The reason HTML and markdown are more widely adopted than latex is because the former are more forgiving.

This seems...high. The levels.fyi data for 68/69 shows them maybe crossing 1M, and the highest available data point at a bit under 1.5 mil. Staff engineers at FAANG are not making 1M a year except in cases of significant stock growth. 700-800K is more reasonable. The AI labs pay a significant premium, yes (2-3x!).

My math on it was basically "I can work for 10-15 more years and retire meeting my financial goals, or work until my 4 year grant at an AI lab finishes and meet all the same goals, with more interesting work".

Again I'll reiterate: the only concern the government has stated is that of Claude refusing to answer things during an active battle.

I respect that decision, I think if the government wants to use OpenAI for that situation it makes sense. They can and should pick suppliers who meet their needs.

Nothing about what the government has said suggests any reason that Claude Code is a risk. So why are you insisting that it is and that the government is justified in banning it?

I'm struggling here because you're basing this determination entirely on things which the government never acused anthropic of, and haven't asked anthropic to address, and which the underlying law probably doesn't support.

The government made clear that it was only designating anthropic a supply chain risk due to their refusal to allow Claude to be used lawfully for autonomous kill chains and domestic surveillance. Why are you bringing up this other stuff when the government never did?

To which you brought up something that still wasn't relevant.

The government is perfectly able to choose not to work with contractors it doesn't like. That's not this.

Thus, it cannot be a supply chain risk.

The DoD's response was to ban anthropic from all DoD facing activity and supply chains, whether or not there is a human in the loop. This does in fact run into exactly GP's issue. (or as I suggest elsewhere, can Boeing use Claude Code for software development?)

The DoD responded not by saying "alright, we will use OpenAI for our kill-chain uses" but by saying "Boeing is no longer allowed to use Claude Code".

I haven't seen any evidence of this. Much of the anger is coming from the unaffiliated researcher. levent (the anthropic employee) has mostly constrained his comments to basically "I would have been happy to collaborate w/ folks from OAI"

there is a well studied scenario where humans are asked to hire people from four groups. These groups will be judged in their performance on a job and the humans rated on their hiring abilities. Unbeknownst to the human participants, all applicants are drawn from a single skill distribution, with groups assigned essentially randomly. Stastically, all groups have identical performance. Despite this, humans generalize over their early experiences, and develop biases towards specific groups.

While not identical, I relate this to the experience I have playing Fire emblem with random growths. A unit can get lucky and favored early despite being overall mediocre (hello Diamant from my first run through engage).

The researchers recreated this experiment with LLMs, and showed that the LLMs reproduce the human behavior of overgeneralizing early and failing to, as the paper says, sufficiently explore the space[0].

[1]: They instead exploit in the technical sense (https://en.wikipedia.org/wiki/Multi-armed_bandit), but exploit based on incomplete information.

First, it's a really compelling roguelite puzzler where you explore and understand the house and it's mechanics. The puzzles are mostly small and straightforward (but not easy!). You can probably win this without taking notes. I found this part compelling and delightful.

Then it's something else where you learn about the world. The puzzles are more complicated and need notebooks and notes to keep the loose threads tracked. This part for me became a slog by the end and I finished mostly to say that I had.

Then its a third thing for sickos.

Others have already said this, but the watermarking is something like "when the model flips a coin picking between two values, always choose heads". It was already flipping a coin. You're not choosing a less good result, you're just using a deterministic process when it was stochastic before.

This will have some impact on outputs, but unless you have some reason to believe that always picking tails was better than always picking heads (in which case, you should be working at one of these companies in model training!) it won't have any impact on output quality.