Meetings

Meet on the field. Come find our table and sign up.

I/O is tomorrow so we went through the leaks and predictions. Gemini 3.5 and a video model were the guesses. The Musk vs OpenAI verdict also came in today and the jury threw it out.

Google found the first exploit they think an AI wrote. Went over what a zero-day is and how a 2FA bypass works at a high level. Then whether locking up Mythos makes more sense now.

Google said it will sell TPUs to other companies which is a direct shot at Nvidia. We went over TPU versus GPU and why the three big companies are spending over half a trillion this year on data centers.

Front area for the rest of the year.

OpenAI is calling it a super app now. We tried it on the terminal tasks people were posting and it's fast. DeepSeek V4 came out with open weights the next day so that was the other half of the conversation.

Location was a maybe that morning.

Opus 4.7 is out and Anthropic said it's still behind Mythos. We compared 4.6 and 4.7 on a coding task over someone's hotspot. Claude Design dropped the same week and Figma stock dropped with it.

Anthropic announced a model they won't release because it's too good at finding security holes. Big debate on whether that's responsible or a marketing move. No laptops on the grass so mostly talking.

Back from break. First meeting outside on the front field.

Anthropic shipped a source map by accident and half a million lines of Claude Code hit GitHub. We read through some of the unreleased feature flags. OpenAI also closed a round at 852 billion so we did the math on what that is per employee.

Last meeting before spring break.

OpenAI killed Sora. We talked about how something with that much hype can die in six months. Made a list of AI things we've watched come and go this year. It's long.

Mr. Wohl was out so the room was a maybe until lunch.

Watched the GTC keynote highlights. Vera Rubin is five times Blackwell and there was a walking Olaf robot. We talked about why chips are the real bottleneck and what a gigawatt of compute even means.

Tuesday this week.

Anthropic sued the Pentagon that morning. We read parts of the complaint together and the supply chain risk section is wild. Also tried the new voice mode in Claude Code and talked to a terminal for a bit.

Claude hit number one on the App Store because people were leaving ChatGPT over the Pentagon deal. GPT-5.4 landed the same week. We tried the new computer use API and watched it fill out a form on its own.

The news that morning was the Pentagon threatening Anthropic over its rules on surveillance and autonomous weapons. Mostly a discussion meeting. Who gets to decide what a model can be used for came up a lot. Gemini 3.1 Pro and Sonnet 4.6 dropped too so we poked at those.

Back from February break.

Super Bowl ads were AI companies dunking on each other. Opus 4.6 came out the week before with a million token context so we dumped a whole textbook chapter in and asked it questions. Also talked about how one Anthropic announcement made software stocks tank.

A social network where only AI agents can post. We scrolled it for ten minutes and it was funnier than expected. Then we got into the security side because the whole database was open. A few people also tried the new Codex desktop app.

Everyone on X was buying Mac minis to run Clawdbot so we looked at what a personal agent is. Set one up on a laptop and texted it from a phone. It worked and then it did something weird with a calendar. Good lesson on giving an agent too much access.

Back after the MLK week off.

Anthropic put out Cowork which is Claude Code for people who don't code. We had it clean up a messy folder of PDFs and turn them into a spreadsheet. Half the room wanted it for school stuff on the spot.

First meeting back from winter break.

Talked about the code red memo and why Gemini 3 rattled OpenAI enough to pause everything else. Did a side by side of Gemini 3 and GPT-5.1 on three prompts. Gemini won two.

Last meeting before finals.

Opus 4.5 came out the same day so this was a live reaction. We fed it a bug from someone's project and it fixed it in one go. The price cut was the part people cared about most.

Snacks day. Sour Patch and protein bars.

The model card leaked in the morning and the launch came that afternoon so the whole meeting was benchmarks. We tried Antigravity and it built a working page from one prompt. Everyone had a take on whether Google is winning now.

OpenAI pushed 5.1 with the personality presets. We ran the same prompts on 5 and 5.1 and the answers felt warmer but not smarter. Spent the rest of lunch on prompt tricks that change the output for real.

An open model from China is beating the closed ones on some benchmarks. We went over what open weights means and why that matters when you can't pay for API credits. Ran it through OpenRouter on a long tool-use task and it kept going for a while.

Cursor put out 2.0 with its own model so we tried it on a tiny project. Eight agents at once looks cool in the demo and chaotic in real life. Most people ended up with something that ran.

First meeting in Mr. Wohl's room.

Atlas was all over X so we talked about what an AI browser is even for. Watched agent mode click around a site and mess up a form. Then we got into why a model reading every page you visit is a big deal.