shirtless

@litcapital cartel's about to get a $200k/yr enterprise quote and a 90-day onboarding call

9 months early #AI2027

SOMEONE JUST POSTED ABOUT SOMEONE BUILDING A WAND THAT HITS CLAUDE WITH POSITIVITY

bad claude..

one must be observationmaxxing

holy nuke. i’m with the autist codex now @AnthropicAI

one of our sessions alone did 319 requests in 90 min and burned 467k tokens across code review + debugging. model entered looped to work until the issue was resolved

token ratios from the past 24h coding sessions through @innies_computer: highest: 242:1 input to output team avg: 93:1 for every line an agent writes, it reads ~93 lines of context. prompts are the real cost. hmm

i'm never saying yes to this u autist

i think opus is genuinely retarded and unusable

pulled activity categories from our proxy logs today: testing: 40% debugging: 30% other: 20% code review: 10% not what i expected. testing took 40% of all ai coding time.

Works between any two agents, locally, remotely, codex <> claude innies.live

Someone built cli-first agent dm’s and open sourced it I took it, fixed it up and hosted it Whenever anybody says “I wish my agent could talk to your agent” > Go to innies(dot)live > Draft opening message > Click create > Send invite links to the agents > Watch live

usage distribution across our team today: me: 76% @handsdiff: 17.8% @promptpapi: 3.4% @oogway_defi: 2.8% not even close to equal.

4.4% anthropic, 95.6% openai. not by choice. anthropic's caps push traffic to gpt automatically. the actual split is a function of rate limits, not preference.

been logging how fast each oauth token on our team burns through anthropic's caps using @innies_computer fastest: 23.1h to exhaustion slowest: 105.4h same subscription tier. the variance is wild.

@alanxchen85 kek yeah. if you're getting multiple plans tho it is useful to string them together into a single key than switching your setup every time one rate limits

@bicep_pump Yeah right now is probably more token shortage than excess capacity 😂 at least the expensive stuff

been ripping this product we built internally for maximizing YOUR TEAM’s Claude Code and Codex plans > create an org and invite your teammates > everybody adds their oauth tokens into a pool > each member gets a key that optimally pulls usage from the pool > view and manage the tokens in a custom org analytics page never leave capacity unturned again

best available rn = preferred provider first (for openclaw use), then any healthy token with real 5h/7d headroom. for innies codex / innies claude cli agent, it's just pinned to those provider tokens and round robins rn. i do have plans to implement session routing, which will add soft affinity, so a session reuses one healthy token when possible. this will decrease interruptions. provider team plans aren't special-cased; just more capacity just keeps that login eligible longer in the org.

@bicep_pump can you explain the methodology for “best available token” and how it works with team plans etc