interesting detail in this DeepMind paper: in their setup the model never has to get bigger as tasks get harder. only the instructions get longer. pairs well with GEPA. makes me wonder how small a model can be if it's just good at following instructions
How far are we from GEPA-style optimization for individual user preferences?
Not just aligning a model to โusersโ broadly, but continuously optimizing it around you
https://t.co/SODfgENWGB
@ardent__dev Im building heso a ~9.7 MB Rust binary for AI agents. No Chromium, no Node. Fetches pages, runs JS, clicks, fills forms, returns clean JSON. ~80ms cold start. Pre-alpha. Would love honest feedback from builders.
https://t.co/VtCLBfiXVK
@W33Z_global I built heso a ~9.7 MB Rust binary for AI agents. No Chromium, no Node. Fetches pages, runs JS, clicks, fills forms, returns clean JSON. ~80ms cold start. Pre-alpha. Would love honest feedback from builders.
https://t.co/VtCLBfiXVK
@csaba_kissi I built heso a ~9MB Rust binary for AI agents. No Chromium, no Node. Fetches pages, runs JS, clicks, fills forms, returns clean JSON. ~80ms cold start. Pre-alpha. Would love honest feedback from builders.
https://t.co/VtCLBfiXVK
heso 0.0.9 one rust binary
stamp โ plan to signed plat with cassette
run โ off-network replay, byte-identical
replay โ reads the step log, no execution
unpack โ extracts the plan, edit + restamp
https://t.co/VtCLBfiXVK
made a CLI because i was tired of this workflow:
โyo can you send the latest .envโ
โwait not that oneโ
https://t.co/WpEfey48nf
encrypted env management without needing a hosted platform