11 January
News
OpenAI does health
OpenAI launched a new dedicated health product, which can connect to (US) medical records and personal health data (eg Apple Health) and give explanations and advice. There are plenty of pretty obvious things to say about this (error rates, unbundling…) but the general issue is that OpenAI is trying to work out how to build a product that will give it sustainable defensibility given that the models remain commodities. Meanwhile, recall that Google has spent something over a decade working on health data, and then read the next story. LINK
AI as a Gmail feature
Google is (finally) adding AI overviews, summaries, and suggestions to Gmail. Incumbents always try to make the new thing a feature, but they’re often right - it makes a lot more sense to integrate this into the product than to try to have ChatGPT look at Gmail from the outside and try to make suggestions. More importantly, though, this is a story of distribution: if the models are commodities, how do you compete? OpenAI is trying to invent products, but Google, Meta, Amazon, Apple, and Microsoft already have products and surface areas. LINK
Grok content problems
A small story last week and a big one this week: the ‘Grok’ chatbot for Elon Musk’s xAI will happily make pornographic images and ‘deep fake’ nudes of real people on request, and some of that is borderline or actual CSAM, which is straightforwardly illegal in many countries. It’s not easy to stop this, and perhaps impossible to prevent entirely, but companies have to try. This is what happens if you presume that any content moderation or safety processes are ‘censorship’, and the fact that half of Elon Musk’s own posts on Twitter are now promoting some flavour of white supremacy means he no longer gets the benefit of the doubt, especially when he now claims this is a free speech issue. LINK
Meta electricity
Power has become the big bottleneck for US AI datacenter deployment, and this week Meta announced $6.6GW of deals with power utilities. LINK
Ads and AI
Google is adding checkout and ads inside Gemini, together with a new proposed standard (obviously). Meanwhile Walmart is also including ads in its AI shopping assistant, joining Amazon’s Rufus. GOOGLE, WALMART, AMAZON
Reddit launched its new AI ad platform around CES. LINK
And Amazon, which sold about $65bn of ads in 2025, has an extended push for more growth. LINK
Amazon’s off-site agents
Amazon has a new agentic commerce project that scrapes small third-party e-commerce sites (especially those using platforms like Squarespace, Shopify, or WooCommerce) and lists the products for sale on Amazon - if you make a purchase, it looks like a normal Amazon purchase, but Amazon bots go and make the purchase for you. Amusingly, Amazon is suing Perplexity for doing more or less the same thing as Amazon.com itself. LINK
Accenture’s AI strategies
Accenture will buy the UK AI company Faculty, apparently for $1bn. Faculty builds enterprise and government tools, and its founder will become CTO of Accenture. Meanwhile, in its most recent quarter, it reported $2.2bn of new ‘Advanced AI’ bookings, about 10% of the total, and said that it will stop reporting this number because all of its projects now touch on generative AI.
There are a lot of different opinions about what generative AI means for Accenture and its peers. Their business is building and running large, complex IT systems, and AI totally changes how code will be written, while agents will probably change how complex systems are run. But very few big companies will make that change by themselves, and meanwhile, all of this (waves hands) means lots of challenges and opportunities that they’ll need help with as well. So do the clients hire Accenture to shrink Accenture TAM? Or is there a new, maybe bigger but quite different TAM? Maybe the Occam’s Razor here is that if you boast that you are ‘nearing the goal of 80,000 AI & data professionals’, but also think you need to buy a startup with 600 ‘AI native professionals’ and make the founder your CTO, you probably haven’t worked it out at all.
Finally, note that though they are all called ‘consulting’, this is a very different business to Bain, BCG, and McKinsey, which have their own AI questions entirely. LINK
Amazon stores
The Information reports that Amazon is planning to open a big-box general retail store, looking something like a Walmart or Target. Amazon physical retail has a long history of experimentation, on the basis that the value of finding a new hybrid model that works is worth all of the failed experiments - expect more of the same. LINK
Apple Card moves
Noted for the record - after a year or two of discussions, the Apple Card credit card will move its back-end from Goldman Sachs to JPMC. Goldman pushed into consumer a few years ago as a potentially cheap source of capital, discovered it was harder and more expensive than expected, and is now backing out. (Also, the Apple Card does a few things that apparently make it more expensive to run, such as giving everyone their monthly statements on the first of the month, which puts more load on support.) LINK
Ideas
Steven Sinofsky’s annual write-up of CES. LINK
A lot of today’s AI use-cases are automating boring functions in the back office of giant industries that most people know nothing about. This week’s example: JP Morgan’s asset management arm has to decide how to vote the shares that it holds in all of its portfolios. That’s thousands of companies, and it doesn’t have the resources to do all of that itself, so historically it and other asset managers used recommendations from a small number of third-party advisors (‘proxy advisors’): now it will replace them with an internal system based on generative AI. LINK
A New Year essay from OpenAI project chief Fidji Simo: they will build lots of different products to close the ‘capability gap’ between what the models are theoretically capable of and how people can actually use them. Yes, but that gap is the entire problem, how much does that mean new features on existing products (eg Google’s AI Overviews) on one hand and thousands of new companies on the other? Can OpenAI invent all the new use cases? There were lots of third-party apps on Windows but Microsoft had Office as the anchor - what’s the analogue here? LINK
WSJ on Tesla’s robot project. There’s a huge wave of excitement about humanoid robots right now, but also plenty of questions. On one side, new batteries and motors make the mechanics much more practical, and AI means they don’t fall over. On the other side, AI also means they might be able to do a lot more tasks that could not previously be automated (or needed many different single-purpose systems). The trouble is, making a motion guidance system that can copy a dance or a back-flip is a radically different kind of problem to ‘go into a strange apartment and make a cup of coffee’ - which actually has very little to do with robotics. A revealing video from CES shows a humanoid robot doing a brilliant and complicated dance sequence… but drifting into a table and being unable to handle that. So… this might turn out like autonomous driving, where a breakthrough in machine learning meant it went from not working to working 90% of the time, but a decade later the last 10% is still not really working.
