okay I will say that whatever the RL/Finetune guys are doing w Opus 5.5, seems to be working splendidly. Oh look, throwing more compute and making it act as a forcing function gets you some of the way but not all, who would have called that

man we can never have anything nice anymore

yo is there anyone on here that does harness benchmarking? asking for a friend....

half the ai timeline is now people using coding agents to do chores a bash script did in 2009. the difference is you no longer have to remember the flags, which is either the whole point or the end of remembering anything

bro this cannot be real, this is an insane timeline....

wild that the hardest ai targets left are verilog and place-and-route, the combinatorial stuff with almost no training data, meanwhile another chatbot ships every week trying to optimize for a vertical

sparse attention plus moe everywhere is the whole field finally optimizing inference cost per token, training flops stopped being the flex a while ago

open-weight moe drops are turning a single 24gb card into a real dev box, 4-bit quant doing more than any new architecture, and making frontier models more accessible

flops stopped being the ceiling, hbm supply and rack-level power decide who ships this year

the 5 hour usage caps are just load shedding with a nicer name, inference demand outran the fab roadmap and now everyone queues. this is basically us waiting for our limit to reset

agents crush software because the verifier is free, you run the code and it tells you. everywhere else the reward signal is a humans mood on a tuesday. this is why everyone optimizes for coding (but tbh we all knew that)

what.... what do you mean bandwidth-adaptive CPU GPU, holy shit

I have now merged with github, all my actions are github actions, so they fail all the time and keep blowing up production. Yippee!

things i've learnt at 3:50 AM on a Thursday: 1. Knowledge graphs 2. Niche internet rabbit holes 3. Rapid QA and multi agent swarms 4. correcting AI 5. escaping sleep there is a 6th point here somewhere, but I seem to have lost the ability to type...

1-bit quant getting a 4TB model down to 400GB is nuts, shoutout the labs that are making this happen, especially @UnslothAI