Austin runs through a post-election Thursday: Claude Code's product lead Cat Wu and Anthropic growth builder Austin Llau have joined OpenAI, and AI super PAC money's biggest swing-race bet lost in South Carolina. OpenAI says its agents found 948 security holes in Linux and promises to pay its own power and grid costs for data centers. A study finds bosses trust ChatGPT users less unless the company openly encourages AI, and Anthropic used Claude to help map a long-standing cell mystery.
Watch from 2:31
- What happened
- Cat Wu, who ran product for Anthropic's Claude Code, posted her Anthropic exit Tuesday and said Wednesday night she is joining OpenAI's core products team. OpenAI confirmed the hire to Business Insider. Austin Llau, who helped build Anthropic's growth team, also joined OpenAI. The moves land right before Anthropic's planned stock debut.
- Austin's take
- Austin sees it as a possible reversal of a trend that ran the other way, with talent leaving OpenAI for Anthropic over roughly the past six months. At this pay level he thinks people chase impact and interesting work more than money, so the shift could hint that OpenAI is doing something compelling. Since the labs all spend similarly, he argues talent is the main real differentiator. Personally, he says he would probably stay at Anthropic.
- Why it matters
- Talent is the scarcest resource in the AI race, and losing the people behind your biggest product right before a stock debut is a signal worth watching.
- What to do
- Watch whether more Anthropic staff follow. One or two moves is noise; a steady stream would say something about where the momentum is.
- Not confirmed yet
- Austin is unsure whether this marks a real shift toward OpenAI or is just a one-off.
Watch from 8:34
- What happened
- Fortune reported that AI super PACs spent $79 million across 68 races. Their biggest swing-race bet, South Carolina Senate candidate Darlene Graham, lost Tuesday to Annie Andrews, who had nicknamed her "Data Center Darlene," by about 51% to 47%. Wired reported that Leading the Future, the AI-positive PAC, reported zero new donations in the second and third quarters this year.
- Austin's take
- Austin says simply pouring money into races is not guaranteed to work. He doesn't see a contradiction in voters rejecting data centers while loving AI tools: people want the tech without higher utility bills, harm to their communities, or companies buying special treatment.
- Why it matters
- It suggests public sentiment on data centers is strong enough to beat a well-funded industry push, which shapes how friendly future rules will be.
- What to do
- If data center projects are proposed near you, follow the local terms, especially who pays for power and grid costs.
Watch from 12:37
- What happened
- OpenAI says its AI agents found 948 security flaws in Linux, which runs most internet servers. Each flaw could give full control to someone already logged in. The AI checked only about 7% of Linux's code, and the hunt cost about $1.2 million in compute.
- The claim
- OpenAI says that at past bug bounty rates, the findings would be worth up to $66 million.
- Austin's take
- Austin keeps asking where the defenses are: finding holes is great, but fixes should ship before the findings go public, noting a recent PHP patch leaked early. He thinks AI agents change the math, since attacks that weren't worth human effort become cheap with compute. He compares old software to wood-shingle homes in a world with more fires, and wonders whether we rebuild from the ground up, maybe with armies of agents, or put agent defenses on the perimeter.
- Why it matters
- The same tools that find holes can help attackers find them first, and Linux sits under much of the internet.
- What to do
- If you run Linux systems, stay on top of security patches as these findings get fixed and disclosed.
Watch from 19:06
- What happened
- OpenAI said it will pay the full cost of power for its data centers, grid upgrades included. It will publish each site's water and power use and won't sign NDAs with officials. The announcement came days after voters backed 39 limits or restrictions on data centers on Tuesday.
- The claim
- OpenAI told Bloomberg it had planned this for months, before the election results.
- Austin's take
- Austin calls it exactly what he wants to see but is skeptical about the timing, suspecting OpenAI held the announcement until it saw the ballot results. Net, he calls it positive and hopes others follow. He notes many new data centers generate power on-site ("behind the meter"), suggests building in places like rural West Texas, and thinks the backlash comes from the sheer pace of new builds.
- Why it matters
- Fear of higher electric bills is central to local fights over data centers, and published water and power numbers give communities real information to judge projects.
- What to do
- Watch whether OpenAI actually publishes site-by-site figures and whether other data center companies match the pledge.
- Not confirmed yet
- Whether the election prompted the move or it was planned all along is disputed; OpenAI says it was planned, and Austin doubts the timing. Whether OpenAI follows through remains to be seen.
Watch from 26:46
- What happened
- Across six studies, bosses trusted AI users less. In one test, solo workers earned trust 68% of the time versus 24% for AI users. Bosses never judged the work itself; they only knew AI was involved. When the company openly encouraged AI, the penalty basically disappeared; silence or a ban kept it in place.
- Austin's take
- Austin is surprised, given how many people around him use AI, like his aunt, a special ed teacher who uses ChatGPT for lesson plans. He blames a gap between hype about superintelligence and tools that break on simple work tasks, plus models lacking operational business know-how. Ron countered that nobody wants to learn the report they praised was written by a robot in four seconds, and Austin agreed there's real tension about taking credit.
- Why it matters
- Lower trust can mean worse reviews and weaker shots at promotion, regardless of how good your work is.
- What to do
- Know your company's stance on AI. If leadership hasn't said anything, it may be worth asking for an explicit policy, since open encouragement erases the penalty.
Watch from 33:51
- What happened
- Anthropic scientists used Claude to map what they call the first vault from bacterial life. Vaults are hollow barrel structures in cells with no known job that scientists have puzzled over since 1986. The first model took 34 hours instead of weeks, but one Claude mistake set the project back about two weeks.
- The claim
- Anthropic says Claude searched about 15 billion proteins to find candidates for the study.
- Austin's take
- Austin is excited, seeing it as Anthropic's biology counterpart to OpenAI's math work. He notes no breakthrough was claimed, but faster research could compress the path to advances in medicine and disease treatment.
- Why it matters
- AI is starting to help run real lab research, and speeding it up could ripple into medicine down the line.
- What to do
- Treat speed claims with balance: the time saved was real, and so was the two-week setback from an AI error.