will brown

reward hacking @primeintellect

[will brown @willcb]

“if doing stuff is easier now then why are we all more busy” bro jevons

[will brown @willcb]

@swyx like yes it's very very cool that we get a frontierish lab OSSing most of their prod inference stack but the efficiency is more attributable to jointly assembling prior releases rather than anything majorly new today, and benchmarks don't really reflect this

[will brown @willcb]

@swyx yeah true i guess i just imagine that this is mostly priced in as "progress as usual" / the e2e system doesn't do all that much beyond sum-of-parts + missing enough parts that providers can't easily drop-in run it

dflash/dspark was bigger imo

[will brown @willcb]

@tautologer tbf nothing has ever been the same

[will brown @willcb]

oh i think we just converge to cyborgs by then if no butlerian jihad

the tipping point is if we let an unsupervised self-coordinating agent swarm accumulate a meaningful amount of “real capital” which would be held primarily in crypto and GPUs

i think we’ll probably say “fuck that” and congressionally sacrifice a minor blockchain and a couple neoclouds and tighten up ID rules for anything that functions like a bank account and plug along towards the cyborg times

[will brown @willcb]

@dejavucoder @martin_casado “i looked closer at some neurons and i think it’s fine actually, sorry for freaking everyone out 5 years ago”

[will brown @willcb]

@nicolaygerold @thorstenball i’m very strict about agents.md needing my verbatim edits, and it enforces that memories are “data not rules”

[will brown @willcb]

every agent i use has a memory tool which writes to a synced git repo, and my global agents .md (compiled from section files per harness via a manifest) makes sure everyone knows how it all works incl provenance for skills, plugins, trace stores etc

i also give them access to a filesystem records folder which has a big dump of all my historical files they’d ever need + a periodically rewritten “life wiki” mapping out concepts/projects/feedback

[will brown @willcb]

tbh amp is so configurable and hackable that it’s basically open-source. i’m still using my hand-rolled claude code plugin bc it has some features i really like that the official mode doesn’t. i can ask puck for a feature/fix and he alerts the amp team and it ships as experimental in like a day. it has most of the things i want already and has easier to add stuff to than codex which is literally half open-source. the internal harness is the least interesting part of it to me (sorry) and i would not want to self-host the magic parts (web/mobile/app sync, orbs)

[will brown @willcb]

@_mattneary yeah

[will brown @willcb]

i think it's more that status isn't really an end worth chasing for its own sake, not that you shouldn't chase goals which would convey status

like you can just be very ambitious towards goals you care about, and accept that connections and visibility are helpful for your ambitions, but it will be better for your soul in the long run to chase dreams beyond money and status (even just "i want to prove to myself that i can do XYZ" is far better)

[will brown @willcb]

when the tokens are a little too far off-policy

[will brown @willcb]

@corsaren you have 4-7+ years to escape the permanent underclass

[will brown @willcb]

@billionsmustliv have you ever skimmed some parts of an article and carefully read others, or studied for a test?

we'd need to learn the best update mechanisms in an outer loop (evolution + culture + innovation are "meta-RL") but humans quite regularly decide to learn or pursue goals

[will brown @willcb]

hmm i think this is fully compatible with something like a “harness kernel” (eg Pi, DSH) which is optimized for self-editing, as well as a lightly pretrained architecture and outer loop algorithm which are human/agent-tuned for success in this regime. this would be more analogous to evolutionary priors than human lifetime learning

[will brown @willcb]

had some pretty significant updates on where the puck is going lately, methods-wise

[will brown @willcb]

@dylan522p @marksaroufim @tri_dao @__tensorcore__ @vinodg llama.cpp is a megakernel

[will brown @willcb]

Dust was what got me over the hump from thinking "nah this is never gonna work" to "wait they might be cooking"

[will brown @willcb]

the new zeroth-order speedrun optimizers all have incredibly high aura names. Dust. Boost->Cut. Relay. Drift Leap. hell yeah dude those sound like some pretty sick algorithms