Even if there's zero malice in what these labs do, the reality is the local models are demonstrating that having the Uber Mensch of models isn't actually the "TAM" they forcast. The fact that a lot of people get along with models on consumer hardware scares them because their large models have a very small actual audience. Which means their future is closer to your utility closet than any magnificence of the singularity.

I have built some ridiculously cool stuff with free tier Gemini Pro handling online research/plans and Muse Glimmer 30B handling "fill in the middle" of Gemini plans; this with a single 64GB+24GB 3090, 2023-2024 i5 CPU (bought right before price fuckery)

Am talking about low level Linux kernel interfacing stuff, my own Google Opal like UI generator...still rough but it's machine state to bang into shape not a bridge. It works but needs the refinement common to early development

Imo due to recent job trope developments a lot of devs expect a git repo that feels like a 3-ring binder with their senior thesis and such tucked within

Local models inability to one shot such is being used as cope. Imo these models are more capable than we want to believe

This all reminds of 20 years ago when sysadmins refused to acknowledge the cloud would take off due to privacy concerns of businesses

Look how that went

Data center needs are inflated to try and capture all the business but future devices capable of optimized battery/telemetry aware local inference would satisfy most people

The data centers serve 3 groups; SaaS workers, data center landlords/CEO/shareholders, politicians being paid off by data center landlords/shareholders

The other billions on the planet do not directly benefit as much. Merely as a side effect of trickle down

I spent about a week as a side project trying to get qwen 2.5b running locally to be a job application filler bot with toolset. It was amazing to me that it was able to apply to about 65% of the jobs, based on certain issues that might arise, sucessfully with that toolset, when given the job app urls as the only starting point info (plus user job data/resume in json).

Of course the Sota models on the same custom harness got to about 99%, but I'm amazed at the performance of such a tiny public free model running on my 3080 laptop gpu. If it was actually finetuned for said task I imagine it could get to 90%+ completion.

Right, my point is quite clear: Even if their "total addressable market" is a real number (big if), that TAM is probably easily divided up by tiers of models, and a significant portion of those tiers is already covered by local models.

Of course, any given person might want their own navier-tokes agent swarm of a solution, but the number times you want a bot to fill out a resume rather than solve a millenium problem is absurdly lopsided.

So it's just kinda, ok guys, are you going to just try to hold us hostage with your magicaly super-duper-intelligence, or are you going to admit that having great models means humanity gets a few smart people to be even more useful for whatever they're working on.

The rest of us arn't going to even see that future to begin with, because food, it's not about how much food humanity grows; it's how it's distributed. Unless this AI stuff solves a problem _no one_ has solved for food, it's just going to be another power chair to fight about who gets to sit there, who has the right to sit there, or like I suggest, it's just a utility thats to be managed by the same governance all other utilties do and has the same problem thereof.

So like, boring stuff; not p(dooom)

Yeah, I'm also amazed how the conversation is always about 'coding', like these models are only good for that, and are going to replace all software engineers while leaving everyone else alone.

IMO these models are best at non-coding tasks. They of course have to use a computer, but from CAD generation, 3d Modeling, emails, legal and academic analysis, etc. these models are incredibly capable, and we are going to see a massive shift in the entire job market due to this, and ironically I suspect one of if not the most protected career as a result of this massive upcoming buildout will actually be software engineers.

Essentially I agree about not reducing them to just coding apps.

Am ultimately whittling everything down to Linux kernel + inference + data model so far captured from my current OS use.

Am ready to be done with syntactic programming of the machine. Rather than label the filesystem with context I can label entities in the viewport and those entities can happen to look like word processors not just video game characters.

An basically making a Blender-clone designed to browse and edit models that replaces the usual desktop metaphor.

Then I can noodle the virtual geometry, label it and understand under the hood its just a function factory mutating geometry of memory address truth table and display truth table.

The last gasp/push to rid ourselves of OOP is upon us; code objects of all kinds gonna be dead sooner than later.

Not sure I agree then software engineers will be insulated when everyone can do state management through visual edits. Maybe visual design work will still be a thing; the how to get machine into desired state is most certainly not going to be with visual syntax structure fir much longer