The AI Reader.
Wednesday, September 9, 202638 min read
Working Knowledge · What Could Go Wrong

Automated trading crashed the market once, in 2010, with no AI involved

A rule-based sell order and a herd of high-frequency traders drove the 2010 Flash Crash, while the case that AI agents will cause the next one still rests on simulations.

9 min read8 sources
Working Knowledge · The Instruments

PaLM's training ran at 46 percent efficiency, or 58, depending on which FLOPs you count

A lab reports it to show a training run used its chips well, and precision plus a sparsity footnote can swing the peak underneath it fourfold on a single card.

8 min read8 sources
Working Knowledge · The Evidence

ReAct searched Wikipedia between thoughts and still trailed chain-of-thought on HotpotQA

Yao's team prompted PaLM-540B with one or two examples, and reading a real Wikipedia line before each thought erased the invented facts that were chain-of-thought's most common failure mode.

7 min read6 sources
Working Knowledge · The Mechanics

The same open model returns identical text two to three times faster on one service

A small draft model guesses the next few tokens and the big model checks them all in a single pass, so the words that would have come one at a time arrive in bursts.

7 min read8 sources
Working Knowledge · When AI Breaks

Cigna doctors denied over 300,000 claims in two months without opening the files

Its PxDx software flagged claims whose procedure did not match an approved diagnosis, and medical directors signed the denials in batches at about 1.2 seconds each.

7 min read8 sources