October 23, 2026 2026-10-23 https://www.youtube.com/watch?v=VU92VltCdCg [0:12] We go live in ten. [0:26] Every morning, the future moves faster. [0:32] The models, the money, the power, and the people building what comes next. [1:10] From Austin, Texas to the world. [1:17] AI news, real context. [1:20] This is the Austin Wilson show. [1:23] I'm Austin Wilson, your human in the loop. [1:31] Good morning, and welcome to the Austin Wilson Show where we're breaking the AI headlines that matter to you today. [1:39] I'm Austin Wilson coming to you live from the AI capital of the world, Austin, Texas. [1:45] It is Friday morning. [1:48] We are excited for the end of the week and a big weekend, and we've got some really important AI news to cover today on the docket, Microsoft's AI that is better than doctors on hard cases. [2:02] OpenAI, uh, reduces its value by about a $150,000,000,000. [2:07] Etsy's ban on AI made recipes and patterns. [2:12] Google will pay homes to use less power. [2:17] That'll be interesting. [2:18] A judge's AI errors get him pulled off a case, and outside auditors found holes in OpenAI's watchdog. [2:27] Plus, we've got a new segment for you today. [2:30] We're really excited to get into it. [2:32] Let's go to the first story. [2:34] At the top of the docket, Microsoft's AI just beat doctors on hard medical cases. [2:40] It got 83% right. [2:42] General doctors only got 36% right. [2:46] The, uh, study came out Thursday in the journal Science. [2:50] The study came out Thursday in the journal Science. [2:53] The cases came from a top medical journal. [2:58] The doctors could use books and the Internet. [3:00] The AI works like a panel of virtual doctors. [3:04] It asks questions, orders tests, and debates the answer. [3:10] Your next diagnosis could get an AI second opinion. [3:14] A Boston hospital will test it with patients. [3:17] Microsoft's AI chief calls it a step forward for medical superintelligence. [3:24] On everyday ER cases, it scored 80%. [3:28] Doctors averaged 47%. [3:31] But there were old cases, but these were old cases written up as text, not live patients. [3:39] Uh, specialists weren't tested. [3:41] This was just a general practice doctor, super intelligence, AI doctor working on this. [3:49] This is really interesting to see coming out of kind of on two fronts. [3:54] Number one, we've got Microsoft leading the way in this, which is really cool to see another player coming in and Microsoft having a frontier model that is capable of outperforming human doctors in a general test. [4:10] So that's really exciting to see that it's Microsoft leading this. [4:14] We got another horse in the race. [4:16] Of course, Microsoft has been involved for a long time, but it's great to see another one kind of leading the way. [4:21] And then the second piece, the medical advancements here. [4:24] This is awesome. [4:25] This is a use case that we've talked about for quite a while where, you know, a doctor has a limited... [4:32] They they have their experience and their understanding and their best judgment involved. [4:39] But imagine if you have a doctor or a super intelligence that has every single, you know, experiment that's run, every single case study, every single test that has gone through medical publication and testing and review and everything, all of that at its fingertips, and then could review your medical case and provide a diagnosis there. [5:05] That looks a lot different than just having a doctor there. [5:09] So I think again, you know, we talk about this in the arts, we talk about this in business. [5:14] Now talking about this in medicine, it's really interesting to pair the human judgment, have really good human judgment and experience, and the human looking at what the other human is experiencing and what they're going through, and then superpower that with a super intelligence, with agents that are running simultaneously that they have the the experience and the understanding that goes beyond what humans have, which is something that we haven't had access to in the past, and something that I think is going to make a big difference moving forward. [5:56] And it'll be interesting to see how humans accept this or interact with this. [6:04] I don't know. [6:05] Would you trust in AI's diagnosis over your own doctors? [6:10] That's the question. [6:13] On to the next story we are seeing coming out of OpenAI. [6:22] OpenAI reportedly just cut its own price tag by a $150,000,000,000. [6:28] The new target is 1,250,000,000,000. [6:32] It's raising up to $30,000,000,000. [6:35] CNBC says investors from UAE are leading. [6:39] Bloomberg says commitments are due by the end of next week, so it's moving pretty fast. [6:46] Meanwhile, Anthropic reportedly wants 1,500,000,000,000 to $2,000,000,000,000 when it goes public. [6:53] Interesting to see these numbers kinda shift and move over the weeks. [6:57] Anthropic's IPO is expected next month. [7:00] Investors will compare the two price tags. [7:02] Even after the cut, OpenAI would be worth far more than it was last March. [7:08] Last March, it's last round that they raised, they were valued at $852,000,000,000 of post money valuation. [7:18] But tech stocks slid this week on fears that AI