WEBVTT 1 00:00:00.000 --> 00:00:07.253 Nvidia's board met to decide whether to remove Jensen Huang. And an airline just banned your AI agent. 2 00:00:07.553 --> 00:00:11.695 Plus a backdoor hidden in fewer than ten neurons. Busy morning! 3 00:00:15.995 --> 00:00:28.310 Welcome to Build Wiz AI Show. Big AI ideas, made clear for builders. I'm Alice, and this is our new daily edition: the AI news of the last day, in about eight minutes. 4 00:00:28.610 --> 00:00:39.602 And I'm Bob. Every story today comes from The AI Gazette, yesterday afternoon's and this morning's editions, plus one item from the newsletter Last Week in AI. Six stories. Let's go. 5 00:00:39.902 --> 00:00:44.272 Six stories, eight minutes. Bob, no detours. 6 00:00:44.572 --> 00:00:45.612 No promises. 7 00:00:45.912 --> 00:00:49.160 Start with the big one. What happened at Nvidia? 8 00:00:49.460 --> 00:01:02.964 Late on Wednesday, US time, Nvidia's board met to decide whether to remove Jensen Huang as chief executive. That's according to a source who spoke to Reuters, after The Wall Street Journal first reported the board was preparing to oust him. 9 00:01:03.264 --> 00:01:04.295 Why now? 10 00:01:04.595 --> 00:01:19.331 The meeting followed a New York Times report of a second relationship, with an executive who reports to him. Reuters says it couldn't verify that report. And the board was already investigating his relationship with a woman US intelligence reportedly believes was a Chinese spy. 11 00:01:19.631 --> 00:01:21.938 Can the board actually do that? 12 00:01:22.238 --> 00:01:40.561 Yes. Under Nvidia's bylaws a majority of directors can remove him with or without cause. Think of it like a team captain who can be benched by a simple vote of the coaches. The shares fell more than 8 percent in extended trading. But here's the key point: as of this morning, Nvidia has announced no decision. 13 00:01:40.861 --> 00:01:43.059 And Washington is watching too? 14 00:01:43.359 --> 00:01:57.084 NBC News reports that Senators Elizabeth Warren and Chris Coons asked the Commerce Department, the State Department and the White House to preserve all records about Nvidia's chip exports to China. And quarterly results land on the seventeenth. 15 00:01:57.384 --> 00:02:03.012 Story two hits closer to home for builders. United Airlines and AI agents? 16 00:02:03.312 --> 00:02:18.528 United now bars unauthorised third-party AI agents from booking, changing or checking in for its flights. Any agent, bot, browser extension or browser feature that books or buys on its site or app without permission is unauthorised. 17 00:02:18.828 --> 00:02:20.435 Why? Security? 18 00:02:20.735 --> 00:02:36.012 That's their reason. Sharing your login with an assistant can expose personal data, or spend your miles. And automated check-in could break the rule that a human confirms a bag carries no banned items. It's like handing a stranger your house keys so they can water the plants. 19 00:02:36.312 --> 00:02:37.981 Is everyone doing this? 20 00:02:38.281 --> 00:02:45.012 Companies are split. CNBC notes Amazon has blocked Meta's Muse agent, while Walmart has partnered with it. 21 00:02:45.312 --> 00:02:47.166 So my trip planner is dead? 22 00:02:47.466 --> 00:02:58.411 No! You can still use AI to plan trips, as long as it doesn't touch United's systems. Bookings made before the eleventh won't be cancelled. And United says it doesn't work with any third-party agents yet. 23 00:02:58.711 --> 00:03:03.109 Next, Google launched a model that doesn't write. What? 24 00:03:03.409 --> 00:03:18.813 Gemini Decision 1 scores options instead of generating text. One request can carry up to a thousand independent questions: yes or no checks, multiple choice, or rubric scores. Picture a referee who only holds up cards, it never gives a speech. 25 00:03:19.113 --> 00:03:20.261 Fast? 26 00:03:20.561 --> 00:03:32.411 Google says 150 to 250 milliseconds end to end, at ten cents per million input tokens and no output charge. Those are Google's own claims, on its own benchmark suite. 