Post

shout out to this imho _incredible_ @himself.bsky.social thread, which i think super nails the cognito-hazard of over-polarizing oneself. we generally OUGHT work to be receptive to many ways the world is happening around us, even admists clear bad. don't affix thineself. bsky.app/profile/hims...

I suspect that this is, maybe, true for classic chatbot inference but not for reasoning models. It comes from the logic of the race too: better incrementally spend 2x on compute than be caught with your pants down being 2% behind on the benchmarks.

Heh I'm not quite so adversarial as the last half of your statement, especially since I think companies go through lifecycles during which their initial growth tends to produce more social value than they capture and so I wouldnt want to interrupt that - yet But yes it's premature atm

(Tongue planted firmly in cheek) if only there were a trustless decentralized mechanism for assuring service (preferably in TEEs) and a means to do global payments between providers and consumers

of course it's not just about jobs

frankly i’m already impressed by the models i can run on my phone. knowledge is an issue, but their ability to perform purely language-related tasks isn’t really all that far off the big models at all

You can run a decent model on 16 gigs of ram on a M series Mac now. Building agents is a lot like web dev, request response and doing things in between. I have enjoyed working with the smaller ones, it’s interesting what can be done.

one underappreciated fact is that there are like 100 “neoclouds” out there right now that you can go to and run open weights models

I think the cognitive load required to make something that conforms to your firms expectations will still require people regardless of how efficient these models are. Similar to how outsourcing your core product never really works I agree the future is less concentrated tech talent

Somebody recently posted here about how people ai coding are able to solve their own needs but not make good software for other people. That tracks to me. That's where I think dedicated professionals could still be very necessary. Again, at the current model abilities.

Universities too. I work as a developer for a mid-sized state university and we do a lot of custom work that would otherwise be prohibitively expensive and ill-suited to our situation.

"They will be able to do anything you can think of/speak/express." At that point SaaS becomes a triviality and we are dealing with the same type of problems we always have: "managing conflict and resources."

We’ve been interested in a hardware co-op model across the board, including regular compute/cloud stuff. Still in the recurring idle thought phase.

I've had similar thoughts www.newscientist.com/article/2516...

I expect the models will become commodities. Open ones are not that far behind. I see the risk as that the companies de-skill huge numbers of people who become dependent on the convenience of their systems. Then the prices will start going up and the companies will manipulate their users.

methinks possible with tools like exo!(I’m not a developer/contributor, just a user) - distributed compute seems like a fantastic way to bring large-model power to local systems; I’m hoping it can scale well

Yeah it’s an interesting concept @scoiattolo.mountainherder.xyz

This may become a necessity. Again, if models become "super intelligent" whatever the heck this means and there are only like 3 providers ... that is very very bad under the current conditions. We need to hedge ... but all the talent is concentrated there.

More than just running open models, there is an opportunity to provide small models trained for a specific task. Agents defined for a specific task, do not need a general purpose LLM. They can use a task specific model. This reduces training and inference costs.

i think it's cool that the billionaires get to decide to shake society and see if it breaks and none of us get to do a damn thing about it