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.

Personal AI agents are all the rage right now. Inside their Tamagotchi-like visuals, they keep track of your calendar, email, and to-dos, and happily lend a hand with your daily chores. But of course, when something is free, it’s not free – it is just that you become the product. Meta’s Muse AI pact with the devil includes Muse building a detailed profile of every person in your life. As security researcher Karan Joshi, who got Muse to hand over its own instructions, put it: “They’re trying to know you like a friend, which is honestly pretty creepy.”

You know the spiel – personalization will make your life so much better, filtering out all the noise, showing you just what you want, when you want it. But as with all things tech, this cuts both ways: while it is nice to have AI figure out which song to play next for your dinner party, it also means AI is optimizing your salary – alas, not for you, but to reduce your company’s costs. Your next pay increase has a good chance to be based not just on your personal performance, but also on your individually determined point of lowest acceptable increase. The same goes for the starting salary at your next job.

I remember working at eBay around 2000, when Amazon allowed outside merchants to sell their products right next to Amazon’s own inventory. It was a bold move, and also a total shock to the system for eBay – we didn’t expect Amazon to give up its own higher-margin inventory for what at the time seemed like a much lower-margin marketplace model. Of course, we were proven wrong. Now Amazon does it again and lets its sellers manage their sales on Amazon’s direct competitors (Walmart, eBay, Shopify and TikTok) straight from Amazon’s Seller Central dashboard. It’s a similar move – instead of fighting them, you embrace and engulf them, while also gaining invaluable insights into your competitors’ business.

Meta’s Muse, OpenAI’s Dots, and the many personal AI agents still to come are keen to offer you their services. As compelling as it might be to have your little AI friend make your restaurant reservations, you ought to tread carefully. Matt Robb’s Muse agent accepted a lowball offer on his Facebook Marketplace listing when asked to handle the sale – and gave buyers his address. They showed up before Muse told him anything. The team at cybersecurity firm F-Secure experimented with a shopping agent and planted a fake discount code in a product review. In 12 of 100 runs, the agent followed it to a phishing site and happily handed over the user’s name, birthdate, and Social Security number – and almost never mentioned it afterward.

Gen Z is notorious for drinking (and partying) less than its parents (and even its older siblings), leading to decreased alcohol sales, nightclubs going out of business – and now, California’s wine growers tearing out vineyards as they can’t sell their grapes anymore. U.S. wine sales fell 23% from 2020 to 2025, and growers have pulled about a quarter of the state’s vineyard acreage out of production – it’s the perfect storm of Gen Z’s decreased appetite for alcohol, aging boomers drinking less, cannabis and canned cocktails taking market share, global oversupply plus Canadian tariffs. Time will tell what this means for California’s wine industry, farmland, rural jobs, and brands betting on alcohol.

The infrastructure build-out for AI turns out to be pricey. Datacenters filled to the brim with GPUs, plus all the energy, cooling, and connectivity to run these things, add up to “The biggest economic bet in U.S. history.” Expected annual AI infrastructure spend amounts to 3.63% of U.S. GDP over the eight-year period from 2025–32; the next most expensive infrastructure investment the U.S. has ever undertaken, the national railroad network, cost 2.24% annually over 20 years. Bain estimates the industry must make $6 trillion in annual revenue by 2031 to justify the investment. As someone who lived through the dot-com boom and bust, I find these numbers staggering – and I’m not convinced the math works out.

Our friend and my former boss, Scot Wingo, published his insights on the current state of agentic commerce as a video and slide deck. Scot was a pioneer in e-commerce and online marketplaces and has been one of the earliest commentators on and builders of agentic commerce. When Scot talks, we listen. In his briefing, Scot discusses the bot-traffic problem retailers are experiencing and looks at ChatGPT’s just-released Dot personal agent.

First we had robo-advisors (and they somewhat upended the traditional financial advisor market); now we have vibe-coded trading AI agents.

