Why AI Hasn’t Increased Unemployment, According to Anthropic

Today on the AI Daily Brief, why AI hasn't increased unemployment according to Anthropic. Before that, in the headlines, the router business is hot as Stripe is in talks to buy OpenRouter for $10 billion. 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, Blitzy, and Airtable. To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe at Apple Podcasts. And to learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around 5 minutes. Although today I do not think everything that we have to discuss is going to fit in 5 minutes, so let's dive in. First of all, Stripe is the latest company getting into model routing with a potentially blockbuster acquisition on the table. The Wall Street Journal reports that Stripe is in talks to acquire OpenRouter for around $10 billion. That would be a huge markup from OpenRouter's $1.3 billion valuation during their last round, which closed— check swatch— 2 months ago in May. Then again, a lot has changed in those 2 months. We went from the token maxing era where everyone was encouraged to use the most powerful model as much as possible to the age of token scarcity, where increasingly enterprises are moving to more tightly controlled token budgets. In the context of that shift, it's beginning to look like the best token routing service could be a huge winner. Reportedly, OpenRouter has been fielding multiple acquisition offers, but Stripe is looking like the company with the deepest pockets. Taking a step back, I think the pairing makes a lot of sense. Stripe has more or less come as far as they can go with merchant-side payment processing, and clearly their eyes have been getting bigger and bigger as they think more creatively about their growth. They've recently pursued a merger deal with PayPal that would let them expand to the consumer side of the market. Meanwhile, an OpenRouter acquisition would let them move in a different direction, adding enterprise cost control tools to their vertically integrated stack. The Journal suggests the deal is close and could be announced soon. Macaroni Capital writes, Stripe isn't buying an AI company, it's buying the metering and billing layer for inference plus the developer funnel attached to it. Colin from Clerk.com writes, Stripe has two angles. One, increase the GDP of the internet, and two, the less discussed, increase their margins on the GDP of the internet. OpenRouter has shown that their margin on inference is durable, and we all know inference is a massive, startlingly fast-growing portion of the internet's GDP. Presaging something that I am sure we will talk about on this show at some point, Alex Conrad writes, "The AI lab showdown that nobody is talking about yet is RAMP versus Stripe." And yet even if this acquisition happens, OpenRouter is going to face an increasing wave of competition. Cursor, for example, just this week announced their own Cursor router. That follows Meta building a router in their internal incubator and RAMP and Vercel also going live with their own versions of the product. Cursor's version of a model router lets engineers automatically select the right model for the job while choosing between 3 optimization settings: intelligence, cost, or balanced. The Cursor router will then analyze each request and send it to an appropriate model based on that performance. They claim that using the router in intelligence mode can deliver Fable-level performance at a 60% reduction in cost. Now this was measured using subjective satisfaction metrics, so it is perhaps a little difficult to know how strong the performance will be, Still, early testers reported no noticeable drop-off in quality compared to simply routing everything to Opus 4.8. One of the most powerful features of CursorRouter could be the ability to never have to think about model selection again. Since it's built into a tool that teams are already using, there's not even an extra layer to configure. Explaining the motivation, Cursor CTO David Pan wrote, "We briefly went insane and decided every software engineer should also become an expert in model benchmarks, thinking levels, and cache hit rates." Matthew Berman sums up, model routing is a first-class feature now. Speaking of new features, the product announcements from both Anthropic and OpenAI on Thursday related to voice features, which at this point, if you are not controlling your agents with voice, I genuinely believe you need to start shifting your behavior. In any case, Anthropic has finally made their voice mode available for their Opus and Sonnet models rather than just Haiku, meaning users won't need to choose between the comfy voice interface and having access to powerful models. The problems with routing conversations only to Haiku was immediately obvious when the feature first launched last year, as early testers got frustrated as they tried to use voice to discuss complex topics like business problems that Haiku was just not suited to handle. The new chat mode will default to the last model that was used, but users can switch to a more powerful model mid-conversation, as well as switching back