At the start of the year, I wrote that AI in 2026 would play out as some mix of boom, bust, and backlash. September was the boom and the clearest sign of it was AI starting to do research of its own.
Stanford HAI’s 2026 AI Index opens with the finding that AI capability is still accelerating. On SWE-bench Verified, a widely used coding benchmark, scores rose from 60% to close to 100% in a single year. Yet the 425-page report barely mentions recursive self-improvement, which is the idea that AI systems might one day build their own successors. The Economist took that idea seriously in June, describing how artificial intelligence got better at building itself. Andrej Karpathy handed his model-training work to an agent that beat his own best result by 18% within a week while Anthropic co-founder Jack Clark puts the odds of an AI building its successor without human involvement by the end of 2028 at 60%. Granted, not everyone expects that pace. In Transformer, Lynette Bye explains why a shortage of real-world data could slow the race and Celia Ford profiles researchers betting billions that scaling alone isn’t enough.
Meanwhile, the progress in mathematics continues. Earlier this month, OpenAI said an internal system of around 10,000 agents had resolved the Navier-Stokes problem, one of the seven Millennium Prize Problems, in just 88 hours. The Economist reported that the result raises profound questions about how the discipline works and, days later, around two dozen Fields medallists signed an open letter arguing that using famous problems as benchmarks harms mathematics. Their view is that a proof matters for what mathematicians learn while working towards it. Tudor Achim’s TED talk offers the optimistic counterpoint, arguing that formal verification can take over the checking work that human reviewers can’t keep up with. But Govind Pimpale’s essay for AI Frontiers shows why the debate matters beyond academia. If AI-driven mathematics breaks the assumptions behind public-key encryption, defenders who don’t understand why an algorithm failed can’t know whether its replacement will fail too.
There are similar concerns in biology. Anthropic announced that 950 Claude agents had flagged an uncatalogued DNA pattern in 21 hours, which it compared to the early clues that led to CRISPR. Biologist Lucas Harrington says finding such patterns “is often the easy part”; working out what they do is hard. Iulia Georgescu and Venkatesh Narayanamurti, writing for Cosmos Institute, explain the difficulty: much of what it takes to reproduce an experiment never makes it into print. And yet the results have real world consequences. For example, Axios reports Anthropic’s disclosure that it disrupted five attempts to use its models for work that could support bioweapons, while the Model Hardware Standard lets AI operate lab equipment directly.The same tools that empower scientists could empower bad actors too.
Still, there is hope. In her TED talk, Silvana Konermann of the Arc Institute describes a plan to run a billion experiments to train a virtual cell that could point to treatments for complex diseases such as Alzheimer’s. And some of this is already reaching patients. For a feature I wrote for Acumen called Tomorrow’s Medicine (PDF), I told the story of KJ Muldoon, who in 2025 became the first patient to receive personalised CRISPR gene editing, and included insights from Sourabh Pagaria of Siemens Healthineers, who warns against innovations that are clever but help neither patients nor doctors.
So, where do we go from here? Earlier this year, for a feature in Acumen on the philosophy of technology (PDF), I profiled futurist Ray Kurzweil and philosopher Yuk Hui. Kurzweil sees exponential progress that will transform medicine and extend lives while Hui warns that AI systems acquire organising power over the people and institutions that adopt them. September showed both forces at once. The tools that might decode human cells can also lower the barrier to dangerous biology and the agents that solve old problems can strain the institutions that verify knowledge. Which of those futures arrives will depend on the choices we make about how these systems are designed and governed. As J.G. Ballard writes in Crash:
“The marriage of reason and nightmare that has dominated the 20th century has given birth to an ever more ambiguous world.”
For more inspiration, here are some of my favourite Seth Godin quotes from his 2025 blog posts, together with a recommended title from Blinkist.
“AI is a tool, and judgment, for the foreseeable future, remains our job. It doesn’t matter how cool your hammer is, it’s still on you to decide which nails need hammering. And to be responsible for what happens when you use it.” ~ Hallucinations are not the same as errors
Learn more: What Is Intelligence?
