12 July 2026
News
Model week
After Anthropic spent a couple of weeks leading the market with Fable, OpenAI caught back up with its new 5.6 models, which are roughly at par. This has been the pattern since the beginning of last year - the top handful of labs jump ahead of each other every month or two, but no-one has a sustainable lead, and there is no clear path to getting one. LINK
Proving the point - Anthropic got buzz in tech with CoWork, a software layer building agentic flows with third-party plug-ins, and now OpenAI launched ChatGPT Work, doing much the same thing. The most interesting part is how incredibly chaotic, confused and messy the app experience is - really bad work. Meanwhile, OpenAI’s AI-enabled web browser, Atlas, has been discontinued: we haven’t found the right UX here yet. WORK, APP, BROWSER
Fidji Simo, OpenAI’s head of product and business (who was a potential replacement for Sam Altman), went on a medical leave of absence earlier this year and has now had to step down entirely. The management structure will change again, again (which might be part of why the new app is such a fiasco). LINK
Meanwhile, SpaceX/xAI launched a new version of its Grok model that put it back up towards the top of the leader boards, and more importantly at the ‘pareto curve’ - it has a leading price/ performance ratio. And Meta’s new AI lab also launched a new version of its rebuilt model, Muse Spark, which isn’t far behind. Neither of these are quite at the cutting edge for model performance, but they’re competitive or nearly so on price/ performance, which is suddenly much more relevant. Equally, Mark Zuckerberg explicitly claimed the the current models are too expensive - and followed through with his proposal to make his models available in the cloud to developers, competing directly with Gemini, OpenAI and Anthropic for the first time.
To my opening point above - neither of these have yet leapfrogged the leaders, but they’re not that far off, and they’ve managed to get back into the game after failing out last year. It remains entirely unclear what structural barriers to entry there are beyond cash. SPACEX, META, META CLOUD
Apple versus OpenAI
In 2024 Jony Ive, Tang Tan (a former senior hardware exec at Apple) and a bunch of other mostly-former Apple people founded Io, a startup planning to build an AI device. In 2025 OpenAI bought Io, for $6.5bn, and made Tang Tan head of hardware. And this week, Apple sued Tang Tan, OpenAI and Chang Liu, another former Apple employee also now at OpenAI.
Apple claims that Tang Tan took Apple proprietary data when he left and told Apple employees with job interviews at OpenAI to take ’actual parts’ from Apple’s office and bring them to OpenAI. It claims that Chang Liu not only took Apple data before leaving, but continued to access Apple systems after he left and took more data (he had access because of a bug), and coached other Apple employees joining OpenAI to take more Apple data before leaving, and coached them on how to avoid Apple security checks.
I’m obviously not a lawyer, and some parts of this seem more blatantly problematic than others (and God knows what will come out in discovery), but some of the externalities are interesting. Apple doesn’t mention Jony Ive at all. It does mention that over 400 former Apple engineers now work at OpenAI. And none of this is good for OpenAI’s plans for a consumer hardware device, assuming that’s still live anyway, nor for an IPO. LINK, IO/LOVEFROM
AI export wars
Reuters says Beijing is considering export controls on open-source models, which have seen a surge in interest in the US give the pricing crunch (OpenRouter, with a sample so skewed I’m generally reluctant to reference it, says Deepseek and Qwen alone are 25% of its usage). As I wrote in the essay linked above, all of the variables for token pricing are in play right now. LINK, OPENROUTER
On the other side of this trade, meanwhile, the FT found OpenAI and Google selling access to their models to the Singapore subsidiaries of Alibaba, Baidu and Tencent, which are blacklisted by the USA. LINK
The week in AI
SK Hynix, which dominates the memory market along with Samsung and Micron, did a secondary listing on Nasdaq that raised $26.5bn. The CEO told Reuters he thinks demand will be ahead of supply until 2030. LINK
The industry continues to reconfigure around the token pricing crunch: apparently Microsoft is starting to use its own models for AI features in Excel and Outlook, replacing OpenAI and Anthropic. Maybe it should dive into open source. LINK
Following last week’s wave of AI DeployCo announcements, this week Warner Discovery is working on a new agentic ads stack with AWS. Go-to-market plus integration plus known partners. LINK
