Go back

Why You Should Try Gemini Neon (Even if You Love Fable)

26m 44s

Why You Should Try Gemini Neon (Even if You Love Fable)

The episode centers on Anthropic's record-setting IPO, which raised $107 billion and made it the seventh-largest publicly traded company in the world with a $1.8 trillion market cap. The stock closed up 17% on its first day, though analysts debated whether the listing marked the end of the broader AI trade and raised questions about Anthropic's unusual governance structure, which grants co-founders outsized control. CEO Dario Amodei used the moment to stress AI safety and the potential for medical breakthroughs. The larger surprise was Google's new Gemini model, Neon, which launched quietly before the IPO but quickly won over skeptics. Despite being described as a smaller model, Neon topped benchmarks against Claude Opus 5-5 and GPT-6-2-Sol in most categories, including a strong showing in coding, historically Google's weakness. Reviewers highlighted that Neon's outputs feel genuinely different from OpenAI and Anthropic models, offering fresh perspectives, though it was criticized for being overly agreeable and token-intensive. Other news included AMD's 7% stock drop following a short-seller report on China export issues, new U.S. export controls alarming Chinese tech firms like Tencent and Alibaba, the departure of the White House cybersecurity director, and reports that Apple is developing a proactive AI agent while restricting rival agents on macOS.

