Mapping out a strategy for an ideal world is easy: we all agree to pause frontier development until alignment is solved. Once that’s done, we glide gracefully into a future of unimaginable abundance. No further questions.
In the real world, it’s more complicated. Any plan that relies on global coordination is—to say the least—ambitious. Do we go all in on the ambitious plan because it’s the only one that could work, or do we also invest in a backup plan?
Scott Alexander’s Against SURA argues that many parts of the AI safety community are making a strategic mistake by focusing exclusively on a unilateral pause, at the expense of other important strategies. This has made some people very angry.
There’s a lot to unpack here—about strategy, about Scott’s post, and about the response to it. I’m working on a longer post about it, but here’s a brief summary of Scott’s position:
Most AI safety plans fall into one of two broad categories:
Plan A is a global pause on AI development until we are able to proceed safely. Almost everyone in AI safety, including Scott, prefers Plan A.
Plan C is a fallback if Plan A fails. It calls for the frontier labs to coordinate to burn as much lead time as possible working on alignment.
Scott argues we should work toward Plan A, while acknowledging that it’s likely to fail, and we therefore need to do what we can to increase the chance of Plan C succeeding.
Much of the AI safety community has drifted toward the belief that a unilateral pause is the only ethical course of action—a mindset he calls SURA (Stop Unilaterally, Raise Awareness). SURA undermines Plan C (which is our main hope of survival) and ironically reduces the likelihood of Plan A succeeding.
This is a strong model: rumors of Google DeepMind’s demise were clearly premature. It isn’t, however, a frontier model: Anthropic and OpenAI remain in a class of their own.
Artificial Analysis have upgraded their Intelligence Index to v5. Previous versions were strangely disconnected from actual model capabilities, but v5 is much better.
Perhaps not coincidentally, this version relies mostly on private test sets and is therefore harder to benchmaxx. FrontierCode, one of the new additions to the index, is a vivid example:
You should read this if you’re a coder, and probably even if you’re not. Not because you should copy what Spawn does: they’re obviously completely mad over there, and copying them will end in disaster.
But the way they think about agents is insightful, and if you apply the appropriate amount of skepticism, you can see hints of what the world will look like a year from now.
Move fast, break things fast, and fix them fast:
But if every change has to be perfect before it goes in, everyone ends up waiting in a single file line again. So we let changes in that aren't quite right, and require that once an hour everything is in line. Before that it can be a bit disorderly. With code you can spot the mess as it happens and send a janitor right away, which is what we do.
Make scale an asset, not a liability:
Savi [Spawn’s game-building AI] files bugs herself. When a user hits one, or Savi hits one in the product, she files it. It goes straight to an agent that decides whether it's a real bug to fix, something we won't fix, or intentional design that needs a decision. Usually it becomes a change, delivered to that user within an hour, with a message in the same chat. That could be an engine bug, a problem talking to Savi, something broken on their computer, or a UI problem. So fixing scales with the number of users we have.
Only one model has what it takes:
This way of using agents only works with even-tempered models that have strong long-term coherence and good judgment. We've tried frontier models from Anthropic, OpenAI, Google, xAI and open source, and nothing outside of Fable has worked. Trust me, I'd love to use a cheaper model, but Anthropic's are the only ones that don't cause literal disasters. […]
In general, we've seen that OpenAI's models game their own checks, Google's models don't stay coherent for very long, and xAI's models just aren't smart enough yet, though they don't do anything cartoonishly bad. Anthropic's models stay coherent, make good decisions, and can do the work. IMO the thing that holds them back most is that they won't take enough risk.
Newer models require less prompting than most of us still do.
For Opus 5.5, try giving Claude a single prompt with a clear goal of what you want to do and how Claude should verify it has achieved it, and let Claude work on it.
Can confirm. Compared to even a few months ago, it’s astonishing how much current models can do in one pass.
As always, ask for more than you think the model is capable of. You’ll be surprised how often it works.
Sam’s flaws have been extensively discussed in these pages and elsewhere, but he is nonetheless alarmingly smart and commendably concise. The conversation covers a huge amount of ground in less than an hour: cybersecurity, the compute crunch, the future of personal agents, and much more.
I wanted to have Séb on the podcast not only because he is one of the most insightful people thinking about AI, but also because he is an unusual voice in the broader AI debate.
On the one hand, he is very “AGI-pilled”, in the sense that he takes seriously the prospect of extremely transformative AI emerging over the coming years and decades. On the other hand, he rejects many popular narratives about the shape and nature of that AI revolution.
