One of OpenAI’s many product announcements in the last six months was direct integration of shopping into ‘Instant Checkout’ - purchasing in the chatbot feed, so that you can ask ChatGPT for a product recommendation and then click ‘buy’ right there, powered under the hood by the ‘Agentic Commerce Protocol’ that it built in partnership with Stripe, with integration (opt-in) to Shopify stores and Etsy. (This is separate from the launch of apps, where third parties can enable an embedded widget of their own experience to show up inside the feed, if suggested by ChatGPT.)
Now, apparently, OpenAI is pulling back from this, while Shopify said that only a dozen of the millions of stores on its platform were enabled so far.
There’s a lot to unpick here, some of it very specialised, and this is all still very early - AI today is where the web was in 1996 and 1997, so there are lots of ideas and acronyms floating around that won’t work (and there’s plenty of competition for agentic commerce standards). But integrating shopping into general-purpose search and discovery layers has been a tar-pit for 20 years: Google and Meta have both failed at this multiple times. It’s very hard to standardise metadata for millions of merchants on a single platform UX, you need to offer merchants a lot of distribution before they’ll give up control, and it’s not clear what the consumer flow should be either. Theoretically, I should be able to give an AI app a photo of my fridge and a recipe and tell it to buy the ingredients I don’t have on Instacart, and theoretically I should be able to tell an app to look at my Instagram feed and order some coats that work in a New York winter and freshen up my look. But it’s far easier to say that than to deliver all of the product, BD, and consumer behaviour around it.
There’s also a broader point - how many products has OpenAI announced lately, how many of them had real strategic depth behind them, and how many of them were cargo cults? I suggested in my essay on OpenAI’s surety that it’s frantically trying to build differentiation on top of a commodity product, but I’m not sure how coherent that effort is yet. LINK
Oracle has a very cash-generative legacy business that’s been losing share for a generation, and in the last two years decided to use those cashflows and borrow heavily again them to buy its way into being a datacentre provider for AI. Last October it said it plans hosting revenue of $225bn in the 12 months to May 2030: in February it said it would raise up to $50bn in new capital, and analysts forecast $75-100bn of cash burn. The risk premium on its bonds has spiked accordingly, and apparently it’s now planning layoffs, perhaps in the tens of thousands, to free up more cash. LINK
Google’s head of infra says it’s possible it will spend at current levels ($185bn planned for 2026) for a decade, and does the maths. LINK
Netflix acquired InterPositive, a startup created by the actor/producer/director Ben Affleck that uses AI to help automate tedious post-production tasks. Generative AI is a transformative enabling technology, but most real-world applications need it to be wrapped in product and use-case. LINK
Barclays is pushing into stablecoins and tokenised deposits. NFTs were silly; programmable payment rails seem to be useful. LINK
Two AWS datacentres in the UAE were hit by Iranian drones (real damage unclear), while it appears that a $300m radar base station for a US THAAD air defence system may also have been destroyed. Meanwhile, a bunch of reporting suggests that the US had not asked Ukraine for advice on dealing with Iran’s Shaheed drones before going to war.
George W. Bush once said “I'm not going to fire a $2 million missile at a $10 empty tent and hit a camel in the butt.” Iran fired a $20,000 drone at a $300m radar. Arms races are often about cost asymmetries, and drones are a perfect example: it’s not sustainable to shoot multi-million dollar missiles with low production volumes at drones that cost a fraction of that and can be mass-manufactured in a workshop and launched by the dozen from trucks. This is what defence-tech is all about, and China is watching closely. AMAZON, UKRAINE, THAAD
45% of the PC market by units is consumer (and another 10% is education). IDC calls half of that ‘mainstream’ with an ASP of about $500. Apple has a ~$1500 Mac ASP and has always refused to make a Mac at anywhere close to $500 (just as it never made a $150 phone), saying, in Steve Jobs’ words, that it didn’t know how to make one that wasn’t a piece of junk (as they are). Now it’s ready. The ‘MacBook Neo’ starts at $599 and uses the A19 chip from the iPhone 16, plus an aluminium body, and has a 16-hour battery life. This will be more than competitive for mainstream tasks than most x86 machines at the price, and the machined aluminium body (free-riding on Apple’s massive investment in tooling) should feel far more premium. If you know how much RAM it has, this isn’t for you, but if you’re in the market for a $600 laptop this will be great, while it will be very hard for PC OEMs to match. Apple has about 10% of the market by volume (up for 2-3% when Steve came back), and this will give it more.