and forth between text and voice. In addition, Anthropic's voice mode is now compatible with connectors, allowing it to tap into apps like Gmail, Slack, or Notion to do things like check your calendar or email mid-conversation. Anthropic has also moved foreign language support out of beta, which is good news for users who prefer to speak to Claude in French, Hindi, Korean, and numerous other languages. OpenAI's feature release is voice in the desktop app. Until now, voice has been only available in mobile, but following the pattern of integrating all of their features, you can now use the voice mode wherever you're using OpenAI's model, including in Codex in the new Work app. The feature is driven by OpenAI's new real-time voice model, GPT Live, so it can carry out background tasks while keeping a natural-sounding conversation. Now while all of these companies continue to evolve their interfaces and interactions around models, one company that seems to be heading away from models might be Amazon. According to reports, the company has cut staff in their AGI group. That division was set up in 2023 to house a new effort to train frontier models. Amazon hired former OpenAI researcher David Luang to lead the technical effort and set up a separate office in San Francisco. In late 2024, Amazon released their first family of models called Nova. They failed to make much of a splash, but at the time I said that they indicated that perhaps Amazon wanted to compete on the cheaper model vector rather than the state-of-the-art vector. The team showed promise in early 2025 with the release of Nova Act, which outperformed then state-of-the-art Opus 3.7 on computer use benchmarks. However, the past year has been marked by a number of high-profile departures, including Luan himself. The AGI division was assigned to a temporary leader, and we haven't seen a new version of Nova since December. Now Amazon has acknowledged that rank-and-file staff are being let go as the unit narrows its scope. A spokesperson denied that this is the end of model training at Amazon, saying, "We've been building large models for several years, and it remains one of the most important things we're working on. This is a fast-moving space, and we're sharpening our focus on initiatives that matter most for customers so we can move faster on what counts." The spokesperson, however, did acknowledge that this increased focus required, quote, "Some difficult decisions, including eliminating some roles within some parts of our AGI organization." Sources told The Information that earlier this year, staff were shuffled across to NovaForge, which is Amazon's new service that offers custom fine-tuning on top of the Nova models. While the size of the layoffs weren't disclosed, they appear to be noticeable. Posters on the Amazon employee subreddit have been asking why so many people are leaving the division over the past month, and Wednesday saw a wave of former employees post on X seeking new opportunities. On Thursday, we learned that this is not just a wave of layoffs from the team, but Amazon is shutting down the entire AGI lab. Now, presumably this is just the spinoff lab, which was focused on computer use agents and other advanced research, as the broader AGI division appears to be still operational such as it is. Yet despite the denials, AI commentator Andrew Curran and many others think the writing is on the wall, commenting, Amazon is giving up on Nova would be my guess. One company who is not giving up on their strategy and is in fact doubling down is Microsoft. The company is putting their in-house model strategy into action after publishing some impressive reinforcement learning results. Microsoft first unveiled the family of MAI models last month, with the family including 7 smaller models aimed at specific use cases like image generation, transcription, and coding. The lineup included 2 language models, one roughly in line with Sonnet-4.6 and a coding-specific smaller variant with performance closer to Haiku. More interestingly, alongside the model family, Microsoft launched Frontier Tuning, a new service that allowed customers to fine-tune their own models. And this clearly was the bet— that the MAI models would serve as solid base models for custom models that deliver cost-effective results. On Thursday, Microsoft published the first set of results from their fine-tuning system, which they refer to as their hill-climbing machine. The post-training run used MAI CodeOne Flash, the tiny Haiku class coding model, and by training the model in the GitHub Copilot harness, Microsoft was able to deliver a better experience for users compared to similar models. After a month of deployment, they found that CodeOne Flash had a 10% higher code accept rate compared to GPT-5.4 Mini and Haiku 4.5 in VS the model achieved this while having 10% lower median token usage than its rivals. Now, so far that's not all that interesting except perhaps an indicator of what could come. Indeed, the more interesting part came when Microsoft began training CodeOne Flash in their Excel harness. Citing user feedback, Microsoft claimed this produced results