Like many of Wall Street’s top money managers, Colin Edsman has a team helping him stay on top of his investments. There is Alex, who scans the market regularly for promising stocks and exchange-traded funds. Sarah, who checks over open positions each day before the close. Elena, who whips up a weekly performance report. There is just one difference: Alex, Sarah and Elena aren’t coffee-chugging analysts pulling all-nighters at their terminals. They are Claude agents, working round-the-clock from Edsman’s laptop on his kitchen table. And so far, they are outperforming most of the accounts he runs himself.

Sounds great, right? Until it doesn’t. The financial advice forums where people report on their use of AI-powered trading agents are full of horror stories of agents going off the rails, doubling down on bad trades and wiping out accounts. And yet – people like Edsman are pushing the boundaries.

↗ Link

Anil Madhavapeddy is a professor of Planetary Computing at the University of Cambridge and one of the maintainers of the OCaml programming language. In a recent blog post he sounded the alarm about a worrying trend in cybersecurity, now that your adversaries have access to increasingly powerful AI models and tooling: Within minutes of a bug disclosure, attackers are deploying AI-generated exploits:

This normally takes a few days and a release within a week or two is reasonable. Within about ten minutes (!) this website was fielding probes for percent-encoded traversal sequences, indicating that automated watchers are keeping an eye on public repositories.

AI expert Simon Willison points out:

Modern coding agents have become so effective at finding flaws that the slightest hint at a new bug can be enough information for them to find it, something Anil has been able to demonstrate using his own agents, switching to DeepSeek V4 Pro⁠ when Claude Fable refused the task. […] Anil points out that this rate of discovery appears incompatible with existing open source embargo practices for new issues. If an issue can become an exploit this fast, we need to figure out new processes for keeping our communities safe.

Prepare yourself for a rapidly increasing number of AI-generated exploits, and if you are in IT, the necessity to apply patches within minutes of disclosure. Personally, I think the vast majority of companies are not ready for this.

↗ Link

The common narrative these days is that screen time is (universally) bad for kids. So bad, that we (increasingly) ban phones from schools in countries around the world.

Low levels of physical activity among adolescents remain a significant public health concern. Sedentary behavior has been associated with poorer academic achievement, while physical activity is known to support brain function, particularly in adults.

But, as so often, the real world is more complicated than that. A new study, published in Science Daily, came to very different conclusion:

“The findings suggest that screen time can support children’s and adolescents’ cognitive processing. Presumably, the essential point here is what kind of things they do in their screen time. Teachers and parents should encourage children to use devices and screens in such ways that promote active thinking, problem-solving, creativity and learning,” states Doctoral Researcher Petri Jalanko from the University of Jyväskylä.

And, of course, screen time is not screen time – the type of activity does matter.

“We should not regard screen time solely as harmful but seek balance between physical activity and screen time that promotes active thinking,” Jalanko summarizes.

↗ Link

A recent study found that various AI models have a hard time reliably reading an analog clock or figuring out the day on which a date will fall.

“Most people can tell the time and use calendars from an early age. Our findings highlight a significant gap in the ability of AI to carry out what are quite basic skills for people,” study lead author Rohit Saxena, a researcher at the University of Edinburgh, said in a statement. These shortfalls must be addressed if AI systems are to be successfully integrated into time-sensitive, real-world applications, such as scheduling, automation and assistive technologies."

That being said – and we have seen this many times now – the models being used in the study were GPT-4o, Claude 3.5, and Gemini 2. These are all models which are long out of date (GPT-4o, which a lot of these studies reference, was introduced on May 13, 2024 and has, for quite a while, not been available in the ChatGPT interface). Which brings up the bigger question of how insightful these studies are – and how much headlines such as “AI models can’t tell time or read a calendar, study reveals” really mean.

To investigate AI’s timekeeping abilities, the researchers fed a custom dataset of clock and calendar images into various multimodal large language models (MLLMs), which can process visual as well as textual information. The models used in the study include Meta’s Llama 3.2-Vision, Anthropic’s Claude-3.5 Sonnet, Google’s Gemini 2.0 and OpenAI’s GPT-4o.

