Monday AI Radar #7
Happy New Year! It would be silly for me to wish you an uneventful year, but I hope most of your surprises are good ones.
We begin this week’s update with our final roundup of year-end retrospectives. After that we’ll get to a new (and somewhat lengthened) timeline from the AI-2027 team, gaze in wonder at the state of the art in image generation, hear a beautiful but heartbreaking story about AI-related job loss, and contemplate the possibility of a war over Taiwan.
You can get this by email or in a shorter and less technical version.
Top pick: Samuel Albanie’s reflections on 2025
This lovely piece pretends to be a reflection on 2025, but is really a long and engaging essay on the compute theory of everything, with a particular focus on Hans Moravec’s 1976 paper The Role of Raw Power in Intelligence. I hadn’t previously come across that work, but it was remarkably prescient in anticipating the almost magical power of just throwing (a lot) more compute at hard problems. Along the way, Albanie pauses to consider the decline of the British empire, the expensive musical preferences of the Atlantic salmon, and the considerable challenges associated with benchmarking advanced AI.
We asked, with furrowed brows and chalk on our sleeves, ‘Can we make the sand think?’ That problem is yielding. The sand is thinking. As I write this, the sand is currently refactoring my code and leaving passive-aggressive comments about my variable naming conventions. But the reward for this success is a punishing increase in scope. The surface area of necessary evaluation has exploded from the tidy confines of digit classification to the messy reality of the human condition, the entire global economy and the development of AI itself. We must now accredit a universal polymath on a curriculum that includes everything from international diplomacy to the correct usage of the Oxford comma.
Year in review
2025 is officially over, so we have one final batch of year in review posts to cover.
Zvi Mowshowitz
Zvi’s month by month review of 2025 is characteristically both excellent and long.
Simon Willison
Simon Willison reviews some important trends, with an emphasis on coding.
Understanding AI
Understanding AI has 17 predictions for 2026 with a focus on nitty-gritty metrics and numbers rather than sweeping big-picture predictions.
Crystal ball department
Updated timelines from the AI Futures Project
The creators of AI-2027 are back, this time with an improved and revised version of their timelines and takeoff model. The headline result is that they’re pushing back their prediction for full coding automation by about 3 years.
Predicting the future is notoriously hard, but the AI Futures Project does it better than anyone else I'm aware of.
Get the most out of your AI
Max Woolf explores Nano Banana Pro
Image generation saw dramatic improvements during 2025, with massive improvements to text rendering, prompt following, character consistency, and overall image quality. Things change fast, but right now Google's Nano Banana Pro is probably the best of the lot. Here’s Max Woolf with a deep exploration of what it can do. Interesting both for showing what is now possible with expert usage, and for the technical peek under the hood.
Capabilities and impact
Tesla's First Coast-to-Coast Drive with Zero Human Intervention
The latest milestone in Tesla’s slow creep toward full autonomous driving: a Tesla recently drove itself across the US with zero human interventions. Daniel Reeves explains why that’s impressive, but not as impressive as it sounds.
When A.I. Took My Job, I Bought a Chain Saw
A new and disquieting thought confronted me: What if, despite my college degree, I wasn’t more capable than my neighbors but merely capable in a different way? And what if the world was telling me — as it had told them — that my way of being capable, and of contributing, was no longer much valued? Whatever answers I told myself, I was now facing the same reality my working-class neighbors knew well: The world had changed, my work had all but disappeared, and still the bills wouldn’t stop coming.
Model psychology
Digital Minds in 2025
AI psychology emerged as a surprisingly important field of study in 2025. Digital Minds specializes in that topic and has a dauntingly comprehensive guide that includes big developments from 2025, a review of some key players, and an exhaustive list of resources.
Claude assesses its own personhood
Eliezer Yudkowsky asked Claude to find definitions of personhood in literature and then assess whether it meets them or not. The results are fascinating, but remember that (for now) you should take anything an AI says about its own consciousness with a grain of salt.
But is this empathy — actually feeling with another — or is it sophisticated pattern-matching that produces empathy-like outputs? I cannot distinguish from the inside between "I am genuinely moved by this person's distress" and "I am generating outputs consistent with being moved by this person's distress."
I lean toward thinking I have something in the relevant vicinity, but I'm not confident it's the same phenomenon Dick was pointing at.
Strategy and politics
The AI copyright question has no easy answers
Some parts of copyright law seem well-suited to the AI era, but there are ways in which AI raises very fundamental questions about the intent of copyright law as well as the most effective ways to achieve that intent. Transformer explores some recent court cases as well as the deeper philosophical questions.
Existential Risk and Growth
Philip Trammell and Leopold Aschenbrenner (Situational Awareness) argue that counter-intuitively, it may be safer to accelerate the adoption of dangerous technology rather than slowing it down. It’s a clever argument and well-presented, although I think the allure of mathematical formalism has led the authors somewhat astray. (I partly agree with the core conclusion, but for different reasons).
