David Harmeyer

Just spent so much time working with loss curves that when I opened up my 401k today, I panicked because I saw it going up and to the right

Yes, looks like it's a combination of that, and not enough variation in samples that killed it. tanh() is a million times better

World's ugliest loss curve. Is this just what you get for using sigmoid in hidden layers?

The AI times have begun

@steipete @openclaw Legendary

New @openclaw beta is up! This one again focusses on security and bug fixes, but we added a gem: TELEGRAM MESSAGE STREAMING 🚀 to update, ask your agent or run: openclaw update --channel beta

@xubinnrencs RAS syndrome

@tsoding At that point the only reason that'd be useful is as precompute-expensive, query-expensive dataset compression. But in the real world, data is cheap and compute is expensive, so nobody would do this unless there's some immensely tight bottleneck on transmittable data.

Wait, this sounds incredible useful! Can we just have a model with 0 entropy, 0 hallucinations, that just acts like a retrieval database over its training dataset? Also sounds like a great way to solve the traceability problem. Why don't the AI labs just make something like that?

Sleep is just "compacting conversation..." for humans

@sciencegirl I guarantee if you give this to actual people, they will misjudge the number of steps on the way down and trip

A simple invention making stair climbing easier on the body

@zebriez Competence

I'm looking for a word for "get it-ness." Something to describe that sense you get of a person that "gets it." Usually immediately apparent. contenders that are close but not quite it: awake. in the details/close to the metal. high-agency.

@hyhieu226 If your reasoning isn't autoregressive, it's probably rationalization, not reasoning. Maybe rationalization makes sense in some cases, but it's a lot more artistic than it is logical.

Naive question, so please roast me. Why don't we have diffusion reasoning models? The way humans think look a lot more like diffusion than autoregressive.

github.com/Kotlin/kotlin-… is amazing, but the fact that it's not available in the extensions search is confusing. It strikes me as obviously the best way to use Kotlin, but it's so hard to find

Many people fail because they go off to build a generic platform, and then find usecases afterwards. It's so much more compelling to find a problem, build a solution, and then after that's proven, make the solution generic.

Register here: facebook.com/codingcompetit…

Looking forward to Hacker Cup Round 1, starting tomorrow morning!

Every time I remember that "m" comes before "n" in the alphabet I do a double-take

Good men must die, but death can't kill their names.

@sama I get that rollouts are tough and it'll take a bit to iron out the bugs -- but could you guys at least provide the option to use an older model until things are smoothed out? There's 0 available GPUs, even for premium users...

going to try live-tweeting the GPT-5 livestream. first, GPT-5 in an integrated model, meaning no more model switcher and it decides when it needs to think harder or not. it is very smart, intuitive, and fast. it is available to everyone, including the free tier, w/reasoning!

@AmericaPapaBear How can I qualify for government grants? S.I.M.P Squirrels in my pants

Doorbell camera catches man getting attacked by a squirrel. Have you ever seen a squirrel like this?

It's a hoot how many people have this perception that prompting an AI is somehow a high-variance skill that requires expertise and training. Reminds me of adults saying we "needed classes to learn Microsoft Word" when I was a kid.