More specifically, Tesla had a very clear thesis for how it could get far more unique driving data than anyone else, and even that thesis remains unproven. I don’t know of an equivalent thesis for how it could have more training data for robots. LINK
An interesting Uber paper on managing demand and driver availability at airports, which are about 15% of gross mobility bookings (so, $13-14bn). LINK
Another marketing AI case study for the library, this time Colgate/Palmolive. LINK
An unintentionally fascinating interview with the creator of the niche social network Mastodon. You wish social networks didn’t do all the things that social networks do to drive growth and engagement, so you build one that doesn’t do any of those things: no virality, no recommendations, not even any search. And then, you’re surprised that you don’t have many users or much engagement. Why could that be? “The users need education”. It’s the same argument of ‘false consciousness’ we heard from open source people over and over again - “one day everyone will wake up and realise they are oppressed”. LINK
The AI food delivery hoax. AI means there will be a lot more of this: journalists will need to be very sceptical of ‘leakers’ and ‘whistleblowers’ now that you can generate a fake 50-page ‘internal document’ in a few minutes. LINK
A good analysis of the Warner takeover battle. LINK
Apparently, Meta is running lots of gambling ads despite their being against its policy. LINK
A detailed analysis of just one of the terrible design decisions in the latest macOS: inscrutable, useless little icons everywhere. LINK
Outside interests
This little story has been everywhere either the last few weeks: the Ford engineer who suggested putting an arrow on the fuel gauge to remind people what side of the car to find the fuel tank. LINK
An unknown, newly discovered sketch by Michelangelo for part of the Sistine Chapel. LINK
A profile of Leica, which has successfully turned cameras into a luxury good as smartphones (mostly) removed their original use-case. Annual revenue is now about €600m. LINK
Data
McKinsey put together a very detailed rundown of time spent, revenue, profit, and profit per hour watched for different kinds of entertainment, from live sports to linear video. This must have been a lot of work and will show up in a lot of PowerPoints. One to bookmark. LINK
Column
AI questions
As I wrote in the last two weeks, conversations about generative AI have become specialist debates at every level of the stack from chips to GUIs and advertising DSPs, and each big tech company now has its own set of questions, often quite different.
Something similar is also happening outside tech. Marketing is the first industry to be submerged by the wave, with automated content creation already starting to reach scale, automation of back-office processes perhaps bigger, and the big advertising networks scrambling (just look at WPP). The same applies to the hundreds or thousands of companies that are using this technology to automate back-office processes and workflows: no one else knows those problems (see the JP Morgan story above, for example).
On the other hand, as I pointed out in the presentation I published at the end of last year, the deployment of important new technologies tends to start with automating the problems we already have and know we have, but the next steps are creating new kinds of products and revenue entirely, and then, sometimes, redefining what the market or product could be (so-called ‘disruption’). Uber doesn’t sell booking software to taxi companies and Airbnb doesn’t sell software to hotels.
Those kinds of changes are, inherently, much harder to predict, and they’ll come one at a time, in different industries - Uber and Airbnb came separately too. There’ll also be many false starts - a lot of what we see now will end up like Netscape, Pointcast or Yahoo.
That said, there are some deterministic levers we can point to, and the big ones I think about now are first that there will be a collapse in the cost of creating generic kinds of content, and in particular video, and secondly that there will be a collapse in the cost of processing, summarising and synthesising information.
If the internet meant you could get that 200-page document in five minutes instead of going to a library or waiting for it to be posted to you, and Google means you could discover it existed, an LLM can read it for you, and 100 others, and give you the answer you were looking for. So what industries are protected because that used to be hard, but now it will be easy? And things were impossible because you’d have needed a million people to read all the stuff, and now an LLM can do it?
Turning this on its head, there was a joke a few years ago that half of AI will be turning three bullet points into emails and the other half will be summarising emails into three bullet points. Brands are now using AI to turn ‘three bullet points’ into hundreds of ads, but won’t the LLMs be reading the ads? Amazon has something close to a billion SKUs, and doesn’t really know what they all are - now the machine can look.
I don’t know how that will play out, and, as I also often suggest, it may be the wrong question - that’s certainly how it feels looking at the questions we asked in 1997 or 2010. But the real end to these kinds of conversations is the current Netflix battle, which is now entirely a conversation for the media industry - Netflix, in that sense, is now entirely a media company, not a tech company. Streaming graduated from ‘tech’.