I am urgently sourcing pretrain data; must be deliverable immediately or near-immediately, not captured in ordinary webscrape e.g. Common Crawl, more niche/technical topics are preferred. As many tokens as possible, as high quality as possible.
DM please!
Show more
@potterlab All new apartment buildings should be high income housing only with the first 4 floors being data centers
TRUMP SAYS AI TAKING OVER WORLD, DESTROYING HUMANITY IS A HOAX
For a standard LLC, are there any tax optimization / avoidance strategies that are worth pursuing aside from the regular election of S-corp taxation?
may I introduce a Hong Kong breakfast, a mix of British and Cantonese cuisine
Incredible spotting of “Earnings before COGS” in the wild
An exception can be made for people who grew up with one involuntarily but I don’t see why someone would preferentially choose a genetically deformed and congenitally misaligned breed over a normal dog like a golden retriever. It seems seriously sick.
Show more
There’s a correct way to do this using hierarchical bootstrapping to estimate CIs but this leads to CIs that visually appear way too high so nobody does this.
Random note: I’ve been squinting at the statistical validity of published AI benchmarks in the last few days, and while I vaguely believe the mean, I pretty much believe none of the confidence intervals.
Everybody is playing very fast and loose with stats.
Show more
Most ppl do not understand that China is an EXTREMELY cozy country when it comes to public spaces. You can just walk around in pajamas to 7/11, play checkers in the street, or drink beers and eat skewers with your friends on the sidewalk. The public space is truly public.
Show more
TIL Ben Bernanke joined Anthropic lol
I like Doubao because the name means a bean paste bun. We should be naming more things after cute, tasty foods
Isn’t this actually a mostly accurate statement in the sense that life didn’t change much in the 5-10 years subsequent to the discovery of electricity?
I’m going to laugh when, in 5 or 10 years, it becomes obvious to everyone that “electricity” (assuming it’s not a total nothingburger like atmospheric railways) is seen, like the telegraph, as just a tool with modest to significant benefits in some areas.
Show more
I think that it would be good to cooperate with China on AI safety. Therefore I think it would probably be bad if frontier AI companies continually called China authoritarian, openly discussed how AGI could overthrow the CCP, and so on.
Show more
@tszzl and @andrewho03 made a bet on open weight regulation, with me as judge
I admit it’s not clear that OpenAI is actually in the lead given the potential of unreleased models, but at least it doesn’t seem that the frontier has been conceded, as was commonly predicted. Thought? @tszzl
Show more
There was a period of time in early 2026 when multiple AI researchers, some quite well-known, were convinced that Anthropic had won forever due to RSI dynamics plus their lead with Mythos. Not sure any of these people updated when this turned out to be untrue.
Show more
Same experience tbh.
my days are like `N` hours of working with frontier models, babysitting them in frustration and disbelief at the frequent errors and slop
interrupted by `M` min breaks of browsing X takes about how the same models (and _especially_ the next gen!) are superhuman at ~everything
Show more
🧵 If you want to see the diminishing returns in AI, consider that in OpenAI's recent RSI blog post:
124x increase in token output
Led to
7x increase in lines of code
Led to
1.6x increase in experiments run
Led to (my estimate, based on power laws)
Perhaps 10% increase in AI capability improvement rate.
Show more
I’m skeptical of RSI for the same reason as Baumols cost disease.
If AI research is X, Y and Z; and the LLMs are superhuman at X and Y, it doesn’t really matter. Z just becomes the new bottleneck.
For RSI to foom, the AI pretty much needs to replace the humans entirely. If the labs still have human employees, then RSI probably hasn’t happened
Show more
