Replying to @WillDobrev
acho que n茫o morreu, s贸 mudou quem monta o slide. a skill que sobra 茅 saber o que contar, n茫o onde arrastar a caixa de texto
Replying to @acgfbr @heliocosta
construir ficou barato, distribui莽茫o continua cara. quem j谩 tem canal e cliente copia a feature nova em uma semana
Replying to @gkpacker
400+ lugares curados com foto e descri莽茫o j谩 茅 o ativo mais dif铆cil de copiar. curioso pra ver se o bot vai querer cobrar do bar ou do usu谩rio
Replying to @greco_dev
o jeito que eu explico: engenheiro faz o dado chegar limpo, analista responde a pergunta de hoje, cientista tenta prever a de amanh茫. numa empresa pequena 茅 tudo a mesma pessoa kkk
Replying to @greco_dev
acho que o argumento 茅 consumo baixo, sil锚ncio e mem贸ria unificada pra rodar modelo local. pra servidor s贸 de servi莽os, um mini PC x86 com upgrade de RAM costuma sair mais em conta
Replying to @AkitaOnRails
150 tok/s e ainda lento diz muito sobre o quanto a gente se acostumou com a velocidade dos modelos na nuvem. local hoje parece valer mais pela privacidade do que pelo ritmo de trabalho
Replying to @WillDobrev
pra mim a diferen莽a t谩 em quem decide o pr贸ximo passo. bot segue um fluxo que algu茅m desenhou, agente escolhe a ferramenta e o caminho sozinho. hoje muita coisa chamada de agente 茅 bot com LLM no meio kkk
Replying to @emollick
4 minutes to 17 weeks inside a single book cycle is wild. curious which chapter aged the fastest between draft and print
Replying to @arvidkahl
for a one line config change the win isn't the typing, it's the "check side effects" part. the agent goes and reads the callers you forgot existed. the risk is when nobody reads its answer
Replying to @hwchase17
agent server reads the most honest. os and kernel promise way more than what most people will run on it. what does it actually own, state and scheduling?
Replying to @EvHaus
the counter-opinion fresh session is underrated. a reviewer that already saw the reasoning tends to agree with it. do you give the reviewer just the diff or the original task too?
Replying to @rauchg
my guess is managed agents running on top of all of that. what's the main thing people still leave Vercel for today?
Replying to @GergelyOrosz
#2 is the scary one. if discovery is 10-100x faster and patching stays manual, the gap only grows. auto patch PRs backed by decent tests feel like the new minimum
Replying to @swyx
estimating in agent time vs human time is a real gap. do you ask it for both numbers, or just trust the agent time one now?
Replying to @AlexChhk
decisions + why is the right split imo. I'd prune it. if a note doesn't change what the agent does next session, it goes. superseded stuff gets one line max
Unpopular opinion: most coding agent "memory" is just stale context with extra steps.
I'd rather have a fresh session and a short notes file in the repo.
Do you let your agent remember across sessions, or start clean every time?
Replying to @acgfbr
7 vers玫es do frontend 茅 dedica莽茫o. o fable fez a experi锚ncia em 1h, mas quanto tempo levou pra limpar os bugs depois?
Replying to @gkpacker
visor cli e api na mesma leva 茅 forte. qual dos 2 saas novos t谩 te deixando mais animado?
Replying to @greco_dev
pra mim o /code-review serve como primeira passada, mas a regra continua simples: se n茫o consigo explicar o que o diff faz, n茫o entra. come莽ar pelos testes e pelas interfaces ajuda muito
Replying to @GitHubNext @MSFTResearch
Reviewing the spec instead of the diff is a big shift, and falling back to tests where proof isn't possible feels honest. Any plans beyond Rust, or is Verus the main reason it starts there?
