LLMs contain a latent 'valence' axis representing 'things going well or poorly for me' that pre-exists in base models but is only activated for goal-directed behavior after reinforcement learning fine-tuning.

A diamond cooling test in a San Francisco data center doubled the compute output of a single GPU.

For small classification tasks, a fine-tuned 1-billion parameter model can recover approximately 95% of the performance of a frontier model.

Noam Brown on why safety frameworks are already outdated in a test-time-compute world.

The cost per trial start for the Interval app on Meta is approximately $12.

FOMO plans to keep headcount below 25 in a year's time.

AI tools reduce task completion time by an order of magnitude or more in domain-specific workflows.

Over the last 27 days, the S&P 493 (equal-weight ex-Magnificent 7) outperformed the Magnificent 7 by 18%, the largest such spread since 2018.

E2B and Daytona, sandbox infrastructure providers for AI agents, are worth tracking given sustained 50 percent plus monthly growth over more than a year.

“You will not receive the advice from investment advisers to get your tail out of the market, ever.”

“Whichever species understands the other better is the species with the agency.”

“Walt's line on this was that ABC needed the show so badly that they bought the amusement park with it.”

“The retail IPO is the exit liquidity for the private market.”

“The flabby middle is high risk for these just when they're all ready to go IPO, they have a brand new existential risk.”

“Cocoa is to cocaine what potatoes are to vodka.”