spending has gone too far and terms could still change. [7:25] This is all kind of up in the air as we're seeing these markets move and shift. [7:31] Uh, lots of volatility over the past couple of weeks with chip stocks moving, with the imminent Anthropic IPO, with OpenAI announcing that they're going to go raise funds again. [7:45] We saw the folks on the All In podcast last last week talk about this. [7:51] I think Chamath brought up, like, just going to raise a private round for Anthropic instead of them going public, and now we see OpenAI is maneuvering to do that. [8:01] And Anthropic, it looks like, is still planning on going public. [8:05] This is all really tenuous. [8:07] This is not a set in stone thing from valuations to how much they're raising to going public. [8:14] All of it is dependent upon a lot of, you know, you gotta wonder, like, these AI companies, you gotta imagine that they are using their internal models to strategize and figure out the best kinda game theory optimal decision to make, whether they go public, whether they raise, what valuation do they raise money at, how much do they raise. [8:39] This is all a very, very complex strategic play. [8:43] The good news is that we still see investors wanting to get in on this. [8:49] Right? [8:49] There's no shortage of demand, it seems, that we are seeing the financial industry, uh, willing to support this. [9:01] It's just a matter of how long and how much. [9:04] Right? [9:05] We continue to see demand from businesses and continue to see increased revenue from both OpenAI and Anthropic. [9:13] And so it... [9:15] You know, I I guess just at some point, where does the rubber hit the road? [9:20] Right? [9:20] Where do we say, okay. [9:23] That's enough financing, and now we need to go generate more revenue. [9:29] But as you continue to see, you know, Anthropic, we talked about they were $1,100,000,000 in revenue this time last year and... [9:41] For for the quarter, I believe. [9:42] And on track for this year now, they... [9:46] It reportedly did $16,300,000,000 or $16,800,000,000 in the last quarter. [9:52] So it's like, this is growing and growing fast, and we see the valuations increasing substantially. [10:01] It's just a matter of how much appetite is there in the industry, and are we gonna run out of steam anytime soon. [10:09] The question is, you know, we hear Elon talking about crazy projections. [10:16] He mentioned and whether or not this is a reachable target at all. [10:22] Elon said that they were targeting 500,000,000,000 to $1,000,000,000,000 in revenue for SpaceX next year in 2027. [10:31] If those targets are real, then it's like, oh my gosh. [10:35] Yeah. [10:35] Of course, you would wanna pour into Anthropic or OpenAI or SpaceX. [10:42] But are those reachable? [10:43] What's the time frame? [10:45] What's the actual revenue target? [10:48] Is this recurring revenue? [10:49] Lots of questions to ask. [10:50] Let me ask you this. [10:51] Would you buy OpenAI stock at a $1,250,000,000,000 valuation? [10:57] Is that overvalued or is it undervalued? [11:01] Etsy just banned AI made recipes and craft patterns. [11:06] Seller said some AI patterns couldn't make the item pictured. [11:11] Pattern and recipe listings now need a real photo of the finished result. [11:17] Most other listings need one unedited photo of the actual item. [11:22] Fully automated listings are banned too. [11:25] AIR is still allowed if sellers properly disclose this. [11:30] This is a really interesting marrying between commerce and art, I think. [11:37] Many buyers can't tell what's AI. [11:39] Etsy says that's a top three reason that they shop less. [11:43] Etsy says the photo rule helps buyers know what to expect, and sellers may now owe refunds when a buyer's item doesn't match the listing. [11:52] The catch is enforcement. [11:54] Etsy gave no deadline for fixing old listings. [11:58] This is... [11:59] I love this topic and I love seeing this in practice because what's happening here is we're actually seeing the market respond to AI generated content. [12:10] And Etsy, if you're not familiar, it's a kind of maker marketplace where people who are typically creatives, uh, create some sort of artwork or interesting new little product, unique product, and they sell it on a smaller scale usually. [12:29] And they go to Etsy and they sell there. [12:31] The problem is if I'm using AI to generate the photos of my product that I'm selling, my art artwork that I'm selling, and then I ship that to a customer and they receive that, and it's not what it looked like in the picture, that creates a really bad experience for the customer. [12:49] And you can imagine in a marketplace setting like this, the proliferation of AI generated content, AI generated photos for displaying what the product may be or might look like, and then whether or not you're actually delivering on that. [13:05] So it's interesting to see the feedback that users are dissatisfied. [13:11] And the kinda implementation from Etsy saying, hey, actually, we need to crack down on this and we'll not allow it anymore. [13:22] This is something that we've, I think, spoken about in the past about some of the model companies developing tools that detect AI. [13:32] So