live footage of me logging in to wandb to check on my model youtube.com/shorts/PKg2ZzP…

easy trick to improve all your model fits

News you can use (assume all the authors did hands on work)

will something finally dethrone adam?

❤️anthropic

AI efficiency is important. Today, Google is sharing a technical paper detailing our comprehensive methodology for measuring the environmental impact of Gemini inference. We estimate that the median Gemini Apps text prompt uses 0.24 watt-hours of energy (equivalent to watching an

if you’re teaching econ101 this quarter here is your bertrand vs cournot example

weaponize your adhd by keeping habits you want to keep up with always in plain sight

a research project proceeds in this order: get drunk with your co-author in a bar, argue about some stuff, write a paper/and-or ship the idea to prod within a few weeks or months Adam Lerer

biology remains a mostly unexplored frontier

gpt5 and claude both strongly believe that torch.sdpa wants attention masks with True meaning masked (rather than the actual True means valid)... is this old behavior of the function? it's weird that they both think this

in flash attention v5 the parts of the gpu that compute attention softmax will be physically moving at relativistic speed relative to the rest of the machine creating a net speedup

this is just more evidence that llms are basically real developers

nice paper. an interesting question here is how much of this is driven by 1) real vs perceived abilities of ai (vibe coding is not as good as everyone on twitter thinks it is) and 2) firms' uncertainty about tomorrow's ai capabilities (maybe gpt6 will do the thing gpt5 can't?)

ai is writing a huge fraction of my boilerplate code, i basically write stuff in a combination of english and python syntax (e.g. train_data = my_dataset(load all the parquets in args.folder, other args that are important)) and its awesome. text2app doesn't really work, obviously

little known fact is that the guy who invented quants was inspired by mentats in dune

how to apply academic ml literature to your practical problem: skim a bunch of the sota papers there will be some simple baseline they’re all barely beating use that but clean the data, scale up the data 10x and the model size 5x