Context Window 100
A milestone today: one hundred of these newsletters. Thank you to everyone who’s read them, and particularly to those of you who respond—always a highlight of my week, so please let me know your thoughts.
Congratulations to subscriber Dan Doyon whose company Readwise (“Strava for reading”—a super piece of positioning) just launched an ebook store, US-only for now. Beyond celebrating success for great people, the interesting aspect is a full range of AI features, including an MCP server to allow agentic access. I also really like Readwise’s ethos of permission over forgiveness in working with publishers on AI features, and I hope it’s rewarded by publishers wanting to work with them and availability in a wider set of markets.
After several weeks of controversy, Meta has backed down on its new commercial terms, which would have allowed crawling of business customers’ websites for AI training. Meta has now promised to consult publishers, including stating that if that consultation runs past the previously announced 20 November effective date, implementation will be delayed until the consultation is complete. It’s not clear which publishers or publisher groups will be consulted—expect that to skew to larger groups and to the news media—but it’s positive that industry and regulatory pressure has brought about a change of thinking.
On the subject of web scraping, Helen King put together a useful summary of robots.txt policies across major academic publishers, as compared with the news media—as a snapshot, some 53% of academic sites blocked at least one AI bot, compared to 86% of news sites. I was curious to see how this compared with trade publishing and asked an AI agent to look at the top 20 trade publishers: you can find the research on my website, and if you’re at a major publisher, yours may well be featured. Across 27 UK and US domains, only 19% blocked at least one bot. This is five minutes of raw research, and the agent makes a good point that there are some rational reasons for a trade house to approach this differently. Incidentally, guess which of the Big Five gets a gold star for being the only publisher to specify Universal Commerce Protocol (UCP) and Model Context Protocol (MCP) endpoints for agentic browsers and shopping…
Those endpoints matter as AI agents become part of the everyday discovery and purchase experience for readers: Meta announced that its Muse agent has 10 million users, of whom 25% are using it daily. Shopping is the biggest use case.
The New York Times added AI features to its cooking app. With knowing irony, Nieman Lab used AI to analyse the sentiment of reader comments on the announcement. It’s a good reminder to cookery publishers—or indeed anyone publishing practical non-fiction—that just because an AI feature is possible doesn’t necessarily make it a good idea.
Toymaker Hasbro came up twice in my research this week, each time with a transferable insight for publishing. Earlier this year, the company launched Sixth Wall, an AI studio to develop and license Hasbro IP. ElevenLabs is one of the launch partners. In a podcast interview this week, Hasbro CEO Chris Cocks talked about the potential and the need for guardrails. The company calls this category Behavioral Licensing, “a new category of IP licensing focused on how characters think, speak, and interact in dynamic experiences.” Food for thought for the biggest character-led fiction brands.
In the same interview, Cocks reflected on the huge ongoing success of collectible card game Magic: The Gathering. A potential wrinkle in that: workers responsible for Magic and Dungeons & Dragons at Hasbro subsidiary Wizards of the Coast will vote in a union ballot after the corporation declined to recognise a trade union. A “robust and enforceable” AI policy is one of the principal demands made by workers—echoing discussions at my old shop, Hachette, and other book publishers this year.
Design platform Canva introduced its own image generation model, Canva Craft, which appears to perform strongly. It’s a really interesting contrary bet to Adobe. Adobe’s strategy is a commercially-safe, middle-of-the-road proprietary model built around photography, with a marketplace giving access to models from frontier labs. Canva is betting on its own model and training on its own templates. For publishers, the choice is partly photos or layout, and partly whether you want tight integration with a particular platform, or the ability to swap out models as winners become clear.
If anyone wants a longer, thoughtful piece of reading for the weekend, a couple of great options. Two of OpenAI’s economists published an essay called The Future of the Firm, hypothesising that AI comes into a company’s org chart at one of three levels: beneath us (we all become managers of agents), beside us (we become its colleagues) or between us (as a coordinating layer between managers and workers). I talk about the first scenario in my book, alongside some of the problems with governance and decisions if people are getting coordinated—let’s be honest, managed—by AI…
And for all round wisdom and perspective, science fiction writer Naomi Alderman’s interview with Microsoft AI is worth five minutes of everyone’s time. This is as good a mission statement for the next hundred of these newsletters as I can think of:
“We find out by seeing how people actually use things, and we are still in the very early days.”
I wanted to finish this week by sharing a practical tip for a scheduled AI task that has saved me time every week. I use Google Meet and get it to transcribe meetings by default. I created a scheduled task in Claude to check the Google Drive folder where those transcripts sit at the end of every working day, review them and add dates and action points to my diary via the Google Calendar MCP. The same principle would work with Teams, with the prompt pointed at SharePoint or OneDrive, or Zoom via Claude’s MCP connector.
The advantage of using Claude for this over Google’s own functionality is that it knows my context outside the transcript and more reliably picks up on mistakes—for example, it found the word ‘ashet’ in a raw file, worked out that it was Google transcribing me pronouncing Hachette with a silent H and corrected it.
This now saves me 5-10 minutes per meeting, and a busy week could be up to 20 meetings, so it’s hours back. Get in touch if you’d like to figure out what this looks like for you.