djoldman

> The simplest, most reliable way to ensure the works you invest in are fully and entirely eligible for copyright is to just use humans for the whole project. Declare a "no-AI policy," and hire human artists to produce the works you hope to sell. Monitor their activity to ensure they're not using AI. Keep good records.

That seems true.

But also: generate all you like and then have humans edit, and keep a record of it, seems plausibly defensible too.

The current state of a lot of AI generated content material, not all / not necessarily program code, seems to really benefit from human editing.

Forget the transistor for now: that's many years/decades away.

Much more likely to actually get to market is diamond as the heat dissipating substrate.

It's almost never like that. Usually some team collaborates with some other team and sets something up with approval from departments A, B, and C with director X signing off. There's rarely one person you can point to. Also, wait a year and all those As, Bs, Cs, and Xs have switched to different roles or don't exist.

> Even if the overall productivity boost due to AI "only" turns out to be 40%, the lowest among your numbers, it's equivalent to 17 years of growth at 2% per year.

Two reasons to question the 40% number:

  - Anthropic probably represents the extreme case of AI adoption, if they're not the absolute leader. I would see 40% as the absolute max for that category without knowing anything else.
  - Anthropic is a tech company, willing to change their workflows rapidly, having maximal enthusiasm for the technology, and maximum knowledge of it. The vast majority of the world's companies don't come close to it for these aspects and therefore adoption and speed should lag significantly compared to Anthropic.

> For those hard theorems where we really want LLMs to help us, these Lean proofs are extremely long, like hundreds of thousands of lines of code long. A seemingly emergent trend is that the longer the Lean version of a proof, the messier and less faithful its machine-generated “natural language” translation seems to be.

This seems to be a common and continuing objection to the OpenAI math repo release. Specifically that at least some, perhaps many of

  * the natural language proofs are low quality or otherwise extremely difficult to understand (although possibly correct?)
  * the lean proofs are too long, convoluted, or otherwise extremely difficult to understand

I wonder if LLMs will be able to generate "high quality" natural language and lean proofs soon.

And specifically: iterate on the OpenAI repo to produce "better" proofs that prove the same things.

  Commits per engineer 5.1x
  Queries per data scientist 3.1×
  Contracts per legal employee 3.1x
  Deals per salesperson 1.7x
  Hires per recruiter 1.4x
  Invoices per AP processor 1.4x

These are the stats that might be hiding the possibly small potential for AI to generate "real" economic activity.

The first 3 stats, commits+queries+contracts, are secondary/meta statistics that could possibly go up 1000x and have very little affect on Anthropic's business, whereas that's much less likely to be the case with the last 3.

AI can make it very easy to do more and more busywork that doesn't move the needle.

If the above is true, AI econ must rely on a second order effect here in that busywork/BS is recognized more quickly and discontinued.

For example, if an organization paid an employee to compile a monthly report and it took them a very long time. But with AI, they can make that report every day. Now the organization realizes that they don't even need it.

> Anthropic has expanded rapidly. But it has hired less per dollar of revenue than an earlier generation of large tech companies. At five years old, Anthropic takes in between 2–4 times as much revenue per employee as Meta and Google as they crested in their own rapid growth.

This may not be saying much at all. So much goes into deciding how fast to hire. Instagram had 13 employees / 1B rev at acquisition?

Also revenue per employee can vary a lot depending on the industry.

A better metric to understand AI effects might be revenue per employee at companies that have been around > 5 years that are AI-enthusiastic.

> ... runs each model in its lab's 1st party CLI harness at the highest reasoning setting it offers.

  Anthropic (Claude models): Claude Code
  OpenAI (GPT models): Codex
  Google (Gemini models): Antigravity CLI
  Meta (Muse models): Muse Code

> The CLIs' web search and web fetch tools are turned off, and the agent has no other internet access. The CLI is installed before the clock starts, and the hour counts from when the agent starts. There is no submit tool: the harness collects the bot when the time runs out.

One thing we're learning is that harness is HUGE. The difference in performance for the same underlying model can change dramatically across:

1. harnesses

2. allowed tools (no web search/fetch - how does the model refer to documentation?)

3. thinking levels