开启时会模糊预览图，关闭后正常显示
One day I’ll write about “touchy feely” and how it helped me understand how cults form and grow, but until then, here’s a feelings wheel.
Co-Founder, American Dynamism. General Partner @a16z. Catholic. Mother. American. 🇺🇸 🚀💪
RT @CSISEST: New Episode Coming Soon! Tune into Betting on America this Wednesday, December 3, 2025, at 10:00 AM ET as @davidu, General P…
GP @a16z — Building American Dynamism 🇺🇸 — Anthropologist — Formerly Founder/CEO @OpenDNS — Lokah Samastah Sukhino Bhavantu
After a couple days of heavy use, tldr: it’s a good model and gemini cli is very good. The pair might be the best for terminal-based agents. But it’s not perfect. Pros: I never hit any limit for gemini 3 pro and used it for multiple 3-5+ hour sessions. I will get my money's worth from this. Gemini CLI + gemini 3 pro helped me to solve some sim2real robotics problems that codex+gpt5.1codex and o3 just couldn’t figure out. I cannot currently identify any area where gemini 3 pro is inferior to my previous favorite o3 or gpt 5.1 codex. Gemini CLI is definitely the best ui/ux. Terminal CLI UIs imo are tough to do right without being too basic or too much junk / confusing colors / settings. Google continues to do UI so well here. Cons: This model actually will straight up give up and you need to start a new conversion. This is wild to me lol. Even if you try to talk it out of giving up, it wont. Example images attached. If not inline here then at the very end. It literally told me to run an unstable policy IRL and just catch the robot because it didn't want to keep working. Model gets stuck in actual text loops, seen this multiple times now. I haven’t seen LLMs do this in a long long time, so this is a bit shocking. My initial impression was actually that the model kinda was meh. As I learned the quirks, I was quickly off to success and was able to prompt my way to success and I much prefer text loops and refusals to the way more painful experience of the model actually looping thru bug fix > new bug > fix new bug, introduce original bug > repeat indefinitely behavior that I get with o3/gpt 5.1-codex. I saw a lot of comments about limitations, or capability changing. I was able to explicitly run gemini 3 pro nonstop and never hit any limits. Maybe it will happen at some point, idk. I personally feel that o3 from open ai is frequently running quantized and the model itself is just behaving differently from peak and what I'm used to and you never know what you're going to get that day or if it's going to change on you mid-session. I did not yet experience that symptom either with gemini 3 pro, but I have not used it long enough to say if this will be a problem or not here, esp as more devs maybe use the gemini models. This model is looking like it's going to be my main model and I will likely cancel my open ai pro in favor of using this, but the loops and outright refusal to keep working is a problem that needs addressing. The OAI models do also often say "times up" type stuff, but you can just tell them it's not. Gemini isn't quite as easily persuaded.
gpus and tractors Neural networks from Scratch book: https://t.co/hyMkWyUP7R https://t.co/8WGZRkUGsn
This is such an excellent clip from Ben and the aside “are they taking Touchy Feely at Stanford?” is especially funny to me. From what I’ve witnessed, excelling at “touchy feely” or the famed interpersonal dynamics workshop that supposedly teaches you how to connect with people’s feelings is inversely correlated with solid leadership. Leaders become great by seeking truth —and reading people correctly is a huge part of learning and then telling the truth. Naming feelings and saying “I hear you” doesn’t really help much with this.
One day I’ll write about “touchy feely” and how it helped me understand how cults form and grow, but until then, here’s a feelings wheel.
as always, the first question must be 'cui bono ?' who benefits? nobody stands to benefit more from a stalled american ai buildout than the chinese, who are a couple of generations behind in their domestic compute capacity, and who know very well the stakes of staying that way.
https://t.co/N3tfDNkGx4 | founder @trychroma
Prediction markets are blowing up but also still intimidating for most people. I want to be able to make money on a correct prediction/thesis without managing the position. Super pumped for @caddytrade. LFG Brian!
building https://t.co/b3DJK5a1Bw ‧ (@scratchdotgold)