Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command. Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec. It's a straightforward way to experiment without needing your own GPU. Runs at roughly $3/hr and drops to $0 when idle. Get the code and instructions here: github.com/taeold/djev-run

Updated:

Updated:

Updated:

All-In Summit: Naveen Rao -- Why AI is extremely inefficient, and how to fix it -- AI is about to run out of power -- Your brain does this on 20 watts, the data center needs a city -- Every computer since 1945 is built wrong -- Why the doomers have the energy story backwards ++ much more! (0:00) Welcome @NaveenGRao (4:45) Is energy really the problem? The cost of a token, power contracts & the gap to close (10:24) Cutting out the middleman: abstractions, dynamical systems & a new kind of machine (19:40) Chamath joins: the path to product, porting existing models & building the team

IQ points per watt: Humans: 5 AI: 0.000001

Ossoff: ​​This week we were warned again of existential risk to the human species. Did you see it? And what fills us with dread is the crushing weakness of our politics against the massive stakes of our times. Tech titans dig bunkers while warning us the new intelligence they're creating risks human extinction. Congress debates youth sports, and the president builds a ballroom. An ancient establishment, barely capable of email, sleepwalks into a technological revolution, while the president dismantles precautions and calls us fools for worrying. Just listen to this: On day one, Donald Trump rescinded the AI safety order meant to guard against catastrophic risks, including the engineering of bioweapons. Meanwhile, Don Jr.'s 1789 Capital bought big money stakes in data centers and AI labs. Four days before inauguration, the Emirati spy chief committed $500 million to the Trump family crypto business, and four months later, the Trump administration granted the Emirates the rights to buy our most sensitive AI chips.

I consider @sama a friend. But let’s think about what this means. It means if you were to do the impossible as a mathematician and prove the Riemann Hypothesis or P=NP, you can’t buy a home with your winnings in Atherton next to VCs & CEOs. Navier-Stokes may be worth Billions.

This model represents a step-function improvement on many benchmarks, and its training is ongoing. Our internal model group arrived at the Navier–Stokes solution in 88 hours, using around 10,000 coordinating AI agents. Throughout the effort, we maintained the strict safeguards—including monitoring and isolation—that we apply to all our frontier evaluations.

I love Smash 64 but I wish I was in it, so I hacked the original game to run in the browser and put myself + 1000 others in Upload yourself at smash dot fun

"openai engineer explains /why/ he didn't need to understand the kernel line by line" 🙃 we're doing compilers 2.0

i agree with this and think that CoT is at best an epiphenomenon of current training methods. it will break (in the future, not now). safety community too often centers thought terminating taboos like “neuralese” and “training on interp”

Letting models use worktrees was a terrible idea. They are all over the place, they are never cleaned up, and they are huge with lots of duplication, clogging up the whole disk in no time. Awful.

Introducing Atlas: The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D. Model the world, move the camera, and simulate space & time.