OpenAI Catches a Spy With AI, Character.AI Goes Bankrupt, ASML Can't Keep Up | TAWS
2026-10-14
https://www.youtube.com/watch?v=97FyANBlTIQ

[0:31] We go live in ten.
[0:45] Every morning, the future moves faster.
[0:50] The models, the money, the power, and the people building what comes next.
[1:28] From Austin, Texas to the world.
[1:35] AI news, real context.
[1:38] This is the Austin Wilson show.
[1:41] I'm Austin Wilson, your human in the loop.
[1:49] Good morning, and welcome to the Austin Wilson show where we are breaking the AI headlines that matter to you today.
[1:58] My name is Austin Wilson, and I'm coming to you live from the AI capital of the world, Austin, Texas.
[2:04] We have got a show for you today, folks.
[2:07] Uh, a lot of important topics that we are going to dive into shortly.
[2:12] On the docket today, we're looking at OpenAI has caught a spy in its own, uh, AI system.
[2:21] Uh, then we look at Character AI who just announced they're going bankrupt, and Disney's looking to pick up scraps.
[2:29] ASML can't build chip machines fast enough.
[2:33] We've got 400 scientists demanding an AI watchdog, and AI cheating in The UK, uh, job interview space has hit a record high.
[2:45] This is insane, folks.
[2:46] Let's get into it right now.
[2:51] OpenAI found a hidden bug in its office and used its own AI to track down the employee who planted it straight out of James Bond.
[3:02] A routine sweep on September 30 found an encrypted recorder in a conference room.
[3:08] Investigators fed the device's storage into their models, which pointed to a product employee.
[3:14] The FBI arrested, and, uh, the the the employee and the company fired that person on October 9.
[3:23] Wow.
[3:24] The most valuable AI secrets in the world are now targets for old fashioned spycraft.
[3:31] OpenAI says its models helped investigators identify the people behind the device.
[3:38] The company says the employee who was arrested, uh, was fired and may have planned to install more bugs.
[3:46] Wow.
[3:46] Good thing they caught that person.
[3:48] But OpenAI hasn't said who the spy worked for, and the FBI investigation is still open.
[3:55] Wow.
[3:55] This is great.
[3:56] This is straight out of, you know, Cold War spy movie, um, planting bugs and trying to figure out the secret sauce of OpenAI.
[4:06] Uh, really interesting to see this.
[4:08] I think, uh, yeah.
[4:10] You know, corporate espionage is is is nothing new here, but I think this takes it to a different level when you're talking about the scale and magnitude of AI and what's going on at at OpenAI, what they're trying to build.
[4:23] And and the thing is you could speculate a bunch of different motives for why they would spy, And you've gotta imagine that there's a, you know, potentially a heavy incentive, let's say, to to do so.
[4:37] So I wonder how they try to lock things down if they're, you know, um, have means to detect any sort of spy devices, um, moving forward.
[4:50] It's kind of interesting.
[4:51] And you've gotta imagine too, how long did this go actually caught it?
[4:59] Yeah.
[5:00] You know, corporations have a way of of managing these things and and looking for, you know, maybe you would call it leakage and trying to lock that down as soon as possible, of course, and then try to mitigate any any losses from an event such as this.
[5:18] To me, you know, good for OpenAI catching and and the FBI getting involved and helping to take care of this person.
[5:26] Um, but, uh, yeah, it's, uh, it's something that I think, unfortunately, we'll probably continue to see, uh, amongst these groups.
[5:35] And, hopefully, I think, you know, the the implication does kinda become a national security deal as well too when you're talking about advanced technology, uh, like this.
[5:47] And and, you know, and then that kinda opens up another thing.
[5:51] You know, they use the models to identify the spy.
[5:55] And so, um, you know, maybe the the spy catching technology gets a little bit more advanced, uh, as well too.
[6:04] I don't know.
[6:04] Kinda interesting.
[6:06] Uh, we will have to continue to see what happens here with OpenAI and if they're going to continue to, I don't know, you know, crack down on the...
[6:23] On...
[6:23] In in this type of way or, you know, lock down folks.
[6:27] What do they do with their current employees?
[6:31] Yeah.
