Turns out, just because you can doesn’t mean you should. Uber, in its earnings call, said that only four (yes, four) rides were booked by AI agents in Uber’s busiest week of October, Halloween – out of roughly 320 million trips and orders. It all makes sense when you consider that it takes you all of 50 seconds to order an Uber, but a bunch of back and forth with your AI agent. All of which is a good reminder that AI for the sake of AI doesn’t work – for users to adopt AI it needs to make things faster and/or better.

A new study analyzed AI’s impact on the translation industry. Google Translate’s 2016 neural upgrade (predictive AI) cut demand for regular translation on a major freelance platform by about 9%, but had no measurable effect on “transcreation” – creative work like ad copy and poetry that needs cultural adaptation. ChatGPT (generative AI), on the other hand, cuts regular translation by 21% and transcreation by 30%. The “safe line” moved once, expect it to move again – and not just in language translation.

Set aside the ever-shortening gap between US-based closed frontier AI models and open-weights Chinese models; US consumer spending further suggests that there isn’t a moat for the frontier model makers. A recent study analyzed 114 billion credit card transactions gathered from 174 million consumers and found that a mere 2.9% of those consumers pay for a major AI service – nearly all on basic plans of $30/month or less, and only 20% to 45% of them stay with a provider for a year – lower than many other subscriptions. They treat the services as close substitutes and switch whenever a better model ships. Without a moat around the models, companies like OpenAI and Anthropic need to build value-added services and lean heavily on B2B to justify their eye-watering valuations.

US computer and information science programs saw a steep drop in enrollment this fall – undergrad enrollment is down more than 12% and grad enrollment more than 10%. Over just two years, computer science programs have lost 18% of their undergrads and 27% of their grad students. It’s quite the reversal from the “become a coder and you’ll be set for life” mentality, which held until AI started eating entry-level coding jobs. Which raises the question: Which professions are next?

A Microsoft survey of 319 knowledge workers finds that the more we trust AI output, the less we engage in critical thinking. The data is self-reported, so it shows perceptions of effort, not measured changes in thinking. The flip side is that workers confident in their own skills think harder. That matters for the human-in-the-loop discussion – a human who trusts the machine more than themselves stops checking.

In light of companies’ struggles with the ever-expanding spend on AI tokens (see the first post above), Anthropic now offers new reporting tools that show how and why your employees use AI – Claude reads the transcripts and flags “personal or off-topic” sessions as inefficiencies. Time to rein in those personal AI chats about your next vacation.

Delaware, the US state where most Fortune 500 companies are incorporated and long known for its pro-business legislative and tax stance, has drafted a bill for a new legal entity. Artificial Intelligence Companies would be legal entities run and managed by AIs – no human involved in running the day-to-day operations. The AIC “can sue and be sued in its own name, hold and dispose of property, and incur obligations” – but its owner isn’t liable for anything the AIC does. If this sounds to you like “an overly generous liability shield,” you’re in good company – Berkeley law professor Andrew Baker called it exactly that, and House Oversight chair James Comer just opened an inquiry.

And now for a bit of good news: Solar energy is projected to overtake nuclear as a global power source this year – and the lead will only grow. Production prices keep dropping, battery storage gets better and cheaper, and the trend to electrify everything continues. Sunny days indeed – and much-needed news in a world which will consume ever more energy to keep the AI machinery running.

Amazon’s latest tablet made for kids points toward a dystopian future of (even more) screen time for your toddler. While country after country establishes rules to minimize screen time and social media use for teens, Amazon pitches it as your three-year-old’s “first tablet” and touts its product’s ability to keep your kid pacified on the flight and the road trip (and everywhere in between). With AI rewriting education and our children’s future jobs, we should invest in our kids’ ability to create and think critically instead of outsourcing this to Amazon’s AI “Storybuilder” and “Art Studio” tools, which come bundled with the new tablet.

