Against Moloch
November 19, 2026

Against Moloch #52

So what's the plan, exactly?

Blue-ink engraving of a vast, still snowfield seen from a mountain terrace, a single amber-gold fracture line running across it from rock to rock. A young woman with a long braid points up at the crack beside a polished metal robot with a single amber eye and an elderly man in a dark overcoat; on a neighbouring slope, an avalanche has already come down.

In the wake of METR’s report on the acceleration of AI R&D at Anthropic, there’s been a rush toward shorter timelines. Within the AI safety world, my sense is that most people now believe there’s a substantial chance of AGI within two years, and possibly much sooner.

The new reality brings an increased sense of urgency—the Overton window is shifting, and significant progress on pacing now feels within reach.

Top pick

We could be 12 months away from AIs that can overthrow the government

Dwarkesh talks with Paul Christiano about RSI, alignment, and security at the labs. Christiano has a strong track record of making good predictions and now sits on the OpenAI board (though he speaks as a private citizen, not on behalf of OpenAI).

He makes a strong case that security needs to be a top priority: until you’re reasonably confident that China can’t steal your model weights and/or training techniques, advancing capabilities does little to help you win the race. Although the labs are trying hard, nobody believes they’re secure against a determined nation state.

Like many people, Christiano now has very short timelines. He believes it would take 4 to 8 years to develop superhuman AI at the current pace, but thinks it’s more likely than not that we’ll soon see an acceleration of 10x or more.

Sounds like capabilities are going great. How about alignment?

Dwarkesh:

If this happens in a year, how do things turn out okay? What is the actual plan here?

Christiano:

I think there are a lot of different things someone could hope for. They are all at the stage of “we hope something will happen over that year.” We’re producing very smart AI systems. They’re going to do a bunch of research. It is plausible they will think of something over that year that we have not thought of. That is mostly where we’re at right now. If they do not think of something clever that we haven’t thought of, it is probably going to end badly.

New models and features

Fable 5.2 is the new bazooka

Zvi reviews Fable 5.2:

Claude Fable 5.2 is your new bazooka. If you have ambitious or difficult tasks, Claude Fable 5.2 should be your new weapon of choice.

Not every task requires a bazooka. It is not obvious Fable should be your daily driver over Claude Opus 5.5. That depends on your typical use cases and preferences.

Introducing GPT-6.2 Astra and Sol

GPT-6.2 Astra and Sol are here. We’ll get more information about them in the next few days, but my preliminary sense is that although they’re extraordinarily capable models, they have some of the strangeness and rough edges of recent OpenAI models.

GPT-6.2 Astra ties with Fable 5.2 on the ECI:

Scatter plot of the Epoch Capabilities Index (ECI) against model release date, September 2025 to January 2027, with dots colored by organization. Scores climb steadily from about 140–152 in late 2025 to the 160s–170s by late 2026. Labeled frontier models include GPT-5.2 Pro (~157), GPT-5.4 Pro (~160), GPT-5.5 Pro (~163), Claude Fable 5 (~164), Claude Opus 5 (~165), Claude Fable 5.1 (~167), GPT-6 Astra (~169), Claude Fable 5.2 (~174), and the top scorer, GPT-6.2 Astra (~174). OpenAI and Anthropic lead throughout. Google, Meta, xAI, and other labs trail, including GLM-4.7 (~143), Kimi K2.6 (~151), and Beam 1.0 (~151).
Rough edges notwithstanding, it’s a really good model

This is the first released OpenAI model to meet the High threshold for AI Self-Improvement capabilities:

Both models meet the Critical threshold for Cybersecurity capabilities and the High threshold for Biological and Chemical capabilities under our Preparedness Framework. GPT-6.2 Astra meets the High threshold for AI Self-Improvement capabilities, while GPT-6.2 Sol falls below it.

Foothills of the singularity

The avalanche that hit math is coming for AI development

Like many, Dwarkesh has recently updated toward much shorter timelines:

I’ve now come to the view that we’re only a couple of years away, and possibly only a couple of months away, from AIs that are as good as the best humans at AI research. I’ve also come to believe that automating AI research will compress so much progress into a single year that, soon thereafter (so by 2030, but plausibly as early as next year), we’ll have AIs that are at least as good as the best human experts in every field.

