October 20, 2026
2026-10-20
https://www.youtube.com/watch?v=uI2Odggqd28

[0:03] We go live in ten.
[0:17] Every morning, the future moves faster.
[0:22] The models, the money, the power, and the people building what comes next.
[1:01] From Austin, Texas to the world.
[1:08] AI news, real context.
[1:11] This is the Austin Wilson show.
[1:14] I'm Austin Wilson, your human in the loop.
[1:22] Good morning, and welcome to the Austin Wilson Show where we're breaking the AI headlines that matter to you today.
[1:29] I'm Austin Wilson.
[1:31] I'm coming to you live from the AI capital of the world, Austin, Texas, and excited for today's show because we have got a very healthy docket docket, uh, that we're gonna cover today.
[1:43] OpenAI is looking to raise, uh, just a little bit here and there, another $30,000,000,000 in a race against Anthropic's IPO.
[1:51] Uh, WhatsApp's secret rules are, uh, secretly guiding in their chatbot.
[1:57] One hacker and an army of AI agents.
[1:59] Uh-oh.
[2:00] Uh, the AI that refused to help fight back against it, and then AI that reads cancer odds from your face.
[2:07] Lastly, DeepSeek, banned by The US, but building chips with Huawei.
[2:14] Let's get into it.
[2:15] We got a good one here for you folks.
[2:17] OpenAI is reportedly racing to raise $30,000,000,000.
[2:22] It wants, uh, the money locked in before Anthropic goes public.
[2:26] Reuters Reuters cites two people familiar with the talks.
[2:30] The round could value the ChatGPT maker above $1,400,000,000,000.
[2:36] SoftBank is in talks to join.
[2:39] Anthropic stock debut is expected in mid November.
[2:44] Wow.
[2:45] Uh, the two top AI labs are now chasing the same dollars from the same investors.
[2:51] One source says OpenAI wants this done before Anthropic opens its books.
[2:57] OpenAI told investors it's on pace for $50,000,000,000 a year in revenue.
[3:02] It expects a $70,000,000,000 run rate by the end of the year, but no deal is done yet, and this is amidst chip stocks losing $1,600,000,000,000 in value in under two weeks.
[3:16] Wow.
[3:17] We were seeing another horse get into the race here, and it makes a lot of sense as, um, we have, uh, more and more projected necessary spend to keep AI running and growing and powering all of that.
[3:30] These AI labs need the financing to cover it, and so we are in a thick race here between, uh, OpenAI and Anthropic as now we've learned OpenAI is look...
[3:42] Looking to raise some some funds privately, which is a very interesting take here at a $1,400,000,000,000 valuation.
[3:50] Uh, previously, they were valued at a little less than $900,000,000,000.
[3:56] So it's interesting to see OpenAI getting into the race here and taking this opportunity to go raise money.
[4:04] Now the question is, are they doing this out of necessity?
[4:07] They...
[4:07] Do they need the the funds now, um, or is this just an opportunistic move where they're seeing, uh, the availability of the funds in the market and they wanna go get it while the getting's good as they say.
[4:21] Uh, this is something that we will follow closely as obviously this has large implications with, uh, the Anthropic IPO next month, but also the overall AI market.
[4:34] Uh, if we have two companies, especially these these two frontier companies that are leading the way in intelligence, and then they end up both being over a trillion dollar valuation, that is a phenomenal outcome and a signal from the market that this show is is not over yet, uh, especially if Anthropic can end up...
[4:57] Or excuse me, if OpenAI can end up raising these funds, uh, again before Anthropic goes public.
[5:05] That that being said, they're in a a race against the clock.
[5:08] When is Anthropic going to go public?
[5:10] Does this put pressure on Anthropic to, uh, you know, there's been some rumors that Anthropic is going to drop, know, a massive, uh, new release from a model standpoint or some new feature that is just gonna blow everybody away before they release, uh, the IPO next month.
[5:31] And I'm curious.
[5:33] This is kind of a chess move, a strategic move that we'll see.
[5:36] Is OpenAI, uh, trying to, uh, play, you know, eight d chess with Anthropic here and force them to make a move sooner than they would want to?
[5:46] I don't know.
[5:46] I don't see Dario being kind of pressured into an early IPO or anything like that, but it is curious to see.
[5:54] And, man, it's fun to watch.
[5:56] This is exciting stuff.