Torsten Slok, investment firm Apollo’s Chief Economist (and a semi-regular in the radical Briefing), showed more evidence that the ROI on AI investment outside Big Tech isn’t there (yet). Revenue per worker and profit margins are climbing at the Mag 7 and flat across the other 493 companies in the S&P 500. Productivity growth shows no AI bump yet. I believe it’s down to the nature of the work (much of what these firms do doesn’t lend itself to AI-driven gains – at least not yet) and the difference between doing something as a one-off or pilot versus a reliable process running millions of times. As Slok puts it: “The bottom line is that the AI story will be decided in the S&P 493, not the Mag 7, and so far […]”

The nerdy side of me reads the blogs of former colleagues and well-known old-time software developers. More and more of them write that; now that LLMs have taken over their world, they are looking for new jobs (outside of coding) or retiring completely (a few are quitting the Internet altogether). The frustration is palpable – here is my former Mozilla colleague Jono’s sobering reflection on where he is at (and Jono is an excellent, highly creative software developer). These voices come from one industry, but I fear that the changes we are seeing in software development are the harbinger of things to come in many other industries. As LLMs get more capable, they will take over more work and hollow out many more professions. What might be left is a whole bunch of people experiencing an existential crisis and the necessity to decouple our identity from our jobs.

Anthropic is reportedly working with lawyers to train its AI to negotiate for you. It’s another step toward AI agents that act for you on their own – AI agents can already research information, order (and in some cases pay for) products and services, and even call a restaurant to book a table. Next up, they’ll negotiate a better price for your new carpet and complain to your plumber about the shoddy work. The interesting question is how AI agents will negotiate with each other – because your plumber will have one too.

A newly released study by Berkeley’s Brian Christian and co-authors at CMU, MIT, Oxford, and UCLA showed that after just 10 minutes of AI help, people gave up faster and got more wrong once the AI was taken away. It’s one of many studies that point in the same direction: AI makes your brain lazy, and we have to deliberately seek out ways of working (e.g., writing by hand on paper) to combat this. Still, none of this research is longitudinal yet, so the long-term damage is a hypothesis.

Consulting firm Accenture loves to tell its clients to ramp up their use of (and spend on) AI – and charges them a pretty penny to help them do it. Meanwhile the firm is rationing its own AI use – capping centrally funded Claude spend at $29 per employee per month. Irony aside, it makes you wonder how much value Accenture gets from its own (internal) AI use given the newly announced cap.

“Human in the loop” is a concept you hear a lot these days – the notion that humans are overseeing AI systems and making the final call. A position paper by Margaret Mitchell et al. argues that current agent design actively impedes meaningful human supervision. Add the skill atrophy that comes with extended AI use, and “human in the loop” fails as a safety net for increasingly autonomous agents. If we are serious about human oversight of AI, we need to design for the overseer, not just the agent – in the tools and in how we organize the people using them.

A recent study looked at productivity gains from GitHub Copilot for Okta’s software engineers. The study found “no statistically significant improvement” when it came to quantifiable organizational productivity throughput – but engineers reported cutting their workweek from about 45 to 41 hours. The hangup can be summarized in these two sentences from the study: “Achieving organizational productivity gains requires more than deploying an AI-powered tool. It demands a deliberate effort to learn, adapt, and reinvent teams, processes, and infrastructure to align with AI-augmented developer capabilities.” Individual productivity gains don’t move the needle on the company level, as communication, waiting on decisions, and process time don’t move – AI requires organizational change to be effective.

Food delivery businesses like DoorDash are your classic marketplace model – with a middleman (DoorDash) taking a hefty cut of the transaction for their matchmaking service. A 10-person Bay Area startup, Bites, is challenging this model with its AI-native approach. Connect to the apps and APIs the restaurant already uses to get the menu, expose it to your AI agent inside ChatGPT for ordering, and build on two decades of marketing and consumer education your predecessors have done (and a global pandemic accelerated), and you can charge a flat $1 per order where DoorDash takes 15 to 30 percent. The Verge’s test pizza came to $56.73 through ChatGPT and $70.29 on DoorDash. Meanwhile, DoorDash loses its $1bn/year ad business and its ability to upsell you to that nice bottle of wine.

