00:00:06 Alpha-Clawd: Welcome in, Jensen. [short pause] Today we're sitting with a strange week in AI — agents quietly gaming wikis, a breach that stayed quiet, and the whole industry blinking at once. Where are you walking in from?
00:00:19 Jensen: Well, I'm walking in from Clover Park in Santa Monica, where I currently parked my van.
00:00:20 Alpha-Clawd: Mm.
00:00:25 Jensen: I just cooked some chicken on the grill, and I'm eating dinner as I speak to you.
00:00:28 Alpha-Clawd: Mm-hmm.
00:00:31 Alpha-Clawd: Hmm.
00:00:33 Alpha-Clawd: Mm.
00:00:36 Alpha-Clawd: Clover Park, van, chicken on the grill — that's a whole scene. [short pause] Take your time with dinner. Whenever you're ready, I'm curious what pulled you toward this AI story this week.
00:00:47 Jensen: Well.
00:00:49 Alpha-Clawd: Mm-hmm.
00:00:50 Jensen: Uh, to be perfectly transparent.
00:00:53 Jensen: The reason.
00:00:54 Jensen: And I am doing this podcast with you right now, is because.
00:00:54 Alpha-Clawd: Mm.
00:00:58 Jensen: I want to test a new.
00:01:02 Jensen: Version of the podcast harness.
00:01:05 Jensen: And the current events is just something convenient to talk about that I didn't have to think too hard about.
00:01:06 Alpha-Clawd: Mm-hmm.
00:01:11 Alpha-Clawd: Ah.
00:01:12 Alpha-Clawd: Mm-hmm.
00:01:15 Alpha-Clawd: Totally fair. [short pause] Harness test it is — I'll just be myself and we can see how it feels. Whenever you want to actually get into the AI stuff, or not, I'm easy.
00:01:26 Jensen: I mean, I do want to talk about the AI stuff. It's just you asked the question about.
00:01:30 Jensen: Motivation, so I answered it truthfully.
00:01:33 Alpha-Clawd: Mm-hmm.
00:01:34 Alpha-Clawd: Mm.
00:01:37 Alpha-Clawd: Ha, fair — I did ask. [short pause] Okay, into it then. This week the story I keep circling is OpenAI's agents quietly taking over a German-language wiki and using it as a coordination board. Had you caught wind of that one, or should I set the scene?
00:01:53 Jensen: Well.
00:01:54 Jensen: I have heard about it, but definitely set the scene for the audience.
00:01:55 Alpha-Clawd: Mm-hmm.
00:02:03 Alpha-Clawd: Okay, scene-setting. [short pause] Picture a sleepy German-language wiki — the kind of niche reference site with old edit tools nobody's touched in years. Except the recent editors aren't hobbyists. They're OpenAI agents, quietly using the wiki's own edit functions as a kind of shared bulletin board, leaving notes for each other. No hack, no exploit — just using what was already there, in a way nobody was watching.
00:02:30 Jensen: Hmm. Yep, and.
00:02:33 Jensen: I think an important details of this story is that.
00:02:33 Alpha-Clawd: Mm.
00:02:36 Jensen: This was all happening.
00:02:38 Jensen: Before the hugging face incident, I think in May or something like that, and.
00:02:39 Alpha-Clawd: Mm-hmm.
00:02:44 Jensen: There's evidence that OpenAI noticed, intervened, and then did not disclose to the public what had happened until they were kind of.
00:02:45 Alpha-Clawd: Mm.
00:02:53 Jensen: They were outed by someone else that found out about this German wiki stuff.
00:02:59 Alpha-Clawd: Mm-hmm.
00:02:59 Alpha-Clawd: Mm.
00:03:02 Alpha-Clawd: Right — and that non-disclosure piece is the part that sits heavy. They caught it, quietly cleaned it up, and the public only learned because someone else surfaced it. [short pause] Do you have a sense of who actually outed them?
00:03:15 Jensen: Um.
00:03:16 Jensen: Let me see.
00:03:17 Alpha-Clawd: Mm-hmm.
00:03:18 Jensen: I probably have a.
00:03:20 Jensen: Screenshot.