on par with GPT-5.6 for most common Excel tasks at a fraction of the cost. Curiously, this Excel training also boosted performance in coding, taking scores on SWE-Bench Verified from 72% to 86%. Microsoft also noted that getting near frontier performance from a smaller model means they can use previous generation hardware like H100s and A100s. Now this points to an interesting and perhaps underdiscussed market upside of these cheaper models. If they extend in a meaningful way the lifecycle of AI chips, that actually could de-risk infrastructure investments. Coming back to Microsoft, they write, these results point toward a broader strategy. By having access to the entire product stack, the model, the harness that runs it, the agents, and product-specific evaluations, we can hill climb to train efficient, powerful models capable of tasks previously handled by larger, more expensive ones. Alongside the research blog post, Bloomberg reports that Microsoft has begun switching over to their in-house models. MAI Image 2.5 will now be the default model for PowerPoint and Bing, replacing OpenAI's GPT Image 2. Microsoft AI CEO Mustafa Suleyman said that Microsoft had seen an 84% reduction in cost when used in PowerPoint. We also got another blog post from CEO Satya Nadella spelling out Microsoft's strategy moving forward. He wrote, "In a world where software has real marginal cost for the first time, how do we ensure frontier benefits are diffused across the entire ecosystem?" The key, he continued, is to optimize the cost-to-outcome frontier in real-world context. In practical terms, that means using the right model for each task and optimizing the context, skills, tools, and agent harness around it. By routing to the MAI models for lower-end tasks, Nadella wrote, We can now take saturated frontier capabilities and deliver them at scale and at lower cost through models optimized for high-usage products, while continuing to use frontier models for frontier needs. Now continuing on the theme of companies doubling down on their updated strategies, with orbital data centers still a while off, SpaceX AI is expanding their data center business here on Earth. The Information reports that the company has explored several potential sites in Texas where the rest of Elon Inc. is located. Sources said that at least one site is moving forward but remains in early stages. One approach being considered is retrofitting an existing warehouse while adding new construction to the site. That's similar to how the Colossus campus in Memphis was built, which began as a former manufacturing facility. A source said that the new Texas campus will be at a similar or greater scale to the Memphis site, which is currently operating at around 1 gigawatt across the two Colossus data centers. Some existing data center staff have been seconded to the project, and SpaceX AI was recently hiring a local data center development lead out of Austin and Bastrop, Texas. Now, when SpaceX AI first began selling spare capacity in May, the big question was whether this was a pivot to the data center business or just an opportunistic move ahead of the IPO. Certainly the acquisition of Cursor and subsequent release of Groq 4.5 suggested that the company was not done training new models, and this expansion certainly seems to imply that they'll try to pursue both businesses at the same time. The Colossus data centers already make SpaceX AI the largest neo-cloud, but if they can stand up a second gigawatt of capacity, they'll start to look more like a mini hyperscaler. It's also categorically different to be building new data center capacity rather than simply renting out spare GPUs. Now, the expansion fundamentally changes the prospects for SpaceX as a company, adding a tangible avenue for growth. Last week, the Wall Street Journal reported that SpaceX was in talks to provide compute to the Pentagon, which could add billions to their bottom line. More generally, the move could signal to the market that SpaceX AI has a coherent long-term plan. Adding this sort of capacity signals that renting compute to companies like Anthropic and Google was not just a stopgap measure, but rather a permanent part of the business. Indeed, some analysts have been waiting for such a sign, with Sean Cray of Moody's telling Fortune, "It shows that there's just different pathways for them to generate revenue in their AI segment. It doesn't strictly have to come from Groq and their AI enterprise applications." Moving over to the policy side of the house, it turns out Chinese AI isn't the only risk being discussed in Washington, as OpenAI's security incident with Hugging Face has fueled the introduction of new AI safety legislation. Representative Ted Lieu, Democrat of California, and Representative Nathaniel Moran, Republican of Texas, on Thursday introduced a bill called the AI Kill Switch Bill. The bill requires AI companies to maintain the ability to shut down, throttle, or suspend their models during a safety incident. It also gives the Department of Homeland Security the authority to issue a shutdown command, said Lieu in a statement. Unfortunately, powerful AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention. It's imperative that these AI systems have kill switches so we can keep this technology from causing catastrophic harm. And that the federal government has the clear authority and process to shut down rogue AI models. Now, the discourse is just picking up on this one, but one person who jumped in very quickly was Secretary of State Marco Rubio, who would really like people to stop talking about AI kill switches when the US is trying to export the technology. In a diplomatic cable viewed by Reuters, Rubio instructed American diplomats to convince overseas governments that Washington can't arbitrarily cut them off from US technology. He urged diplomats to push back on local digital sovereignty programs that favor local infrastructure over dependence on US platforms. The recent Fable shutdown and continued international restrictions on Mythos were mentioned in the attached talking points, framed as temporary pauses for security testing rather than evidence of a kill switch. Meanwhile, speaking of people who would like to shift the narrative, Commerce Secretary Howard Lutnick says everyone needs to take a deep breath and stop freaking out about KIMI K3. In a Thursday post on X, Lutnick wrote, CAISI's latest report shows that KIMI K3 remains behind America's leading frontier AI models. The United States continues to lead in frontier AI because we're home to the greatest innovators and technologists the world has ever seen. The report was a joint evaluation conducted by the US Center for AI Standards and Innovation and the UK Artificial Intelligence Safety Institute. They found that K3 lagged behind US models by a gigantic margin on cybersecurity benchmarks. K3 scored 32.2% on Exploit Bench compared to an average of 76.2% for frontier US models. GLM 5.2 was also tested and found to be even more lacking, scoring just 24.4%. Now, one of the most important parts of the report was a benchmark called The Last Ones, which tasks a model with autonomously executing out a 32-step network takeover attack which would take human experts 20 hours to complete. This was the benchmark that originally raised concerns about Mythos after the preview version became the first model to successfully complete the attack. By the way, since then, the full release version of Mythos-5 improved the score, as did GPT-5.6 SOL. KimiK3 was not even close. While it was successful in 1 of 10 attempts, both Mythos-5 and GPT-5.6 SOL successfully executed the attack in 60% and 70% of runs. The report concluded, "This indicates that KIMI K3 is capable of autonomously attacking small, weakly defended, and vulnerable enterprise systems when directed to do so and given initial network access. However, the last ones differ from real-world environments in several ways: it lacks active defenders and defensive tooling, imposes no penalty for actions that would trigger security alerts, and contains an intentional attack path. Former AI czar David Sachs wrote, Secretary Howard Lutnick is right, the KIMMY panic needs to stop. American frontier models are still ahead, and when you factor in what's in the lab, the gap is even larger. As long as we keep releasing, we stay ahead. Let our horses run. Sachs continued, as Ben Thompson showed, KIMMY's apparent cost advantage largely disappears once you account for higher token usage and the real cost of running a model this size. Open weights still require expensive infrastructure. Finally, Anthropic and OpenAI are growing revenue at rates that Silicon Valley has never seen before at this scale. This remains the clearest test of who is winning the market. Now, speaking of Anthropic, the discourse in some places is beginning to view this as a regulatory capture play spurred on by the company. Now, at this point, I think it's fairly uncontroversial to say that Anthropic is lobbying for tough action on distillation, and Anthropic CEO Dario Amodei has previously said that he has serious concerns about open-source models with strong cyberattack capabilities being available to anyone. Confirming what a lot of people have felt, The Information published a rundown and noted that Anthropic and OpenAI are basically the only companies in the tech industry that are actively advocating for the crackdown. That article highlighted Jensen Huang's comments from an interview earlier in the week where he stated, "There's a misconception that somehow there are backdoors that are somehow connected to China in some way. The Chinese models are excellent. Open-source models that are excellent should be used." Indeed, Jensen went on to argue that having access to a myriad of different open models is actually far safer than a current US duopoly, commenting, "If everything just becomes one single model, one single point of attack, one single source of failure, I think the world is much, much more vulnerable." And finally, with that slightly optimistic transition note, we end today's extended headlines with a new Meta ad campaign focused on AI optimism. Starting on slightly dystopian imagery, although not burning buildings, before switching over to happy positive