Remember when VW partnered with Bosch to work on self-driving? How well was that likely to work? Yep. LINK
The gaming slump
Microsoft’s Xbox division will lay off 20% of its staff and restructure. The hardcore games industry is deep in a slump - earlier this year GDC found that a third of US developers had been laid off in the last two years. But this also has something to do with the mixed success of Microsoft’s big push into streaming games, supported by close to $100bn (!) of M&A (remember the $75bn Activision deal in 2023?). LINK, STREAMING, GDC
Rocket ships
SpaceX has filed with the FCC to operate a new constellation of 100k satellites, each one 2,000-2,500kg, for “for consumers, enterprises, and government users and billions of AI-powered devices around the world”. For context, the current model (V2 mini) is 800kg: Starlink operates about 11k of them with approval for 4k more, and launches them 26 at a time in the Falcon 9. Putting up 100k of the v3 at more than double the weight probably needs the in-development ‘Starship’ rocket, with a planned payload on 100-200 tonnes, implying several thousand launches where SpaceX has done close to 700 so far, all of which sounds practical up to point. (Note also that this system so far would only use frequencies about 10GHz, which doesn’t work for mobile phones.) LINK, FILING
China successfully recovered a re-usable rocket booster (though catching it in a net rather than landing by itself or using SpaceX’s ‘chopsticks’). LINK
Jeff Bezos’s Blue Origin is raising outside capital for the first time, targeting $10bn at a $130bn valuation. LINK
Ideas
Software is dead? Bloomberg says Starbucks has a project to migrate away from Microsoft and IBM to more home-built-with-AI software. It would be easy to see this as the SaaS apocalypse, but it’s also a cycle that’s been going back and forth since the 1980s. LINK
Vibe-coding has produced a flood of new contributions to open-source projects, many of them misguided and all adding to the workload. LINK, DATA
Unsurprisingly, there’s an explosion in the use of AI to make more convincing fraudulent insurance claims - just add damage to a photo of your van! In the UK, Aviva caught 18k fake claims worth £233m. LINK
This week the University of Chicago’s Law School has a new AI policy. Steven Sinofsky (who ran Office and then Windows) looks back to the 1980s, when the same institutions worried about whether ‘word processors’ would destroy critical thinking. 2026, SINOFSKY
The FT discusses the business problems caused by the outspoken right-wing politics of Palantir’s founders. This was an effective technique to break into the market as a startup, but today, gloating on social media about killing people is a problem in your next sales call with a children’s hospital, and positioning yourself as a MAGA company is a problem outside the USA, and even inside the USA when the government changes. Meanwhile, a majority of the actual staff appear to vote Democrat, and the AI researchers would rather work somewhere else. LINK
America’s attack on Iran has repeated the Ukrainian lessons on the need for ‘affordable mass’ - the US uses Reaper drones for reconnaissance at $30-50m each, but Iran has shot down 30 so far and the US only has 135 left. LINK
The Nigerian Islamist group ‘Boko Haram’ uses AI for planning and training. The content restrictions ought to block any dangerous questions, but apparently, “boys that have received extensive training bypass the restrictions - they say they need it for a movie or something like that”. LINK
An argument that China’s lead in power for data centres is more theoretical than real. LINK
On the ethics of people with the right friends (Elon Musk, Sam Altman) getting allocations in hot IPOs. What’s old is new. THIS TIME, LAST TIME
Outside interests
The berry king. LINK
A Mercator atlas from the 1580s, for £1m. LINK
Sotheby’s is selling an Apple Macintosh Plus system from 1986… that’s still in the original boxes, unopened. LINK
10k books in a Manhattan studio apartment. LINK
RIP David Porter, founder of Psion (which I now know stood for Potter Scientific Instruments Or Nothing) and Symbian. LINK
Data
Chart of the week - a Brown University professor ran his first class test remotely and then suddenly switched the next test from remote to in-person. The score for all except two students collapsed (Student 22 is a hero). LINK
Detailed analysis of willingness to pay for news, using data from a ‘major US regional newspaper’. LINK
