AI's Great Divergence

Today on the AI Daily Brief: AI's Great Divergence. Before that, in the headlines: One of the weirdest AI pivots yet. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Blitzy, ZenCoder, and Granola. To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts. If you are interested in sponsoring the show, send us a note at sponsors@aidailybrief.ai. As I finished recording this episode, Anthropic dropped Claude Opus 4.7. The show was produced before we got that announcement. Come back tomorrow for that episode, but for now, let's talk about that weird pivot. Yesterday, AI made waves on Wall Street once again. Although the context for it may be the most absurd yet. You might remember briefly popular sneaker company Allbirds. They were beloved by many in the tech sector, and in 2021 when they went public, the company was worth over $4 billion. Their stock has since cratered 99%, and earlier this month they sold their assets and intellectual property for $39 million to a holding company called American Exchange Group, which is known for acquiring fashion brands like Ed Hardy. That left Allbirds as a largely valueless shell company, a blank canvas if you will. And on Wednesday, the company announced that their next chapter would be, drum roll please, an AI neocloud provider. They said they would be raising $50 million to fund the pivot and would be changing the company name to Newbird AI. Now, rebirthing a dying company to chase a hot new trend is not nearly as uncommon as you would think. In 2017, a beverage company called Long Island Iced Tea changed their name to Long Island Blockchain and saw a huge pop. Just kidding, the company was later delisted and charges were filed against it for insider trading. The crypto industry saw similar plays with Kodak, RadioShack, and of course, Enron. Now in the AI domain, more recently, a former karaoke machine company announced that they would be releasing AI logistics software. Cynical though the analysis may be, usually these rebrands have very little substance beyond pumping the stock, and Allbirds certainly received a solid pump. The stock soared by as much as 875% yesterday, But whether they can actually do anything, most people are fairly dubious on. The Wall Street Journal notes that $50 million doesn't get you far in the AI race, with neoclouds like CoreWeave and Nevious planning to spend tens of billions on infrastructure this year. Matt Levine sums it up: Of course, there are two levels of analysis here. One is, sure, Allbirds is pivoting its business to AI compute infrastructure. That seems like a competitive and capital-intensive business in which Allbirds has no obvious expertise, but whatever, nostalgic fondness for the sneakers, maybe it'll work out. The other level is that Allbirds is pivoting its stock to being an AI meme stock. That definitely worked out. I would say that that is a story we can safely leave behind and move into something that is much more relevant, which is that OpenAI has updated their agents SDK with a host of new features that make it easier to build enterprise-grade agents. The software developer kit now includes native sandbox integration, allowing developers to keep agents contained to particular systems and workflows. The basic gist here is that the harness is now separated from the compute layer. Meaning data can live in the sandbox rather than being jammed into context. Interestingly, this is not dissimilar from what we talked about on our Harness Engineering show in terms of Anthropic's managed agents. Both companies independently arrived at a similar architectural move— Anthropic called it decoupling the brain from the hands, while OpenAI called it separating the harness from compute. Both, however, cite the same reasons: security, i.e., credentials shouldn't live where model-generated code runs, durability, i.e., losing a sandbox shouldn't kill the session, and scale, spinning up many sandboxes per agent as needed. The new Agents SDK for OpenAI also delivers significant upgrades to the in-distribution harness, improving file access tools as well as adding memory and compaction. Overall, the release brings OpenAI's infrastructure closer to the way agents need to operate within secure systems. Karen Sharma, a member of the product team, said, "This launch at its core is about taking our existing Agents SDK and making it so it's compatible with all of these sandbox providers." Together with performance upgrades, Sharma said the goal is to allow companies to "go build these long-horizon agents using our harness with whatever infrastructure they have." One way to look at this is another example of the mad dash to translate prosumer AI products into enterprise products that can conform to security and operational standards. Steve Coffey from OpenAI writes, "This is the direction I'm excited about for agents: open harnesses that give you the flexibility to deploy your agents at scale, with your own data, on your own terms." Armand Sidhu writes, "Agents can now run in controlled environments where their access to resources, APIs, and data can be scoped precisely. This isn't for consumer chatbots. This is for enterprise deployments where you need to let an AI loose on real systems without letting it break things." Now in a very different part of OpenAI's business, The Information reports that the company is shifting their ad revenue model to pay-per-click. One of the frustrations with the early version of ChatGPT ads was that advertisers complained that they couldn't properly track performance. OpenAI's ad data was less developed than Google or Meta, so advertisers were left guessing on how well their ads were converting. OpenAI was also charging a high premium for those who wanted to participate in the early trial. The Information now reports that OpenAI will now charge only when users click on an ad. That's opposed to charging per view. They're also looking at other action-based pricing, including