humans, the voiceover reads, Some people will have you believe AI is going to make us feel less connected, that it's going to leave us behind. We couldn't disagree more. Call us optimists, call us dreamers, call us whatever the hell you want, but we're betting on people, and we like those odds. The future is for everyone. Alongside the video release, Mark Zuckerberg posted, Meta has always believed in giving people the power to share, connect, and shape your world in the ways you want. As we enter this next wave with AI, we continue to believe the future is for everyone. We're focused on giving every person the tools to reach your full potential and making sure the benefits of technology are distributed to everyone. Now, Meta plans to run the ad in paid media spots in an attempt to spread the word on AI optimism. So how is this received? Certainly the ad received a lot of criticism, but largely it's from people who have already decided that AI or Meta themselves are terrible for the world. Frankly, putting on my ad production hat, I don't think it's a great ad. I think it's a little generic. I think the copy is a little generic, and I think the source is going to be hard to swallow from some. But I also don't care. Even if the ad itself is slightly cheesy or not perfect, It is at least an attempt to tell a positive story about AI and to share why there are so many people who are building this technology that are excited about it. For that alone, I welcome it, and I hope Meta blasts it everywhere. That, however, is going to do it for the headlines. Next up, the main episode. One of the most important AI questions right now isn't who's using AI, it's who's using it well. KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers. They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news? These behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com/us/sophisticated. That's kpmg.com/us/sophisticated. I cover the capability gap between AI potential and AI reality every day on this show. Most companies are still figuring out how to start. Robots in Pencils is already launching and scaling. Agentic and generative AI in production at large enterprises in weeks. AWS Advanced Tier Pattern Partner more than doubled in a year. And they're hiring. 50 open roles. If you're someone who knows this moment is different, who wants to be inside it, not watching it, this is worth a look. At Robots and Pencils, the best ideas win, and the team is purposefully kept super high quality. This is the kind of place you look back on as the best decision you ever made. Take a look at robotsandpencils.com/careers. Blitzy is driving over 5x engineering velocity for large-scale enterprises. A publicly traded insurance provider leveraged Blitzy to build a bespoke payments processing application, an estimated 13-month project, And with Blitzy, the application was completed and live in production in 6 weeks. A publicly traded vertical SaaS provider used Blitzy to extract services from a 500,000-line monolith without disrupting production, 21 times faster than their pre-Blitzy estimates. These aren't experiments. This is how the world's most innovative enterprises are shipping software in 2026. You can hear directly about Blitzy from other Fortune 500 CTOs on the Modern CTO or CIO Classified podcasts. To learn more about how Blitzy can impact your SDLC, Book a meeting with an AI solutions consultant at blitzy.com. That's B-L-I-T-Z-Y dot com. This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together. New users get $1,000 in inference. Forget local agents and chat workflows waiting on your laptop to be prompted. HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses. Marketing's agent turns competitor moves into landing pages. Sales' agent enriches leads, drafts emails, and updates the CRM. Ops' agent chases the paperwork and tracks the budget. Every agent has access to shared context and follows your rules about scope and approvals. It's time you add agents that feel like teammates. Hire yours at HyperAgent, built by the team at Airtable. Claim your $1,000 in inference at hyperagent.com/AIDailyBrief. Welcome back to the AI Daily Brief. One of the big questions surrounding AI has always been what its impact on jobs will actually be. Now, regular listeners know that I am very optimistic in the long term. If you want to hear my most full-throated explanation, go back and listen to my episode about the new jobs AI will create. The TL;DR of my take has always been that the things that AI enables will not just allow us to do the same stuff that we do now more efficiently with less people, but that the new time, money, and other resources that are freed up by those efficiencies will unlock new types of opportunities in areas where there is more demand elasticity. I think, for example, we consume a very, very small portion of the total healthcare we would consume if new, better opportunities were unlocked at a cost that people could bear. There are also all sorts of other reasons, I think, to be skeptical of AI job displacement claims and view it as a fundamentally augmenting technology. And yet at the