🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Gemini-cli has improved at an astonishing pace. I'm impressed.

this is actually a pretty reasonable plain english summary of modern(ish) recsys. matrix factorization -> collaborative filtering -> iterative candidate generation and filtering -> a/b testing -> cohort filtering and expansion -> iterative signal based rerank

very good mech interp research. excited to see real focused attention here and have been very impressed with the work from goodfire generally

this plus the zurich crew is even more impressive than the thinky departures. strong focus on multi-modal and reasoning

im trying to build tools that support dark mode and light mode to a reasonable degree. as a result i've started using 'system theme' which switches me from light -> dark at sunset. This is my personal hell. i do no recommend this to anyone ...

i'm telling everyone, giving up on MCTS is a bad idea (this is an interesting paper)

this is what peak nerd "for the love of the game" looks like and i could not love it more. just guys being dudes running Omarchy, nix and Ghostty and building

my doubt has nothing to do with the state of AI more the state of the internal tooling, security and bureaucracy at these places

Wow ... Based on the popularity of claude code this is a pretty impressive get for anysphere

This is one of my "lol this app is free" moments

i will happily give zaz more GPUs. Meta FAIR not supporting this researcher and this work while also starting MSL feels ... odd ...

no. just do MCP with reasonable auth. this is silly (as is A2A)

replace the vision encoder with SigLIP and this might actually be a banger model

muting 🧵and 🚀

this looks solid. good work team

This is one of the most interesting papers ive read in a long time. not only in terms of token efficiency but also in terms of potential interesting latent interactions with the higher order trilinear representations. arxiv.org/abs/2507.02754

2017 - 2022 is when interesting things were still being published (but not yet tested at scale) and then most were summarily forgotten in the race to clone chatgpt starting in 2023

they are also different to the reference implementations (at least for llama4 and DeepSeek)

If the cursor pricing drama has highlighted anything for the community, it should be that the model is the product. The economics for wrapping and aggregating tokens just doesn't make sense long term

this is happening more and more often and i dont like it