Funny numbers
Late last month, Axios and other media outlets reported that OpenAI’s annualised monthly revenue had passed $70bn. This week, it emerged that this was a mistake and the actual number was ‘close to $50bn’. The confusion seems to be that in general OpenAI has been reporting revenue net of payments to companies that resell its models (Microsoft, AWS) where Anthropic has been reporting gross, and both OpenAI and investors were also trying to gross-up the numbers to make them comparable, without making it clear which numbers they were using when leaking to the media. See this week’s column. LINK
Agent incentives
Agents remain super hot, even if Muse and especially Instinct look an awful lot like the fever-brain excitement around Clubhouse in the pandemic. But if they are going to work, lots of other things need to be plugged together, most obviously how they actually talk to other websites and each other. Meta has launched a new standard (‘Personal Agent Protocol’) with the bunch of logos on the slide. Standards, of course, are one of the key strategic weapons when tech ecosystems try to go to war, and it’s always notable who isn’t on the list, and important to remember that getting the partnership person to sign a partnership deal is not the same as getting the company to use the partnership. LINK
Meanwhile, Amazon and a bunch of other retailers are blocking agents entirely. This is partly short-term concern around fraud chargebacks and liability, and it’s not irrational for them to press pause until they worked out how to manage this more generally. There’s a 20-year-old and 150-year-old question of customer ownership versus distribution. If you’re big and already have traffic, you want to own the experience and talk to the customer directly, and if you’re small, you’re willing to take distribution instead. The additional wrinkle, of course, is that Amazon did $75 billion of ad revenue in the last 12 months, and the agent doesn’t see the ads. LINK
Open outside China
Since Meta abandoned Llama, most of the credible open source models have come from China, which skews the debate about open a fair bit, but now American startup Reflection has launched a pretty credible open model and Mistral, which everyone had forgotten about, has come back from France with a credible model of its own. The FT reports that Nvidia is looking at buying Reflection, as part of the Bank of Jensen strategy to accelerate AI adoption while fragmenting the customer base. REFLECTION, NVIDIA, MISTRAL, ANALYSIS
IPO overhang
I think there were two fundamental changes in AI this year: agentic coding achieved explosive product-market fit, and AI infrastructure capex outgrew the cash flows of the hyperscalers and had to go to the financial markets. There’s been an enormous amount of capital raising of every kind this year, and markets are starting to get uncomfortable: in the last quarter a whole bunch of IPOs were pulled, partly also due to the imminent Anthropic IPO crowding everything out. FIRMUS, BACKGROUND
The week in AI
Anthropic quietly updated its terms and conditions to say you’re not allowed to be “cruel” to its models. This is more of the anthropomorphism that seems like sensible speculation at a minimum to some AI researchers and insane to a lot of other people, as I wrote in my column last week. Meanwhile, it creates circularity problems - if you tell a model it has X characteristic, it will try to look like that. LINK
OpenAI followed Anthropic in building a text watermarking system to comply with EU requirements. Unlike Anthropic, this is only applied in the EU, and OpenAI is also a lot clearer that they don’t think this is really a good idea: the tech is not reliable enough and too easy to work-around. LINK
Google launched Playground, an experimental space to use AI to generate your own games. AI is clearly going to transform both the economics of creating games (especially the hugely-expensive content creation of AAA games) and the mechanics of how games work (the NPCs can actually be intelligent now), as well as letting people spin-up improvised games like this. Compare mobile and app stores, and 3D. LINK
Google launched a native Mac app that can connect to all of your software and take notes on what you’re working on, all running locally on-device. I don’t think it’s clear yet whether things like this are a feature or product, but meanwhile it is both predictable and hilarious that this looks nothing like a native Mac app, the download link doesn’t work in Safari, and the app is called “Google AI Edge Foresight”, which is exactly how Microsoft names products. LINK
SpaceX Mobile
Following its earlier comments, SpaceX bought a portfolio of low-frequency cellular spectrum in the USA that would be needed to build an actual terrestrial network. US mobile operator stocks fell sharply. As I’ve I pointed out before, while SpaceX technology is extremely cool, satellite is either impractical or useless for the vast majority of real cellular use, and so if SpaceX actually wants to build a US mobile service, it will have to build what everyone else has built, spend more or less what anybody else would spend, charge more or less the same prices, and provide probably a mostly inferior service (the best sites are already taken). The most cynical interpretations remain that this is really leverage to get a good MVNO deal. LINK
Apple as shadow ads regulator
Last month Apple briefly blocked one of the big ad network domains on iOS, and now there’s a wave of speculation that it’s working on a blacklist of a huge number of third-party targeting services, all of which allow advertisers to do terrible evil anti-democratic things to violate your human rights, like show you products you might be interested in and make money for independent publishers.
Apple here is an interesting example of what’s sometimes called a shadow regulator: it’s doing things that are really public policy rather than engineering or product decisions, but doing them without any due process, accountability or debate on devices used by well over half of Americans. LINK
|