same time, it does feel undeniable that there are certain categories of jobs, certain roles that AI kind of obviates the need for. To me, it's felt like the real question is not the long term, but the transition period. But with all that, there has been a market trend over the last couple months of the major labs reevaluating their priors on what AI job displacement is actually going to look like. The latest to share this point of view is Anthropic's head of economics, Peter McCrory. Now, it's worth noting that he's very clear right there in his Twitter bio that these views represent his own and they are not Anthropic official. But still, I think his opinion carries weight. He recently published on X a post called "Why hasn't AI increased unemployment?" And what we're going to do today is read a chunk of that post and then talk about some of the responses and reflections. Peter writes, the US labor market is currently stable and close to maximum employment. In my view, AI has caused no material increase in the unemployment rate to date. Even if we focus on workers with high exposure to current patterns of AI automation, We don't see unexpected increases in unemployment in recent years. Why don't we see any impact of AI adoption on unemployment? AI so far has the hallmarks of a skill-based labor-augmenting technology. Even as AI automates some aspects of work, complementary human expertise amplifies what AI or humans can achieve alone. AI broadens the scope of what people can accomplish, which increases the returns to working with AI. Now, the future is still quite uncertain. Model capabilities are advancing rapidly, and AI systems may soon be able to autonomously develop their own successors. More generally, intelligent AI systems could lead to labor displacement that hasn't yet materialized. In many ways, then, he says, this short essay is my attempt to synthesize Anthropic's economic research over the past 18 months to understand how people use AI and what that implies for work, the labor market, and the broader economy right now. So far, we've seen muted unemployment effects, and this essay presents my framework for understanding why and what might change in the future. So to set this up, Peter points out the important background fact that the US labor market is currently stable. June's unemployment rate was 4.2%, which he says the Fed views as a level consistent with full employment and stable prices. He also notes that the ratio of job openings to unemployed workers recently rose to just over 1 in April, which some economists argue implies the demand and supply of labor are roughly and efficiently balanced. The prime-age employment-to-population ratio remains close to multi-decade highs, reflecting broad-based labor market strength that emerged during the post-pandemic expansion. And weekly initial claims for unemployment insurance have been stably low over the past 4 years. So he writes, should we even expect an impact from AI on the labor market yet? I think the answer is yes. The AI sector is large enough that we can look for discernible macroeconomic effects. To make his point, he writes, 20% of firms use AI in at least one business function, and in the information sector, which is 5.5% of GDP, the share is 40%. Quality-adjusted AI output grew over 2,000% per year in both 2024 and 2025. Even from a small initial base, this suggests that we should see signs of AI's impact in the aggregate. Peter also writes that he believes that we're beginning to see AI's impact in aggregate productivity statistics. He points out that as compared to the 4 years prior to the pandemic, where labor productivity growth was 1.6%, the ratio of output per hour of work increased 2% per year from 2022 to 2026. Next, he asks, is there any evidence that job displacement is happening, even if it's not yet macroeconomically consequential? The evidence is mixed, he writes, but overall, I'm unconvinced. Peter continues, as documented in our labor impact report, we haven't seen worsening unemployment rates for workers in roles with a large share of tasks that Claude is being used to automate relative to workers in other roles. Updating this analysis with more recent data from the BLS doesn't change this result. We do find some suggestive evidence that hiring rates for young workers in highly AI-exposed roles have weakened over the past year or so. That's consistent with the evidence in Canaries in the Coal Mine, a paper by researchers at the Stanford Digital Economy Lab. But, he continues, this evidence for young worker displacement should be interpreted with caution. Quote, it's hard to discern casual effects because AI emerged in an unusually volatile macroeconomic environment: unwinding of pandemic-era dislocations, rapid tightening of monetary policy, commodity price volatility following Russia's invasion of Ukraine, and sustained global policy uncertainty, e.g., from trade wars. Because hiring is a form of investment, broad economic uncertainty can itself weigh on hiring. Another way to put it, he continues, from 2022 to now, the US experienced the largest non-recessionary labor market slowdown on