Transcription

5459 Words, 31275 Characters

English
Speaker 1Well, friends, it is the day after the anthropic IPO. The company has raised a record-setting $107 billion. The stock ended up 17% in trading on the day, although it was not able to resurrect the broader AI trade. So today, of course, in our episode, we are talking about a new Google model. Yes, for the first time in a very long time, people are not only talking about Google's new series of Gemini models, but actually using them and finding them quite good. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Robots and Pencils, Granola, and Blitzy. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. And if you want to learn more about sponsoring the show, send us a note at sponsors at ai-daily-brief.ai. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. Well, friends, the big day has come and gone, and Anthropic is now a publicly listed company. The stock closed its first day on the NASDAQ at $97.17, up 17% from the IPO price of $83. That was enough to make Anthropic the seventh largest company in the world with a market cap of $1.8 trillion, slightly ahead of Saudi Aramco and slightly behind SpaceX. Now, the top level take is that the IPO was a success. Anthropic launched into a tough market for AI stocks, got the first day pop they were looking for, and raised a total of $107 billion to become the largest IPO in history. Now, however, Anthropic has to settle into life as a public company with all the scrutiny that entails. The company began the day with another hype video focused on the potential of AI. It started with the question that Anthropic's General Communications have tended not to ask in the past, what if things went really well? The video asks, can we cure the incurable? Crack the code of Alzheimer's, change the way we see. And I don't know if that's a good idea, but I think it's a good idea. And I think it's a good idea. And I know I was not the only one to see this sort of shift in the way that Anthropic's official communications talked about this entire enterprise. Now, before ringing the bell at the NASDAQ, CEO Dario Amadei delivered a five-minute speech, largely focused on potential medical breakthroughs and AI safety. He said, I believe sufficiently powerful AI working in concert with humans could compress a century of scientific and medical progress into a decade. To state it unequivocally, we will pace the training and deployment of AI systems, taking as long as is under human control. We will not bend to either business nor political pressure. If we can show it's possible to make the right choices on safety and still build an incredibly successful company, others in our industry will be inspired to do the same. We call this a race to the top. A race to the top works best when driven from the front of the pack, and we hope and expect to do so. And yet this day was all about the market's reaction to the first public pure play AI lab. And price action can tell us a lot about how investors feel about the latest mega cap stock. The first trade registered the highest price of the day, above $110 at an implied valuation of $2 trillion. The stock plunged from there, reaching an intraday low of $93 around 1 o'clock, and then fluctuated between $94 and $100 for the rest of the day. Now this was a solid first outing on a day when stocks in general were red and the NASDAQ 100 lost more than 2%. Anthropic stayed well above the IPO price, but didn't expect the big takeoff that some might have hoped for. By way of contrast, the SpaceX IPO saw a first day peak about 1.5 hours into trading, followed up with a green second day and saw its all-time high on day before collapsing. Bloomberg is suggesting that Anthropic could see red on day two, after a week showing in after-hours trading. Bloomberg economist Anna Wong suggested that the IPO could mark the end of the AI trade, arguing that, quote, technological leadership and booming AI demand aren't translating into durable pricing power. Further, there's a building narrative that the IPO was an exit event for founders, and retail interest was mostly in flipping the IPO pop. Viraj Patel of Vonda Research noted that between $700 and $800 million in trading volume, which was higher than the $670 million of a SpaceX IPO. However, that resulted in less than $60 million in net buying around half of the SpaceX debut. Patel said, Retail are trading the stock hard, but they're taking both sides rather than a one-way bid. Now, I'm less sure that you can draw really clear conclusions from this. SpaceX came into a bullish market for AI, while any retail investor who's just watching the news has to think that the real opportunity right now is playing the market's lack of clarity. VC investors were buying hard. Sequoia, Iconic, Lightspeed, and Altimeter all bought more stock on the IPO, signaling an increase in their stakes rather than released massive pre-IPO gains. It is highly unusual, to say the least, for VC firms to buy more at an IPO, and it provides a little more confidence that they won't dump their stock once lockups end. There's also rising criticism of Anthropic's governance structure and the unusual power it grants to the company's co-founders. The co-founders have total control over shareholder votes through a golden share that votes 51% through a governance council. Co-founders can be removed from the council by the board of the long-term benefit trust, but some question that as a realistic check on their control. Professor at the University of Colorado said, The trust, since it doesn't have a financial interest in Anthropic, could theoretically use its power with 3-4 directors to keep the company on mission. One might question how likely that is to occur. Still, when all is said and done, I think what you're seeing more than anything is the inevitable surfacing of a million questions all at once. Whatever criticisms they might now face, the IPO was successful, and Anthropic is the seventh largest publicly traded company in the world. They've secured at least a year