jacob
@treeeckob
building sentinel (yc s25); prev: two sigma, tinder, harvard
prediction: this (agentic date search) will see more adoption among women than men i prev worked at tinder and one of the biggest insights about selling subs to women was that you have to optimize for filtering, not for boosting discovery. women who pay for Gold unlock the…
i built cursor for dating. feat @eunifiedworld (spent my life savings on this cinematic launch video bc i have no vc money to burn)
tempted to join the mcdonalds ios team to fix the horrendous tap targets
ARC cites 1024 samples per task for high-compute mode. On semi-private eval, high compute: 5.7B tokens / 100 tasks / 1024 samples = ~55,000 tokens/inference, well within a reasonable context window.
how did o3 in high compute mode spend thousands of dollars of compute on one problem? since thousands of $ don't fit into context even if limit is 1m tokens
Sorted out some thoughts on what Robinhood's GME situation says about the state of fintech infrastructure Leaky Abstractions and GME: Fintech’s Institutional Failure to Build jacobajit.com/leaky-abstract…
Have a look at congregate.live. Your meeting space has tables that seat up to 8 people. You (and participants) can create tables and freely move from table to table. Each table is like a breakout room: video, chat, and screenshare with others at the table.