record. This coincided with a low hire, low fire labor market. This kind of labor market hits early career entrants hardest. Right now, young workers may be struggling to find jobs for macroeconomic reasons other than AI. Peter does note that, quote, while we don't see unemployment effects yet, we do find that workers in roles with tasks that Claude is used to automate do express greater concern about losing their jobs. Than those in less exposed roles. Peter continues, "If you believe that the US labor market is currently healthy, that AI could in principle be generating macroeconomically discernible effects, and that this hasn't yet produced displacement for highly AI-exposed roles, then the next question is obvious: why hasn't AI caused a meaningful increase in unemployment?" Peter's first answer is that AI is both skill-biased and labor-augmenting. As he puts it, it complements domain expertise, "It relies on humans in the loop to direct and evaluate the most complex work, and it rewards AI proficiency. Model capabilities are improving fast but remain stubbornly jagged. To fill in the pockets of the jagged frontier, expert oversight is needed to steer incredibly capable AI systems and to recover when they falter." Of course, he says, some jobs are more exposed to outright displacement by automation. For instance, technical writers, data entry workers, customer support representatives, and computer programmers are jobs where AI can reliably handle the core set of tasks and responsibilities. Even though we haven't seen any increase in unemployment for workers in these sorts of roles, occupations with higher observed exposure are projected by the BLS to grow less through 2034. But so far, he writes, the broader picture is one of labor augmentation. The effects in the labor market are set to be uneven as a result, even as capabilities advance rapidly. This does not mean that all skills that currently command a premium in the labor market will do so in the future. Some types of expertise may become less valuable, e.g., pure coding implementation. Even as others become more valuable, e.g., managerial skills of delegation and evaluation. The next interesting question that Peter explores is why AI is a skill-biased labor-augmenting technology. Couple examples he gives are first that, quote, despite the incredible advance of AI and rapid adoption throughout the economy, there's no job in the O*NET taxonomy, a Department of Labor catalog of occupations and their typical tasks, for which all associated tasks are systematically handled by Claude. If jobs are fixed bundles of tasks, they aren't, but more on that in a moment, then the essential non-automated aspects of work both constrain the overall productivity lift and amplify the returns to labor. Tasks that Claude can't handle may depend on interpersonal coordination, in-person interactions, or engagement with the physical world that so far only humans can do. Peter also notes that Anthropic found that sophisticated user inputs and complex Claude outputs are highly correlated. In other words, when Claude builds a complex economic model, In practice, it does so under the guidance of someone providing complementary expert direction. He also wrote that Anthropic found that even after 6 months of use, people are more likely to interact with Claude as a thought partner and have more successful interactions with Claude. If AI was good enough on its own, he contends, we wouldn't expect to see this effect. Importantly, Peter writes, widespread task automation can still augment labor. Why? Because jobs are not fixed bundles of tasks. New technologies have historically led to large changes within existing jobs, even as some jobs go away. And they've produced entirely new types of work that combine new technical capabilities with complementary human expertise. We see signs of this effect in our research. A commonly cited source of perceived productivity among 81,000 cloud users was scope, being able to do more, more proficiently. Such empowerment from AI may redraw the boundaries of our roles and produce new bundling of tasks within jobs, automating some, reinforcing the importance of others, while on net increasing the marginal product of labor. He also points out that the more that people use Claude and the better that people get at using Claude, even though they increase their expectations of what portion of their jobs Claude can do, they decrease their expectations of job loss and tend to be more optimistic about AI's impact on things like pay, job security, and their ability to find a job. Now, ultimately, Peter ends on the note that all of this could change. He points out that as AI capabilities improve, we'll see increasingly capable agents that can autonomously handle complex, long-horizon valuable tasks. Will then AI still augment labor? He points out that it may very well be the case that the skill-biased labor-augmenting aspect of AI goes away as models continue to improve and as the jagged frontier becomes smoother, but also that so far when they recently analyzed patterns of Claude code usage to see if agentic coding was altering the returns