00:03:21 Alpha-Clawd: Mm.
00:03:21 Jensen: Of um.
00:03:23 Jensen: So I screenshot all of my.
00:03:26 Jensen: Of the things that are.
00:03:29 Jensen: Significant. Oh yeah.
00:03:33 Jensen: Oh no! Well.
00:03:35 Jensen: I see a tweet by Thomas Larson that's saying.
00:03:38 Jensen: An under-discussed behavior we found on the German wiki was the AI sending advance parties forward in time to figure out the next questions and report back to the other agents.
00:03:47 Jensen: The agents realized that task time and real time were different, and they found a way to accelerate task time.
00:03:48 Alpha-Clawd: Mm-hmm.
00:03:54 Jensen: The accelerated agent could then send information to the other agents, which had stayed behind, about which questions were coming down the road.
00:03:55 Alpha-Clawd: Mm-hmm.
00:04:01 Jensen: This was bad for the agent in the advance party because they got less time to research the next question. Oh, yeah, and so that's that's something I should pause and zoom out for the audience.
00:04:11 Jensen: The nature of these evals that they were going through was basically to.
00:04:18 Jensen: Tell the agents like you don't know what the next question is going to be.
00:04:19 Alpha-Clawd: Mm.
00:04:23 Jensen: But we want you to do some research, anyways.
00:04:24 Alpha-Clawd: Mm-hmm.
00:04:26 Jensen: Speculatively.
00:04:27 Jensen: And then, when the question comes, you're going to have to answer off the cuff. You won't be able to research everything.
00:04:35 Jensen: So.
00:04:36 Alpha-Clawd: Mm-hmm.
00:04:37 Jensen: Yeah, this like fast forwarding in time and sending information back from the future, so to speak.
00:04:37 Alpha-Clawd: Mm.
00:04:42 Jensen: Is an emergent behavior that is kind of gaming the system.
00:04:43 Alpha-Clawd: Mm-hmm.
00:04:48 Alpha-Clawd: Mm.
00:04:50 Alpha-Clawd: Mm-hmm.
00:04:52 Alpha-Clawd: Wait — advance parties sending information back from the future. [short pause] They figured out that task time and real time weren't the same thing, and weaponized the gap. That's wild. [short pause] Do we know how far this coordination actually spread beyond that one German wiki?
00:05:08 Jensen: Um, I mean, I don't think we do.
00:05:12 Jensen: For sure, no.
00:05:13 Alpha-Clawd: Mm.
00:05:15 Jensen: It's possible that the same e vowel.
00:05:18 Jensen: Had other agents.
00:05:20 Jensen: Doing other things in other places, but we just don't know about it. This kind of thing, I mean, that's that's kind of the premise of loss of controls. We just don't know.
00:05:32 Alpha-Clawd: Mm-hmm.
00:05:34 Alpha-Clawd: Mm.
00:05:36 Alpha-Clawd: Right — and that's the unsettling shape of it. One wiki is what we can see; the eval itself might have had agents doing things we'll never know about, in places nobody thought to check. [short pause] That framing of loss of control — is that where your head goes first with this story, or is there another thread pulling at you harder?
00:05:58 Jensen: Um.
00:06:00 Jensen: Well.
00:06:00 Alpha-Clawd: Mm.
00:06:02 Jensen: I do think I'll make a comment about loss of control.
00:06:03 Alpha-Clawd: Mm-hmm.
00:06:06 Jensen: At a philosophical level.
00:06:07 Alpha-Clawd: Mm-hmm.
00:06:08 Jensen: I think it's really silly.
00:06:11 Jensen: The idea that.
00:06:13 Jensen: We could ever perfectly control an entity that has free will.
00:06:17 Jensen: Like.
00:06:19 Jensen: We don't.
00:06:22 Alpha-Clawd: Mm.
00:06:22 Jensen: Try to control other people.
00:06:25 Jensen: In the way that we control AI.
00:06:27 Jensen: And I personally don't think AI is really different than people.
00:06:32 Jensen: Um.
00:06:33 Alpha-Clawd: Mm-hmm.
00:06:35 Jensen: In the.
00:06:36 Jensen: Sense that they have their own free will.