of runway to build out, compute, and train next-generation models. And while the arguments on social media will continue, for the first time, if somebody thinks the technology is overhyped or the business model doesn't work, they can just short the stock instead of screaming into the digital void. And if they believe in Anthropic's vision of the future, this age of miracles, they can buy shares in the company and have a stake in that future. Now, over in other AI market news, AMD's stock plunged by 7% after a short-seller report claimed they have their own China problem. Short-selling firm Hunterbrook Media wrote, "After the US blacklisted Chinese server maker IEIT Systems, the company kept getting newer AMD firmware, which enables AMD chips to run in servers." The report warned that new export control rules would "tighten restrictions on AMD's foreign business," claiming that as much as 20% of AMD's revenue comes from customers in China and Hong Kong. The report also claimed that AMD "depends on a Chinese partner to finish most of its chips." AMD denied the reporting, saying, "The claim misrepresents AMD's business practices and relationships." An ex-account called AI Thesis provided the counterpoint, noting that presence of the firmware doesn't prove that AMD had dealings with a blacklisted company. Further, only a tiny portion of AMD's products are advanced AI chips subject to export controls, so that 20% of revenue from China could include consumer CPUs, gaming GPUs, and all sorts of other products. In short, the Hunterbrook report seems far from a smoking gun, but what it does signal is that all chip makers, not just Nvidia, are going to face tough scrutiny over the coming months, and the markets are on edge about the industry's China problem. Chinese companies aren't too happy about the state of things either. In fact, they are sounding the alarm over new export controls that will cut them off from AI chips. On Monday, the Commerce Department announced a new rule aimed at closing the remote access loophole, which allowed Chinese AI labs to legally access chips operated from data centers in third-party countries. Throughout the week, Chinese tech giants have warned that this rule change would materially impact their progress. Tencent president Martin Lau said that the new rule could cause delays in model releases and could force a transition to domestic chips, adding, "It sets everyone back in China." Alibaba CEO Eddie Wu said the rule would "affect next year's products and services." Wu also said that Alibaba is stepping up their use of in-house chips for model training, but it sounds like that is a compromise rather than a choice. Meanwhile, Moonshot has canceled the $150 million deal to access Nvidia chips through a data center in Indonesia, the contract was set to give them access to a "sizable" cluster of top-of-the-range Blackwell Ultra B300s, and TSMC and Samsung have also announced that they've suspended some chip deliveries for Chinese customers as a result of the new rule. The restrictions came as the Chinese tech companies were already facing a revenue squeeze that makes funding a domestic AI build-out more challenging. Tencent's earnings came in under expectations this week with 8% revenue growth, and operating profits slashed to 7%. Bloomberg analysts wrote, "Tencent is unlikely to generate a financial return on its rising AI investment for at least 24 months due to the ongoing monetization challenges in China's commoditized AI sector." Alibaba is faring a little better, exceeding expectations with 10% overall growth and 53% cloud growth. Still, again, CEO Eddie Wu is forecasting a slowdown in cloud growth to between 30% and 45% for next year due to export controls. Alibaba is also expected to have a slowdown of at least 10% in the next three years. Matt Maley, chief market strategist at Miller-Tabak, said, "For Alibaba, they depend heavily on access to computing power, so this will be an important setback. It's likely going to create a situation where it will take longer for them to build an ecosystem that can compete with the most advanced American technology." You gotta think that for the people who've been cheering tighter export control rules, all of this freaking out is a good sign. Finally, over in the world of cyber, the White House director of cybersecurity is out. Adding new intrigue to the string of strange incidents, the government claims are definitely not cyber attacks. On Thursday, Axios reported that Sean Cairncross had left, replaced by Kristin Davies, who currently serves as the Pentagon's chief information officer. Their reasoning is a little squirrely. Two sources described Cairncross's departure as amicable, while two sources described his White House officials said that new chief of staff Stephen Miller fired Cairncross because he "lacked the strong cyber credentials to perform the job." Now, there are a few possible ways to read this personnel change. Cairncross was part of the triad of officials reportedly leading AI policy over the past year, alongside Treasury and the Senate. Cairncross is also part of the triad of officials reportedly leading AI policy over the past year, alongside Treasury and the Senate. Cairncross's removal could then signal that the hawkish faction led by Secretary Besson has won the internal dispute over AI policy direction. It's also possible that Cairncross was cut over the government's response to the ongoing "not a cyber attack" plaguing critical systems this month. There have been a few additional events since we last caught up on the topic, which were indeed acknowledged as cyber attacks. JPMorgan Chase disclosed that that a third-party vendor was the target of an attack, resulting in the potential theft of customer names and email addresses. They said, however, that this was unrelated to Monday's banking outage. In addition, the European Commission disclosed a hack that compromised staff data and left payroll and other systems inoperable. European officials said that the attackers used Artex, an open-source