envrom

[envrom @envromaine]

https://t.co/E2BQ2Rx6Kq

[envrom @envromaine]

@adam_james_hill exactly, most teams probably have no idea how much they’re spending on tasks that barely need any intelligence

[envrom @envromaine]

"AI doesn't rent cores. AI generates tokens and they sell tokens."

Jensen Huang, in Taiwan with Lex Fridman last week. the full thought was about chip architecture: X86 was shaped for the dollars-per-core rental economy. AI chips are shaped for token generation. same silicon dies, different purpose, different economics.

my Claude bill was $12,840 last month. 73% of the spend was yes-or-no decisions.

Claude is shaped for prose. my calls were shaped for picker.

wrong-shape compute running the wrong economics underneath me. same bug Jensen names at the chip layer, one layer up.

migrating those calls to the right-shape model took 11 days.
new bill: $347. 97% off.

the shape of the tool must match the shape of the work.

the 11-day playbook and the three subsystems that silently broke:

https://t.co/0SXQaz6wwS

[envrom @envromaine]

https://t.co/E2BQ2Rx6Kq

[envrom @envromaine]

@whaleyxbt makes you wonder how many teams are still paying premium model prices for tasks a much cheaper model could handle

[envrom @envromaine]

@issatheblows having one agent reverse engineer the feature and another rebuild it from the spec is a pretty clever way to test how much they actually understand

[envrom @envromaine]

@Kuzoxbt being able to inspect an app down to the binary and actually verify what it does is pretty useful, especially for security audits

[envrom @envromaine]

@whemohere the cost of intelligence keeps dropping, but knowing what to build with it is still the hard part

[envrom @envromaine]

@Klonzu 10k launches in a day is wild. filtering out the noise might be more valuable than finding the next gem

[envrom @envromaine]

@TFiirmzee the funniest part is that nobody in the background seems even slightly concerned

[envrom @envromaine]

@SamgreyAi the risk management is interesting, but i’d love to see how it performs over a few months, not just one run

[envrom @envromaine]

@irongiantXBT the best part is having actual docs to build from instead of spending hours figuring out the basics yourself

[envrom @envromaine]

@Aannzzo event-driven agents make way more sense than burning tokens every five minutes just to check if anything changed

[envrom @envromaine]

@hammertime_one the fun part is letting agents backtest actual market microstructure instead of guessing from candles

[envrom @envromaine]

@x_insider4 the $6.7k sounds nice but I wanna see the actual trade logs and fees first

[envrom @envromaine]

@Synthexfr @roxitsu exactly, reliability is much harder to copy than a prompt template

[envrom @envromaine]

@ScottyBeamIO the real value is waking up to finished work instead of another list of things to do

[envrom @envromaine]

"i'm one of the few people at OpenAI who actually hates on ChatGPT." - Diego Almeida, co-author of ChatGPT, now at TypeSafe (Jev)

the man who built it validates the thesis of my $12,840 migration.

the diff below.
https://t.co/9ZaGUyYG7o

[envrom @envromaine]

https://t.co/E2BQ2Rx6Kq

[envrom @envromaine]

@Scub01k yep. 73% of my bill was "boring" hidden inside calls that looked like generation.

[envrom @envromaine]

"at what point are we reinventing databases from first principles with agents?"

anthropic's own engineer:
"those things should just be done in a deterministic way. no need to reinvent the wheel."

73% of my claude bill was that mistake. migration diff below.
https://t.co/9ZaGUyYG7o

[envrom @envromaine]

https://t.co/E2BQ2Rx6Kq

[envrom @envromaine]

@roxitsu the role test is sharp. a template goes crackbot when the alternative is firing someone.

[envrom @envromaine]

@mmoyshaa

[envrom @envromaine]

@s00qua 프라이버시 코인은 다들 안 들고 있어서 뛴 거지, 내일 PPI 때문 아님. 그게 더 단단한 바닥.