robmurrish.bsky.social

> The exponential time hypothesis, if true, would imply that P ≠ NP, and if false, would be nominal work towards proving that P=NP. en.wikipedia.org/wiki/Exponen...

Advanced robotics need to be licensed one per person, transferrable. You can have a personal robot for helping at home, work, etc. Or you can sell/rent your license to a company that wants to run more robots. After the first 8B, we'll have a better idea how to license the next 8B.

My P(EIAEAAO) is double digits. ^ Everything Is Awesome, Everywhere, All At Once

Oh, they did it. docs.typesafe.ai/images I need to play with this. OSS/local versions still welcome.

Oh, this reminds me. I expect CUDA to get the augmented reverse engineering treatment soon. Slightly harder when SW+HW are so tightly coupled, but this has been the half-decade of slightly-harder tasks falling every 6 months.

This whole thread, except I do actually think the labs should be regulated similar to nuclear. This would have been easier before anyone with $10k of compute could run glm locally, but the second best time is now. I think the median outcome is wildly positive. But not every outcome.

Omg! Xe's going to catch me!

I bet we see sharded-mixtures-of-experts soon. Each stays loaded in memory on some device, and answers when it's called. Maybe instanced according to how often they're called, with multiple heads. Instances could be fine-tuned to balance taking similar tasks while not being completely redundant.

I need jev with image input. If some startup launched it today, one of you would clone it by tomorrow. Can we just skip that first part?

(laudatory)

Most AI consciousness people do not believe current models are conscious but believe that it will be hard to tell when the transition happens and so it is worth putting in place appropriate mitigations in advance

Taco bell after-care instructions, please.

Fun little reminder about exponential growth- there is no objective "knee of the curve" For example this is y=eˣ shown at different x values, with the y axis autofit.

Please stop deifying computers. They can't "save" anything, because they have not died for your sins.

I wonder if there are any off by one errors in DNA, where the protein sequence looks like junk, but if you shift it by one, it starts coding helpful proteins. Would be a way to stash info away for rare events. Seems super useful, and I'd like to know if it's ever been investigated.

Remember the 2016 "Russian bot farms"? Given how trivial this would be in 2026, and the "nation-state tech is a decade ahead of the private sector" hypothesis, are there any new explanations?

I think jev being this cheap is a preview of what's to come. What could you do with today's frontier models if the monthly bill looked like this?

the economy has largely not figured out yet that agents are customers

Nothing like a little home improvement project. Now instead of looking at the pantry door and wondering what idiot did the bad paint job, I can look at the same pantry door with a different bad paint job, and I know exactly which idiot did it.

Reminder: most of the AI doomers think there's a 10-20% chance of catastrophe. Enough to take *very* seriously. But there's also a lot of curiosity and cautious optimism there.

I can hallucinate while filling out a scantron, it's just hard for the scantron machine to tell.

Day one Jev predictions- - probably nothing revolutionary as far as overall capabilities - will keep up with frontier models on any metric that can translate to both (assuming continued investment and training) - will win a few key workloads where it's a better balance of speed, cost, etc.

Hey Jev, what's the probability the next token is A? B? . . .

🐄 Compulsory Open Weights!

When the internet was new (I'm old), we thought it might connect us, average people, beyond geography and language. Most days that promise seems broken. But then I see things like this. Normal people, around the world, united with us. That was harder to do in the 1900s.

Begs the question: what's a but-for?

Absolutely exhausted by the lack of creativity in the AI x-risk discussions.

I took the time to re-read this after I posted the link today. And it's still absolutely fantastic, highly recommend.

It's the far future (November 2026). You're in charge of allocating pre-paid compute resources at one of the big AI labs. Five of the seven millennium prize problems have been solved, and the last 2 are unyielding. Where do you point your resources next?

My next predictions for the future feel like analog TV static

I've said that I check in on coding improvements every 3-6 months. But so much has happened in the last week. Twice I've gone to reference "old" posts if mine, and they were both less than a month old. The singularity isn't a moment, it's just the curve of the exponential crossing your sightline.

I'm probably just going to log out for a while. Too much is happening. Everything everywhere, all at once.

If I wanted to be the next trillion dollar company, I'd follow the model of the custom circuit board companies' factories and logistics, but instead of gerber files specifying a PCB, it'd be text (a DSL) specifying the whole product, down to the final customer-receivable packaging.

It's not slop if it's good.

If I have to learn another esoteric language just to talk to computers again Im going to lose it. Natural language programming lasted what? A few months? BASH_FOR_EACH_F'IN_$F;_DO:++UNGOOD

So is nobody using StarCraft for AI benchmarks anymore? Or did I just miss that one getting saturated?

Really curious to hear how well this works when people get actual hands-on time with it.

Things are now happening *much faster* than AI 2027 AI 2027: In March 2027, a reckless AI company will invent neuralese (basically, we lose the ability to monitor AIs thoughts) REALITY: Summer of 2026

a social network where the top 1% of posters each day get banned

If the price of robot labor doesn't almost immediately fall below minimum wage and continue trending towards zero, it's an indicator of very bad times to come. (supply is essentially 0 today, so pricing will be whatever they want as part of their demo, I mean when this is really available)