“So. we stack layers where each does Wx + b followed by a nonlinearity e.g. ReLU. This builds a deep function. We train it by minimizing loss using backpropagation and gradient descen-“

My boss is cool btw you should come report to her and build cool stuff

Real

notable imo that the world's most valuable companies *are not* casinos, slop mobile games, or even tiktok/youtube instead they're apple (tools that ehance creativity), meta (originally about connecting people), and nvidia (enabling ai to accelerate science)

If you’re working in quant and are tired of the lack of meaning in your work, we’re always hiring! I promise you, we have more than enough compute. P.S. we’re chaotic because building anything new is messy. Do it with us!

getting started with evals doesn't require too much. the pattern that we've seen work for small teams looks a lot like test‑driven development applied to AI engineering: 1/ anchor evals in user stories, not in abstract benchmarks: sit down with your product/design counterpart

I think we’re in a timeline where > US builds general robots for manufacturing in the next two years > doesn’t matter because of our land use regulation

can’t believe we got jony just to fix our model naming scheme

opus not being great at benchmarks (but having really good user testimonials) is further confirmation of the deep deep eval crisis we’re in you can max out any benchmark with enough RL, and that doesn’t translate into a good product. you can optimize for DAUs and glaze-hack it

another banger hit the TL

Now that this is out I can finally say that we were building a vape but it turned out to be too dangerous, the tricks we prompted it to do were too sick.

A recent clarity that I gained is viewing AI research as a “max-performance domain”, which means that you can be world-class by being very good at only one part of your job. As long as you can create seminal impact (e.g., train the best model, start a new paradigm, or create

From the same admin trying to cancel nuclear subsidies which could be powering the datacenters lol

I'm a bit skeptical of the papers mentioned below, but I think it's a good thing to rerun the eval yourself and report whatever number you get. It's hard to replicate someones eval setup!

Me after using more than one pipe in a command

The admin wants to be good at AI until it means either recognizing > foreign talent > forms of energy they don’t find masculine or something idk

It feels somewhat astonishing that a third of the strategic bomber fleet of a great power can now be taken out in a daring drone attack - the foundations of war appear to be changing at speed. I don't know what this will lead to.

We must manufacture batteries at scale here in the US if we want to have any shot at the very near future of defense, robotics, mobility, and energy. Basically everything that matters right now goes back to batteries.

No, it’s like if neuroscience had unlimited resolution imaging and perfect intervention. You know, two key bottlenecks in that field.