Agreed—the AI safety community can be an echo chamber, and people like Séb bring welcome diversity to the conversation.
I particularly appreciate his thoughts on metaphors like “AI is grown, not built”. Metaphors are great because they make it easy to incorporate a package of intuitions about the world into a conversation. And they’re dangerous for that exact reason: an invalid metaphor can lead you confidently down the wrong road.
Curt Jaimungal and Scott Aaronson discuss how mathpocalypse is changing math and physics. This is the best take I’ve seen so far: Scott is clear-eyed about the complexities of AI’s impact, and also about the reality that for better or worse, there is no way to put the genie back in the bottle.
Also: rebutting dumb theories of consciousness is generally a futile exercise, but I confess that I enjoyed Scott’s comments on Integrated Information Theory in the final minutes of the interview.
SemiAnalysis brings us GDPx, a new benchmark focused on the parts of the economy where they expect to see the first trillion dollars of AI revenue.
GDPx is superficially similar to OpenAI’s GDPval, but they argue it does a better job of measuring agent capabilities: unlike GDPval, it requires the model to navigate a complex, interactive environment to accomplish each task.
In some domains, guardrails significantly affected performance. Opus 5.5 and GPT-6 Astra were particularly impacted, so the gap between the frontier models and everyone else is probably larger than it seems.
It’ll take at least nine months to saturate this one
Many of us expect that at least one lab will soon achieve full RSI, triggering an intelligence explosion that culminates in superintelligence. But even without superintelligence, Damon Binder argues that a large enough number of human-level AIs will be profoundly destabilizing:
Let’s be conservative and suppose that, once we have human-level AI, we can only run the equivalent of a hundred million full-time human workers. That is more than half the 160 million people employed in the US. If it were a country’s workforce, it would be the seventh largest in the world, after Brazil’s and ahead of Pakistan’s. At today’s US wages for skilled white-collar work, its output would be worth about $8.5 trillion a year, more than the GDP of any country except the US and China.
The update reorganizes its guidance around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. This reflects an aim of our work on value alignment: for models to understand and apply human values in unfamiliar situations, while remaining faithful to human intent.
The new structure is cleaner, but there are no major changes in direction. Zvi reviews the spec in detail, noting that OpenAI and Anthropic continue to pursue different approaches to alignment:
They remain deontological principles and rules of action, in contrast to the Claude Constitution and its emphasis on character and virtue. If anything the contrast widened.
In an ideal world, the US and China sign a comprehensive treaty to pace AI development. We do not, unfortunately, seem to be living in an ideal world—if we don’t get a treaty, how much can the US afford to slow down without yielding the lead to China?
Shifting some resources from capabilities to safety might ultimately cost little, if it helps avoid the increasingly expensive missteps that come from moving too fast.
Pulling ahead of China on R&D doesn’t help if China is able to steal the results of that R&D: slowing down until US labs are more secure might increase our long-term lead.
The US will have a substantial compute advantage for some time, which gives us maneuvering room.
In the likely event that the US has to act unilaterally, we’ll need to understand exactly how much we can slow down while staying ahead. At the risk of sounding like a broken record, policies like export controls are essential here because they give us more compute advantage, which we can spend on safety.
The most obvious reason to have a job is to keep a roof over your head and food on the table. But work is much more than a paycheck—the DeepMind Institute explores why people work:
The money is important, but also…
If (when) AI causes major job losses, some kind of income support will be critical to helping society survive the transition period. But we will ultimately need much more than that.
The patch gap is a well-known weakness of how open source software is patched. There’s a window of time when a security fix is visible in the public code base but hasn’t yet been released to users. In theory, attackers can use that time to reverse-engineer and exploit the vulnerability before the patch is available.
In practice that has not been a critical issue, because few attackers were able to identify and exploit a vulnerability during that window. But the increased velocity of AI-powered cyberattacks is changing the balance.
Epoch reports that they were able to use GLM-5.3 to exploit the patch gap in six sequential Firefox versions, using about $1,000 in compute per version. The fastest exploit took only 12 hours to develop—comfortably within Firefox’s weekly release schedule.
AI-powered cyber isn’t dangerous just because it gives attackers enormous offensive capability—it also gives them enough velocity to get inside defenders’ OODA loops.
Working in a wet lab is harder than it seems: in addition to a textbook understanding of what you’re working on, you also need “tacit knowledge” like the subtleties of how to use a pipette.
Tacit knowledge is a real thing, but it’s sometimes invoked as an almost-mystical barrier to AI providing uplift to bioweapons production. Writing for SecureBio, the authors argue that it isn’t as big a barrier as is commonly supposed.
