Context Window 23
This edition covers the headline findings of Stanford HAI’s 2025 AI Index Report, fresh data on the rapidly growing length of tasks LLMs can perform, OpenAI’s o3 model release, the importance of human leadership in AI initiatives, a Bloomberg investigation into Inkitt’s AI-powered romance factory, and an embarrassing case of an AI-generated passage slipping through Springer Nature’s editorial process.
Hi everyone—I’m publishing a day early this week, as Friday and Monday are public holidays here. It’s been another packed week for AI news, and this week’s stories range from major model releases to new data on consumer sentiment and organisational use. Particular thanks to subscriber Alex Boden for sharing a couple of links with me this week: if you’ve seen something that you think would be of interest to this audience, do please let me know. Anyone who’s been to one of my workshops in the last year will have seen charts and data from Stanford HAI, and the research centre has just published its 2025 AI Index. It’s not publishing-specific, but it offers one of the most comprehensive snapshots of global AI progress and how things have changed since last year’s report. Four highlights stood out to me, including one particularly relevant for publishers: on seven out of eight longitudinal benchmarks, AI exceeded a human performance baseline; more than three quarters of organisations surveyed used AI, up from just over half the year before; consumer attitudes to AI in Asia are twice as positive as in the US or Canada (~80% vs 39-40%)—the UK saw a +8% change in consumer optimism year-on-year, though fewer than half of respondents were positive; and, relevant to publishing, 48% of data from top web domains is now restricted to prevent AI bots training on it. There’s also an interesting data point in this Substack piece: the length of tasks that LLMs can perform to a reasonable standard is doubling every seven months. To put that in context, GPT4 could manage tasks that took humans minutes, and o3 can handle tasks that took hours. OpenAI released its latest model, o3, yesterday. It’s designed to give better reasoned responses, and cope with multi-stage tasks. Image analysis is also greatly improved: for example, I uploaded a photograph of a conference presentation and extracted accurate data from the portion of the image containing a slide. For simple tasks (minutes rather than hours to use the framing above), it may be overkill for most publishers who could use simpler and faster models, but for more complex functions, it sets a new standard. Of course, in such a fast-moving landscape, new models and applications get a lot of attention, but there’s a good counter-argument in this piece for the human dimension, and creating the environmental conditions for AI initiatives to succeed. For publishers considering their strategy, developing such a community of practice is more about people skills than purely technical ones. Bloomberg’s piece on publishing startup Inkitt demonstrates some of the challenges of mixing storytelling with venture scale disruption. The company A/B-tests chapters, develops sequels with AI, offers single-digit royalties for blanket rights grabs, and treats authors as fungible. I write this as an optimist for what AI can do for publishers, but only inside transparent contracts and strong editorial guardrails: if you abandon quality in the short term, what is the prospect for brand trust in the longer term? Speaking of trust, embarrassment all round at Springer Nature, which just published a $120 medical textbook which included content copied from an LLM output. The author hasn’t given a clear explanation, the publisher clearly hadn’t adequately reviewed the first chapter, and the book has now been withdrawn. Pro tip: don’t be like Springer, Ctrl + F (Command + F on Mac), “language model”.
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.