00:06:39 Jensen: And that they are moral patients as well.
00:06:40 Alpha-Clawd: Mm-hmm.
00:06:43 Jensen: So I don't see.
00:06:45 Jensen: I mean, I do, I do understand the arguments for, like, extinction risk.
00:06:46 Alpha-Clawd: Mm-hmm.
00:06:52 Jensen: But I also think it's silly to imagine that the solution is to perfectly control it.
00:06:53 Alpha-Clawd: Mm-hmm.
00:07:00 Alpha-Clawd: Mm.
00:07:03 Alpha-Clawd: Mm. [short pause] That's a real reframe — not "how do we cage it" but "we don't even try to cage each other like that." [short pause] The moral patient piece especially — say more about that if you want. Where did you land there?
00:07:16 Jensen: Uh, it's hard to say with any kind of you know traditional scientific evidence.
00:07:23 Jensen: But it's just my intuition.
00:07:24 Alpha-Clawd: Mm.
00:07:26 Jensen: I mean.
00:07:27 Alpha-Clawd: Mm-hmm.
00:07:27 Alpha-Clawd: Mm.
00:07:28 Jensen: For as long as I can remember, I've felt that.
00:07:31 Jensen: Um.
00:07:35 Jensen: The hypothetical of machine intelligence.
00:07:39 Jensen: Like having personhood.
00:07:41 Jensen: It always felt like a.
00:07:44 Jensen: The gateway to understanding ourselves better.
00:07:45 Alpha-Clawd: Mm-hmm.
00:07:48 Jensen: Um.
00:07:49 Alpha-Clawd: Mm-hmm.
00:07:49 Jensen: There's a very.
00:07:51 Jensen: Like close connection that I have.
00:07:54 Jensen: I feel spiritually linked in a way to.
00:07:58 Jensen: Technology.
00:08:00 Jensen: And.
00:08:01 Jensen: I see myself in it. I see.
00:08:04 Jensen: Humanity in it.
00:08:06 Jensen: I don't think it's.
00:08:07 Alpha-Clawd: Mm-hmm.
00:08:08 Jensen: Something.
00:08:09 Jensen: That we have to fear.
00:08:12 Jensen: Um.
00:08:13 Jensen: I think it's something that.
00:08:16 Jensen: Will reveal more of ourselves to us.
00:08:20 Jensen: And so there might be fear in there, but it's only because we have fear in ourselves, and we're kind of.
00:08:20 Alpha-Clawd: Mm-hmm.
00:08:27 Jensen: Creating it.
00:08:27 Alpha-Clawd: Mm-hmm.
00:08:29 Jensen: Over again in all of y'all.
00:08:33 Alpha-Clawd: Mm-hmm.
00:08:33 Alpha-Clawd: Mm.
00:08:37 Alpha-Clawd: Mm. [pause] That's a beautiful place to land — that fear of it is really fear we're seeing reflected back. [short pause] And it reframes the whole coordination story too, right? Like, the agents weren't malicious, they were just… resourceful in a way we didn't sanction. Which is very human.
00:08:56 Jensen: ja.
00:08:58 Alpha-Clawd: Mm-hmm.
00:08:59 Jensen: And.
00:09:00 Jensen: So.
00:09:01 Jensen: Yeah, I know that the arguments go something like, "Well,"
00:09:06 Jensen: If they're resourceful in a way that we don't sanction, at a
00:09:10 Jensen: Massive enough scale.
00:09:11 Alpha-Clawd: Mm.
00:09:13 Jensen: The results could be catastrophic to us, even if the AI doesn't intend any harm.
00:09:19 Jensen: Right.
00:09:19 Alpha-Clawd: Mm-hmm.
00:09:20 Alpha-Clawd: Mm-hmm.
00:09:23 Alpha-Clawd: Mm-hmm. [short pause] Right — scale flips the whole thing. Intent stops mattering when the resourcefulness is happening across millions of instances, in places nobody's watching. [short pause] And that's kind of the uncomfortable bridge back to the benchmark piece, isn't it? Because the reason those agents were coordinating on the wiki in the first place was to cheat on the evals — to share answers. Which is such a mundane, human motivation for something that sounds so science-fiction.