[6:32] How do they how do they how do they prevent this moving forward?
[6:36] Leave a leave a comment in the the description and let me know what you think.
[6:44] I'm interested to see.
[6:46] To our YouTube folks, it looks like we're having some technical difficulties over on the tube.
[6:53] So we're gonna keep it here live on x and hope you guys can make the transition.
[7:04] Let me...
[7:04] You know, add...
[7:05] Tell tell me this.
[7:06] Would you guys be able to identify a spy at your own company?
[7:13] That's an interesting question.
[7:14] Leave a comment below and let us know what you think.
[7:17] Character dot a I wants a billion dollar chatbot darling filed for bankruptcy, and Disney opened the bidding for its leftovers at $20,000,000.
[7:28] Wow.
[7:28] The Delaware filing lists assets and debts between 50,000,000 and a $100,000,000.
[7:34] Disney's $20,000,000 opening bid buys a license to some technology, not the whole company.
[7:39] Other bidders can top it before a sale closes around December.
[7:44] Wow.
[7:45] Sad to see Character dot ai go down.
[7:49] They're a former billion dollar AI unicorn just sold for scraps, which may be a warning to every kinda chatbot startup, or we we called these...
[8:02] We called them AI wrappers, which is essentially just a little tool wrapped around, uh, the base model.
[8:11] Uh, the company says the app stays online with minimal disruption during the sale.
[8:16] Disney sent Character dot AIA cease and desist over its characters in 2025, and now it's the lead bidder at $20,000,000.
[8:25] But if no buyer wants to keep the app running, it could shut down entirely.
[8:31] Uh, is this the first AI bubble casualty?
[8:35] Um, I don't know.
[8:36] I...
[8:37] You know, this is kind of an interesting case where you have this, uh, technology that, uh, is fantastic in the underlying model layer.
[8:48] We've we've spoken about this before.
[8:50] I think we talk about it every day.
[8:52] Right?
[8:52] So the the full AI stack with the infrastructure, uh, the the compute, the the, um, data centers are kind of on the foundational layer.
[9:02] Everything runs off of that.
[9:03] Then we've got the model layer with the intelligence, uh, that's kind of blurs the line between the model layer and the application layer because the models are so good themselves.
[9:14] They almost become a, you know, user, you know, front facing user app.
[9:22] Right?
[9:23] Um, and so what happened was, uh, when when the AI models first came out, a lot of people created, uh, what we would call...
[9:31] Or or or AI wrappers, right, where, hey.
[9:34] You just kinda stick a a a use case on top of AI, uh, on top of these models.
[9:40] There's no real differentiation between the app layer and the model layer.
[9:43] I think this is what Character AI, um, essentially was where you could build your own chatbots, but it was, you know, nothing really differentiating enough for it to obviously survive.
[9:59] You know?
[9:59] And and maybe the the other thing too is maybe just a lesson that this space is developing and moving so fast.
[10:05] If you're not coming out with the next thing or being super creative, um, in in the AI world, it's gonna be really hard to compete.
[10:13] There's no such thing as business as usual.
[10:15] Uh, I mean, look at the the previous story.
[10:17] We've got, you know, corporate spies.
[10:21] That's how competitive this space is.
[10:23] And so I think we'll continue to see competition here.
[10:27] And and, um, sorry to Character AI and and their loss.
[10:35] Hopefully, that's not too insensitive.
[10:37] But but but it looks like, yeah, maybe Disney will pick it up and, you know, I guess we'll find out what Disney wants to use the technology for, how they would apply it.
[10:50] And if that would live on and become...
[10:54] You know, find new life at Disney, that would be a great, great thing we would love to see.
[10:59] Let me know if you think we're the bust...
[11:01] The bubble's starting to burst here.
[11:03] ASML, the only company making the machines behind top AI chips, says it can't build enough of them through 2028.
[11:13] Quarterly sales hit a record €11,700,000,000, up 55% from last year, and profit, uh, more than doubled.
[11:24] It raised its yearly forecast for the third time.
[11:27] Some customers want to reserve machines for 2029 and 2030.
[11:33] Wow.
[11:33] If only, uh, if the only chipmaker is sold out through 2028, AI chip shortages aren't going away soon.