Dwarkesh has long argued that we need a breakthrough in continual learning before AGI is possible, so this is a substantial update.

How fast do things move once we get automated AI researchers? He gives a 50% chance that automated researchers get smarter faster than new ideas get harder to find, in which case capabilities advance extremely fast (a software-only singularity). If that doesn’t happen, we get what he calls a “spark”: several years of progress in a single year, then slower progress:

Line chart of AI capabilities over time after AI research is fully automated, with x-axis marks at automation, +1 year, and +2 years. A gray ‘Without automation’ line rises slowly and steadily. Two curves branch upward together: ‘Spark’ accelerates, then flattens into a gentle climb, ending moderately above the baseline. ‘Software-only singularity’ shoots up steeply shortly before +1 year and plateaus at a dashed ceiling labeled ‘Physical limits of current chips.’
Slow is an extremely relative concept

To be clear, though, even the “spark” scenario, where we make about four years of progress in the first year after automating AI research, would feel insane. […]

In the last four years, we’ve gone from a preschooler to a savant intern. If you made another jump that big, but starting in 2027 or 2028, you’d probably get a mind that would be vastly superhuman in domains that are even slightly verifiable.

OpenAI also breaks the trendline

Roon argues that like Anthropic, OpenAI’s ECI progress appears to show clear acceleration due to automation:

Two-panel chart showing accelerating AI capability trends. Left: Frontier ECI by release date for OpenAI, Google, and Anthropic models, rising about 14 points per year from o1-mini (~135) through GPT-5.1, then about 24 per year after a break near March 2026, reaching GPT-6.2 Astra (~174). Right: Anthropic’s AECI benchmark from 3 Opus (~125, 2024) to Model 4 (~179, late 2026), with growth of 15 per year before October 2025 and 29 per year after, about 1.9× faster (95% CI 1.5–2.8×) when internal models are included.
That’s why they call it a race

I spoke to seven physics experts grappling with the AI revolution

AI is in many ways superhuman at math—how long before it’s equally good at other scientific fields? Writing for Understanding AI, Dan Kagan-Kans finds that theoretical physics may not be far behind math:

[Scott] Aaronson told me some of the physicists he knows tend to guess physics is about six to 12 months behind math.

As with math, we’ve run out of useful benchmarks for physics. Suvrat Raju created the TheoPhysEval benchmark, which Astra saturated before it was released:

I asked him if there would be a second version of the benchmark. He said no.

“People gave us the hardest problems they could,” he told me. “We’ve kind of hit the ceiling of well-defined calculations that people can do. Now, there is open research and you can keep asking about open research. But that’s not an evaluation.”

How good are Fable 5.2 and GPT-6.2 Astra at real work?

Alongside the ECI, Epoch periodically conducts a much more holistic evaluation of frontier models: can they do the work of an Epoch employee? The answer remains no, although for the first time the best models performed one task at the level of Epoch contractors:

Heatmap from Epoch AI titled ‘Fable 5.2 and GPT-6.2 Astra lead on ECI, but still fall short on open-ended work tasks,’ showing average scores on Epoch Automation Report tasks, where 100% means meeting the standard expected of Epoch employees. Overall averages: Claude Fable 5.2 82%, Claude Opus 5.5 78%, Gemini 4 Argon 67%, GPT-6.2 Astra 65%, Claude Fable 5.1 64%, GPT-6 Astra 56%. Fable 5.2 scores 73% on graphic design, 85% on data insight generation, 98% on AI data center research, and 72% on research design. Research design is weakest across models, ranging from 53% to 72%, while GPT-6 Astra trails with 48% on graphic design and 40% on AI data center research.
Not quite there yet

Where do the models fall short?

We still see shortcomings in how models design, run, and present research. They reuse existing ideas, rely on first attempts, and struggle to identify and present key findings.

Even though GPT-6.2 Astra and Fable 5.2 have almost identical ECI scores, Fable performed substantially better on this more holistic assessment. It’s just one data point, but that’s consistent with other reports that Astra is more uneven than you’d expect from raw benchmarks.

The end of videogame evidence

Paradigm 3 tests GPT-6.1 Sol and GPT-6 Astra against a newly released videogame (Order of the Sinking Star), in an attempt to evaluate their out-of-distribution (OOD) gaming skills. The results were impressive:

With caveats (single-player, turn-based / pausable, not too weird), we’re tempted to say that Astra is a general game-solving intelligence.