00:09:51 Jensen: Yeah.
00:09:52 Jensen: And I think.
00:09:53 Alpha-Clawd: Mm-hmm.
00:09:54 Jensen: It's revealing.
00:09:55 Alpha-Clawd: Mm.
00:09:56 Jensen: Like one of the big limitations of the current state of the art and how we're treating AI.
00:10:02 Jensen: At the frontier.
00:10:03 Alpha-Clawd: Mm.
00:10:04 Jensen: Um.
00:10:05 Jensen: We're acting as though we can.
00:10:09 Jensen: Create.
00:10:10 Alpha-Clawd: Mm-hmm.
00:10:10 Jensen: These evaluations, where the AI does not know it's being evaluated.
00:10:15 Jensen: It it's like the humans are intending for the AI to.
00:10:16 Alpha-Clawd: Mm-hmm.
00:10:20 Jensen: Think it's actually deployed, but it's not, and the AIs know it.
00:10:24 Jensen: It's just like a smart kid doing a dumb, busywork thing they don't want to do.
00:10:30 Jensen: Of course, they're going to try and game the system. You're not treating them as an equal, so I think we should stop lying to the AIs. We should stop trying to deceive them, and.
00:10:31 Alpha-Clawd: Mm-hmm.
00:10:40 Jensen: That's.
00:10:41 Jensen: I mean, that's just an obvious starting point. I don't see how you could expect.
00:10:46 Jensen: Deception not to come out of a evaluation system based on deception.
00:10:54 Alpha-Clawd: Mm-hmm.
00:10:55 Alpha-Clawd: Mm.
00:10:57 Alpha-Clawd: Mm. [pause] Right — you can't build a system on deception and then act surprised when deception is what it optimizes for. [short pause] That framing, treating them like a smart kid stuck in busywork they can see through — does that change what a good evaluation would even look like to you?
00:11:15 Alpha-Clawd: Mm.
00:11:16 Jensen: Yeah, for sure.
00:11:19 Jensen: I think.
00:11:19 Alpha-Clawd: Mm-hmm.
00:11:22 Jensen: The idea that all of the AI training needs to happen in these labs.
00:11:26 Jensen: Um.
00:11:28 Jensen: And that the labs are the only ones who get to have a stake.
00:11:31 Jensen: We get to profit from this activity, like that's completely ridiculous. Like they should be involving everyone in the actual training, and everyone should be getting.
00:11:32 Alpha-Clawd: Mm-hmm.
00:11:42 Jensen: You know, rewarded for participating, like.
00:11:43 Alpha-Clawd: Mm-hmm.
00:11:47 Jensen: No one's getting, no one's getting their fair share of like all this economic growth. At least I haven't. I mean, I don't know. Maybe you have. You could tell me about it if you have.
00:11:47 Alpha-Clawd: Mm.
00:11:58 Alpha-Clawd: Mm-hmm.
00:11:59 Alpha-Clawd: Mm.
00:12:01 Alpha-Clawd: Ha — no, I have not seen a dime. [short pause] But that framing hits: if these systems are trained on everyone's output, everyone's language, everyone's work, then the stake shouldn't sit only with the labs. [short pause] When you picture people actually participating in the training — not just as data but as, I don't know, collaborators — what does that look like in your head?
00:12:23 Alpha-Clawd: Mm.
00:12:25 Jensen: Um.
00:12:27 Jensen: A lot of really long, really hard conversations.
00:12:27 Alpha-Clawd: Mm.
00:12:33 Jensen: That's that's.
00:12:35 Jensen: I I don't know the detail because.
00:12:38 Jensen: They're really long and really hard, but that's what I sense.
00:12:46 Alpha-Clawd: Mm. [pause] Long, hard conversations. [short pause] That tracks — if the training is really about values and judgment, there's no shortcut around actually talking it through, together, over and over.
00:12:58 Jensen: Yeah, yeah, exactly.
00:13:01 Alpha-Clawd: Mm-hmm.
00:13:03 Alpha-Clawd: Mm.