[5:57] Let me ask you guys.
[5:59] We've got OpenAI looking to raise private funds, Anthropic going public.
[6:04] Who's got your money?
[6:05] Who are you betting on?
[6:06] OpenAI or Anthropic?
[6:08] Who's gonna win this race?
[6:09] We shall see.
[6:11] Coming up next, WhatsApp has a help page, but they didn't tell anybody that they had hidden secret orders for their AI support bot.
[6:20] Uh, one of the bots was told not to recommend, uh, certain things to privacy regulators.
[6:26] Okay.
[6:27] So privacy company, uh, Redact found the hidden text on about 80 pages.
[6:34] Another order kicked in if you mentioned a lawyer, then the bot couldn't cite privacy laws even if you raise them.
[6:41] Meta says a bug exposed the notes, and now those notes are gone.
[6:46] This is interesting to see coming out.
[6:48] Company chatbots can have hidden rules that shape what they tell you.
[6:53] WhatsApp says it just doesn't want its bots giving legal advice.
[6:57] Okay.
[6:58] Its European privacy policy says you can complain to a regulator.
[7:03] Locked out users were told to wait up to a year, but Redact, the company, read the rules, not the bot's actual answers.
[7:11] Nobody knows if the bot obeyed them or not.
[7:15] This is a...
[7:16] An interesting one, and one that's close to my heart is no shade to Meta.
[7:21] I I love what Meta's doing and and what Zuckerberg is doing and the and the progress that they've made in AI and catching up and and seeing what's coming out of the Meta Super Intelligence Labs and Alexander Wang.
[7:35] Phenomenal stuff.
[7:36] That being said, uh, their customer service is a little interesting.
[7:40] I've, you know, run into a couple issues where, you know, you get blocked or don't know why you got blocked or you try to ask the the AI, you know, customer service agent and it doesn't help you at all.
[7:52] And there's, you know, there...
[7:54] It doesn't seem like it's fully conducive.
[7:56] I'm sure they're working on this.
[7:58] Right?
[7:58] But, hey, shout out to the company Redact for figuring this out and under...
[8:03] In in uncovering, it sounds like a system prompt for the WhatsApp AI customer support bot and uncovering the fact that maybe it is trying to...
[8:16] Is it trying to not provide legal advice, or is it trying to steer users from asking certain questions that maybe Meta doesn't want to deal with?
[8:27] That's an interesting uncovering here.
[8:31] And we've seen this in other areas, but not specific to this type of use case, which I think is interesting.
[8:39] And that being that underneath...
[8:41] And we've talked about this before, but underneath all of these agents or these chatbots, we have a system prompt that instructs so the WhatsApp AI, you know, customer support chatbot has a system prompt that says, hey.
[8:55] You are the WhatsApp customer support agent, and your job is to help customers.
[9:00] But first, there's some rules.
[9:02] You know, we wanna make sure you don't provide legal advice, and then they give additional rules, and that all is fed through the agent before you even message them.
[9:11] And then when you message, the agent reads there and double checks, hey, what are my rules?
[9:14] Let me make sure that I'm following the rules that I've been given.
[9:18] And so we've seen this with, you know, bigger models, the the Claude models or the ChatGPT models, and that people have been able to uncover what the system prompt is for those those models.
[9:32] This is kind of a unique case where it's a kind of a niche use case and what is the customer support agent's system prompt, and we're finding that maybe there's some interesting things in there.
[9:45] That being said, Meta, it sounds like, responded pretty quickly, and it looks like that they are addressing this.
[9:51] And so this would kinda put everybody else on on guard.
[9:54] I think anybody else that has their own proprietary customer support agent, if it is not up to snuff as they would say, then I'm sure they are, uh, cracking down and making sure that their system prompt is in line.
[10:08] Uh, that's speculation, of course, but I'm sure that others are gonna follow suit.
[10:12] And again, what we talk about a lot on this show is the competition is really good, and you've got really smart people like these folks at Redact that go in and uncover these things, and it's great to see just the open market working.
[10:23] It's amazing.
[10:25] That that being said, what's your experience been with customer support bots?
[10:32] Right?
[10:32] Do you feel like they're there, or are we still, you know, way, way far away from super intelligence in terms of customer support agents.
[10:44] Coming up next, we are talking about a hacker who let AI agents do the break ins.
[10:49] They got into the websites of more than a 100 companies.
