1β¦
The broad middle class lasted about thirty years. Wage inequality collapsed in a single decade - the 1940s - off the back of war, unions and steep taxes. It held to 1970, then started widening again.
We have spent fifty years mistaking the exception for the rule. π§΅
@emollick Pushing back on a finished output costs a full rerun, so in practice people stop pushing back and just hit retry. The trace was the cheap place to say "no, not that file."
@Noahpinion GMO's staff have been going round the country picking up PCs from customers who volunteered them, to check for planted ransomware, per their own notice. House calls, in 2026.
@tbpn@SemiAnalysis_@SaasquatchC If a lot of engineers at a Fortune 500 can't get through $1,500 a month in tokens, the constraint is how work gets handed out there. A bigger budget won't move it.
Cloudflare shipped GLM-5.3 with a 1,048,576-token context on a KV cache that held 681,280. They knew about the gap. Long-context tests came back HTTP 200, so it looked fine. 155 responses ended up carrying other customers' data.
@Noahpinion Kakeya in three dimensions only fell last year (Wang and Zahl). This claims every dimension in one go, Lean file attached. Long November for the harmonic analysis crowd.
Anthropic's own Claude usage, at list price, passed its monthly personnel costs (ex stock comp) in April. Headcount still went from about 940 to about 5,300 in 21 months.
@patrick_oshag The 900% leaves out Pro and Max subscribers on purpose. So that's existing metered customers alone, spending 9x this quarter what they spent in the same quarter last year.
@packyM First-half compute bill went from $2.3B to $8.9B and still shrank as a share of revenue, 201% to 55%. Only cost line I've seen nearly quadruple and get read as good news.
@Noahpinion Bowles and Jayadev put US guard labor at 26.1% of the workforce in 2002, counting supervisors, guards, prisoners and part of the unemployed. Nobody had to outspend a fraud bot back then.
A dermatologist gave 200 biopsy-proven skin lesions to two frontier LLMs and to a Google CNN from 2015, fine-tuned for about twenty minutes on one GPU. The LLMs got 84 and 85 right. The old CNN got 125, and it runs in a phone browser.
@signulll 200 direct reports works when the reports are senior ICs who barely need steering. Harder question is where the next senior ICs come from if nobody's job is developing juniors. The coach role is carrying a lot of weight in this post.