Eliezer

It's the reinforcement learning rather than supervised learning.

> even with Fable, I've been at the point personally where I am reasonably confident that there's essentially nothing that I am better than Fable at despite generally being substantively above average on human benchmarks

If I asked you to write fiction, you'd be much better at keeping track of which characters knew which facts.

Why on Earth would anyone take anything this man said to be informative about anything, other than what he thinks will currently sound best that minute to the current Zeitgeist that day?

You can turn that off in settings.

The safety teams are trivial expenses for them. They fire the safety team because explicit failure makes them look bad, or because the safety team doesn't go along with a party line and gets labeled disloyal.

Came here to say the same.

There's intrinsic limits to vanilla transformer stacks. Nobody knows where they are. We don't know how unvanilla Opus 4.6 or GPT 5.3 are. We don't know what's in development or which new ideas will pan out. But it will still probably be called an "LLM".

Every time somebody writes an article like this without any dates and without saying which model they used, my guess is that they've simply failed to internalize the idea that "AI" is a moving target; nor understood that they saw a capability level from a fleeting moment of time, rather than an Eternal Verity about the Forever Limits of AI.

"2024 AI was a horse". People really like to imagine that the last 6 months constitute their true observation of the new eternal state of the future.

It would not surprise me at all for bb7 to exceed Graham's number. Just a Kirby-Paris hydra or a Goodstein sequence gets you to epsilon zero in the fast-growing hierarchy, where Graham is around omega+2.

Thanks to OpenAI for voluntarily sharing these important and valuable statistics. I think these ought to be mandatory government statistics, but until they are or it becomes an industry standard, I will not criticize the first company to helpfully share them, on the basis of what they shared. Incentives.

Privacy.com has always been a working general solution to this, for me. Disposable CC aliases with spending caps.

My god, how long has it been since you tried to use an AI model?

Doubling the productivity of 20% of workers, in cases where a lower price doesn't increase demand, can shift prices in the whole system as unemployed artists compete with other artists for wages. AI won't take your job, someone else unemployed by AI will take your job. (NGDPLT partially solves this but that's a higher competence level than civilization has.)