Stepping back, though - when’s the last time we talked about a new computer? Apple treads its own path, and follows long-term strategies - while everyone else works on AI, it just came for a market that was created 30 years ago. LINK
Tim Sweeney’s quixotic battle against app stores and commissions (except commissions that he charges) is finally winding down with a settlement with Google, in which, hilariously, he agrees not to criticise Google’s politics until 2030. LINK
More fallout from the squabble between Anthropic and the US Department of Defence: the US military is relying on AI tools, including Claude (embedded in Palantir) for targeting and mission analysis. LINK
Emil Michel, a famously aggressive former Uber exec now the Pentagon’s head of AI, gave an interview outlining their view of the problem: you can’t have a civilian supplier trying to second-guess what commanders can and can’t do with a tool. For example, he claims the original contract said you could not use Claude to plan targeting, which seems like a pretty basic use-case. LINK
Dario Amodei of Anthropic, unsurprisingly, takes a different view. LINK
On the other hand, OpenAI’s head of Robotics quit over Sam Altman’s willingness to take over the contract. There are lots of culture clashes here, but the core issue, I think, is over what you’re actually doing when you sell to the military. They will use your stuff to kill people, they will do so knowing it’s not 100% reliable, because nothing they use is, and they’ve been using ‘autonomous systems’ to kill people since the 1950s. The problem here is that this technology is clearly essential to the military of the next decade, but there are only 2-3 possible suppliers, so a government as pugnacious and single-minded (also, simple-minded) as this one will not feel able to take no for an answer. LINK
Cursor, the hot AI coding startup this time last year, thinks that Anthropic and OpenAI are massively subsidising their coding tools - here it claims that a $200/month Claude Code subscription might consume up to $5000 in compute. LINK
China’s OpenClaw craze. LINK
What happens when people never see your app, your website or your marketing? LINK
A fascinating piece on some of the ways that generative AI will democratise and accelerate refund fraud and charge-back fraud. Content marketing, obviously, but good even so. LINK
While the US debates what its existing copyright law means for LLMs, other countries can change their laws, and the UK has been running a big consultation on what that should look like. Unfortunately, that posed so many difficult questions that they postponed the whole exercise indefinitely (avoiding difficult decisions is the defining characteristic of the current UK government). The questions are still there, though - isn’t as simple as ‘piracy’ or theft’, but you’re still building a trillion-dollar industry on a foundation of analysing other people’s work, and what does that mean? LINK
The BBC’s paper arguing for a new funding relationship with the UK government has a lot of interesting discussion of the TV landscape. If you think it’s tough being a US broadcaster handling streaming, imagine being a European broadcaster that used to buy the hits from the USA and now has to compete head-on for audience with people that have 10x your budget. LINK
Adweek on six GEO startups. Less interesting for what they’re doing than for how fast every layer of the AI stack is specialising and fragmenting. LINK
Om Malik looks at the clever engineering behind Apple’s latest high-end chips. LINK
Maybe something, maybe nothing: Alibaba’s Qwen model is currently the best open source LLM (probably), but it looks like there’s a lot of turmoil in the team. LINK
GPS jamming and alternatives to satellites. LINK
BYD’s high-speed EV charging is now almost as fast as gasoline. LINK
Fishy pulpits. LINK
Richard Hell’s tenement. LINK
Gender biases in LLMs. Some of this isn’t surprising (“stereotypically feminine sentences are consistently attributed to female writers“), but this is more interesting: “forms of violence central to the gender parity debate are deemed less acceptable than objectively worse forms of violence”. The more we all agree on something, the less text there is about it in the training data (because you don’t need to write about that), so the less clear, perhaps, it is to the model. Google had similar issues in search 20 years ago: most of the text on the internet about ‘were the Moon Landings faked?’ was written by people who claim it was, so that looked like the best match. LINK
A single-dose treatment for sleeping sickness. LINK
A big survey of British attitudes to technology. The more prosperous and better educated you are, the more optimistic you’ll be. Lots of other data shows similar (eg the Edelman trust surveys). LINK
Dan Fromer’s latest consumer trends deck: food and bev. LINK