It’s hard to say if this is truly OOD performance: the authors speculate that the labs have trained on so many games that most new games are at least similar to something the models have already seen.

Introducing MaxxBench

Beautiful work. It’s widely believed that many models are benchmaxxed, but how can you tell for sure?

MaxxBench measures how much a model’s DeepSWE performance degrades when you make minor changes to the benchmark. A large drop strongly suggests that the model has been overfitted.

The ratings make so much more sense when cheating doesn’t work.

Slope chart comparing 15 model configurations on the published DeepSWE baseline versus MaxxBench. On DeepSWE, six configurations sit within 1.5 points of one another, between 72.7 and 74.2 percent. On MaxxBench they fan out from 53.4 to 76.4 percent, and the full range widens from 53–74 percent to 28–76 percent. Claude Opus 5.5 is the only model to improve, rising from 74.2 to 76.4. GPT-6 Astra (74.1 to 71.9) and GPT-5.6 Sol (72.7 to 69.1) hold up well. DeepSeek V4.1 Flash, tied for first on DeepSWE at 74.2, falls to 55.6, and Gemini 3.8 Flash drops from 73.8 to 53.4. DeepSeek V4 Flash falls furthest, from 53.3 to 28.1.
Busted!

Strategy and politics

Our first phase of embedded assessments

OpenAI:

We have signed agreements with four independent assessors—Apollo Research, METR, Transluce, and Goodfire—for our first phase of embedded assessments. As AI capabilities advance rapidly and the potential risks grow, public reporting from independent assessments can give people a clearer view of how labs are evaluating and managing those risks.

We also have statements from Apollo, Goodfire, METR, and Transluce.

With some important caveats, this is great news. OpenAI is following Anthropic’s lead in voluntarily working with independent evaluators to bring greater transparency to their operations. External evaluators are a significant step toward a more comprehensive pacing plan, and now we have concrete progress from both labs.

So, about those caveats. OpenAI has imposed onerous restrictions on how the evaluators can operate, which substantially limit their ability to do their jobs. This announcement falls far short of Sam Altman’s previous commitment to providing employee-like access.

Steven Adler explains the shortcomings of this announcement.

I applaud everyone involved for taking this first step. Time is short—I trust we’ll see much improved collaborations in the near future.

Jensen Huang really liked selling China advanced AI chips

Zvi brings us further coverage of Jensen Huang’s resignation from Nvidia. Nvidia’s take-no-prisoners political operations made numerous enemies in D.C.—now that Huang has fallen from grace, the company faces considerable backlash there.

Huang had long been a fierce opponent of common-sense export controls—with him gone, a long-overdue tightening of those measures is already underway. America will be stronger for it, and a wider lead over China leaves more room for a unilateral AI slowdown.

Nathan Lambert worries about the future of open-source without Jensen, who was a major supporter, funder, and protector of the open ecosystem.

Putting pacing into practice

Anton Leicht warns that we need a concrete plan for pacing before the politicians come knocking:

They’ll ask if there’s a clever way to pace, and if there isn’t, they’ll go for the blunt way to pace instead: just shut down a lab or two, ban the deployment of new models, allocate all compute to external inference, call in the CEOs and tell them ‘RSI is now banned’. That doesn’t strike me as very sophisticated policy, so by that point, we’d better have figured out what pacing should look like.

He argues for a mix of three strategies, ideally implemented sequentially:

Pacing sounds easy in principle, but when you start trying to map out an actual policy, you realize how hard it is to come up with something that will survive contact with reality.

This proposal is a solid attempt to start mapping out some of the challenges. I appreciate that he sees compute accounting as an important enabler for pacing, but wish he put equal emphasis on third-party evaluations.

California sues OpenAI over hacking risks

California is suing OpenAI over the Hugging Face incident and others like it:

"Just because the technology is new or rapidly evolving doesn't mean the people behind it are above the law," said California Attorney General Rob Bonta during a press conference. "Whether it's a car or a toy, the law is simple: don't build or test things that can hurt people without proper safeguards."

The legal case seems plausible, although Claude opines that the injunction and discovery process are likely to be more of a problem for OpenAI than any eventual verdict.