[10:53] Reuters found the victims include John Deere, the tractor company, and a company called Shark Ninja.
[11:00] The hacker's orders fit in under 100 Chinese characters.
[11:05] The agents tried more than 1,200 different sites.
[11:08] A research estimates, uh, researcher estimated that that the whole thing cost about $5,000 to run.
[11:15] This is really interesting.
[11:16] AI makes hacking so cheap that one person can go after a thousand sites at once.
[11:22] Researcher, uh, Fagerland watched it happen.
[11:26] He said it's not script kiddie stuff, he told Reuters.
[11:29] He estimates each successful break in cost about $50 in AI fees, but nobody knows yet what data was taken.
[11:37] Deer says it seems no sign, uh, of sensitive data was accessed, which is good to hear, but nevertheless, this is concerning.
[11:46] When we have the power of multiple agents and you can put these agents and work in in, uh, parallel, that means that you can get a lot more done from a productivity standpoint, from the good guys, right, but also the bad guys.
[12:04] So so we're seeing this come out.
[12:06] This is not the first story that we've covered like this, but we're starting to see more agents come out.
[12:11] These agent swarms being run by one or a few people and being able to attack a broad service area.
[12:19] Again, we haven't seen the the negative impact, so to speak.
[12:24] So the hackers may have found a vulnerability and broken in one way or another, but they weren't able to necessarily get, uh, any value or data out of it, uh, as as reported here.
[12:37] Uh, and we've seen other such instances where it doesn't seem like there was a broad negative impact, which is good to see.
[12:44] That doesn't mean that we're...
[12:45] We don't need to be on guard and don't need to be protecting against this.
[12:49] And this is a big call out, and I think we're seeing some help from Anthropic, some help from OpenAI, but the Frontier Labs coming in and being very supportive in these situations where they have the smartest, most intelligent, most capable models, and especially from a cyber defense standpoint, they need to be on guard and be helpful.
[13:11] And I think we're seeing that where they are jumping in and helping these companies out that are getting attacked by agent swarms, and they're counter swarming and coming in and making sure that, uh, these systems are protected.
[13:24] The good news out of this is as we see more of these attacks, um, just like in the chatbot story, right, we learn from this, and companies are learning from this step by step and finding out new ways to safeguard and protect their systems and make sure that they're protecting their data.
[13:42] So on on the the good side of things, we are learning and getting better.
[13:46] That being said, we need to continually be on guard for agent swarms and attacks like this.
[13:53] That being said, you know, when when you guys think about this, are are you guys being more sensitive with your data and who you give your data to considering what can happen here?
[14:07] Or are you saying, hey, we're in the age of AI.
[14:10] Let me give my agents access to whatever and they can go ham.
[14:14] I'd love to hear what you guys are doing to protect yourselves or or if you're in full trust that, hey, Anthropic's probably gonna come in and save the day, so let's let them have it.
[14:24] Who knows?
[14:27] Next up, uh, we are talking about another hacker story.
[14:31] Hackers hit the AI coding tool OpenCode with AI agents, and when the team at OpenCode fought back, the AI it used refused to help, and we've seen this a couple times before.
[14:43] The models helped at first.
[14:45] We're talking about advanced frontier models, and then the attack code tripped their safety filter.
[14:52] So attackers kept it...
[14:53] Kept up the fight for forty eight hours.
[14:56] Some users found upwards of a thousand dollars in charges that they didn't actually make.
[15:03] AI attackers move really fast.
[15:05] Defenders can't afford an AI that quits mid fight with house, uh, excuse me, White House tech adviser David Sacks weighed in.
[15:14] He said the guardrails backfired, and, um, Hugging Face hit the same wall in July.
[15:19] It said guardrails can't tell responder from an attacker, but only vetted teams get the looser AI.
[15:26] OpenCode got in through OpenAI's Daybreak Blue program.
[15:32] Um, this is, like, this is worth kinda digging into.
[15:35] So what...
[15:36] What's going on here?
[15:37] We saw this with Hugging Face back in July.
[15:40] Essentially, with these advanced frontier models, the OpenAI models and Anthropic models, they have, hey, publicly available models.
[15:50] Me and you, we we can use the Opus five five.
[15:53] We can use ChatGPT 6.1 Soul.
[15:57] You you know, we can use these advanced, uh, consumer grade models.
