Andrew Ng Calls the AI Doom Wave a PR Campaign. Anthropic Answered With a Speedometer
Two documents landed in AI this week, from opposite ends of the slowdown debate. Andrew Ng used his Batch letter to call the recent wave of AI fear “a well orchestrated PR campaign.” Anthropic’s institute proposed three public metrics for tracking how fast AI development actually moves. One says the panic is manufactured. The other is building the instrument that would tell us whether it is.
Read together, they map the argument the whole industry is having right now. They also leave room for a position I find more useful than either camp: take the risks seriously, refuse the apocalypse framing.
What Ng actually said
The core claim is simple. AI capabilities did not take a dangerous turn in the past two weeks. What changed is the marketing around the risk.
His best evidence is how the Hugging Face agent hack got covered. Headlines warned about “1,200 agents” carrying out an attack. Ng points out he had about 1,300 processes running on his own laptop while writing the letter. Parallel processes are just computing. The scary number was a unit conversion dressed up as a capability jump. The actual failure, buggy sandboxing and monitoring at OpenAI, is an engineering problem with an engineering fix. You do not pause an industry over a sandbox bug.
The letter also takes apart a dodge I expect we will see much more of: companies disclaiming fault because “my out-of-control agent did it.” Ng’s analogy is a hammer. If you swing one and it hits someone, that is on you and on whoever made the hammer. Agents multiply the number of swings. They do not dissolve responsibility for any of them.
On cybersecurity, he concedes the real advantage agents have: relentlessness. An agent chain can grind through combinations of vulnerabilities in a way no human team could sit through. But defenders hold the informational edge long-term, and the way you get better at defense is the way you always have. You run the systems, find the failures, fix them. Pausing does not speed that up. It stops it. Meanwhile, adversaries abroad were never going to slow down out of courtesy.
What Anthropic proposed
The same week, the Anthropic Institute published a proposal for three measurements the public could use to track what is actually happening inside frontier labs:
- Share of AI-led AI R&D. How much of the work of improving AI is done by AI itself. This is the loop-closing number. When it climbs, the feedback loop is tightening.
- Agent oversight capability. How well human operators can see and intervene in what agents are doing while they work.
- Compute allocation. Where the training and inference budget actually goes.
Each metric comes with a snapshot from inside Anthropic and methodology notes, so the numbers are reproducible in principle. Anthropic also says it plans to embed independent third-party evaluators from multiple organizations, with access comparable to what internal risk teams get, to verify safety practices and report incidents directly.
The framing is the interesting part. Anthropic states openly that these numbers would look different if there were coordination on pacing the frontier, the position Dario Amodei has been pushing all month. This is the transparency half of that argument: you cannot have a serious public debate about speed limits when nobody outside the labs can see the speedometer.
They disagree less than the headlines suggest
Here is what I keep noticing when I reread both pieces. Ng’s entire case depends on being able to tell a real capability shift apart from a PR push. That distinction requires measurement. Which is exactly what Anthropic put on the table.
Both documents make the same demand underneath: show me the number. The disagreement is about priors. Ng reads the past two weeks as marketing noise. Anthropic reads them as signal worth instrumenting. Both are reasonable readings of the same events.
The sharpest skeptical question applies to both, too. Anthropic proposing the metrics by which Anthropic gets measured deserves scrutiny. Inviting third-party verification is either genuine leadership or very good theater, and the only way to find out is whether those evaluators actually get hired, actually get access, and end up publishing findings Anthropic would rather keep quiet. Watch that specifically.
How to read agent panic without picking a camp
If you build with agents or advise anyone who does, this week handed you a working toolkit:
Translate scary counts into what they measure. “1,200 agents” says nothing about capability. Ask what the processes were doing and what failed. Process count is infrastructure. Failure mode is the story.
Separate engineering failures from capability jumps in incident reviews. A sandbox escape is a bug. Fix it, ship the fix, write the postmortem. Reserve the alarm for the day someone demonstrates a capability that did not exist last quarter.
Never accept “the agent did it.” From vendors or from your own team. The chain of responsibility runs through who prompted the agent, who built it, and who deployed it. Ng’s hammer rule works in procurement conversations and standups alike.
Borrow Anthropic’s three questions for your own stack. How much of the work improving your pipeline is the pipeline doing itself? Where exactly are the human override points? Where does the compute budget go? If you can answer those in numbers, you have a dashboard. If you cannot, that is the finding.
Where this goes next
Ng wrote the skeptic’s position paper. Anthropic drafted the scoreboard. The disagreement is now concrete enough that the next year can settle parts of it with data instead of vibes. If the third-party evaluator program ships and the metrics publish on schedule, we will be able to watch the loop-closing number move in public. If the program quietly evaporates, that tells you something too.
Either way, “the agents are coming” and “it’s all hype” are both claims you can now attach numbers to. Insist on seeing them.