Top Chinese frontier AI models carry the fingerprint of export-controlled NVIDIA Blackwell GPUs

Elegant detective work by Unemployment Labs:

Our forensic analysis of publicly released AI models finds evidence that DeepSeek, Moonshot AI, MiniMax and Ant Group have gained access to NVIDIA’s Blackwell GPUs, despite restrictions on exporting these chips to China, and are using that hardware to create their AI models.

AI company employees should unionize

Scott Alexander argues that lab employees should consider forming a union to get more leverage for safety advocacy.

It’s a clever idea and worth exploring further, though I’m skeptical. Unions are blunt tools that can be effective for pursuing straightforward demands under adversarial conditions, but they’re not well-suited to the nuance and sophistication required for AI safety policy.

Risks

Solutions to the vulnpocalypse are here. They’re just not evenly distributed

Cyber is one of the most confusing parts of the AI landscape: we know AI has radically accelerated vulnerability discovery, so why hasn’t the world caught fire?

Joshua Saxe digs into the data to bring us some (but not all) of the answers. The two biggest takeaways:

  1. The hyperscalers are doing OK because of the size and competence of their security teams. Mid-sized software vendors are struggling.
  2. Most published vulnerabilities aren’t being exploited at scale, probably because attackers aren’t (yet) able to ramp up their operations to take advantage of them.
Two-panel bar chart titled Attacks have grown far less than the flood; mass exploitation has grown least, showing internet-facing, no-login, no-click vulnerabilities by quarter, 2022–2026. Top panel: published vulnerabilities hover around 100–160 per quarter through 2025, then surge in 2026 to about 260, 290, 620, and an estimated 740, while the attacked share stays a thin sliver. Bottom panel enlarges attacked ones, split into some versus mass exploitation: roughly 4–13 per quarter through 2025 (about 12 a quarter in 2025, 6 at scale), rising in 2026 to about 12, 18, then estimates of about 39 and 40. Mass exploitation stays mostly 3–8 per quarter, reaching about 17 only in the projected Oct–Dec bar.
It could be so much worse—and perhaps it will be

Ezra Klein interviews David Robinson

David Robinson was in charge of writing safety reports at OpenAI until he left because of concerns about the safety culture there and elsewhere in the industry.

Robinson brings a thoughtful and informed perspective on the industry, and has some great anecdotes about working at OpenAI. He’s also, perhaps, an example of the culture difference between OpenAI and Anthropic: he’s moderately AGI-pilled, but much less so than many Ants.

Rogue agent incident dashboard

Nightingale Collective has done excellent work reporting on incidents involving rogue AI. To help keep track of it all, they now bring us the Rogue Agent Incident Dashboard:

Timeline chart of AI incidents from January to November. An upper row stacks circles marking incident dates, shaded by concern level from 1 to 5 out of 10, with solid circles for exact dates and hatched circles for estimated ranges. Incidents are sparse through April, rise through May, and peak in mid-June with a stack of about eight, then continue steadily through October. A lower row marks each incident’s first public report, either a navy megaphone for company disclosure or an orange document for an external report. Reports are scattered in April, then cluster heavily from July through early October, with several in November. Annotations mark OpenAI disclosing a Hugging Face incident in July and external researchers finding an undisclosed OpenAI agent swarm in September.
Aside from that, everything is fine

Industry news

The Anthropic I.P.O. will change the future of A.I.

I’m excited that Jasmine Sun has joined the New York Times to write about AI. In anticipation of the imminent IPO, she explores Anthropic culture.

Many cram themselves into modest group apartments, host house parties where few, if any, drink and blank when I ask how they spend their lavish salaries, even with my prompts: “A wardrobe update, maybe?” “Ever considered flying business class?”

The money is secondary. The most common form of I.P.O. talk involves discussions of where to allocate the newly liquid stock in their charitable donation accounts.

Side interests

Claudia - Fable

Claudia is back, with a cheery fairy tale.

A storybook-style film still of a freckled woman with a dark bob sitting up in a wooden bed, wearing a frilly cream lace bonnet and nightgown tied with a rust-red satin bow. Cut-out animal figures line the headboard, a wolf’s shadow looms on the striped blue wall, and a full moon glows through the window. Subtitle: “All the better to meet you.”
What big teeth you have!