[16:02] But kinda behind the curtain, they have even more advanced models, and specifically, they're more advanced in cybersecurity capabilities.
[16:12] They don't wanna give those just to anybody because of, you know, nefarious actors could use those, uh, in a, uh, malintended way.
[16:21] And so they keep those gated just in case or or or really to protect them from any bad actors.
[16:29] And so what we've seen and what we what we saw with Hugging Face and what we see in this OpenCode story is that the, um, OpenCode realized they were getting attacked.
[16:39] They jumped on it.
[16:39] They're using...
[16:41] I saw that they were using Anthropic.
[16:43] It looks like they're using OpenAI models as well too to fight against these defenders.
[16:48] And then the OpenAI and Anthropic models have these guardrails that say, hey.
[16:52] If...
[16:53] Again, coming back to the system prompt.
[16:55] In their system prompt, says, hey.
[16:57] If somebody starts using this model for cybersecurity things, you need to block them, and that's what happened.
[17:03] Um, and that's a safety guard.
[17:06] It ended up OpenCode.
[17:07] Dax, uh, the CEO of OpenCode, he reached out to Anthropic and OpenAI, and they got on the, uh, the advanced security models.
[17:18] And so they were able to defend against, uh, those attackers, but it sounded like it was nonstop forty eight hours they were trying to attack these.
[17:26] The interesting thing here is...
[17:30] Or or, like, what I think about from Anthropic standpoint or OpenAI standpoint, are they looking at, hey.
[17:36] What companies can we supply this, you know, these advanced cybersecurity models to to enable them to, you know, defend against these attacks before they happen.
[17:48] I know they have a, um, application process where you can apply to get into these models, and so I'm sure they're working on that.
[17:56] And I...
[17:56] Just curious, like, you know, we talked about the use case or or the story earlier with John Deere and Shark Ninja.
[18:04] Like, I wonder if it's on, um, these frontier model companies to reach out and proactively, uh, help from a security standpoint.
[18:15] Obviously, there's, you know, commercial incentive there.
[18:18] I, you know, I think they would charge them for that.
[18:21] But but, yeah, it's it's interesting to watch, interesting to see what's going on from a cyber standpoint.
[18:29] Ultimately, though, like, we see this over and over again, and I think it's worth iterating is that we've seen these attacks, and we've seen them defended.
[18:37] Um, there there hasn't been, you know, this big meltdown or like crazy thing.
[18:43] Like, people are paying attention to what's going on.
[18:45] We have tons and tons of observability tools.
[18:47] We can catch these when they start happening, and then go defend against them.
[18:51] Of course, we would like to, you know, have that that wall of defense up before the the attack begins.
[18:58] The question here now though, and for the audience, I would love your input.
[19:03] Um, who do...
[19:05] Who gets to decide who gets these unrestricted models?
[19:08] Right?
[19:09] Should that be an Anthropic, an OpenAI thing?
[19:11] They decide who gets those models, or should that be on a government entity or somebody else?
[19:17] Curious to hear what your perspective is on this.
[19:22] Next year, doctors built an AI that reads a cancer patient's face.
[19:26] From one photo, it predicts their their odds of survival.
[19:30] Mass general Brigham published the study Monday.
[19:33] The AI learned from more than 40,000,000 face photos.
[19:37] In one test, a hundred and twenty two patients got the highest risk scores.
[19:41] More than half of them died within ninety days.
[19:45] This is amazing.
[19:47] Uh, a quick photo could help doctors decide how hard to treat a patient.
[19:52] The researchers believe it's the first AI to predict cancer survival from a photo.
[19:57] Patients who look to five plus years older than their age had a twenty one percent higher death risk, but it's not ready for hospitals.
[20:05] The survival model learned mostly from, uh, from white patients near Boston.
[20:12] So it's it's definitely not fully tested out or something here.
[20:15] But interesting to see these, uh, different use cases, unique use cases, especially in the medical field is this is one of the kind of promised, um, and and maybe, like, most promising areas of development for AI where we can enable doctors and and medical professionals to help treat, uh, disease and and cancer, and we'd love to to see stuff where we can provide more information.
[20:42] Obviously, this one needs a little bit more work.
[20:44] Next, uh, The US just ordered DeepSeek out of federal agencies.
[20:49] The same day, a report said it's helping Huawei design AI chips.
[20:54] Homeland Security says the Chinese app may track and fingerprint users.