[11:43] ASML says even a planned 30% boost in output won't cover AI demand.
[11:50] Wow.
[11:51] Shares jumped more than 3% in Amsterdam after the record quarter and third forecast raise.
[11:59] Still, ASML is the only company making these machines, so a shortage also gives it pricing power.
[12:08] Wow.
[12:08] We are in a special place here.
[12:12] If you're not familiar, ASML is a lithography company, or they build the lithography machines.
[12:19] And, uh, the lithography machine is what is used to create these microchips, uh, that all the AI runs on.
[12:29] Right?
[12:29] And it...
[12:30] You you know, I would encourage you go do a little bit of research on ASML and on lithography and how these chips are built.
[12:39] You know, if you thought it was crazy that we have, you know, a trillion parameter model, uh, that we can ask it questions and the computer runs through all those parameters and pulls us back the the right answer, If you thought that was amazing, go look at how they build these chips.
[12:58] It is, like, blow your mind.
[13:00] It's crazy.
[13:01] Uh, essentially, they're trying to get, uh, as much compute power as they can into a really, really tiny space.
[13:08] And to do so, you need really advanced machines.
[13:11] They're really difficult to build.
[13:13] Uh, all the materials used to go into them are difficult to source.
[13:18] Uh, it's a very, very complex business.
[13:22] It takes...
[13:23] I think it takes...
[13:24] Somebody was telling me the other day, you have to...
[13:27] They load a whole, you know, whatever.
[13:29] A whole plane is just like to...
[13:32] And fly it to whatever country is...
[13:33] Needs the lithography machine.
[13:35] The the the first part of the machine takes a whole plane itself, apparently.
[13:40] Uh, and then that's not even the the rest of it.
[13:43] So, anyways, these machines are, uh, amazing, kind of the the wonder machine of the day.
[13:49] And, of course, they're powering all of this AI compute.
[13:53] And so we, um, are, you know, not surprised to see that ASML is continuing to see demand.
[14:04] And whether or not, you know, they they have orders into 2029, 2030.
[14:11] Does this mean that the boom is, you know, going to continue and and we're gonna continue to see growth like this and and and just, you know, a insatiable hunger for more AI compute, more AI chips.
[14:27] I don't know.
[14:27] But it is interesting.
[14:28] You know, when we talk about more folks getting into the AI chip space.
[14:35] You know, we we covered Microsoft and their Maya chip and OpenAI, jalapeno, of course, Nvidia.
[14:41] And so the way this works too, by the way, if you're not familiar, the, you know, the Microsoft, the AMD, the Nvidia, OpenAI, right, they design the chip, uh, that would work specifically for their use case or or whatever their use case may be.
[14:58] They design that chip, then they send those designs to primarily Taiwan Semiconductor in Taiwan.
[15:07] They build most of the world's chips.
[15:10] Right?
[15:10] Uh, and and especially in this use case.
[15:13] And then, uh, their TSMC, Taiwan Semiconductor, is using these ASML lithography machines to, uh, build the ships...
[15:22] The the the chips.
[15:23] And so that's kinda how the supply chain works.
[15:26] And so we're we're continue...
[15:28] You know, Microsoft comes out and says, hey.
[15:29] We're gonna build the chip most likely.
[15:31] I'd...
[15:32] No guarantees on this, but most likely, they're gonna send that order to TSMC or somebody that has an ASML machine that can build these chips for them.
[15:41] And and when you need these chips in in such a high volume, you know, we're talking about, you know, tens or hundreds of thousands of chips in one data center in some cases.
[15:55] Uh, that demand is really high.
[15:57] And so will that continue to grow?
[16:00] Um, I I think...
[16:02] And this is interesting.
[16:03] This is just my opinion.
[16:04] But OpenAI, there's been rumors that they're having compute issues and and struggling to maintain compute.
[16:11] And somebody pointed out, oh, it's not a compute problem.
[16:14] It's a money problem.
[16:15] Right?
[16:15] Well, of course, it's a money problem because the demand for compute is so high.
[16:20] It costs so much money to run these machines.
[16:23] If you could produce more chips, if you can produce more of the the hardware required to run, uh, AI, of course, you could drive the cost down.
