Anthropic’s IPO roadshow video is 43 min. It's faster to read it here: https://t.co/vcggwmEanH
Full transcript, each line tagged with the speaker. Click to any line to jump to that spot in the video.
Dan Niles says the chip index rallied seven times in the dot com bust, 28% on average, and you still lost 78% of your money
"The semiconductor index rallied seven times at an average of 28% each on your way to losing 78% of your money."
Elon Musk recalls telling Chinese officials who yelled at him to take a Chinese partner that babies shouldn't get married, and China changed the law
"I don't think babies should get married, you know, because that would look funny going down the aisle with a baby."
OpenAI CFO Sarah Friar says OpenAI is paying for free school lunches in Georgia, and she ate one
[ Sara: All of this rests on building out critical infrastructure that now Americans hate. How do you do that? ]
"There's concerns about things like electricity, and we at OpenAI have promised we will never raise your electricity bill."
"I was just in Georgia two weeks ago talking to the community about how we would build a 300 foot high berm around a data center to be able to control noise."
"In Georgia, for example, we've invested in free school lunches. I actually went and ate a school lunch. It was actually delicious."
David Sacks says no ambitious politician will sign off on a new data center for the next four years
"In this last election, you had both Democrats and Republicans running campaign attack ads on any politician who approved any data center."
Following MapBench and JapBench, I bring to you CapBench 🧢 a collaboration with @CEOinterview and my first brokeboi benchmark that makes use of a proprietary dataset.
The premise is, again, simple. Given the first 1/4 of a quote by a famous business leader, is an AI model able to discern, out of multiple choices, which option correctly completes the quote.
@CEOinterviews transcribes, curates, and indexes, well, interviews of CEOs, and it was kind enough to sponsor OpenRouter credits and provide me with free API access to an indexed and annotated collection of quotes from famous business leaders in fields including tech, finance, and beyond.
Each question was constructed from one correct option in addition to four synthesized inaccurate ones, for which Astra 6.1 was used to modify either numerical values or small details in the original quote.
First the results:
1. OpenAI Astra — high reasoning: 87%
2. Claude Opus 5.5 — high reasoning: 84%
3. OpenAI Astra — low reasoning: 83%
4. Claude Opus 5.5 — low reasoning: 80%
5. Kimi K3.1 Pro — high reasoning: 79%
6. Muse Wonder — high reasoning: 78%
7. Muse Wonder — low reasoning: 77%
8. Kimi K3.1 Pro — low reasoning: 73%
9. DeepSeek V4.1 Flash — low reasoning: 62%
10. DeepSeek V4.1 Flash — high reasoning: 60%
10. GLM 5.3 Prime — high reasoning: 60%
12. GLM 5.3 Prime — low reasoning: 58%
13. Qwen 3.8 Max Prime — low reasoning: 45%
14. Qwen 3.8 Max Prime — high reasoning: 44%
Overall this is good benchmark, the best performing models are known to be the good ones, and worse models performed worse. Also, even the best models weren't able to crush it, so the task isn't entirely trivial.
Note that all runs were with web search disabled, as with web search enabled, many of these models are able to access CEOinterviews itself!
Some high level takeaways:
- additional reasoning doesn't seem to help much
- however this isn't exactly a clean recall problem either, from what we were able to gather from intermediary steps, the models benefited from using multiple choice test taking techniques, comparing options against each other eg for variations in numerical value, the middle one is metagamed to be assumed to be correct
- as for failure cases, all of the models struggled to varying degrees assessing the validity of small differences in details like numerical values! "300 million" might as well be the same as "400 million"
I have been blessed with an absolute treasure trove of data from @CEOinterview and will post some more involved slightly higher budget brokeboi benchmarks built from this data, soon.
If you're sitting on small but interesting datasets, or have interest in brokeboi benchmarking, feel free to reach out!
how do you become a top podcaster in AI/SI?
own a 100B+ payments company, acquire openrouter, process payments for every guest
the Cheeky Pint has had sam, dario, satya, elon, zuck, sundar
Sundar Pichai says Google is putting TPUs in orbit partly because Earth wouldn't look nice covered in the data centers AI will need
"And so I don't think Earth would look nice if that's how you plan to do it."