charging when a user makes a purchase. The goal is to de-risk trying out this new advertising medium by having the payment structure better align with the outcomes. Moving to a very different topic, the Manus investigation is casting a chilling effect over China's startup scene as founders are forced to pick a side. Earlier this year, reports circulated that the CCP was taking a closer look at Meta's acquisition of Manus. In particular, there was some suggestion that Manus' relocation to Singapore last year was a bid to circumvent Chinese tech export controls. In late March, two Manus co-founders met with Chinese officials and were informed they would not be allowed to leave the country until the investigation concluded. According to The Information's China-based reporter Jing Yang, this move has spooked Chinese founders and neutered hopes of international success. Hank Yuan, a founder working on an AI agent company, said, "If you want to build AI products for markets outside China now, you will have to think even more carefully about which markets to target, how to structure your business, and whether to raise money in Chinese yuan or US dollars." He added, "All the AI startup founders I know are paying attention to Manus." Now, until this point, there had sort of been a tacit truce between Beijing and Shenzhen. Founders could freely travel to the US to seek funding, and there was an implicit understanding that tech success mattered more than strict nationalism. Now, of course, no official policy exists, so there's no policy to change, but it still appears that founders have gotten the message. A co-founder of an AI video startup said, "Originally, we thought we had many options for exits, but now the takeaway from Manus is: if your startup is acquired by other companies, don't get acquired by US companies. If you are acquired by Alibaba or Tencent, that's fine." Now interestingly, the result isn't a total halt to Chinese founders heading to US markets—they seemingly just need to commit to picking a side. One Chinese-born founder working in San Francisco, for example, said he has now pivoted to hiring devs in Singapore rather than China. He commented, "Having a team in Singapore costs more and the quality isn't as good as having a China team, but I still don't want to build a China team. It's too risky." Which I think is interesting context for our last story, which is recent comments from Nvidia founder Jensen Huang, about the need for dialogue between the U.S. and China. Jensen was the latest high-profile guest to appear on the Dwarkesh podcast this week. And in the show, he dug in on why he believes cooperation rather than export controls is the right way to navigate the rise of AI in geopolitics. Dwarkesh framed the question around a scenario where China gets access to enough advanced chips to train a Mythos-level model and can run cyberattacks using millions of agents. Huang rejected the premise, commenting, Mythos was trained on fairly mundane capacity and a fairly mundane amount of it by a fairly exceptional company. So the amount of capacity it was trained on is abundantly available in China. You first need to realize that chips exist in China. Huang went on to explain that China has around half the AI researchers in the world, abundant energy, and chip manufacturing that is swiftly ramping up. Reframing the question then, Huang asked, "If you're worried about them, what is the best way to create a safe world? Victimize them, turning them into an enemy likely isn't the best answer." He continued, "They are an adversary. We want the United States to win." But I think having a dialogue and having research dialogue is probably the safest thing to do. This is an area that is glaringly missing because of our current attitude about China as an adversary. It is essential that our AI researchers and their AI researchers are actually talking. It is essential that we try to both agree on what not to use AI for. Now, for some, this was just Jensen talking his book, as Beth Jezos put it, securing the bag for GPU sales to China. But I think Ed Elson's more nuanced take is closer to right. He writes that he thought that what Jensen was basically trying to say is that the question isn't whether China achieves mythos-level AI, because they will. Ed writes it's whether they will use it to try to destroy America. Bringing up the nuclear comparison, Ed says, "The same question goes for nukes. China has nukes and yet they haven't nuked us. Why? Because they don't want to." The interview is certainly worth a watch, if for no other reason that Dwarkesh seems to be one of very few people who is actually willing to ask CEOs hard questions, but I will say that I don't think it's nearly as contentious and simple as social media is making it out to be. Shocker, right? If nothing else, it did give us a meme video quote, which I will use forever now. You're not talking to somebody who woke up a loser. And that loser attitude, that loser premise makes no sense to me. But with that moment of glory, that's going to do it for today's AI Daily Brief headlines. Next up, the main episode. All right, folks, quick pause. Here's the uncomfortable truth. If your enterprise AI strategy is, we bought some tools, you don't actually have a strategy. KPMG took the harder route and became their own client zero. They embedded AI and agents across the enterprise—how work gets done, how teams collaborate, how decisions move—not as a tech initiative, but as a total operating model shift. And here's the real unlock: that shift raised the ceiling on what people could do. Humans stayed firmly at the center while AI reduced friction, surfaced insight, and accelerated momentum. The outcome was a more capable, more empowered workforce. If you want to understand what that actually looks like in the real world, go to www.kpmg.us/ai. That's www.kpmg.us/ai. Want to accelerate enterprise software development velocity by 5x? You need