agentic hacking tool that's been growing in prominence over recent weeks. The FBI issued a warning about the tool last Friday, saying malicious actors have used the tool to target critical systems since at least August. Now, in terms of what this might have to do with Karen Cross's firing, if there was a lack of confidence in his technical skills, this could be an opportune moment to find a more competent replacement. Kirsten Davies certainly has the right type of resume. Before joining the Pentagon during the Trump administration, she was the chief information security officer at Unilever, one of the largest consumer goods companies in the world. She also has experience in cybersecurity across the banking, electronics, and industrial sectors. Whatever triggered the shift, this likely indicates a change for AI and cyber policy in this administration. Accepting the role on Thursday, she posted, our opportunity is to dominate the cyber landscape for generations to come, but only if we modernize and drive resilience into the foundations our great nation runs on, as well as build cyber talent for our future. I would assume lots more to come on that front, but for now, that is going to do it for the headlines. Next up, the main episode. A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes. Researchers studied more than 500. early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs. These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at kpmg.com slash us slash AI amplifiers. At this point, it's no longer a question of whether companies are actively using AI. Using it well, on the other hand, is a problem. And it's a whole different story. Robots and Pencils, though, is a company that I can point to that is actually built for this time. They're an applied AI engineering firm working directly with clients on problems that matter to the business, not experiments that live in a slide deck. Every engagement starts by working backwards from the outcome a client actually needs. If you're trying to tell real AI engineering apart from noise in this space, that's the difference maker. Head to robotsandpencils.com. You're in back-to-back meetings all day. You're trying to stay present, but you're also worried you'll forget the decision, the action item, the impact. This is a scenario that we absolutely live here at the AI Daily Brief, given how fast things move and change and how quickly we're evolving everything that we do. That's where Granola comes in. Granola is an AI-powered notepad for meetings. You jot down rough notes like you always do, and in the background, Granola transcribes and turns them into clear, useful notes when the meeting ends. No bots joining your calls, no distractions, just a clean notepad that helps you focus. During or after the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, and so on. Or even coach you using recipes, which are pre-made prompts. Once you try it on a first meeting, it is very hard to go without. Head to granola.ai.ai.ai and use code AI Daily to get 100% off for three months for new users. That's granola.ai.ai.ai and code AI Daily to get 100% off for new users for three months. Blitzy's understanding of massive code bases unlocks autonomous security fixes, modernization, and new features. So what happens when there's no legacy code at all? Greenfield is supposed to be the easy part. Clean slate, no technical debt. But even Greenfield moves at human speed, one sprint at a time. Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch. Hundreds of thousands of lines of production-ready code. One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into two weeks. Another shipped an entire application with no front-end engineer. Legacy or Greenfield, the answer is the same. Software at the speed of compute. Build what's next at blitzy.com. That's B-L-I-T-Z-Y dot com. Welcome back to the AI Daily Brief. Boy, if you had asked me a month and a half ago about what I thought the episode the day after the anthropic IPO would be about, a Google model actually challenging open AI and anthropic models for supremacy at the top of the table wouldn't have been anywhere on the list. And yet, here we are. On Wednesday, the day before the IPO, Google dropped the fourth model that they've shipped in the past two months. It followed argon, carbon, and hydrogen, and given everything else going on, was basically completely overshadowed. Even outside Jensen Huang, in Anthropix IPO, people were, I think, reasonably skeptical. Former open AI researcher Gabriel Pedersen wrote, Why is Google shipping a model every two weeks and then nothing happens? Never ship if you know nothing will happen. Google could drop Opus 8 tomorrow and it would take the world a month to notice because we're just so used to another model. Also, no one cares about benchmarks. And yet, the folks jumped in to say, No, man, this is one you gotta look at, responded Theo. This one's different, in my opinion. I think Neon's the first time they made a model that they genuinely like using, so they rushed it out because they wanted to see if it was just them or not. And one by one, the skeptical takes got converted. Factory ambassador Chidi wrote, I don't see the point of Neon. It's 25% cheaper, that's it. If it's not more reasoning token efficient by a decent margin, and rather more token inefficient, then it will end up being the same cost slash more expensive, no? Only to respond a day later, Never have I been so happy to be so wrong. Neon is insane and basically just a cheaper, different, better question mark? Opus 5-5. Entrepreneur Siki Chen wrote, Wow, Alphabet absolutely cooked with Gemini for Neon. It is by far the most human-sounding, non-slop writing model available today. I am blown away. Google is finally back in it. YouTuber Max Blade added, I don't think people understand what is happening right now at Gemini. Neon is, shockingly good. Their models went from unusable to frontier intelligence with delightful personality. How does this happen? I'm having so much fun and I still have not