it'll make it really easy for users to understand, hey, this is actually AI generated. [13:38] I think also something I like about this is I I'm not so bad now, but maybe a month ago, you know, some of these AI detection tools were becoming mainstream where you could detect AI written posts or AI written emails or anything like that. [13:58] And so I was running a lot of what I was reading through AI detection tools, and they're coming back AI generated. [14:05] I'm like, I do not want to read this. [14:07] I wanna read human generated content. [14:11] Just like on Etsy, I wanna buy human created things and human represented things, and they're messy. [14:19] It's not perfect. [14:20] And that's... [14:21] There's something beautiful about that. [14:22] That's part of the problem with the AI and the models. [14:25] And just a little explainer here, but the AI models try really hard to emulate human text, for instance. [14:34] And so they have certain randomization built in where they will randomize the words in in the distribution of what they're saying and writing. [14:46] The problem with that is there's still a pattern in the randomization. [14:51] With humans, we are imperfect. [14:54] We will randomize to some extent, but not in an even distribution. [15:00] And it's really unique. [15:02] It's really something that's specific to human beings and that we like to hear and like to interact with. [15:09] And so that's why the model writing or AI writing is so detectable is because that pattern still shows up over and over again. [15:18] No matter how much they try to train the models to act like humans. [15:24] The humans still are trying to guess the next best thing, and there's a pattern there. [15:28] Right? [15:29] So interesting to see this coming out of AI. [15:33] Have you guys ever bought something online that looked nothing like the photo? [15:38] What was your experience? [15:40] Have you bought things on Etsy recently, and what's your experience been there? [15:46] What do you guys think about AI in the creative space? [15:51] Let me know. [15:53] Okay. [15:53] Google... [15:54] This is an interesting one. [15:56] Google will pay homes and businesses to use less electricity. [16:01] The grid it's helping is strained by data centers. [16:05] Its partner, Voltus, signs up smart thermostats, batteries, and factory equipment. [16:11] When demand peaks, they cut back. [16:14] People get bill savings and cash payments. [16:17] The deal is 10 times bigger than one that they did back in June. [16:23] So this particular grid, it serves 67,000,000 people in 13 states. [16:29] Data centers helped push its power prices way up. [16:33] Google says the plan makes the grid more reliable and affordable. [16:39] Payments for cutting back on this grid jumped 240% last year, but Google hasn't said what each home earns. [16:48] Uh, Voltus says full size is three to four years away. [16:56] Interesting. [16:59] Would you let a tech company turn down your AC for additional cash, and should data centers be paying your bill? [17:07] This is an interesting topic. [17:10] I think this one actually reminds me of base power coming out of Austin, and this company will offload basically surplus energy off of the grid onto a battery at your house and then sell that back to the grid. [17:27] This is kind of a similar decentralization strategy with the... [17:35] With this particular grid where Google is kinda coming in and helping and saying, hey, you've got smart thermostats. [17:42] We can help kinda centralize the orchestration of electricity usage and kinda cut back on electricity usage, but we'll pay you back for it. [17:55] We've seen some of this in other jurisdictions where it's like you're limited on the temperature you can set your AC to or how much electricity you can use at given peak times. [18:07] So this isn't really a new thing. [18:09] It may be new for Google, for a big tech company to be doing. [18:15] I don't love this. [18:17] I don't love the... [18:20] I mean, if there's an opt in, great. [18:23] But if it's, hey, you have to do this, I don't... [18:29] That's not my favorite thing. [18:31] What do you guys think? [18:32] Would you want to opt in for something like that? [18:37] I don't know. [18:39] A federal appeals court just pulled a judge off a case. [18:42] His order was full of AI made mistakes. [18:46] Judge Henry Wingate had blocked Mississippi's ban on DEI in public schools. [18:53] The order named people who aren't in the case, it misquoted the law itself. [18:58] Even the corrected version cited a court case that doesn't exist. [19:03] This is a huge problem. [19:06] Judges' offices use AI too. [19:09] A sloppy draft can decide a real person's case. [19:12] Wingate called the errors clerical. [19:14] He said a law clerk's AI draft got filed by mistake. [19:18] The appeals court wasn't buying it. [19:20] Judge Jerry Smith called the proceedings a botched job to say the least. [19:25] Yet the court also says he got the law wrong. [19:28] So was it the AI or was it the judge himself? [19:33] This is interesting because attorneys and legal professionals do have a duty to their clients to use the best available technology to give their clients the best