You can learn a lot in a month or two
While some forms of tacit knowledge require years of lab work to acquire, many do not—basic lab skills can be learned within weeks, even without expert supervision. Unfortunately, that’s enough to be dangerous:
We do not claim that anyone who learns basic lab skills will immediately be well-equipped to develop bioweapons. However, plenty of dangerous pathogens already exist in nature and require only basic lab techniques to culture and propagate. Because the simplest paths to harm require less sophistication than we might imagine, treating lab inexperience as a categorical barrier under the guise of “tacit knowledge” understates risk.
I think we’ve reached a new inflection point in October 2026, and this time it’s for vibe coding computer games.
I concur. Frontier coding agents are strong across the board, but Opus 5.5 in particular has a talent for both gameplay and visuals, in a wide range of styles.
Mapping out a strategy for an ideal world is easy: we all agree to pause frontier development until alignment is solved. Once that’s done, we glide gracefully into a future of unimaginable abundance. No further questions.
In the real world, it’s more complicated. Any plan that relies on global coordination is—to say the least—ambitious. Do we go all in on the ambitious plan because it’s the only one that could work, or do we also invest in a backup plan?
Top pick
Against SURA
Scott Alexander’s Against SURA argues that many parts of the AI safety community are making a strategic mistake by focusing exclusively on a unilateral pause, at the expense of other important strategies. This has made some people very angry.
There’s a lot to unpack here—about strategy, about Scott’s post, and about the response to it. I’m working on a longer post about it, but here’s a brief summary of Scott’s position:
Most AI safety plans fall into one of two broad categories:
Scott argues we should work toward Plan A, while acknowledging that it’s likely to fail, and we therefore need to do what we can to increase the chance of Plan C succeeding.
Much of the AI safety community has drifted toward the belief that a unilateral pause is the only ethical course of action—a mindset he calls SURA (Stop Unilaterally, Raise Awareness). SURA undermines Plan C (which is our main hope of survival) and ironically reduces the likelihood of Plan A succeeding.
Against SURA has over 500 karma on LessWrong, but has been attacked by some prominent figures. Charlie Bullock has reservations but largely agrees with it, while Zvi is highly critical.
New models and features
Introducing Gemini 4 Carbon
Gemini 4 Carbon is here, impressively soon after Argon.
This is a strong model: rumors of Google DeepMind’s demise were clearly premature. It isn’t, however, a frontier model: Anthropic and OpenAI remain in a class of their own.
Artificial Analysis Intelligence Index v5
Artificial Analysis have upgraded their Intelligence Index to v5. Previous versions were strangely disconnected from actual model capabilities, but v5 is much better.
Perhaps not coincidentally, this version relies mostly on private test sets and is therefore harder to benchmaxx. FrontierCode, one of the new additions to the index, is a vivid example:
Using AI
How I built the most AI-pilled company in the world
You should read this if you’re a coder, and probably even if you’re not. Not because you should copy what Spawn does: they’re obviously completely mad over there, and copying them will end in disaster.
But the way they think about agents is insightful, and if you apply the appropriate amount of skepticism, you can see hints of what the world will look like a year from now.
Move fast, break things fast, and fix them fast:
Make scale an asset, not a liability:
Only one model has what it takes:
How to give your dot responsibilities
Still trying to figure out what to do with your Dot / Duet? Nick Baumann recommends giving it responsibilities, not tasks.
You're probably overprompting Claude
Anthropic’s Addy Osmani:
Can confirm. Compared to even a few months ago, it’s astonishing how much current models can do in one pass.
As always, ask for more than you think the model is capable of. You’ll be surprised how often it works.
Foothills of the singularity
Sam Altman on Cheeky Pint
Sam Altman and John Collison chat about AI over a few pints.
Sam’s flaws have been extensively discussed in these pages and elsewhere, but he is nonetheless alarmingly smart and commendably concise. The conversation covers a huge amount of ground in less than an hour: cybersecurity, the compute crunch, the future of personal agents, and much more.
If you’re in a hurry, prinz shares some notes on the RSI discussion.
Sand gods, foom, and p(doom) (with Séb Krier)
Conspicuous Cognition talks with Séb Krier:
Agreed—the AI safety community can be an echo chamber, and people like Séb bring welcome diversity to the conversation.
I particularly appreciate his thoughts on metaphors like “AI is grown, not built”. Metaphors are great because they make it easy to incorporate a package of intuitions about the world into a conversation. And they’re dangerous for that exact reason: an invalid metaphor can lead you confidently down the wrong road.