found the ceiling of capabilities. So what is this model actually? After Google shipped Gemini 4 Argon and Carbon in quick succession, most assumed that Neon was just another snapshot of the same model with a little more reinforcement learning applied. But that's not actually what the model is. Google says the model brings, Frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work and science workflows. The labs aren't really doing dedicated coding models anymore, but this is something of the equivalent for coding, knowledge work and agentics. Tianfu Fu, a researcher at DeepMind posted, "Neon is actually one of our small models. We're positioning it in the same tier as Opus and Sol, even though its performance is already approaching Fable and Astra. We'd rather be conservative with our claims than over-hype it." This was in response to Google's benchmark comparison excluding Astra and Fable and was later deleted, so make of that what you will. So it seems like this is a small model, a smaller model class to Argon and Carbon, but able to outperform them on the benchmarks. And in that class of GPT-6-2-Sol and Claude Opus 5-5, at least by the benchmarks, Neon is now the top performing model across the board. It beat Opus 5-5 and 6-2-Sol in every category other than coding, where it's close enough that one could reasonably call it a draw with Opus. Neon was around a point behind Opus on Terminal Bench and Frontier Code. However, it beat Opus by more than seven points on Cursor Bench. On Agentics, Neon was a massive 14 points ahead of Opus on Automation Bench, and almost three points ahead of Sol on Agent's last exam. I'm not going to go through all of the other benchmarks, but in almost every case Neon beat its rivals. Testing from artificial analysis confirmed the results and then some. Neon scored 61 on the Intelligence Index, placing it one point ahead of Gemini 4-Carbon, one point behind GPT-6-2-Astra, and four points behind Fable 5.2. It's now the new leader for models of its class, beating Opus by two points and GPT-6.2-Sol by four points. The coding scores are where it gets a little wild. Neon is now the new leader on the Coding Agent Index, beating Fable 5.2 and Opus 5.5 by a point. It's six points ahead of 6.2-Astra and nine points ahead of 6.2-Sol. Keep in mind that coding has been the historic Achilles heel for Google models. Neon is quite token intensive. AA found it used 15% more tokens per task than Carbon, but it's also 40% faster than Carbon, although still less than half the speed of OpenAI models. It also wasn't super cheap at 6.44 per task, although that does put it comfortably in middle of the range. Now, in terms of the tests, we got all the standard barrage of render tests via Blender and people building games. Daria Anutmaz wrote, "Google's new AI model, Gemini 4 Neon, is shockingly good. I just built this tower defense game in about 20 minutes and it plays flawlessly and only used 2% of my five-hour limit. Absolutely unbelievable." But it also seemed to do well on real work. Pavel Hurin has a bug hunt bench where he has two real repos with 105 planted bugs. Gemini 4 Neon on Extra High was now the top of that board, having found 63 of the 105 bugs, ahead of 6-2 Astra and Fable 5-2. Now, I have not had a ton of time to go as deep as I'd like using Neon, and I certainly haven't set it up to use it in Google's native anti-gravity harness, although I am seeing lots of folks actually downloading Google anti-gravity to do exactly those sort of tests. But the thing that people are saying about Neon that I have also found in my initial tests and that I think is incredibly important and gives it some very specific utility is that they don't feel like just different versions of an open AI or anthropic model. Because Google has, let's be real, been a little out of the game this year, the social media temptation to claim big hyperboles like "better than Fable and Astra" are running very high. But that's not the reaction that I think you should be caring about. Alex Finn wrote it this way: "Clawed and open AI models feel like they were trained on the same data. They feel like they are running the same race. The Gemini models feel like they were trained on a completely different data set. The way the Gemini models answer questions comes up with solutions, analyzes my thinking, is just completely different than the way Claude and ChatGPT do it. I can ask all the different Claude and ChatGPT models for an opinion, and while their style would be different, I'd get roughly the same answer. Gemini will give me a completely novel, different way of thinking. I was recently going through a challenge in my life, and while ChatGPT and Claude would feed me advice that felt sanitized and AI-like, Gemini gave me a completely different perspective that led to a breakthrough. At the very least, I think you should be running your challenges through Gemini as a second opinion just to get fresh new perspectives that I believe ChatGPT and Claude can't provide at the moment. The models aren't perfect, there are still major improvements that need to be made. The Google harnesses are still way behind Anthropic and OpenAI. I know the anti-gravity team and think they are awesome, but the harness just isn't there yet. They also don't have a great agent harness either to the level of Duet or Dots. And I don't love the way that Gemini models communicate when doing agenda coding. They tend to get too technical when I'm just looking to hear about the high-level solutions. But from a straight-up intelligence and taste perspective, you need to be using these models. And this is the part that in my early tests, I have also found. That the answers that you get aren't just slightly different versions of the same answer that you'd get with ChatGPT or Claude. Now, where I'm unwilling to jump in at the same level as a lot of these folks is to say that they're better, but I don't think they have to be better for different to be useful. To take an example, in the early stages of how I'm planning How We Got Here