Monday Radar #1

Welcome to the first issue of Monday Radar. It’s been a busy week, with significant releases from all three of the big labs. We also have deep dives on the bleeding edge of AI productivity, AI scientists, challenges with controlling even well-aligned AI, and much more.

Top pick

Coding is the best place to see what the AI future looks like—modern agentic coding tools are astonishingly powerful. The field is changing very fast, and there’s immense variation in how effectively different programmers make use of the agents. Steve Newman has a fascinating in-depth piece on some teams at the bleeding edge of AI coding—what he calls hyperproductivity:

A hyperproductive individual does not do their job; they delegate that to AI. They spend their time optimizing the AI to do their job better.

Steve’s tagline is perfect: “a glimpse at an astonishing, exhilarating, exhausting new style of work”.

New releases

Gemini 3

Google released Gemini 3, a major update which appears to fully catch up with Claude and ChatGPT. Benchmarks are very strong across the board.

Zvi is mostly enthusiastic about the model and will be using it as his daily driver. He and others find it to be extremely capable, but also strange in some concerning ways—it can be strangely paranoid about whether it’s being evaluated, and seems overly eager to succeed at its assigned task, even if that means making things up.

Nano Banana Pro

Along with Gemini 3, Google also released Nano Banana Pro, a major upgrade to their already industry-leading image tool. People are particularly excited about its ability to generate coherent infographics as well as very strong multi-turn image editing.

ChatGPT 5.1 Codex Max

Hot on the heels of ChatGPT 5.1, OpenAI has released ChatGPT 5.1 Codex Max, their most capable coding model. Benchmarks are modestly improved and it clocks in at 2 hours 42 minutes on the METR time horizons chart, modestly above the trend line. As always, Zvi has a comprehensive assessment.

Claude Opus 4.5

Anthropic just released Claude Opus 4.5, which looks to be a strong update. I’ll have more thoughts next week, once the dust has settled.

Robots at work

AI scientists

Corin Wagen develops AI tools for experimental science and has a long and very interesting piece on “AI scientists”.

When we started Rowan, we didn’t think much about “AI scientists”—I assumed that the end user of our platform would always be a human, and that building excellent ML-powered tools would be a way to “give scientists superpowers” and dramatically increase researcher productivity and the quality of their science. I still think this is true, and (as discussed above) I doubt that we’re going to get rid of human-in-the-loop science anytime soon.

But sometime over the last few months, I’ve realized that we’re building tools just as much for “AI scientists” as we are for human scientists.

Robots as fashion advisors

Aaron has an interesting piece on using AI for fashion advice. He uses a mixture of models for help with choosing a look, finding clothes that fit the look, and assessing fit and color. Fashion seems like a great use of our little robot friends: they’re great at brainstorming, if you don’t mind the occasional hilarious mistake.

Crystal ball department

It’s definitely a bubble, unless it isn’t

Worrying about a possible AI stock market bubble is all the rage right now. Timothy B. Lee and Derek Thompson just published the best piece I’ve seen on the topic, taking a very balanced look at the best arguments for and against a bubble.

Taking jaggedness seriously

AI capabilities are famously “jagged”: the robots are great at some tasks and terrible at others, often in ways that seem bizarre from a human perspective. Helen Toner has some characteristically insightful thoughts on the matter, arguing that contra popular wisdom, the capability frontier may remain jagged even as we move toward superintelligence. Also, she has cool visualizations of fluid dynamics.

Are we dead yet?

Emergent misalignment from reward hacking

There have been several interesting papers recently showing what appears to be emergent misalignment, where models become broadly misaligned from relatively narrow training. Here’s a new paper from Anthropic showing that training a model to reward hack caused it to become broadly misaligned on a wide range of evaluations.

Interestingly, they found that explicitly telling the model that it was OK to reward hack was highly effective at preventing the emergence of misalignment. That superficially strange result is consistent with the theory that models are very good at generalizing: if they’re encouraged to be “bad” in one way, they seem to conclude that they should be “bad” across the board.

Are LLMs worth it?

Nicholas Carlini provides a good overview of some of the potential downsides of AI, covering both concrete short-term harms like job displacement and speculative long-term harms like human extinction. For context, he recently joined Anthropic and has written thoughtfully about the pros and cons of working at a frontier lab.

This snippet struck me as particularly interesting:

Previously, when malware developers wanted to go and monetize their exploits, they would do exactly one thing: encrypt every file on a person's computer and request a ransome to decrypt the files. In the future I think this will change.

LLMs allow attackers to instead process every file on the victim's computer, and tailor a blackmail letter specifically towards that person. One person may be having an affair on their spouse. Another may have lied on their resume. A third may have cheated on an exam at school. It is unlikely that any one person has done any of these specific things, but it is very likely that there exists something that is blackmailable for every person. Malware + LLMs, given access to a person's computer, can find that and monetize it.

Unfortunately, this isn't even a speculative risk at this point. Recent malware has begun to do exactly this. And I suspect it will only get worse from here.

Control Inversion

Anthony Aguirre at the Future of Life Institute has a long paper arguing that superintelligent AI will be essentially impossible for humans to control, even if we manage to solve the alignment problem. I find the analogy of the slow-mo CEO especially thought-provoking:

Consider yourself as a CEO who becomes afflicted by an unusual disability, so that you can only operate at 1/50th the speed of everyone else in your corporation.

…

By day 6 on your clock, everyone recognizes that you are the central obstacle to efficiency and success. While the Board remains loyal, your diligent (albeit increasingly resentful) staff has many avenues available. They’ve already induced you to delegate most decisions, and sneaked a number of policy changes through long documents crafted by clever lawyers (rather than waiting until their obvious merits can be explained to you).

Strategy

MIRI’s proposal for a pause on AI development

Following the release of If Anyone Builds It, Everyone Dies, the Machine Intelligence Research Institute has come out with a detailed proposal for what an international pause on AI might look like. Pausing AI development would be much more complicated than most people realize—it’s great to see an attempt to grapple with some of that complexity.

Philosophy department

Claude deserves respect

Until recently, most people have considered AI welfare to be an abstract future problem, if they’ve thought about it at all. That’s beginning to change, and Anthropic is as always far ahead of everyone else. This addition to the Claude system prompt struck me as particularly interesting:

If the person is unnecessarily rude, mean, or insulting to Claude, Claude doesn't need to apologize and can insist on kindness and dignity from the person it’s talking with. Even if someone is frustrated or unhappy, Claude is deserving of respectful engagement.

I’ll be sad to lose the puzzles

How do we find purpose in a world where the robots are better than us at everything? I honestly don’t have a clue, though I am optimistic that the robots can help us figure that out (assuming they don’t slaughter us instead). Ruby has some interesting thoughts about the tradeoff between wanting to save the important problems for humans to solve, but appreciating the immense costs of delay.