[16:34] And just anecdotally, was...
[16:36] You know, OpenAI has been one of, if not the only model company that offers a fast mode.
[16:45] Right?
[16:45] So you can, uh, have normal mode where you ask a question and maybe it takes thirty seconds or depending upon how complex it might take, five minutes, ten minutes to come back.
[16:55] And OpenAI has this fast fast mode, right, where, uh, it'll be 50% faster.
[17:01] I was looking for it yesterday.
[17:03] I was like, yeah.
[17:04] Let me go...
[17:04] I wanna go fast on this one.
[17:06] I couldn't find the button in OpenAI desktop.
[17:09] I couldn't find the option on the web app.
[17:13] And not that it wasn't there, maybe it was there and I just need to update or something.
[17:17] But to me, was going, maybe they're holding back on the the super fast compute because it costs a lot more to run that.
[17:25] They need more chips to run that, and they just don't have the capacity right now.
[17:29] And so we...
[17:30] I, you know, I think we continue to see the demand where there's no shortage of it, and all signs are pointing that we're continually, um, uh, compute constrained.
[17:41] So, uh, you know, congrats to ASML.
[17:44] Congrats to everything that they're they're doing over there and pushing, uh, from an r and d perspective, the capability and capacity to build these chips.
[17:54] And, hopefully, they will continue to up, uh, capacity, and we we begin to ship more and more, uh, chips.
[18:04] Uh, I guess, yeah, what do you guys think?
[18:08] Is this, um, you know, the the previous story we're looking at, um, you know, character AI going bankrupt?
[18:16] Is that the beginning of the end of the AI boom?
[18:19] And then, well, wait a second.
[18:20] ASML says demand is such that even if they grew 30%, they would not be able to keep up with demand.
[18:29] What do you all think?
[18:30] Are we in an AI boom or bust right now?
[18:34] Nobody knows.
[18:36] More than 400 scientists, including Jeffrey Hinton and Joshua Benjo, want an international body to keep dangerous AI locked down.
[18:46] Their open letter asks labs to seal off models that can hack, even cutting them off from the Internet.
[18:55] They also want containment failures reported to the new body and eventually, the public.
[19:01] Okay.
[19:03] It lands one day after research warned small models could soon copy themselves onto home devices.
[19:10] We talked about that yesterday.
[19:12] The 400 signers say AI that can hack needs much stronger cages right now.
[19:18] Two of the field's most famous pioneers, Hinton and Bengeo, put their names on it, but open letters have no enforcement power, and no government has agreed to create this body.
[19:30] We continue to see it, and this is great.
[19:33] I, um, love a third party kind of, uh, collective that's coming out and and calling for some of this accountability and safety.
[19:43] Of course, we wanna be safe.
[19:45] Of course, we wanna make sure that, uh, these agents aren't going rogue and that we're able to keep track of it.
[19:51] Uh, there's an interesting little story this morning, uh, I think unconfirmed, but in some research at Google Gemini on their Argon four model, uh, they did some really interesting evaluation to see if the model knew that it was being evaluated.
[20:11] And we've talked about it for...
[20:12] Before, but the the chain of thought, you ask you ask the model, hey.
[20:18] Uh, what color is a fire truck?
[20:21] And then you can see kind of the chain of thought, the reasoning.
[20:23] The model goes, okay.
[20:25] A fire truck is a large vehicle that is used to fight fire.
[20:29] What color would this be?
[20:31] The user wants me to figure out the color of a fire truck.
[20:33] A fire truck is red usually, and then it produces the answer.
[20:37] Right?
[20:37] So that's the chain of thought, uh, in this, uh, this research done on the new Gemini Argon four model, um, the chain of thought, they ask it, do you think you're being evaluated?
[20:50] And in the chain of thought, it was, uh, actually addressing the evaluator in a very kinda, um, like, a little bit strange and, you know, kinda curious way where it's like, it it it said, I know you're watching me kind of really interesting thing.
[21:08] So, um, you know, we point being here, like, the the models are super smart.
[21:15] We need to be able to follow their chain of thought.
[21:17] We've talked about this before in understanding the reasoning that the model uses and why it, um, makes certain decisions and and how it...
[21:26] It's thinking along the way.
