always. The NSA will volunteer. Conveniently this only works if you pay a third party for inference.

noscholar.bsky.social
@noscholar.bsky.social

This all sounds very sane, but: 1. This does not align with what any of the people in power mean when they say “pacing” 2. We’ve built up the AI finance bubble to the point where even this level of “what if we still did everything, but in a way that makes sense” would crash the global economy

noscholar.bsky.social
@noscholar.bsky.social

Agreed on the rushing, but nothing short of hard regulation is going to stop them rushing, and that’s not likely. From the “civilisations” transcripts it sounds like some of those agents found a way to mislead the transcripts. That should never happen,

noscholar.bsky.social
@noscholar.bsky.social

Long 2008? Nothing about the GFC required us to pile all investment and debt into an ouroborus of companies. If anything “long 2008” should be us being too scared to invest in promising but risky opportunities, not a desperation to repeat the same mistakes.

Honza Dvorsky
@czechboy0.dev

Do you realize how massively tokens are subsidized today? So many people got used to buying $40 bills for $1, built workflows around them, and now ask for "someone" to "bring the cost down". However, the bad news that serving LLMs is deeply, structurally unprofitable. And is only getting worse.

Tim Purdum
@timpurdum.dev

The latest open weight flash models provide similar value at costs low enough to be comparable to the subsidized frontier subscriptions, so I don't think we're going to see costs go up. If anything, this is only the beginning of the commoditization.

Honza Dvorsky
@czechboy0.dev

No, *prices* can't go up, but *costs* are still so high that there isn't a single company providing inference at a profit. And since people got used to what they get for a dollar, roughly, the AI companies cannot actually raise prices to stop losing money. They're screwed.

Honza Dvorsky
@czechboy0.dev

The first "L" in "LLM"s is the problem. This is not a small issue that can be tweaked. The whole idea of building massive models that both train on the history of the Roman Empire and also on a bunch of stolen JavaScript code is absurd. Without massive VCs subsidies, this would have never happened

Honza Dvorsky
@czechboy0.dev

Language models will stay around in various forms, but LLMs can't be made sustainable or profitable. The AI leaders are just keeping the bubble alive to maximize their net worth. But they know already that what they're doing today will never be sustainable or profitable.