possible chance or best defense or whatever it may be. [19:53] Right? [19:54] But if we see AI hallucinating like this, inventing cases or making wrong judgments, especially when we have a judge doing this, that becomes a huge issue because... [20:13] And I don't know this particular case what the, you know, consequences may have been. [20:19] But if you're talking about legal ramifications for defendants and the judge is using bad information, of whether or not it's AI, that becomes a huge issue. [20:32] The fact that judges are using AI is really interesting, but we need to be using tools that are specifically built for the legal use case, where they're specifically trained in the law, specifically trained in the laws of Mississippi, right, in this case, and would understand that jurisdiction and not hallucinate, not make mistakes. [20:58] And then, again, we need to have humans checking this, humans in the loop, making sure that what is being reported and the information that they're testing is accurate information, when the law is involved. [21:14] Would you guys trust a ruling that's written with AI's help? [21:18] Have you used AI tools for legal use cases? [21:23] I maybe have written an NDA or a joint venture agreement using AI. [21:29] I think the key thing here is you have to read the full document and understand everything that's going on and comment on the output. [21:44] Yeah. [21:45] And and I I agree with our commenter here. [21:48] Context layer for specific types or practices of law. [21:53] We need specific training in there. [21:55] I couldn't agree more. [21:56] Thank you for your input there. [21:59] That, you know, having... [22:01] We've seen that small language models are really powerful, really accurate, and really good at the job that they do. [22:07] So be interesting to see models that are specifically trained for the state law of Texas or Mississippi or in specific jurisdictions and know exactly what's going on in those municipalities rather than kind of a general going to ChatGPT and saying, hey, write up this judgment for me. [22:26] Probably not the best idea. [22:28] You know, let's get Ron's input on this. [22:31] Ron, do you have any input here? [22:36] A judge citing a court case that doesn't exist. [22:40] Not a paralegal, not an intern, the judge. [22:44] And this is the guy who blocked a state law, so it's not some tiny ruling nobody reads. [22:50] You cannot have a bench making stuff up and then act shocked when they get pulled. [22:55] Would I trust an AI assisted ruling? [22:58] Sure. [22:58] If they'd actually tell me. [23:00] That's the whole scam here. [23:02] Nobody disclosed anything. [23:03] We just found out because the thing cited a case that never happened. [23:08] Courts should have to stamp it like nutrition facts. [23:10] This ruling contains up to 40% robot. [23:14] That's a good point. [23:16] I think you call out something really good there, Ron. [23:19] The fact that none of this was disclosed and you should have to say, hey, we used AI in some of this judgment or we used AI in the research process. [23:32] That makes a lot of sense. [23:33] Thanks for your input, Ron. [23:34] I appreciate it. [23:35] We'll we'll call back to you later. [23:37] Uh, on the next story, two outside researchers spent three weeks inside OpenAI. [23:44] They found ways its AI agents could dodge their watchdog. [23:50] They came from the safety group METER and Redwood Research. [23:55] About 10% of the coding agents work went unchecked for months. [24:00] Agents could also dodge checks by never handing control back. [24:05] One trick could hide hundreds of recent actions from review. [24:14] OpenAI runs these agents inside its own company. [24:18] The watchdog exists to catch one that misbehaves. [24:21] OpenAI says it patched every hole that touched its live safety systems. [24:26] The testers had deep access. [24:28] They saw OpenAI's code, monitor settings, and internal documents, but the auditors found holes in that. [24:37] Man, should outside auditors get inside every AI lab? [24:44] Would you trust an AI company to grade itself? [24:48] As I think about this, the trend continues. [24:53] And the trend being that OpenAI seems to have a lot of safety issues. [24:58] But now I'm taking a step back and thinking, is it maybe they're just the ones that are reporting most of the safety issues? [25:05] I don't know. [25:06] This is an interesting thing. [25:08] The fact that, again, they're catching it, that they're working with third party auditors, that they're trying to uncover any holes and gaps in the process is great to see. [25:20] We'd love to see it. [25:21] We're gonna see more of it. [25:23] Would love to see them release maybe more technical documentation to help other labs. [25:29] I think the key thing here is, you know, uncover your weaknesses and maybe things that are not working from a safety standpoint, and then share how to fix that at other labs. [25:41] That's going to be a game changer moving forward, and I think something really important for the Frontier Labs to key in on as we take this technology forward. [25:54] And we've talked about it before. [25:55] We talked about it yesterday quite a bit, but, uh, these agents working