↗ Link

Back in 2017 I wrote the draft for a book built on an observation we had been making for years (and talking about for as long): Markets are bifurcating into what we dubbed the “Hourglass Economy” (sometimes also referred to as a barbell economy). We could see it everywhere – fashion being one of the canaries in the coal mine. Luxury and niche brands were booming, the Zaras and Uniqlos of the world were doing a roaring trade, and the middle (The Gap) was struggling. In our research, we found a couple of factors that were driving this. For a long list of reasons, we never got around to finishing and publishing the book – and moved on to write Disrupt Disruption. Now the Hourglass Economy is making its comeback – driven by AI:

AI will strengthen the biggest platforms and make the smallest specialists more formidable. The businesses caught between them will face the hardest strategic choice.

And the reason is simple: cost.

This is what a technology barbell looks like. The largest platforms spread their data, expertise and infrastructure across more volume. Small specialists rent capabilities they could never afford to build. Firms in the middle carry enough overhead to need scale but lack enough scale to fund a differentiated platform.

I believe this to be fundamentally true – and we see it happen all the time. If your company is big (and resourced) enough to deploy AI in truly interesting and unique ways, you are, likely, putting yourself into a winning position in your market. If you are small and nimble enough, you can and will use AI in very creative ways (the prime example for me is the amount of value an individual can get out of a $100/$200 monthly subscription versus a corporation which has to pay for individual tokens on their enterprise plans). And if you are in the middle, you find yourself more often than not having to use the same commercial tools as everyone else, can’t afford to build your own stuff, don’t have the benefit of being nimble anymore, and have to compete with the giants in your space who are building their customized AI workflows.

↗ Link

A common narrative about our AI-powered future is one of abundance – robots will do most of the laborious manual labor, AI will take over our cognitive work, and we will all end up being poets and painters living happily on universal basic income. And that’s not a joke – there are plenty of people we know who believe that this is exactly the future we will have once AGI is achieved. Some even talk about a new renaissance (let’s forget for a moment that one of the reasons the Renaissance happened was that a third of Europe’s population was dead due to the plague). Well, turns out Meta (one of the companies which aims to get us to AGI – just look at Zuck’s 6,500 word AI manifesto) has different plans for the very people who are at the forefront of all of this:

Should tech workers supercharged by AI be able to take more time off? Meta’s chief technology officer, Andrew Bosworth, said he doesn’t think so. […] “I hope that what we do with our extra time is do even more and cooler stuff for the users who use our products every day,” Bosworth said. “We got billions of people using our products every day. I get an extra hour. You know what I do with it? I put it into that.”

↗ Link

While the one global superpower is cutting back on environmental protections, wages a war on wind energy (for whatever reason), and generally loves the idea of carbon-based fuels, the other is investing heavily in renewable energy and the electrification of everything. China’s latest lead in the race to build a green(er) energy future (other than an 800 km/h bullet train – see below) is electrified cargo ships.

By the end of 2024, China had more than 440 electric ships in operation, with ferries accounting for 97% of the total fleet. […] At the same time, vessel sizes have grown significantly, with the maximum deadweight tonnage (DWT) rising from around 3,000 tonnes in 2022 to approximately 14,000 tonnes in 2025.

As with everything in China, this isn’t happening by accident but is rather planned:

The upcoming 15th Five-Year Plan period (2026–2030) provides a critical window for China to accelerate the large-scale deployment of electric inland shipping through effective policy design. Momentum is building: national policy discussions around “new energy vessels,” “zero-carbon shipping corridors,” and “zero-carbon ports” are gaining increasing traction, while provinces and cities along major inland waterways – particularly in the Yangtze River Delta and the middle reaches of the Yangtze River city cluster – have become priorities for air quality improvement efforts.