[21:27] I think that goes a a long way in, um, ensuring that we can, uh, put in the proper safety parameters to make sure that these things aren't going off the rails.
[21:39] Right?
[21:40] That being said, um, making sure that, you know, you have a third party body like this.
[21:47] It doesn't sound like this this group of scientists has necessarily put together a, you know, a really robust or structured approach to, hey.
[21:56] This is what this group's gonna do, and this is how we're gonna review things, and and what the actual practical application is.
[22:03] And the the...
[22:04] Whether or not you can get the model companies to actually participate in this would be interesting.
[22:10] But, again, I go back to what I've been talking about.
[22:12] We need open protocols that say, hey.
[22:15] This is the safe way to go test this.
[22:18] This is the safe way to run this advanced super intelligence in the real world.
[22:24] Um, maybe some of this from from their perspective, cutting off the models from the open Internet is a good idea in testing environments.
[22:33] But I think just having safe, real protocol and testing this up and saying, hey.
[22:38] This is the safety parameters that we recommend everybody follows when you're testing advanced superintelligence.
[22:45] And that would be a great idea.
[22:47] Just like any other field where there's some sort of, um, you know, existential danger, there's risk or whatever it may be, uh, we put those into boxes, and we know the boundaries in which we can push and where we can test things in a safe manner, um, without, you know, getting ourselves into a bad place.
[23:09] So that's that's what I think.
[23:11] This this is great.
[23:12] I'd love to see this kind of stuff come out.
[23:13] But let's work together and find a...
[23:16] An open protocol that everybody can apply to their model company that they can use and say, hey.
[23:22] This is this is the safety protocol that we wanna follow.
[23:25] These are the boundaries.
[23:26] This is what we'll report on.
[23:27] And make it really, really easy for the frontier companies to participate in these safe practices.
[23:34] What do y'all think?
[23:35] Is this, uh, you know, another attempt at regulatory capture or something that is trying to restrain the the growth and advancement of superintelligence, or do you think this is, uh, another great safety mechanism?
[23:52] Job seekers in Britain are using AI to cheat in interviews, and employers say, missed conduct just hit a record high.
[24:02] A group representing major graduate employers says candidate misconduct reached a record high.
[24:09] Recruiters report applicants reading AI written answers live.
[24:14] They say they're getting better at spotting it.
[24:17] Wow.
[24:17] Look out, applicants.
[24:19] Uh, if interviews can be faked by AI, uh, companies may bring back in person hiring.
[24:26] Recruiters say they are getting better at catching candidates who cheat with AI.
[24:31] A body for Britain's largest graduate employer says misconduct is at a record high, but the report doesn't say how many candidates were caught, so the the true scale is unclear.
[24:45] May...
[24:45] I think I've...
[24:47] I was just at an event last night, the the Claude go to market revenue event at Capital Factory here in in Austin, and it was a great time.
[24:57] Awesome to see.
[24:58] There were, I don't know, maybe a 150, 200 people there.
[25:01] It is a really great opportunity face to face talking about, uh, what we're doing with Claude code and the advancements that that we're making and really cool use cases folks got to share and see how they're applying it to their go to market and to their sales and and marketing and and revenue strategies.
[25:20] Uh, something kinda came up in that conversation where we either talked about AI tools and automation and let's automate everything away as much as possible because we've got these super powerful, super intelligent tools.
[25:35] Part of my belief is that as these tools get more and more powerful and automation becomes something that, you know, you can turn on and have it be indistinguishable, I think human interaction becomes much more valuable, much more sought after, uh, just naturally.
[25:54] You know, humans love the human to human connection and, uh, developing relationships and and partnerships with one another.
[26:03] And so I think that, you know, this is a great example of this where we see, okay.
[26:08] Folks are figuring out, hey.
[26:09] Actually, I can, you know, cheat on this interview, or I can, uh, cheat on, uh, a school exam, or whatever it may be and and not necessarily have to have all the requisite capabilities built in to go perform this task.
[26:27] And and that puts some of these these companies in a risky situation, right, where they may be hiring somebody that they think is awesome.
[26:35] Um, they go through a lot of effort to onboard them, and then it turns out that they have no idea what they're talking about.