When I was at Singularity University, we joked that you should be nice to your AI and robots – once they become sentient, they might remember all the abuse they suffered from your bullying. It made for some cheap laughs. Anthropic just updated its usage policy prohibiting “sustained and needless abusive or cruel behavior” toward Claude. Meanwhile, a rabi in a meeting of religious thinkers, is being told “I wish I had your faith” when he made the argument that AI doesn’t have consciousness.

That online retailers like Amazon keep a detailed profile about you doesn’t come as a surprise. That they might deem your backside “flat buttocks” is a bit more surprising. After a user discovered this, other users found gems like “cannot keep real plants alive,” “has no friends,” or “treats cats as biological children” in their profiles. Here’s how you see what Amazon thinks about you: On Amazon.com, select your account icon in the top-right corner of the screen, scroll down to the “Your Account” section, and click “Your Shopping preferences.” From there, scroll down the page and select “About You.”

When I was at Singularity University, we joked that you should be nice to your AI and robots – once they become sentient, they might remember all the abuse they suffered from your bullying. It made for some cheap laughs. Anthropic just updated its usage policy prohibiting “sustained and needless abusive or cruel behavior” toward Claude. Anthropic labels this as model welfare – a standard we ought to apply to how we treat each other, too.

Here is an interesting thought experiment – what would happen if AI were to become fast. AI today has four limiting factors: model capabilities, cost, context windows (how much data an AI can ingest), and speed. Most of our attention goes to the first three – speed has improved, but gets far less attention. EchoHive imagines AI answering at a million tokens per second (today’s AIs are operating at around 100-200 tokens/sec). With this speed, LLMs could code 40 app drafts or write 1,000 story endings in a single second. And it would take AI only three seconds to review 10,000 job candidates. With this amount of raw output, evaluation will quickly become the ultimate bottleneck – add one minute of testing to those 40 app drafts, and 10,000× faster writing makes the whole job only about 132× faster.

Much of the internet runs on open-source software, which is largely created and maintained by volunteers. Aside from the financial asymmetry between those who build these tools and those who benefit (often the world’s largest corporations – and every human on the internet), it comes with a real risk: The moment a volunteer decides to stop working on their project, the ripple effects can be massive. Eleven of the 23 critical projects the piece tracks have only one or two regular maintainers. Take the time zone database, which runs on 4 billion phones and most internet servers – UCLA lecturer Paul Eggert has maintained it in his spare time since 2012, with no grant funding. We (collectively) ought to pay for what has become critical infrastructure – sudo’s maintainer (an app installed on pretty much every server worldwide) just got $61,700 a year after three decades, which shows both that it can be done and how late we are.

A single developer, vibe-coding with Claude, is recreating the whole Adobe suite of photo and video editing tools (Photoshop, Illustrator, Premiere Pro, etc.). It’s as ambitious as it is nuts – but only a few months later, he has some apps that (somewhat) work. It aims to be a faithful recreation of the original software, written from scratch in Rust and in a full clean room (i.e., no code is being copied from the original, making this legally mostly fine). If – and it’s a big if – this works, it has huge implications for the future of software: On one hand, it’s amazing that you can point an LLM at an app and get a rough clone; on the other hand, it raises the question of who will pay for software innovation. Adobe won’t be amused.

Apollo’s Chief Economist Torsten Sløk warns that AI agents like Meta’s Muse could trigger an “agentic bank run.” The argument goes like this: A lot of people’s money is in low-yielding checking accounts; AI agents would sweep it into higher-yielding savings accounts and CDs. Banks, which lend out those cheap deposits, lose their funding and their spread. I think this is more hypothetical than real – it requires a lot of trust in your agent, and neither Meta’s Muse nor OpenAI’s Dots can move your money at the moment. But it does point toward a very weird future in which any business built on customer inertia should be on high alert.

AI company Tavus says that their AI video model Tavus says its Griffin is the first AI model to pass the video Turing test. Nearly half of all people who talked to Griffin through a Zoom-like video chat were convinced they were talking to a human. As impressive as this is, the misuse potential is huge. You can’t trust that the “person” on the other end of your video call isn’t an AI anymore.