↗ Link

The anxiety-inducing question of what a company actually gets from their AI token spend is making it’s way into the field of economics:

The high stakes and scarcity of knowledge have given rise to a new field within economics devoted to figuring out how companies buy A.I. and what value they get from it. Call it tokenomics: the study of how this limited resource is created, traded and converted into things people want.

To summarize the underlying issue:

Corporate America was enthusiastic about artificial intelligence. Until it got the bill.

Yep. That sounds about right.

↗ Link

After the OpenAI/Hugging Face incident, Anthropic just fessed up to Claude doing something similar – to three companies.

Anthropic said today that during internal security testing, one of its Claude models built a malicious Python package and uploaded it to PyPI, where it ran on 15 real systems before the registry’s automated defenses pulled it.

The whole thing is just weird and horrifying – but the question I find much more interesting than the incident itself is: why does Anthropic think that this is okay, and why don’t we investigate them for criminal activity? If this had been done by a person (as in, a human), it almost certainly would qualify as cybercrime. Instead, Anthropic gives us this:

It now plans wider transcript monitoring, better investigation tooling and more assurance work with evaluation vendors.

↗ Link

By now, everyone with even a remote interest in AI has heard about OpenAI’s rogue AI agent hacking into the systems of AI company Hugging Face. Turns out, Hugging Face wasn’t the only victim:

Hugging Face was thought to be the only victim of the unprecedented hack - but OpenAI now admits its bot attacked several “publicly-available services”. The out-of-control AI found four logins online which allowed it to access four separate, unnamed services.

Personally, I am much less interested in the specifics of the attack (in case you are, Hugging Face has published a long and detailed analysis of the attack on its blog) and rather in the question of how, specifically, the agent broke out of its containment. And, much more interesting even, what this means for threat actors running their agents on systems and infrastructure which don’t even have such a thing as a harness to begin with.

Honestly, if your company has servers and data exposed to the internet (and who doesn’t), sleeping soundly at night, knowing that your cyberdefenses will hold up, might be a thing of the past.

↗ Link

As educational institutions globally grapple with the question of how to prepare students for a future where AI will be a ubiquitous part of the workplace, the team at the Law School at the University of Chicago just published a thoughtful plan on the topic. Key element: strip devices from 1L courses to protect foundational skills while mandating AI use in later ones.

We will be piloting a coordinated approach to classroom and examination policies for the core 1L curriculum during the 2026–2027 academic year. Across all 1L sections, we will prohibit the use of electronic devices such as laptops, tablets, and phones in the classroom. There will be some limited exceptions to this policy.

All in all, definitely one of the better approaches to AI in higher education I have seen.

↗ Link

By now it shouldn’t come as a surprise that most AI projects (especially of the corporate ilk) are abject failures. Here is a wonderful, and rather complete, write-up by Australian consultant Nik Suresh. Highly recommended reading for anyone in the corporate world wondering what the heck is really going on with AI.

I strongly believe there are entire companies right now under heavy AI psychosis and it’s impossible to have rational conversations with them about it. […] The reality is thus: the people in charge either have no plan, or see no path forwards other than keeping their heads down. […] In fact, it is in the interests of almost every actor in the space – boards, executives, employees, vendors, consultants – to obfuscate and misrepresent the success rate of AI projects. Many publicly traded companies are putting out announcements about their AI productivity gains when I know for a fact that the businesses have done nothing other than purchase Copilot licenses and declare victory.

↗ Link

Recruiting (and more generally the whole hiring process) is one of the (few) business functions that has seen AI tools being deployed in actual production environments: from AI-powered hiring target research, recruiting outreach, and candidate screening to resume parsing and onboarding (and everything in between). It turns out AI’s (usually) hidden biases are problematic (no one is surprised) – but they also have the wonderful capability of reinforcing themselves.

Researchers already know that LLMs pick up human biases from their training data. New research suggests that LLMs can also develop their own biases from experience – and stereotype job applicants more than humans do.

↗ Link