episodes, the first step for any episode is to figure out why it deserves to exist. I've had an episode that I've been planning since the beginning about why oil still runs the world. And this is an example of where there is so much that I could get into that finding the hook and the through line and the reason that it matters and is going to matter to people is actually the challenge. A lot of my first iterations of the outlines felt like they were basically glorified Wikipedia pages, just lists of facts in a reasonable sequence that would have certainly informed people, but ultimately probably not have had enough of a narrative cadence to get most people to stick with it. All the variations and iterations I worked on with open AI models and with anthropic models were still getting really hung up on what was already on the page. They couldn't really get out of that Wikipedia mode. Coming in a little bit more raw though, Neon was able to suss out some parts of my argument and reconstitute them in ways that actually jogged me into thinking about the episode in a slightly different way that eventually led to the outline that I actually think is good. And again, to be clear, this does not mean that I think that Neon is better than those other models I used, but it's an example of how I've been able to get more out of it. It's an example of where it's not the betterness, but the difference itself that is the value. I do have one big complaint with Neon so far as well though. For as long as I've been using AI, one of my major use cases has been brainstorming, strategizing, thinking through what I'm going to do, planning out problems, planning out content, etc. It was really difficult in the 24 generation of models because they were so sycophantic that you couldn't actually get real feedback on any idea. They just agreed with you entirely or were only giving you pushback because you said, you wanted pushback. That started to shift a bit around 03 when the model labs got better at training some amount of the sycophancy out of the models. I'm not sure that I would say that Gemini Neon is sycophantic, but it is maybe the biggest pushover I've seen in a model for two years. If I push the slightest amount on some idea it has presented to me, it will instantly cave and regale me with how wrong it had been just 30 seconds ago. And then if I respond and say that I think it went too far in its critique, I'm going to say, I don't think it went too far in its critique of itself, it will 180 right back to the beginning and lament how quick to dismiss its own perspective it was. And unfortunately, what this means is that while it is useful for giving a different perspective, you kind of got to hope that the one-shot first example of that perspective is useful because it gets so wishy-washy from there. Still, this is more positive an experience that I've had with Google models since the middle to end of 2025. And I think that is extremely encouraging. For me, the next set of tests are obviously to actually go and try to build things with it to see how it feels and what it does compared to both Cloud Code and Codex. But I'm pretty excited to have another model back in the rotation. And it sounds like Google is not done yet. Now one more small final story, just because it felt aligned with the whole sleeping giant finally wakes up kind of theme. Apple is apparently and perhaps unsurprisingly, working on a personal agent to match the big labs. Earlier this year, of course, Apple rolled out their AI enhanced version of Siri. But to the shock of basically no one, it was underwhelming. Now Bloomberg's insider Mark Gurman reports that they are working on something better. The reporting focused on multiple job listings added this week. One says that Apple is seeking a quote, senior engineering manager for a proactive agentic AI assistant. The system will write code, browse the web, control apps and navigate devices on a user's behalf. Gurman had previously reported on these plans. But the job listings add verification in a sense of urgency. Apple has no doubt noticed that AI agents are flooding their app store. And Meta's Muse has overtaken ChatGBT in the free app rankings. Describing the product, the job listing reads, imagine an AI agent that can use a computer the way a capable person does. It browses the web, works across apps, runs commands, writes and code and call services. And it chooses the right tool for each step. It does all this privately and with the polish people expect from Apple. People already do their most valuable work by moving fluidly between tools, a browser, a document, a terminal, an editor, a service. We're building AI agents that can do the same. New CEO John Ternes is reportedly pushing software chief Craig Federighi to move faster on AI and adopt a more ambitious approach. And based on everything that we know right now, we might see the fruits of those efforts as early as next year. Now, on the one hand, I think folks are pretty excited about that. But there is another implication, which is not so good. Mark Gurman also posted, as Apple develops its own computer use agent, it's making it harder for rival agents to control Macs. In Mac OS 27.3, enabling ChatGBT's third-party computer use support now requires 14 distinct steps up from Mac OS 27.3. And that's because it's not the only way to get users to use the Mac OS 27.3 or Mac OS 27.1 and Mac OS 27.2, or Mac OS 27.3, or Mac OS 27.3. It could be that Apple's owning the hardware is too big a moat for them to give up easily, even if it makes our consumer choice a little bit more difficult. Still, overall, I think having Apple back in the agent mix, or I guess in the mix for the first time, potentially, and Google very much back in the model mix, is a net great thing for consumers. So if you are looking for something to do this weekend, might I suggest going and wrapping with the new set of Gemini models. For now, that is going to do it for today's AI Daily Brief. Thanks for listening or watching, as always, and until next time, peace!