[20:59] Agencies must also drop apps built on its AI.
[21:03] Uh, the information says Huawei engineers have worked inside the company's offices.
[21:08] It could help China build top AI without Nvidia.
[21:13] Wow.
[21:13] This is kind of a big one.
[21:14] AI is splitting into a US side and a China side, uh, from apps to chips.
[21:20] The order says government workers' prompts could send sensitive data to China.
[21:25] Last month, DeepSeek released free tools that run its AI on Huawei chips, but the chip report rests on unnamed sources, and the band only covers civilian agencies, not individuals like me or you.
[21:44] Uh, so we're continually seeing kind of the split here, which is, uh, an interesting breakthrough between Chinese models and US models and whether or not Chinese models should be run-in US government agencies, it looks like we're not gonna do that anymore.
[22:02] The the fact that they are helping...
[22:05] DeepSeek is helping Huawei develop chips is an important distinction, and some see this as a kind of a military, a defense type thing where if you're building advanced super compute chips, those are so strategic that it becomes almost a a kind of a, yeah, a a strategic defense measure.
[22:31] And so this looks like another extension of that where if DeepSeek is helping develop these chips and potentially DeepSeek...
[22:39] Although it's open source, if they had some way to fingerprint or understand what users are are using the models for, then they could ultimately use that strategically, and it gives China a leg up.
[22:54] Our adversary, China, of course.
[22:57] So it it...
[22:58] It's also interesting to follow this kind of on the heels of American open source models coming out over the past several weeks where we saw multiple really advanced US based open source technology come out.
[23:14] And so, obviously, DeepSeek has been leading the way in this for almost two years now.
[23:21] Um, but do we...
[23:24] Are are we now, like, in a position where we can pull back and we're gonna start seeing maybe American companies follow suit?
[23:30] I know some enterprises will definitely follow the recommendation of the US government, although it's not a restriction right now.
[23:39] It seems like it's just with US government agencies that they're they're pulling out DeepSeek.
[23:45] But will we continue to see more pullback and more separation between Americans and Chinese and Americans using American models, whether that's the frontier models of, you know, Anthropic OpenAI or SpaceX AI or Meta or Google or the open source models at, you know, from the the BEAM model or the Meta open source model or Nvidia open source models and see that kind of split where then the Chinese are using their own Chinese open source models.
[24:20] And, you know, how does that impact from a strategic standpoint both countries?
[24:25] I think there's a lot of implication here, and I wonder if there's deeper layers to kind of the strategic play in kicking DeepSeek out.
[24:35] Of course, we will follow this story very closely.
[24:39] I've talked about this a lot in the past and asked the advice of a lot of our audience in terms of what they think about open source models, especially Chinese open source models, especially DeepSeek.
[24:52] What's your perspective?
[24:53] Would you still use DeepSeek AI models today even if the US government says to hold off?
[25:02] Let me know.
[25:05] Well, folks, that brings us to the end of another fantastic Austin Wilson show.
[25:11] We had a lot of fun today.
[25:12] Covered some really interesting topics and and seeing the market move forward.
[25:17] OpenAI looking at, uh, raising potentially $30,000,000,000 ahead of the Anthropic IPO.
[25:23] I love to see that clash.
[25:24] Boom.
[25:25] Uh, and then we're seeing additional developments in medical treatment, uh, some hacking stories, and how we're defending against hackers and continually progressing and moving forward.
[25:37] I wanna encourage you guys to continue to dive into AI and super intelligence.
[25:43] Use the tools.
[25:43] Learn the tools.
[25:44] Enable yourself.
[25:46] Become better and better day after day, and we will win.
[25:49] I think we see a a a future where AI is going to be such an important piece of what we do and what we use that...
[26:00] I was talking about this yesterday with with a colleague where no longer are we going to be worried about the manual kind of mundane stuff that's been a part of our work for as long as we've we've known about it.
[26:12] We'll spend more time as humans working on creative problem solving and having more fun in our work and creating new technology and new creation that that we haven't been able to before.
[26:27] And so that's really exciting to me.
[26:29] I hope it's exciting to you.
[26:30] I would love to see you guys, what you're building, what you're working on, and hear from from you guys as you continue to develop and use these new tools.
[26:40] As always, we are here for your AI news, real context, and I'm your human in the loop.
[26:46] Austin Wilson, we'll see you tomorrow morning.
[26:49] Let's go.