Anthropic monthly revenue has doubled since the beginning of the year: “$19bn annualised” (last month multiplied by 12) up from $9bn in December. I would bet that almost all the growth is in agentic coding. LINK
An NBER study finds that the number of ‘books’ produced annually has tripled since ChatGPT launched, almost entirely as ebooks on Amazon. Interestingly, the biggest growth is in travel guides. How much of this is spam, how much is generic material automatically repurposed for sale, how much is augmentation and acceleration? LINK
A good Deloitte study on AI adoption in the enterprise. LINK
McKinsey survey data on consumer use of and trust in AI tools for shopping. LINK
There are lots of micro studies on AI use now - this one from HBR says that people work more and harder if they have AI tools, which is exactly what happened with every other tool. Excel didn’t result in junior investment bankers working shorter hours. LINK
Every now and then, a chart goes viral, generally from a group of economists, that tries to quantify how far a range of industries and professions are exposed to AI, and map how far AI can already ‘do’ that job. These charts are very effective as content marketing, but close to worthless as forecasts or analysis.
The conceptual problem, I think, is that this approach is attempting to treat a job as a rules-based system, in the same way that generations of failed AI research tried to do, and this is impossible at a practical level for the same reasons.
AI researchers wanted to recognise a picture of a dog, and so you start writing rules for how you would do that, and you look for edges, and fur, and eyes… and five years later, you've got a thousand rules and it still doesn't work. Equally, now, you try to quantify a lawyer's job, and so you break it down into a list, and then ask how far an LLM can do each thing on the list, and that gives you two nice neat numbers, but that list of rules has only a very vague and imprecise connection to what lawyers actually do. You’re fooling yourself with spurious precision, because you don’t have any way to collect data with the kind of granularity that you’d need.
Further, you also risk presuming that the job won’t change around the tool. Imagine you'd done this exercise in 1980 looking at spreadsheets. You would have said, well, X percent of a CPA's time is compiling and totalling numbers, and a spreadsheet can do that in a thousand times faster, so X percent of an accountant's job will be automated. But that wasn't a good analysis of what the accountant was actually doing, and once accountants had spreadsheets, the job changed, and so the number of people working as accountants and auditors has grown every decade since 1900, even as the profession has gone through adding machines, mainframes, data processing, PCs, spreadsheets, ERPs, and now cloud.
Part of this is the Jevons paradox, which everyone discovered last year, and which is really just applied price elasticity: if you make it cheaper to do a piece of work, do you do it for less money, or do you do more of it?
The most important part of this is that past a certain point, you don’t just do more of the same work, but you do entirely new things that just would not have been possible before. You couldn’t run an express train pulled by horses no matter how many horses you bought, and you couldn’t run a quant fund without PCs no matter how many accountants you hired.
The counter to this is to say, well, yes, the precise numbers are wrong, but this is directionally correct: AI will affect accountants more than fitness instructors. You could say the same about the internet: you could have made a numeric score for which industries would be most affected, and the scores would be ‘wrong’, but they’d tell the right story.
The trouble is, maybe AI will mean we have more accountants because they can do more and better analysis and meanwhile AI video analysis will replace most fitness instructors. Backtesting against the internet, imagining you’d tried to analyse the likely impact of the internet on hotels in 1997 - you’d have said that the internet would not change the need to own physical assets, and then Airbnb changed that. Half the point of disruption is that you change the market and change the question - remember the finance professor who claimed Uber’s TAM was limited to the size of the legacy taxi industry? The ‘lump of labour’ fallacy ‘human needs are infinite’ - doesn’t just mean that new jobs and new industries get created - it also means that existing jobs and industries get new things to do.
There’s an old joke that a physicist (or economist) is asked to predict which horse will win a race, and they say “first, I presume the horse is a perfect sphere.” This is the same problem. The equations might be flawless, but the inputs don’t capture the thing you’re trying to measure, and your assumptions don’t capture how that will change.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。