EthanHeilman

I've found it to be useful for prototyping. You can throw a game mechanic up, have it build some AI players, tell you any balance issues it runs into, play against the AI yourself to see if the mechanic feels fun. Then iterate.

A few times I've made a tweak to a mechanic and the LLM balance tests immediately found a major issue. Sometimes the AI players it creates just suck at the game. You still have to think, but just having balance tests and then having them cheap to update when I experiment, it extremely valuable. I tested this on a game I had mocked up with cardboard and forced my friends to play. It found all sorts of stuff I missed in those gameplay tests. It didn't replace those sorts of tests with humans, but it let find issues faster so I got more out of human interaction tests.

Agreed, that was what I was trying to say. You can build a lego structure that you can not infer by just looking at from the top, bottom and sides.

Build an object out of legos and then reconstruct the shape of what you built by only looking at it from the sides and above.

That's what 95% is intended to capture. That some level of expertise in an area is should be captured by 95% sample of the population, you could push it to 99% of 99.9%, but you want it to be quantified by a number to avoid arguments that something is not AGI because obscure field or expert exists that AI can not do.

95% better at 95% of the population is already approaching ASI. One could even argue that AGI is 50% better than 50% of the population.

All that said, this over rotation on benchmarking misses something critical. What is general intelligence? We assume that humans have it and we assume it is captured by benchmarks on "intellectual tasks", but it is probably the case that general intelligence is based displayed by judgement on uncertain outcomes. Benchmarks by their very nature have certain outcomes, they have a wrong and right answer.

Test AIs on questions we don't have the answers to and there is no clear right answer, but there will be at some point in the future. What will the economy do? Which US senators will be be re-elected that polling correctly suggests will not be re-elected. Which US senators, currently not in office, will actually pass bills representing the wishes of their voting base? What published papers will be seen are groundbreaking in 5, 10 15 years? What approach to unifying physics should be taken?

Does RSI actually mean anything specific anymore? RSI, AGI, at this point seem like buzzwords. Sure AGI has definition that are measurable, say "better than 95% of humans on 95% of intellectual tasks" but if we used that definition we already have AGI and almost no one thinks we have achieved AGI. We use AIs to train AIs which we use to train AIs, why is that not RSI? How much human intervention means that is not RSI?

Oh nice, thanks for sharing

I believe jrflo was saying that OpenAI watches the chats of everyone from Anthropic because watching what Anthropic employees type into their personal ChatGPT accounts is a critical source of intelligence on is happening inside of Anthropic.

I would be surprised if OpenAI isn't doing that. OpenAI will take any advantage they can get. If an employee at their primary adversary is typing useful intelligence into OpenAIs website, a website that does not promise privacy from OpenAI, the only reason they wouldn't weaponize that information against Anthropic is ethics or fair play.

A careful reading of "we cannot rule out that de-identified data derived from their usage of our products helped improve our models" could be saying that yes they trained on it but they don't know if that training data resulted in an "improvement" to the model. That is, they can't rule out that the only reason the model found this solution was because it had been trained on this approach.

The term ruled out is very open ended and gives them significant flexibility of meaning. They may have the information to determine exactly what happened, but they haven't looked so they can't "rule it out".

I am surprised that this happened to liquid. Many of the other hacks seem to the sort of mistakes that I expect projects to make. RNG-failure resulting in bad keying material is a mistake that people have been making since at least the last world war almost 100 years ago.

> You have accurately described a email mailing list. Where in the value-add here?

Pre-prints are basically a mailinglist where you post your paper prior to peer review.

The value over a simple mailinglist is:

1. Stable URL and citation to enable other work and discussions to cite and reference it.

2. Versioning of the paper, allowing updates to be made without having mail out the paper, while allowing everyone to find all prior versions

3. Host for a PDF and data that might be quite large

4. Centralized searchable long term archive of scientific papers

5. Scalability, arxiv gets 30,000 submissions a day, no one wants to receive 30,000 PDFs in their inbox everyday