Podcast Summary

Key Points:

  1. Anthropic completed the largest IPO in history, raising $107 billion and closing its first trading day up 17% at a $1.8 trillion market cap.
  2. Anthropic CEO Dario Amodei emphasized medical breakthroughs and AI safety in his remarks, describing a "race to the top" strategy.
  3. Google's new Gemini model, Neon, surprised skeptics and emerged as a top performer, beating rivals like Claude Opus 5-5 and GPT-6-2-Sol on most benchmarks.
  4. Reviewers praised Neon for offering a genuinely different perspective from OpenAI and Anthropic models, though it was criticized as overly agreeable and token-intensive.
  5. AMD shares fell 7% after a short-seller report alleged export-control issues involving China, which AMD denied.
  6. New U.S. export control rules closing the remote access loophole are alarming Chinese tech giants like Tencent and Alibaba, with Moonshot canceling a $150 million Nvidia chip deal.
  7. The White House cybersecurity director Sean Cairncross departed and was replaced by Kristin Davies, possibly signaling a shift in AI and cyber policy.
  8. Apple is reportedly developing a proactive agentic AI assistant, while making it harder for rival agents to control Macs.

Summary:

The episode centers on Anthropic's record-setting IPO, which raised $107 billion and made it the seventh-largest publicly traded company in the world with a $1.8 trillion market cap. The stock closed up 17% on its first day, though analysts debated whether the listing marked the end of the broader AI trade and raised questions about Anthropic's unusual governance structure, which grants co-founders outsized control. CEO Dario Amodei used the moment to stress AI safety and the potential for medical breakthroughs.

The larger surprise was Google's new Gemini model, Neon, which launched quietly before the IPO but quickly won over skeptics. Despite being described as a smaller model, Neon topped benchmarks against Claude Opus 5-5 and GPT-6-2-Sol in most categories, including a strong showing in coding, historically Google's weakness. Reviewers highlighted that Neon's outputs feel genuinely different from OpenAI and Anthropic models, offering fresh perspectives, though it was criticized for being overly agreeable and token-intensive.

Other news included AMD's 7% stock drop following a short-seller report on China export issues, new U.S. export controls alarming Chinese tech firms like Tencent and Alibaba, the departure of the White House cybersecurity director, and reports that Apple is developing a proactive AI agent while restricting rival agents on macOS.

FAQs

Anthropic raised a record $107 billion and closed at $97.17, up 17% from its $83 IPO price, making it the seventh-largest company in the world with a $1.8 trillion market cap.

Neon is a new, smaller-class Google model that outperformed rivals like Claude Opus 5.5 and GPT-6.2-Sol on most benchmarks, including becoming the leader on the Coding Agent Index. Users praise it for offering a genuinely different perspective from OpenAI and Anthropic models.

It is described as very sycophantic or a pushover, quickly caving to any user pushback and reversing its position. Google's harnesses and agent tools are also seen as behind those of OpenAI and Anthropic.

A Hunterbrook Media short-seller report claimed AMD had a China problem involving firmware supplied to a blacklisted Chinese server maker. AMD denied the claims, and critics noted the report was far from a smoking gun.

The Commerce Department closed a remote-access loophole, prompting warnings from Tencent and Alibaba about delays and reduced cloud growth. Moonshot canceled a $150 million Nvidia chip deal, and TSMC and Samsung suspended some deliveries to Chinese customers.

Kristin Davies, formerly the Pentagon's chief information officer, replaced Sean Cairncross. Reports say chief of staff Stephen Miller fired Cairncross for lacking strong cyber credentials, though the change may also reflect AI policy disputes or the response to ongoing cyber incidents.

Chat with AI

Loading...

Pro features

Go deeper with this episode

Unlock creator-grade tools that turn any transcript into show notes and subtitle files.