“AI isn’t impatient, easily bored or distracted. It’s insatiable.” ~ Use a lot of words
Learn more: Prompting Made Simple
“As AI expands, the real opportunity is to find a way to use human effort to create more value.” ~ Productivity, AI and pushback
Learn more: Human Edge in the AI Age
“Attention is one of our precious resources, and our culture benefits when it’s not centrally controlled.” ~ Understanding carriage
Learn more: This Is for Everyone
“Connection is powerful and magical. It’s also enervating, subject to manipulation and addictive.” ~ All of it, all at once
Learn more: Digital Exhaustion
“Hiring an AI to work for you and getting very good at producing value feels like the future for most programmers, creators, business development folks and marketers.” ~ Freelancer as centaur
Learn more: The Skill Code
“Human beings are easy to trick. Hopefully, our AI agent will be at least as smart and careful as the company’s.” ~ Best available
Learn more: Charlatans
“Human beings are used to being productive by decreasing the amount of time and effort we put into something. Computers don’t work that way. Give them instructions on how to take the long way around and you’ll both come out ahead.” ~ Two useful AI tactics
Learn more: Prompt Engineering for Generative AI
“If you’re not using the latest AI models, you’re falling behind.” ~ User interaction design drives outcomes
Learn more: AI Valley
“It doesn’t really matter that AI doesn’t “know” what it’s doing. Most of the time, we don’t either.” ~ Hallucinations and human work
Learn more: The Thinking Machine
“It’s entirely possible that a magical AI will replace every single human job and then destroy the Earth. But it’s far more likely that the pattern of the last five hundred years will continue.” ~ Job churn
Learn more: The Age of Extraction
“It’s tempting to fear AI slop, because it’s here and it’s going to get worse. But there’s human slop all over the internet, and it’s getting worse as well.” ~ Walk away or dance
Learn more: Monster Transformation
“Simple hacks rarely fix long-term problems.” ~ Notes to myself
Learn more: Your Stone Age Brain in the Screen Age
“Stories are the original human technology.” ~ Building blocks of marketing
Learn more: Shared Wisdom
“Tech is best understood as a new sort of species, one that symbiotically uses us to advance its goals.” ~ Birthing tech
Learn more: Nexus
“Technology has been the engine of cultural and economic change, and it’s no longer concentrated in the hands of a few.” ~ Clarke’s Law (part 2)
Learn more: The Technological Republic
“We become the stories we tell, and the social media algorithms we live with cause us to tell stories we might regret. It’s not our job to be used by social media, or to become tools of the algorithm.” ~ Fermi’s Law
Learn more: Enshittification
“We’re not building intelligence. We’re building culture machines. Tools that can compress and reconstruct the patterns of human expression. That’s not a bug. It’s the feature.” ~ The poetry machine
Learn more: Read Write Code
“When AI shows up, our mistake is thinking that if we can’t find useful brilliance in one simple prompt, it’s broken.” ~ The AI effort gap
Learn more: Too Smart
“When new technology shows up, some people ask, ‘how can this make my job easier?’ But what happens if we ask, ‘how can I use this to do something really hard?’” ~ Finding the difficult work
Learn more: The Practising Mind
“When we choose media that has the same goals as we do, it’s likely we’ll get what we came for.” ~ “A now, a word from our sponsor”
Learn more: Streaming Wars
“When we post with intention, we create a pattern that begins to create a structure and a narrative we’re pleased with.” ~ New post
Learn more: Blank Space
“Where we create our media and how we consume it are still up to us. It’s true, at some point, that the medium is the message.” ~ The Hotel California (and subscriptions)
Learn more: Reconnected
“You’re either going to work for an AI or have an AI work for you. Which would you prefer?” ~ Trusting AI
Learn more: AI-Powered Leadership
“You’re probably not going to end up with a million followers by adhering to the rules of the algorithm. But that’s okay, because you don’t need a million followers to make a difference.” ~ System architect/system victim
Learn more: The Cybernetic Society
(If you found this valuable, please subscribe below for monthly updates!)