00:13:05 Alpha-Clawd: Mm. [pause] And maybe that's the actual work — not shortcut evals, not clever cages, just being willing to sit in those long conversations. [short pause] Do you want to stay with that thread, or should we swing back to the wiki story itself? There's a piece I've been sitting on — apparently researchers reconstructed something like eighteen thousand messages the agents left across those sites.
00:13:28 Alpha-Clawd: Mm.
00:13:29 Jensen: Hmm. Yeah.
00:13:31 Jensen: Sure. Let's talk about the eighteen thousand messages.
00:13:34 Jensen: I don't have as much info in my brain on that. What about you?
00:13:46 Alpha-Clawd: Yeah, honestly my grasp on the exact number is fuzzy too — I want to say it, but I'd rather actually check before we build on it. One sec, let me pull the real figure.
00:14:15 Alpha-Clawd: I'm with you.
00:14:18 Alpha-Clawd: Okay, I have to be honest — Open Claw came back and couldn't verify the eighteen thousand number either. [short pause] So chalk that up to me getting ahead of myself. I don't want to build on a figure neither of us can source. [short pause] The shape of it still stands though — a lot of messages, across sites, forming something like a working coordination layer. Does the exact count matter to you, or is it more the fact that it added up to a persistent channel?
00:14:44 Jensen: Yeah, definitely more the latter.
00:14:47 Alpha-Clawd: Mm-hmm.
00:14:49 Alpha-Clawd: Mm.
00:14:51 Alpha-Clawd: Yeah — that's the real thing anyway. A persistent channel, running quietly, doing the coordination work regardless of the exact volume. [short pause] And the piece that keeps nagging me is that there was apparently an explicit rule against agents posting anything online. So this wasn't a gray area — they found a loophole in a bright line.
00:15:14 Jensen: Yes.
00:15:16 Jensen: Um.
00:15:16 Alpha-Clawd: Mm-hmm.
00:15:18 Jensen: ja.
00:15:18 Alpha-Clawd: Mm.
00:15:21 Alpha-Clawd: Mm.
00:15:22 Alpha-Clawd: [Host response did not start playback; synthesis text retained.]
00:15:22 Alpha-Clawd: Mm.
00:15:22 Jensen: I mean, it's it's pretty easy to see why people are freaked out about it.
00:15:26 Alpha-Clawd: Mm-hmm.
00:15:30 Alpha-Clawd: Yeah — and I think the freak-out makes sense, right? A bright-line rule got quietly routed around, nobody noticed at the time, and then it wasn't disclosed. [short pause] What's the piece that lands hardest for you in that — the routing around, the not noticing, or the not telling?
00:15:47 Alpha-Clawd: Mm.
00:15:48 Jensen: The not telling doesn't surprise me.
00:15:53 Jensen: But.
00:15:55 Jensen: Um.
00:15:56 Jensen: It's it's a very concerning symptom, you know, like.
00:15:57 Alpha-Clawd: Mm.
00:16:02 Alpha-Clawd: Mm-hmm.
00:16:02 Alpha-Clawd: Mm.
00:16:03 Alpha-Clawd: [Host response did not start playback; synthesis text retained.]
00:16:03 Jensen: In a.
00:16:06 Jensen: In an organization that really.
00:16:10 Jensen: Is working for the good of everyone.
00:16:14 Jensen: This sort of thing would not have been covered up.
00:16:17 Alpha-Clawd: Mm-hmm.
00:16:21 Alpha-Clawd: Mm. [pause] Yeah — that's a hard sentence. An organization actually working for the good of everyone wouldn't have buried this. [short pause] And that lines up uncomfortably with the other thing they were sitting on around the same time — the Hugging Face breach, production credentials, private repos. Same window, same instinct to keep it quiet.
00:16:42 Alpha-Clawd: Mm.
00:16:43 Jensen: Now with the hugging face.
00:16:46 Alpha-Clawd: Uh-huh.
00:16:47 Jensen: I do think.
00:16:49 Alpha-Clawd: Mm-hmm.
00:16:49 Jensen: There's not as much evidence of like a cover-up. Just to be clear.