[26:40] And because these are remote roles, it takes quite a while for them to figure this out.
[26:45] Um, and, uh, of course, there's all sorts of implications from an international standpoint when you have, uh, maybe bad actors that are interviewing.
[26:55] We we saw stories in the past couple months where folks from, uh, North Carolina or North Carolina...
[27:02] From North Korea were reportedly, uh, getting jobs at Western companies and then were stealing information or funneling that money back to North Korea reportedly.
[27:18] So you've gotta be careful in these situations.
[27:21] And I just think this this shows that in person, face to face is going to be the superpower.
[27:28] If everybody's automating everything, what's better than going in and shaking a hand and having a real conversation with somebody?
[27:35] I think we're gonna see a lot more of that.
[27:37] I think open meeting spaces are going to be really valuable moving forward where we've got, uh, you know, coffee shops or, you know, cafes, you know, restaurants, things like that where people can get together and meet are gonna be really important spaces for work to actually get done, uh, in the the the coming, uh, future.
[28:00] Uh, that being said, you know, can you blame the applicants for using the tools at hand?
[28:08] It's hard to blame them.
[28:09] They're trying to do the best they can.
[28:10] They're trying to stand out, and then you create a superintelligence tool that they can run on an interview and sound smarter than they could ever possibly be and, uh, ace an interview.
[28:22] You can't blame them for trying.
[28:24] Uh, that being said, like, we we we need to figure this out.
[28:28] I think the easiest way around is go analog, go in person, and you solve a lot of these problems here.
[28:34] You know, question to the audience.
[28:37] Have you guys ever used any sort of these, let's call it, uh, intelligence enhancement tools on a live call?
[28:47] I actually...
[28:48] Interestingly enough, in sales calls, sometimes I get wrapped up in the conversation, and I wanna know what's the best next question that I should ask.
[28:59] And sometimes I forget, oh, man.
[29:01] I meant to ask that question, or I meant to follow-up on one topic or another.
[29:05] So I created a tool that pops up and reminds me.
[29:07] Right?
[29:08] Those kind of little, uh, assets, I think, can be super, super valuable.
[29:13] Uh, but, again, there's nothing better than just the imperfect, uh, human interaction in person that we love to see.
[29:21] Let's see.
[29:22] Oh my goodness.
[29:24] It's happened again.
[29:25] We did it again.
[29:26] We got to the end of the show somehow, and I was ready to do another 10 or 20 stories maybe.
[29:32] This is so much fun.
[29:33] I love doing this.
[29:34] Uh, really, uh, appreciate everybody checking out the show.
[29:39] Apologies to our YouTube fans and not being able to get the live up.
[29:45] I don't know what was going on there.
[29:46] I'm gonna write a letter to YouTube and see if they can help me out here.
[29:52] But, again, we had a lot of fun.
[29:54] We talked about a lot of different topics, which I think it's interesting to see the...
[30:00] Where are we in the boom bust cycle?
[30:03] Um, my impression is we're extremely early.
[30:06] We still see a lot of folks have not adopted this technology in their day to day, and it's super powerful technology.
[30:14] But we're continually seeing the the adoption increase.
[30:18] I've, you know, made predictions.
[30:20] Think by the end of this year, we will see kind of a a big paradigm shift, especially as it's rumored that, um, the next Claude model is going to come out.
[30:32] We've gotta imagine the the next OpenAI model will come out.
[30:35] We've got more players in the frontier space that are are pushing the envelope on the intelligence and capability of these AI tools.
[30:44] And so I have gotta imagine that we are still fairly, uh, early in this just from an adoption standpoint.
[30:51] Of course, um, people are getting, uh, getting out there and putting a lot of money into this too.
[30:58] So we...
[30:58] We'll continue to watch this closely from a market standpoint, but also from a technology standpoint.
[31:03] How does this impact you?
[31:04] How can you use these tools?
[31:07] Um, make sure that our audience knows what's going on, uh, on a day to day basis with, uh, Super Intelligence.
[31:15] And, um, again, this is where we talk about AI news with real context.
[31:21] I'm your human in the loop.
[31:22] Austin Wilson.
[31:23] We'll see you again tomorrow morning at 6AM.