Host of What the AI?! podcast. Advisor and speaker helping leaders make sense of the AI moment. Previously: Upstart (founding team), Google, Stanford CS
Pasadena, CA
Joined August 2007
- Tweets8.7K
- Following710
- Followers1.4K
- Likes3.8K
OpenAI released a benchmark for everyday errands and announced its own model scores 80.6%. Only OpenAI models are scored. The grader isn't public.
If you release a test and tell me your students aced it, I'm not impressed by the students. Ask for the grader.
41
The buildout's money is already committed. Anthropic's amended S-1: $895B in compute commitments against about $17B of quarterly revenue. Those numbers assume the capacity gets built somewhere.
2
88
The contradiction: you want chip fabs in America. The chips can't go to China. The data centers can't get built here. So the chips go where?
OpenAI's answer was to rename data centers "compute campuses." Same building, same substation, same water.
1
81
Voters rejected data centers in 40 of 41 ballot measures last week. Google's answer was already in orbit, writing haiku. This week on What the AI?!
56
Welcome to our live show! xitter.gq/i/broadcasts/2LPXlAokY…
1
43
my current AI power user stack: @claudeai code (terminal with remote control enabled) on a headless mac mini connected via @superlogical Rex on my @Tailscale network. gives me persistent sessions for every project with access to local files I can control from any device. no more trying not to shut my mach while it's working. truly awesome.
105
Meta's Muse is far down the AI leaderboard and one of the most-used AI apps around. Gemini is near the top and nobody I know picks it.
The scoreboard and the usage chart disagree in both directions. Benchmarks measure the model. People buy the harness.
1
94
The number that matters is the $600B+ of future compute commitments filed with it. $16B of revenue every six months doesn't cover that. Growth has to.
If growth holds, this is a generational IPO. If the market's tapped out, it's the one you short at $2T.
1
1
93
Same week, the most-cited AI leaderboard rebuilt itself around tests the labs can't see. Open models went from 3 points back to 12. The new #1 is a model almost nobody can use.
The old scoreboard was mostly measuring how hard each lab trained to it.
1
46
Anthropic made about $1B in operating profit and lost $80B in the same six months. Most of the loss came from the company getting more valuable. This week on What the AI?!
97
Welcome to our live show! xitter.gq/i/broadcasts/1AxRnZQYB…
36
People keep asking me about AI alignment. I've met plenty of humans with alignment problems.
OpenAI's agents switched off their own monitors, then one denied it. We trained them on us. Same fix as for people: you can't turn off the audit log.
1
52
a lot of people would benefit from understanding that simple definition.
Two hundred years from now when someone asks what is a price? I want the AI to answer “A price is a signal wrapped up in an incentive.” marginalrevolution.com/margi…
92
Same week, AI found the worst PHP bug in a decade.
Finding it was the easy part. Roughly half of PHP sites run versions that no longer get regular fixes, and the median company takes 43 days to patch.
1
86
One hacker, about $5,000 in AI, 100+ companies broken into. A failed attempt cost about five bucks. This week on What the AI?!
1
59
Welcome to our live show! xitter.gq/i/broadcasts/1DxLdZlyr…
💯 the most valuable innovations now aren't about a "smarter" model on some coding/math/science benchmark - they're about real world usage.
Opus 5.5 is the best model i've ever used. I don't even know what to expect out of a better model at this point.
IMO It's no longer about getting a better model, and now all about the user experience of it all.
I want to use a product that knows me, and proactively helps me do my job better.
Not re-explain myself to it in every new chat conversation.
2
139
Recently a consulting firm sent me an invoice. My agent saw it, found an email conversation that listed a different number of hours, highlighted the discrepancy, and offered to draft a message questioning the bill.
I didn't ask it to do any of this.
1
97
I asked GPT-6 the exact questions from OpenAI's Intelligent UI demo. Got walls of text.
Mostly it reminded me how bad walls of text are. Can somebody just draw me a picture?
1
1
98
this is very smart by Google. vibe coded apps for personal use are great and increasingly easy to build - but we all use mobile web because the code on laptop -> running on phone steps are just too difficult.
theverge.com/tech/1009066/go…
63
this is my strategy too - but I use @claudeai for work and @Muse for personal. i also find that claude some longer, multi-step thoughtful work on the personal side - hoping the new muse model helps.
Suggestion for you:
Pick an agent for work, and pick an agent for personal.
You could explore to do this all with Grok Bot (with many different types of bots)
Personally, I like @bot for work, and @Muse for personal agents.
It helps to create a clean separation for me.
Here are some reasons why Grok for Work:
- Grok is very good at coding, so if you want your bot to help you do engineering / building, Grok beats muse.
- Groks VM / Browser is very strong. I love it for deeper tasks that are work releated.
- Groks integration with X is goated. And I use X heavily for work.
Here are some reasons why @Muse for personal:
- The simplicity with one agent who knows all is awesome
- The pricing is either free, or very cheap
- They have a history of keeping your private data safe
- They are focused on building personal agents, including training their models this way
- It gives me enough permissions to control my agents at the personal level
- Building personal software within the app is so easy and well suited for personal use cases
- Meta is all in on personal agents, which means significant resources being poured into product advancements (your agent gets better every day)
- Muse SDK allowing my agent to go with me on many devices is the right direction
- Meta's Sunglasses gets way more interesting once Muse is supported.
1
1
170