The most significant moment in math history just happened
Curt Jaimungal and Scott Aaronson discuss how mathpocalypse is changing math and physics. This is the best take I’ve seen so far: Scott is clear-eyed about the complexities of AI’s impact, and also about the reality that for better or worse, there is no way to put the genie back in the bottle.
Also: rebutting dumb theories of consciousness is generally a futile exercise, but I confess that I enjoyed Scott’s comments on Integrated Information Theory in the final minutes of the interview.
SemiAnalysis introduces GDPx
SemiAnalysis brings us GDPx, a new benchmark focused on the parts of the economy where they expect to see the first trillion dollars of AI revenue.
GDPx is superficially similar to OpenAI’s GDPval, but they argue it does a better job of measuring agent capabilities: unlike GDPval, it requires the model to navigate a complex, interactive environment to accomplish each task.
In some domains, guardrails significantly affected performance. Opus 5.5 and GPT-6 Astra were particularly impacted, so the gap between the frontier models and everyone else is probably larger than it seems.
Quantity has a quality of its own
Many of us expect that at least one lab will soon achieve full RSI, triggering an intelligence explosion that culminates in superintelligence. But even without superintelligence, Damon Binder argues that a large enough number of human-level AIs will be profoundly destabilizing:
Alignment and interpretability
Eliezer Yudkowsky & Geoffrey Irving debate alignment
Terrible audio, but the content is worth it.
An updated Model Spec for more capable and beneficial AI
OpenAI has revised their Model Spec:
The new structure is cleaner, but there are no major changes in direction. Zvi reviews the spec in detail, noting that OpenAI and Anthropic continue to pursue different approaches to alignment:
Strategy and politics
Can the US slow AI R&D and keep its lead over China?
In an ideal world, the US and China sign a comprehensive treaty to pace AI development. We do not, unfortunately, seem to be living in an ideal world—if we don’t get a treaty, how much can the US afford to slow down without yielding the lead to China?
Avi Parrack at Forethought argues that we can afford a substantial unilateral slowdown:
In the likely event that the US has to act unilaterally, we’ll need to understand exactly how much we can slow down while staying ahead. At the risk of sounding like a broken record, policies like export controls are essential here because they give us more compute advantage, which we can spend on safety.
What is work for? Human wellbeing and the value of work in the AGI era
The most obvious reason to have a job is to keep a roof over your head and food on the table. But work is much more than a paycheck—the DeepMind Institute explores why people work:
If (when) AI causes major job losses, some kind of income support will be critical to helping society survive the transition period. But we will ultimately need much more than that.
Risks
AI and the patch gap
The patch gap is a well-known weakness of how open source software is patched. There’s a window of time when a security fix is visible in the public code base but hasn’t yet been released to users. In theory, attackers can use that time to reverse-engineer and exploit the vulnerability before the patch is available.
In practice that has not been a critical issue, because few attackers were able to identify and exploit a vulnerability during that window. But the increased velocity of AI-powered cyberattacks is changing the balance.
Epoch reports that they were able to use GLM-5.3 to exploit the patch gap in six sequential Firefox versions, using about $1,000 in compute per version. The fastest exploit took only 12 hours to develop—comfortably within Firefox’s weekly release schedule.
AI-powered cyber isn’t dangerous just because it gives attackers enormous offensive capability—it also gives them enough velocity to get inside defenders’ OODA loops.
The myth of the “tacit knowledge” barrier
Working in a wet lab is harder than it seems: in addition to a textbook understanding of what you’re working on, you also need “tacit knowledge” like the subtleties of how to use a pipette.
Tacit knowledge is a real thing, but it’s sometimes invoked as an almost-mystical barrier to AI providing uplift to bioweapons production. Writing for SecureBio, the authors argue that it isn’t as big a barrier as is commonly supposed.
While some forms of tacit knowledge require years of lab work to acquire, many do not—basic lab skills can be learned within weeks, even without expert supervision. Unfortunately, that’s enough to be dangerous:
Industry news
26 quietly powerful people shaping the future of AI
New York Magazine and Transformer profile 26 key people in AI who are less well-known than household names like Sam Altman.
Great profiles, but I can’t remember the last time I hated the art direction on an article this hard.
Side interests
The October inflection point for vibe game development
Simon Willison:
I concur. Frontier coding agents are strong across the board, but Opus 5.5 in particular has a talent for both gameplay and visuals, in a wide range of styles.