Blitzy, the only autonomous software development platform built for enterprise codebases. Your engineers define the project, a new feature, refactor, or greenfield build. Blitzy agents first ingest and map your entire codebase, then the platform generates a bespoke agent action plan for your team to review and approve. Once approved, Blitzy gets to work autonomously generating hundreds of thousands of lines of validated, end-to-end tested code. More than 80% of the work completed in a single run. Blitzy is not generating code, it's developing software at the speed of compute. Your engineers review, refine, and ship. This is how Fortune 500 companies are compressing multi-month projects into a single sprint, accelerating engineering velocity by 5x. Experience Blitzy firsthand at blitzy.com. That's B-L-I-T-Z-Y dot com. So coding agents are basically solved at this point. They're incredible at writing code. But here's the thing nobody talks about: coding is maybe a quarter of an engineer's actual day. The rest is standups, stakeholder updates, meeting prep, chasing context across 6 different tools. And it's not just engineers. Sales spends more time assembling proposals than selling. Finance is manually chasing subscription requests. Marketing finds out what shipped 2 weeks after it merged. ZenCoder just launched ZenFlow Work. It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools. Jira, Gmail, Google Docs, Linear, Calendar, Notion. It runs goal-driven workflows that actually finish. Your standup brief is written before you sit down. Review cycle coming up? It pulls 6 months of tickets and writes the prep doc. Now you might be thinking, didn't OpenClaw try to do this? It did, but it has come with a whole host of security and functional issues which can take a huge amount of time to resolve. ZenCoder took a different approach. SOC 2 Type 2 certified, curated integrations, tighter security perimeter. Enterprise-grade from day one. Model agnostic, and works from Slack or Telegram. Try it at zenflow.free. Today's episode is brought to you by Granola. Granola is the AI notepad for people in back-to-back meetings. You've probably heard people raving about Granola. It's just one of those products that people love to talk about. I myself have been using Granola for well over a year now, and honestly, it's one of the tools that changed the way I work. Granola takes meeting notes for you without any intrusive bots joining your calls. 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, or even coach you using recipes, which are pre-made prompts. Once you try it on a first meeting, it's hard to go without. Head to granola.ai/aidaily and use code AIDAILY. New users get 100% off for the first 3 months. Again, that's granola.ai/aidaily. Welcome back to the AI Daily Brief. One of the big themes of the year is the heightened stakes around everything with AI. Obviously, we're seeing that from a technology perspective as agents come online, and then the implication of agents coming online is that it raises the stakes from a work perspective. And then of course, as the stakes get raised from a work perspective, we have the stakes raised on the politics of AI as well. And that's even before we get into all of the other AI politics issues, even beyond implications for jobs, which are becoming more and more a part of the public discourse. Now, in all of this raised stakes, part of the impact is greater divides between people who sit in different spaces relative to all of these changes. And by that, I mean everything from the difference between leaders and laggards in the corporate sphere to optimists and pessimists in the public sphere. And if you look carefully, this great divergence is showing up in all sorts of different places. We're looking at two of them today in recent studies that have come out, with the first being the annual Stanford Artificial Intelligence Index report. This annual report comes out of the Stanford HAI, or their Center for Human-Centered Artificial Intelligence, is generally seen as a very comprehensive and high-level look at the state of AI, both internal to the industry as well as where it sits in society. And this here tells the divergence story in very clear terms. The report itself is massive, something like 420 pages long. And all across the headliner topics, you see this divergence. On their website summary, one of the big themes that they point to is AI experts and the public having very different perspectives on the technology's future. So let's talk about some of these gaps. A representative gap that they point to is the difference in the way that experts versus the general public view AI's likely impact on how people do their jobs. When asked how AI would impact how people would do their jobs, 73% of experts expect a positive impact compared with just 23% of the public. When expanded out, this gap between experts and the general public shows up all over the place. In addition to that gap we just heard about in terms of how people do their jobs, the economy more broadly sees a similar gap. 69% of AI experts say that AI will have a positive impact on the economy over the next 20 years compared to just 21% of US adults. Medical care is where the general US public is the most optimistic, with 44% saying that AI will have a positive impact, but that is still vanishingly smaller than the 84% of AI experts who say that. On K-12 education, it's 61% optimism for the experts versus 24% for US adults. And pretty much everyone thinks it's going to be bad for elections, with just 11% of AI experts saying that AI will have a positive impact on elections, which is their closest number to the general US public, of whom only 9% think that it will have a positive impact. And other parts of the study show pessimism in more acute ways. When asked whether AI will create or eliminate jobs, almost a full two-thirds of US adults believe that it will lead to fewer jobs. Although perhaps surprisingly, 