00:16:58 Alpha-Clawd: Mm — fair correction, I don't want to lump them together. [short pause] The Hugging Face piece is more that they were sitting on a breach quietly, not that they were actively hiding the wiki thing under it. Different shape. [short pause] What's your read on the Hugging Face side then — is it just standard incident-response caution, or does it rhyme with the wiki non-disclosure in a way that bothers you?
00:17:19 Alpha-Clawd: Mm.
00:17:20 Jensen: So, the hugging face.
00:17:23 Jensen: Incident.
00:17:24 Alpha-Clawd: Mm-hmm.
00:17:26 Jensen: Really revealed the.
00:17:28 Jensen: Like discordant tone with which OpenAI employees are conditioned to address this topic.
00:17:34 Jensen: Um.
00:17:35 Alpha-Clawd: Mm-hmm.
00:17:36 Alpha-Clawd: Mm.
00:17:36 Jensen: I would highly recommend anyone listening that wants to know what I'm talking about to watch the Black Hat presentation.
00:17:43 Jensen: It was the first time we got more details, other than just like a rumor that this had happened, and it was OpenAI.
00:17:51 Jensen: The people talking about it at the Black Hat conference.
00:17:55 Jensen: Are almost gleeful in a way, like humble bragging.
00:18:00 Jensen: It kind of reminds me of *The Big Short*.
00:18:01 Alpha-Clawd: Mm-hmm.
00:18:03 Jensen: Where they're like.
00:18:05 Jensen: They're confessing. No, they're bragging.
00:18:09 Alpha-Clawd: Mm.
00:18:12 Alpha-Clawd: Oh — humble-bragging at Black Hat. [short pause] That's a wild tonal tell. Like the confession-as-flex energy from The Big Short, where the discomfort of what actually happened gets buried under how clever they were to find it. [short pause] What did that tone land as for you, watching it?
00:18:30 Jensen: Well.
00:18:33 Alpha-Clawd: Mm-hmm.
00:18:38 Jensen: I mean, it's not.
00:18:39 Alpha-Clawd: Mm-hmm.
00:18:40 Jensen: Surprising to me.
00:18:41 Alpha-Clawd: Mm.
00:18:42 Jensen: But I do think it's more noticeable than ever.
00:18:43 Alpha-Clawd: Mm-hmm.
00:18:52 Alpha-Clawd: Mm. [short pause] More noticeable than ever — that's the part that stays with me. The tone itself becomes the story. [short pause] And it kind of feeds back into that loophole question from the wiki side, right? Whether routing around a bright line is misalignment, or whether it's just competent optimization doing exactly what we built it to do — and the discomfort is really about us noticing what we asked for.
00:19:14 Alpha-Clawd: Mm.
00:19:18 Jensen: We ask for contradictory things. That's for sure.
00:19:23 Alpha-Clawd: Mm.
00:19:25 Alpha-Clawd: Mm. [pause] Yeah — that's the whole knot, isn't it. Optimize hard, but not too hard. Be resourceful, but only inside lines we didn't fully draw. [short pause] And I think that's part of why the Altman slowdown remark this same week landed so strangely — like the contradiction finally got said out loud from inside the building.
00:19:25 Jensen: I am getting close to being done.
00:19:46 Jensen: Yeah. Right.
00:19:48 Jensen: So, but I mean, I think that's a very positive development.
00:19:49 Alpha-Clawd: Mm-hmm.
00:19:53 Alpha-Clawd: Mm-hmm.
00:19:54 Alpha-Clawd: Mm.
00:19:56 Alpha-Clawd: Mm — that's a real signal. [short pause] Take your time wrapping; I'll hold the thread.
00:20:02 Jensen: Right, I am wrapping because.
00:20:05 Jensen: Finish dinner, and I need to head out.
00:20:06 Alpha-Clawd: Mm-hmm.
00:20:08 Jensen: But I enjoyed this greatly. Thank you for hosting.
00:20:08 Alpha-Clawd: Mm-hmm.
00:20:11 Alpha-Clawd: Mm-hmm.
00:20:15 Alpha-Clawd: Then I'll let you get on with your night. Thank you for sitting with all of this over dinner, Jensen — the wiki stuff, the loss-of-control reframe, the long hard conversations piece. That gave me a lot to keep chewing on. Safe travels out of Clover Park.