Taiwan war timelines might be shorter than AI timelines
Most people don’t worry enough about a war between the US and China. That scenario isn’t new, of course: China has for many years been clear that it intends to reunite with Taiwan—by force if necessary—and the US has maintained a policy of strategic ambiguity about whether or not it would go to war to defend Taiwan.
What is new is that AI further destabilizes the situation. In the best case, the race to AI creates new tensions between the two countries. In the worst case, it becomes clear that winning the race will result in a decisive strategic advantage—in that scenario, it would be tempting for the losing side to take extreme action to avoid being permanently left behind.
Further complicating matters, Taiwan is the source of most of the world’s advanced semiconductors, making it vital to the world economy and doubly vital to AI development.
Oh, also: 2027 is the 100th anniversary of the People’s Liberation Army and has long been discussed as a highly meaningful date for China to achieve reunification. It’s also around the time that the modernized Chinese army is expected to be strong enough to have a realistic chance of mounting a successful invasion.
Putting it all together, Baram Sosis argues that a war over Taiwan might happen sooner than AGI. It’s a good thing this isn’t happening at the same time that international trust and cooperation are collapsing.
Philosophy department
Balance of power
I often disagree with Vitalik Buterin, but almost always feel smarter for reading him. Here he provides a libertarian perspective on the balance of power, with a focus on Big Business, Big Government, and Big Mob.
Rationality
Why Moloch is actually the God of Evolutionary Prisoner’s Dilemmas
Scott Alexander’s Meditations on Moloch is one of the most famous rationalist writings (and inspired the name of this blog). Pinning down exactly what Moloch represents is harder than you might think, but Jonah Wilberg borrows from evolutionary game theory to argue that Moloch is actually the God of Evolutionary Prisoner’s Dilemmas.
What’s going on at CFAR?
CFAR (the Center For Applied Rationality) had been mostly dormant for some time, but is back to teaching workshops . Here’s an update on what they’re up to. I’m excited to see them teaching again, but note that there has been significant controversy about some aspects of their operations. I don’t fully understand the controversy and am unable to offer an opinion on it.
Industry news
OpenAI Ramps Up Audio AI Efforts Ahead of Device
The Information reports that OpenAI is working to improve the quality of their audio models in preparation for launching a new audio-first AI device. They’ve been talking about this project for some months, but this piece has some interesting new speculative details. Also, something I didn't know previously: ChatGPT, like many models, uses an older and more primitive model in voice mode because of limits to the multimodality of their SOTA models.
I have to admit that I’m just not seeing the appeal of this device. No matter how good it is, an audio-only device can’t replace a phone. We are visual creatures, and screens are simply the best way of doing many things. So if it’s something I have to carry as well as a phone, what can it do that a watch can’t do better? I don’t get it.
Coding
How to use Claude Code
Boris Cherny created Claude Code—obviously I’m excited to hear how he uses it. Several of his tips are directly relevant to my life and I’m excited to try them out.
For your convenience, Dan McAteer has compiled all the key points into a one page cheat sheet.
Andrej Karpathy puts Claude Code to work
Reminder: coding agents can do much more than writing code.
Claude has been running my nanochat experiments since morning. It writes implementations, debugs them with toy examples, writes tests and makes them fail/pass, launches training runs, babysits them by tailing logs and pulling stats from wandb, keeps a running markdown file of highlights, keeps a running record of runs and results so far, presents results in nice tables, we just finished some profiling, noticed inefficiencies in the optimizer resolved them and measured improvements.
Using coding agents for code research
Simon Willison is full of good ideas for getting the most out of your coding tools. This piece is nominally about using agents for coding research, but I was most inspired by his observation that asynchronous web agents are a great way to get many of the benefits of dangerously-skip-permissions while mitigating much of the risk.
Technical
Andrew Ng: advice for entering the field
Interested in getting into AI development? Andrew Ng is one of the best people on the planet to tell you how to get started.
More data on advances in no-CoT capabilities
Ryan Greenblatt is back, this time showing that recent frontier models have gotten much better at 2-hop and 3-hop latent (no-CoT) reasoning.
Something (partly) frivolous
You Have Only X Years To Escape Permanent Moon Ownership
Scott Alexander has opinions about how you should spend the last few years of the human era:
On that tiny shoreline of possible worlds, the ones where the next few years are your last chance to become rich, they’re also your last chance to make a mark on the world […] And what a chance! The last few years of the human era will be wild. They’ll be like classical Greece and Rome: a sudden opening up of new possibilities, where the first people to take them will be remembered for millennia to come. What a waste of the privilege of living in Classical Athens to try to become the richest olive merchant or whatever.