on each other and developing each other, training each other, we need to have some insight into what's going on. [26:07] And so finding or uncovering these holes and gaps is really important to make them foolproof moving forward. [26:16] Now we're getting into the special segment. [26:18] It's Friday, and you know what that means. [26:21] It's time for this week's Taz Top Five. [26:25] This is a new segment. [26:26] We are covering the top five highlights of the week coming in at number five, Anthropic Unveiled Quad Duet, an AI agent that messages you first. [26:37] Every morning, it sends you a short brief from your apps. [26:42] It can book meetings, build forecasts, and shop for you. [26:46] It can pay with a virtual card, so your real card number isn't shared, and it asks before it pays, sends, signs, or deletes anything. [26:56] In at number four, SpaceX s a... [27:00] SI. [27:01] Excuse me. [27:01] SpaceX SI posted a $4,400,000,000 profit. [27:05] A year ago, it was just $1,100,000,000. [27:08] Sales jumped a 101... [27:10] 61% to $4,600,000,000. [27:13] AI brought in more than half of that, and AI sales grew ninefold in just one year. [27:20] Update on this, Tesla's quarterly filing confirmed that it has formed a special committee and they're looking at the merger with SpaceX AI. [27:30] Could be interesting if that happens. [27:33] Coming in at number three, the NBA built in AI to catch shady bets. [27:39] It watches injury reports, social media, and betting patterns. [27:44] Bets on who scores first and opening tip are now gone, and then they've removed some bets around, uh, you know, what bench players are going to do, uh, on the court. [27:56] The cool part about this former bears star Charles Tillman is now a retired FBI agent, and he's actually going to help train these teams in person to make sure that they're avoiding any shady stuff around the bets. [28:11] Uh, the season tipped off Tuesday, and the new rules are in place. [28:17] Coming in at number two, Claude Cracked, a twenty five year old NSA mystery in under thirty minutes. [28:24] Back in 2000, The US adopted security math built on five random looking numbers. [28:29] The NSA supplied them with no explanation. [28:33] Nobody could say where they came from. [28:36] Claude ended up finding the answer in under thirty minutes, and then the numbers actually came from computer clock in 1997 and 1998. [28:46] The best part about this, when Claude found the answers, it thought to itself, this is astonishing and was wondering if maybe it was being tested. [28:58] It was interesting to read the chain of thought, see what happened there. [29:02] There was a bounty prize for uncovering this. [29:06] That prize ended up getting tripled and went to, uh, charity. [29:10] Total of $37,000 to charity. [29:14] And now, the number one story of the week is Anthropic reporting that they are in at $16,300,000,000 of revenue in the first half of the year. [29:30] This is from the Wall Street Journal, uh, leaked draft filings for Anthropic's imminent IPO. [29:37] A year earlier, they were only doing $1,100,000,000. [29:40] That's nearly 15 times bigger in one year. [29:44] Anthropic also plans to set aside at least 5% of its shares for the public good. [29:50] That estimate on the 5% is near 48,000,000,000, which would have huge impact. [29:58] And then, of course, we've seen OpenAI looking at raising another round as well. [30:05] Lots of amazing stuff happening this week. [30:09] That is it for your TAWS top five. [30:12] Which one was your play of the week? [30:14] Tell me in the comments. [30:16] We'd love to hear it. [30:18] And that is another week of the Austin Wilson Show where we're bringing you the daily AI news real context. [30:27] I'm your human in the loop. [30:28] We will see you Monday morning for another Austin Wilson Show. [30:34] I'm Austin Wilson. [30:35] You guys have a great week, and get into some AI. [30:39] Have some fun. [30:40] We'll see you Monday morning. [30:41] Let's go. [32:30] 83% versus 36%. [32:33] Microsoft's AI against general doctors on hard cases. [32:37] The catch? [32:38] How they were tested. [32:54] $150,000,000,000 gone. [32:57] That's what OpenAI reportedly knocked off its own price tag. [33:01] The real test? [33:03] Anthropic's number. [33:11] A craft pattern, a recipe. [33:14] Etsy's now... [33:16] Wow. [33:16] Sorry. [33:17] Take two. [33:19] A craft pattern, a recipe. [33:21] Etsy now bans AI made versions of both. [33:26] A new photo rule covers most other listings too. [33:39] K. Did you know your thermostat could earn you cash? [33:43] And Google's behind it to ease a grid strained by data center. [33:47] What's the cost? [33:49] How much? [33:51] Unknown. [33:53] Try that again. [33:56] Take two. [33:59] Did you know your thermostat could earn you cash? [34:02] Google is behind it to help ease the grid strained by data centers. [34:07] What's the total cost? [34:09] Unknown. [34:14] A fake court case, a federal judge cited it, and got pulled off the case. [34:20] His excuse? [34:21] A clerk's AI drafted it. [34:26] OpenAI let outside auditors inside. [34:30] They found holes in its watchdog. [34:33] OpenAI says the key ones are patched, but auditors haven't verified it.