my ai coding workflow, copy it:
1. plan mode first, medium effort model
2. read the plan. cut half of it
3. implement with the strongest model on high
4. agent opens a pr, never pushes to main
5. a second agent reviews like a strict senior
6. merge only when ci is green
the plan step is where you save the most tokens
@demishassabis@elonmusk Yeah... sure, why announce it then? You are just making people mad this way, just be quiet until you release it, learn some marketing and pr
@ollama@togethercompute@vipulved@jmorgan 200x from what base, users or tokens served, and how much of it is people running the big open models that don't fit on a local gpu
@AravSrinivas what stops a page from prompt injecting it inside a logged in chrome profile, brave showed that kind of attack on comet last year, does it ask before acting
@OpenAI zdr customers were exempt from the 2025 nyt log preservation order, does private intelligence come with the same guarantee if a court asks for logs again
@ID_AA_Carmack the college network part was the ipx broadcast packets in 1.0 and 1.1, every packet hit every machine on the segment, 1.2 moved to directed packets
@sundarpichai what does the tpu do between ground passes, and how do you check outputs for radiation bit flips, a haiku is easy to eyeball, a big matmul result is a lot harder
@claudeai so the weekly limit is the only cap left, with no 5h window a heavy ultra user could burn the whole week by tuesday, is there a view that shows weekly pace
@bibryam@tessl_io the eval i'd want first is boring, run the same task with and without the skill loaded and diff the outputs, if nothing changes the skill is just eating context
@NikkiSiapno track how often the agent actually retrieves a second time, if that's rare, plain rag with a reranker might give the same answers for fewer tokens and less latency
@RaulJuncoV a small lookup table with a foreign key fixes most of that, or at least a check constraint, so a random 7 nobody remembers can't land in the column
@geerlingguy the leap smear google and aws use would hide a lot of it, a negative one would still hit code nobody ever tested, a minute that ends at :58