Something Has Gone Terribly Wrong at Anthropic

The core argument here is simple: exponentials don’t get to cheat physics forever. Chip counts, model sizes, GDP, benchmark scores — they can all shoot upward for a while, but sooner or later they run into a constraint and flatten out. The author spends most of the piece making that case with examples from nature, Moore’s law, and the usual Silicon Valley faith in endless upward lines.

That faith, the argument goes, survives by stacking one S-curve on top of another. Transistors improve, then chip performance improves, then model scale, then data quality, then inference-time compute. Zoomed out, that can look like one long straight climb. But the essay says that only postpones the problem. A stack of sigmoids is still made of sigmoids, and physics eventually shows up to spoil the party.

The punchline is Anthropic. The author points to its revenue growth and puts a trendline over it, treating that as a warning sign rather than a victory lap. The chart is used to argue that the company may be running into the same basic problem that catches every runaway curve: growth that once looked inevitable starts looking suspiciously finite.

The piece doesn’t stop at the company as a company. It uses Anthropic as a proxy for a bigger story about AI’s favorite assumption, which is that progress can keep compounding cleanly if you just keep adding more of everything. The author’s answer is blunt: sometimes yes, for a while. But “for a while” is doing a lot of work there. And when the underlying curve bends, all the confidence in the room won’t straighten it back out.