After last week’s data from Ramp showing increased hiring at companies adopting AI, this week Indeed’s data shows a surge in software development jobs since last summer, albeit up from a low base after the post-covid pullback. LINK
In the last 12 months there’s been a surge in new professional services business creation in the USA. Hypothesis: people starting their own businesses with AI. LINK
Column
Most people aren’t tool-builders
A decade ago, there was a fashionable idea in Silicon Valley that SaaS meant enterprise software could go direct. Instead of struggling through IT procurement processes and 18-month sales cycles, you could go direct, bottom up, to the grassroots, and your actual users would drive demand. This kind of worked, up to a point, for a handful of companies (Slack, say, or Notion), and ‘shadow IT’ is a real thing. But more generally, this ran into three pretty basic problems, that apply in exactly the same way to generative AI today: most people don’t see the problem to be solved, most people are not tool builders, and most people aren't in a position to build the tool anyway.
First, think for a minute about how often you just didn’t get it the first time you saw something that now you use every day. And you’re probably a product-minded early-adopter - everyone else struggles even more. A large proportion of productivity software solves a problem that the people who are using it didn't realise that they had. This is half of the job of sales.
Second, most people aren't tool builders. There's a difference between being the right customer for a tool and being the right person to build it. There's a difference between seeing the problem and knowing the right way to create a tool and the right things it would do - the processes, the functions, the networks, and the workflows that would make it a delight. This isn't about technology and programming -
it's about having the experience and mindset of building tools and building software, and that's a different set of skills and experience in a different person. Somebody who's a really good lawyer, a really good salesperson or a really good video editor is almost certainly not the right person to see the problem that a new piece of legal, sales or video software is solving, and even if they are, they’re not the right person to build it. Being a really great live event producer is a different skill to being really good at designing software for live production.
Third, even if you see the problem, and can think about what the solution might be, you're probably not in a position to build it.
Most of the processes where some new tool and some new automation would be useful inside an industry or a big company touch not just you but hundreds or thousands of other people. They touch proprietary or regulated data. They need to interconnect to carefully protected systems of record. That has to be controlled centrally. Who maintains it, who supports it, who does the compliance audit… and who is that gets everybody else to use it?
I often think here of a company I looked at a decade ago called frame.io, which built a collaboration platform for professional video projects. For any given project, there might be half a dozen actual video professionals, but then there might be 50 people in four different companies who need to see it, and touch it, and comment, and see every face, and mark up a frame or two. Frame.io took this from a bunch of shared Dropbox or Vimeo links, Google Sheets and email threads into a platform. But which of the 50 people involved in touching that piece of video was the right person to say not just “we need a piece of software for this” but to get the other 49 people to use it? Slack started as a collaboration tool inside Stewart Butterfield’s games startup (as did Flickr), and he turned it into a company, but most people aren’t that person.
It's easy, of course, to argue the counter to this. There are plenty of 50 and 100 person company departments that do run on a five meg Excel spreadsheet built by someone who left years ago, and no-one is quite sure what that macro does. The typical big company has several thousand applications accumulated over the last 30 years, some of which are really databases, and some of the databases are really spreadsheets, or should be. There is a very broad, fuzzy continuum from a shared folder to SAP. AI will move the thresholds and demarcations within that.
But the core of this is that giving everyone a new way to make tools doesn’t meant that everyone will make tools. Writing the code isn't the hard part, and making the tool isn’t the hard part - the hard part is knowing that it should exist, and knowing how it should exist, and that’s a different person.