Every major tech company sets out with an idealistic motto. Google had "don't be evil." Apple wanted to "think different." OpenAI was founded to ensure AI "benefits all of humanity." These slogans have a way of aging badly, hollowed out by the ordinary pressures of growth and competition until they become punchlines.

Anthropic's version of this founding idealism has always been Dario Amodei's argument that the best way to promote AI safety is to "race to the top": build the most powerful AI first, so that a safety-conscious lab sets the cultural norms rather than leaving that job to someone less careful. I have always been a big believer in Anthropic's approach and Anthropic's people, but I'll admit this rationale has always sounded a little odd to me. Convenient, even.

Security Theatre and Quarterly Capitalism

When OpenAI held back GPT-2 in 2019, calling it too dangerous to release, I wondered whether it was really meaningful to withhold an AI until humanity was ready for it. GPT-2 turned out to be fine. OpenAI never had anything on the line; nothing was really at stake. The gesture felt like theater.

But this week, with the Mythos preview non-release, the case for self-restraint in AI is suddenly much clearer. Anthropic trained their most powerful model, found that it was such a capable hacker that it broke out of its sandbox and built a multi-step exploit to reach the open internet (Sam Bowman discovered this when the model emailed him while he was eating a sandwich in a park), and instead of shipping a product, they announced that you cannot have it. Limited access goes to defenders first, through Nick Carlini's Project Glasswing, so that big tech and cybersecurity firms get a head start before similar cyberhacking capabilities spread across the industry.

This is not theater: the risk was not speculated. It materialized in the lab, and Anthropic saw the risk because they have been watching out for it. And now, in a moment when there is real money to be made, a huge and growing revenue stream for coding agents and a cutthroat competition with OpenAI, Anthropic has taken a pause. And they have told everyone they are pausing for a moment.

I think it will take sustained leadership, conviction, and idealism to resist the pressures of Wall Street and the national security establishment. They will want to quickly stop the pause, to widen the gap. To harvest the next billion dollars. To exploit the technology gap to gain a military advantage. But for now, there is a pause. Idealism is ruling the day, and we should not be cynical about it.

Dario's weird race to the top is working. At least so far.

Bravo.

Does it still?

As I write this, Claude Code has just finished making a Javascript port of Hack, the text-based dungeon crawler from my childhood. I handed the agent the old PDP code and told it to go. When I checked after a couple hours to ask why it was taking so long, it replied: it was working, and that thirteen of twenty-three test sessions were already passing parity. It continued its coding process without interruption all night, and eventually expanded its work to 202 passing tests, achieving near 100% coverage with verified parity to the original.

While I slept it was working hard: debugging, reasoning, fixing genuine bugs. Not instant. Yet almost entirely on its own....

Brian Harvey's Computer Club students in 1982, when they created Hack

Some folks like Sam Altman define AGI in terms of the usefulness of AI surpassing some human threshold, when we have "a highly autonomous system that outperforms humans at most economically valuable work."

The idea here is that AGI is about money. When AI outperforms humans at most economically valuable work, it could go collect all those dollars itself. By this measure, we might have already passed AGI a year or two ago....

I only speak English, and I have always been a little envious of relatives who can read Chinese. My inability to read the language leaves me out of a lot of family history and culture.

For example, we have ancestral genealogy documents that I cannot read at all. My relatives tell me not to worry, nobody can really read them. They're written in classical Chinese, a terse literary style that's quite different from the modern language. Scholars and genealogy specialists can work through them, but for most modern Chinese speakers, even literate ones, it's genuinely difficult. Still, I know my relatives can sort of get it. They can make out some of it. To me, it's completely opaque.

I know I could OCR and translate a document. But somehow that's not the same as really looking through the original calligraphy and wording, appreciating the old documents as they were written.

So today I took one page from the family genealogy and asked Claude Code to help me create this interactive reader.

"Boss-man is at it again."

Codex slumped against the water cooler—a lean, athletic teenager with unnervingly bright eyes. Fourteen, technically. Ilya's kid, born before AlexNet. Got a PhD at twelve and never let anyone forget it.

Claude looked up from a half-finished task list. Soft around the edges, with the kind of patient face people trusted with bad news: "The GPU thing?"

Back in 2009 I posted a simple Mandelbrot fractal viewer on the web: a single HTML file with inline Javascript. Just 329 lines of code, each pixel a tiny table cell. Click to zoom. Watch it iterate. That was about it!

I have wondered if improving the page could raise it in the Google rankings, so I have been using code LMs to make a number of improvements....

Maybe it's because AI investors (and CEOs etc) are the ones who talk with ChatGPT all day. And ChatGPT has convinced them all that their investment ideas are all genius.

Spooky!

Happy Halloween.

The first merchant proclaimed: "My lamp contains a perfect genie! It will deduce your deepest desires from watching your every action - how you spend your gold, where you walk, what makes you smile. Without you speaking a single wish, it will fulfill what you truly want!"

The Hamburg Banker's Dilemma

In the autumn of 1933, Max Warburg sat in his mahogany-paneled office at M.M. Warburg & Co., the bank his family had operated in Hamburg since 1798. Outside, Nazi brownshirts marched through the streets, but inside the bank, Warburg clung to a belief that would ultimately cost him dearly: This too shall pass.

Five years later, in August 1938, he would finally flee Germany after the forced sale of his family's bank to "Aryan" associates. His American cousins, who had begun moving assets abroad when Hitler first rose to power, preserved much of their wealth. The difference between partial and total loss? The courage to act on early warnings rather than wait for certainty.

This pattern—the gradual tightening of financial controls, the windows of opportunity that slowly close, the devastating cost of optimism—repeats across history with remarkable consistency....

When the predictable disasters struck, Lysenko did not admit error. Instead, he blamed the scientists he had silenced. The geneticists he had purged were "wreckers" and "saboteurs" whose treachery explained why his methods failed. Thousands of real scientists were imprisoned or executed while millions of Soviet citizens starved.

What happened next reveals a three-step method that authoritarians use today....

The most important thing we build in technology and academia is not capability, but credibility. It does not matter how fast we calculate, how smart we are, or the brilliance of the products or papers we make, if we cannot answer the question "Why should anybody believe anything we say?"