Zuck is singlehandedly boosting every AI lab's attrition rates, including his own.

Benchmark performance diffs <5% are meaningless.

Every AI code assist user I know experienced this. My advice - dont stop, keep trying. Eventually you get a feel for what these models are good at and what they are not. And then its magic. Many boring repetitive tasks, done in an instant. Massive refactors and translations,

Ok I thought GPT5 is a good direction. But this is really bearish.

Is there any better feeling that waking up in the morning to see your loss has converged...

One of the great bold technical bets made. Can vision be enough for fsd? After all humans do it, right? Imo it's possible cameras are enough. But I'd never try to constrain a machine to human senses. Humans always cheat, using whatever they can to achieve their goals.

It's a nice sentiment, but that's what they said about pirated music. If all AI gets banned, pretty sure we will forget about them in a year

Had my agent moment a month ago. I let Cursor take the wheel and it was... magical. I don't even know what model is being used. I'm quite convinced the app layer will accrue more consumer value than the foundation layer.

Reviewing dozens of resumes, Langchain really is a negative signal

Oh snap. We built this city. Join us to build foundation models that understand sleep and physiology.

Prediction : this will end badly for somebody. Buying your customer is not a good incentive, the only question is who will be left holding the bag.

I find Claude/GPT's use of emojis cute and endearing.

sometimes gpt fixes the issue in 2 loc. and goes on to think for another 10k tokens and writing another 1000 loc to test it.

I haven't googled anything in a month

Forget cows farting, this is why the earth will be much hotter in 5 years.

Every successful startup has at least 1 Bob. If your startup doesn't have one, ngmi

Something clicked in the last month. What I spent the majority of the last 10 years doing -coding- is now 90% vomited by an omniscient dumdum. My expertise has been replaced. My taste hasn't... yet.

what is the name for > train loss is stable > val loss decreases > not in sudden grokking jumps > just steadily decreasing...

Model learning to distill embeddings faster than logits, wat mean?

ok i admit, i yell at my GPT when it tries to gaslight me.