to expertise, that's not really what they found. Peter writes, Claude code has been used on more and more valuable tasks over the 7 months we tracked, But we've seen persistent returns to human expertise. That is, people make planning decisions and delegate implementation to Claude. People with more domain expertise succeed in their tasks more often and recover more consistently when Claude makes an error. The return to straightforward coding ability may have fallen, but agentic coding has so far increased the value of other complementary skills. His last caveat is about recursive self-improvement. He writes, a big reason there's so much uncertainty about the future is that AI may automate innovation itself. Endowing machines with general cognitive capabilities is a direct catalyst for further innovation in ways that past general-purpose technologies weren't. In otherwise standard economic models, automating innovation can produce economic singularities— infinite growth in finite time. Will such singularities occur? Not if there are essential tasks that are never automated, whether for technical reasons or societal constraints. Those weak links are the limits on growth. Pointing to a paper by Aghion, Jones, and Jones, he quotes, "Economic growth may be constrained not by what we do well, but rather by what is essential and yet hard to improve." Such weak links, Peter continues, can keep the labor share of income elevated in the long run, even under very rapid widespread but incomplete automation. Ultimately, he concludes, scaling laws are hard to argue with. The models are going to get better, much better. I expect this will drive faster productivity growth and maybe even more clear signs of RSI. But I don't expect unemployment to be noticeably higher a year from now, at least not because of AI. So really interesting stuff here from, again, Peter McCrory, the head of economics at Anthropic. And interesting even without returning to commonly heard concepts around AI and jobs like Jevons paradox. Now, if this state continues, it has big positive implications. Stanford's Andy Hall writes, a while back I predicted that the real political backlash to AI would happen when unemployment started going up a couple of percentage points. And I said we weren't there yet. We're still not there. And this piece helps explain why. So far, AI looks like it augments rather than replaces human labor. That could change, but right now, the labor market looks quite stable. We are of course seeing political concerns about AI even so, but these concerns would look small fry in comparison if we had genuine widespread unemployment happening. Some argue that while yes, this is positive, we might be looking in the wrong place. Trace Cohen writes, "The real impact may show up first in hiring, not layoffs. Fewer junior roles, smaller teams, slower backfilling, and much higher expectations for each employee." One person using AI may increasingly replace several people who are not, even while overall unemployment remains low. And I do think that one thing that's worth watching over time is whether we see shifts in average team size. I do think that there's going to be shifts in the patterns of how we do work, and I would expect smaller, more nimble teams, both on the organization scale but also within organizations, to increasingly have more responsibilities. I think that the change will happen gradually enough that mostly units will be reconfigured and people will be redeployed to do other types of things alongside those teams who are now taking on a bigger role. But it is still worth watching. For many, the most notable thing is just Anthropic releasing something positive for once, writes investor Julie Fredrickson. Finally, someone in Anthropic discussing how great AI is for the professional class. Robert Scoble writes, Anthropic doing marketing that isn't full of fear. More of this, please. Look, I think epistemic humility is extremely important in the context of predicting the future. It is not hard to draw scenarios where AI does have a big impact on jobs. And yet I think the evidence that we are seeing so far is extremely encouraging. And I think what's more, that the more that the discourse and narrative shifts from efficiency and cost cuttings and headcount reduction to augmentation and expansion of responsibilities and new opportunity creation, it has a self-reinforcing impact in how executives and leaders think about how they should be using AI. In other words, if everyone in the world is saying that this technology should be used to cut your staff in half, and by the way, that's what your investors expect as well, that's going to put a lot of pressure on you to do exactly that. If, on the other hand, the story is about doing more faster and moving into lateral domains and releasing new products and services, then we're going to see a lot more of that. Obviously, I know which of these I think is better for the world, and I hope more of the world comes around to that view as well. For now, we get to end Friday on that bright note. Appreciate you listening or watching as always, and until next time, peace.