CosmicExplorer

The best indicator that we have AGI might unironically be mass layoffs.

@Jirh I think it'll be saturated in 8-9 months.

Astra is still only 65.6% on ClockBench. AGI cancelled.

If it's technically not AGI yet, it's damn close to it. It's very hard to find anything right now that GPT-6 is worse at than an average human.

ARC-AGI-3 has just been solved by GPT-6 Astra. Insane. Humans on this benchmark get ~50%. Astra, even without harness (if I understand it correctly), gets 62.7%. You can not just brute force this benchmark. It very heavily penalizes taking too many steps. I honestly thought it'll take a minimum of 2 years for AI without harness to get better on it than humans are, it took 5 months and a week.

Bernardo Kastrup about AI

It just jumped up to 13.2% Well, because humans are not as widely knowledgeable, but still, to completely replace, let's say, an accountant, AI would need to have at least 95%+ pass rate on something like agents last exam. I think either there will be some bottleneck and it will not happen on a large scale, or when it happens, it will be fast, like if AI crosses some reliability threshold in those hard job tasks, I can't imagine capitalism will not lead to white-collar job losses at a massive scale quickly.

"In psychotherapy, practically the only consistent finding is that whatever kind of psychotherapy the person running the study likes is most effective. Thirty different meta-analyses have confirmed this" https://slatestarcodex.com/2014/04/28/the-control-group-is-out-of-control/ https://jonathanshedler.com/wp-content/uploads/2018/05/Shedler-2018-Where-is-the-evidence-for-evidence-based-therapy.pdf

@Leo Gura Installing a mosquito net on my windows completely solved that problem for me. It was one of my best purchases ever, great cost to usefulness ratio.

That's the main reason why I think current AI will not get to AGI, nor replace white-collar jobs, reliability. Humans are much more reliable than even the current best models. Even if current LLM based AI reaches something that can be called AGI for average scores, the reliability will lag behind. It's extremely hard for this architecture to reach 99% reliability score, even on simpler tasks, and humans are usually more reliable than 99% In the Agents Last Exam benchmark, current AI has a 10.5% "pass rate"/reliability on real-world, economically valuable tasks. Not even close to enough.

Holy shit AI is so stupid sometimes

What about a benchmark like ARC-AGI-3? They give AI a private set of games that AI has not seen before, with no instructions or rules on how to play a game, they have to adapt on the fly to novel interactive environments. It's not just brute-forcing it as they very heavily penalise it in their scoring system. Humans are scoring ~50% on it, current best AI is 7.8%. It still sucks compared to humans but it's not 0. AI is able to solve some levels.

All mental disorders exist on a continuum (shades of gray), and the exact line dividing "normal" from "disordered" is fundamentally a human made construct. In a sense, everybody experiences micro manic and micro depressive episodes because everybody's mood is variable.

If I would have to bet my money on it, I would bet that Connor, during those last 6 years, was more happy than the average person. I don't think it was a suicide. I think it's likely that his death was actually caused by not enough neuroticism rather than too much.

His last(?) interview from 3 weeks ago. Seems completely normal here.