39% of AI experts also think that it will lead to fewer jobs. Another interesting area of divergence is the gap between formal education for AI and informal education for AI. Stanford points out that while over 80% of US high school and college students now use AI for school-related tasks, only half of middle and high schools have AI policies in place, and just 6% of teachers say that those policies are clear. Basically, everyone is getting their AI skills outside of the formal classroom setting, and of course reporting them on LinkedIn. One area where AI is not diverging is in the performance of top US versus Chinese models. In fact, it would be much more accurate to call that a convergence, although we'll have to see if that remains once we actually get Anthropic's Mythos and OpenAI's SpUD. Staying on AI's performance for a moment, Ethan Mollick has often referred to AI as having a jagged frontier. Basically, it can be massively good in some things, including really hard things, and be just pathetically awful at some other things that it seems like it should be good at at the same time. This is actually one of Stanford's big takeaways as well, where AI models can win a gold medal at the International Math Olympiad, but not reliably tell time. Now, this jagged capability frontier can also lead to jagged adoption, especially inside the enterprise, as organizations have to individually figure out where AI does and doesn't fit within what they do. One important area of divergence that is obviously very top of mind for people Stanford sums up as productivity gains from AI appearing in many of the same fields where entry-level employment is starting to decline. They write, studies show productivity gains of 14 to 26% in customer support and software development. And in areas like software development, where AI's measured productivity gains are clearest, US developers ages 22 to 25 saw employment fall nearly 20% from 2024, even as the headcount for older developers continues to grow. And so here we're seeing not just divergence between productivity gains and employment, but actually divergence between different types of employment with early-stage employees going one direction and older employees going the other. Now, if Stanford is showing this story of divergence on the very biggest macro levels, AI's great divergence is also very acutely captured at the enterprise level by a new study from PwC. The study is PwC's annual AI performance study, and the headliner stat is that around 75% of AI's economic gains are being captured by just the top fifth of companies. This is one of the clearest indicators I've seen yet of the difference between leaders and laggards when it comes to corporate AI adoption. This comes from a study that interviewed more than 1,200 senior executives who PwC says are primarily at large publicly listed companies. And what's really interesting about this study is that the difference between efficiency AI and opportunity AI, which we talk about fairly regularly on this show, is on full display. And now by way of reminder, efficiency AI is my term for companies that view AI as a way to do the same with less, basically whose primary interest is in having the same amount of output with less resource input. Opportunity AI, on the other hand, is the idea not of doing the same with less, but of doing more with the same or way more with a little more. Basically, that recognizes that the real opportunity with AI is to go harness new opportunities, do things that weren't possible before, get into new orthogonal fields, release new products, do more R&D. Grow towards the future rather than make the present more efficient. And boy, is that on display in this PwC study. They found that leading organizations were twice as likely to redesign workflows to incorporate AI rather than simply adding AI tools. They found that leading companies were approximately 2 to 3 times more likely to use AI to identify and pursue growth opportunities and reinvent their business model. They sum up, the research shows that these top-performing companies are not simply deploying more AI tools. Instead, they are using AI as a catalyst for growth and business reinvention, particularly by pursuing new revenue opportunities created as industries converge while building strong foundations around data governance and trust. Now, interestingly, one might think that this is all about just using AI for more, and certainly that's part of it. The companies in their survey that had the best AI-driven financial outcomes were twice as likely to be executing multiple tasks within guardrails and about twice as likely to be allowing AI to operate in autonomous, self-optimizing ways. They were increasing the number of decisions made without human intervention at almost 3 times the rate of their peers. And yet the story is a combination of automation but also governance. These leaders were 1.7 times as likely to have mechanisms such as responsible AI frameworks and 1.5 times more likely to have cross-functional AI governance boards. In addition to doing more with AI, the employees of these leaders are twice as likely to trust AI outputs as those from the laggards. Overall, PwC found that the companies that were the most AI fit in their research delivered AI-driven financial performance that was 7.2 times higher than other respondents' performance. As AI continues to proliferate through society, we're going to continue to see these kinds of divergences. In some cases, particularly in the areas of policy, divergence can actually be helpful. It can inspire better debate, and if we have the right systems in place, better, more considered action. In some areas, however, the divergence is dangerous. Divergence, which turns into underperformance, can threaten individual employees and organizations as a whole. That's gonna do it for today's AI Daily Brief. Appreciate you listening or watching, as always, and until next time, peace!