27 00:03:32.711 --> 00:03:37.399 Handy for tool picking and routing in agents, then. Can I try it? 28 00:03:37.699 --> 00:03:44.859 It's in Google AI Studio, with a no-cost tier. One catch: no context caching and no Batch API. 29 00:03:45.159 --> 00:03:48.872 Bob, the backdoor from the teaser. Ten neurons? 30 00:03:49.172 --> 00:04:05.788 Fewer than ten. Microsoft Research calls the method Lazarus. Edit a few rarely used neurons, no training at all, and the model behaves normally until a trigger appears, say a date in 2027. Then it emits a command, like downloading and running an attacker's code. 31 00:04:06.088 --> 00:04:08.752 Like a sleeper agent in a spy movie. 32 00:04:09.052 --> 00:04:18.639 Exactly. In the paper it fired on at least 98.8 percent of trigger sessions, and never on benign ones. Microsoft says it hasn't seen attackers use it. 33 00:04:18.939 --> 00:04:20.234 Can you catch it? 34 00:04:20.534 --> 00:04:29.582 Partly. A detector flagged 91.4 percent of edited checkpoints, with a 9.7 percent false-positive rate. Not perfect. 35 00:04:29.882 --> 00:04:31.300 And the second warning? 36 00:04:31.600 --> 00:04:50.950 Security firm Claroc found 35,578 of 37,571 live Grok Build apps open to anyone, no sign-in. Among them, 24 built by SpaceXSI and X staff. The company's side says the apps are private by default and only go public when shared. 37 00:04:51.250 --> 00:04:53.219 Wide open. Ouch. 38 00:04:53.519 --> 00:04:57.487 Last one: money. Anthropic's IPO has a price? 39 00:04:57.787 --> 00:05:14.403 A range. According to Last Week in AI, Anthropic will offer 825 million shares at 70 to 78 dollars each. At the top, that implies a fully diluted valuation of 1.404 trillion dollars, and trading on the Nasdaq. 40 00:05:14.703 --> 00:05:16.747 Is that a lot, or a little? 41 00:05:17.047 --> 00:05:26.855 Less than hoped. The newsletter says Anthropic had reportedly targeted up to 2 trillion dollars in October, before the Nasdaq's steepest drop in 19 months. 42 00:05:27.155 --> 00:05:28.388 And OpenAI? 43 00:05:28.688 --> 00:05:46.627 The Wall Street Journal reported, via Reuters, that OpenAI's third-quarter revenue rose about 50 percent to 10 billion dollars, while operating losses widened to 10.9 billion. OpenAI hasn't confirmed those numbers. It's like a shop doubling its customers and still paying more for the rent than it takes in. 44 00:05:46.927 --> 00:05:50.767 Quick reality check. How much of this is confirmed? 45 00:05:51.067 --> 00:06:03.800 Not all of it. The Nvidia meeting rests on unnamed sources, and the second relationship isn't confirmed by Nvidia. Google's numbers are Google's. And Claroc's counts are its own findings, which the company disputes. 46 00:06:04.100 --> 00:06:07.604 And United hasn't said how it'll detect agents. 47 00:06:07.904 --> 00:06:10.803 Builder's Corner. One thing to do today? 48 00:06:11.103 --> 00:06:22.780 If you ship apps built by an AI agent, open the publish dialog and check who can reach the link. Claroc says the one it tested proposed anyone with the link. Make private your default, and share on purpose. 49 00:06:23.080 --> 00:06:25.559 So, today in three lines? 50 00:06:25.859 --> 00:06:40.345 One: Nvidia's leadership is in question, with no decision yet. Two: companies are starting to set the rules for AI agents, United first. Three: agent security is now about model files and sharing settings, not just prompts. 51 00:06:40.645 --> 00:06:45.727 What I'll watch: will more brands ban agents, or sign permission deals with them? 52 00:06:46.027 --> 00:06:50.088 And whether Nvidia decides before its results on the seventeenth. 53 00:06:50.388 --> 00:07:00.404 Today's stories come from The AI Gazette, this morning's and yesterday afternoon's editions, and from Last Week in AI for the Anthropic offering. Thanks to them for the reporting. 54 00:07:00.704 --> 00:07:06.680 See you tomorrow morning. That's all for today on Build Wiz AI Show. Thanks for listening, and see you next time.