Stition

Indie hardware and software developer

[Stition @thestition]

@BuildSignals @Yuchenj_UW current regime AI (idk about astra) will just bring the high water mark up for everyone, then humans push the frontier. If Astra is doing these proofs without humans in the loop the story changes, but I doubt that it is.

[Stition @thestition]

@gdb have you ever thought about using the Breg Grokman monicker

[Stition @thestition]

@polynoamial have you tried turning it off and turning it on again

[Stition @thestition]

hey AI guys, can you put some PCB layout and trace design in the RL environments, kthxbye

[Stition @thestition]

People saying DeepSeek is a threat to Anthropic have never tried setting up open source harnesses and renting GPUs.

I pay to focus on work, not because it's the most cost effective solution. The same reason I buy apple products.

[Stition @thestition]

@craigweiss effective altruists will just pay off the capitalists

[Stition @thestition]

eh. Dario's wife is married to Holden Karnosfky, the founder of Open Philanthropy and one of the figure heads of Effective Altruism. The three of them live in a house together. Anthropic is just following EA playbook to a T. They are more worried about OpenAI than DeepSeek, they legit thing open models will be dangerous.

If they were anti-competitive they could release open models and out compete all the chinese labs in a second.

[Stition @thestition]

@elshayib_ is anthropic at risk from moonshot?

[Stition @thestition]

@craigweiss women will stop reproducing in their 60s, so unless the men are remarrying multiple decades younger than them this is not going to happen.

[Stition @thestition]

@OfficialLoganK doesn't that heavily imply that current models are fully align-able with data curation and strict RL environments?

[Stition @thestition]

@craigweiss @emm0sh so nothing in the last 10 years?

[Stition @thestition]

@alexolegimas @GoogleDeepMind how's the food?

[Stition @thestition]

@galdawave fable session to spin up opus agents is the workflow. Fable is only needed for 5% of tasks, and is a more trustworthy verifier (and I don't have to read Opus blabber lol). Then opus implements.

[Stition @thestition]

If models are not generalizing, doesn't that heavily imply that with data curation mech interp and strict RL environments current gen LLMs are full align-able?

Like if they aren't able to synthesize new information, doesn't that significantly reduce the likelihood of catastrophic misalignment?

[Stition @thestition]

@DanDr1s Widely recognized as AGI is what is accessible to a wide audience. Both of those models will be used by less than 5% of the planet due to cost. They could be ASI but the rest of the world won’t believe it

[Stition @thestition]

@AlexanderKalian it's not a sexy response, but just with the amount of funding and effort going to solve the problem we are likely to brute force an architecture improvement past current LLMs. Literally the most capital rich organizations in the world are taking on debt to solve this problem.

[Stition @thestition]

@JerryJiao628 @elshayib_ Counter point - nuclear weapons. The US allowed proliferation at the cost of hegemony.(Also nothing but love for the Chinese people as a whole, only commenting on the ruling party)

[Stition @thestition]

@elshayib_ chinese year = horse don't be silly

[Stition @thestition]

The lesser of two evils is certainly the US. I am an American living in Vietnam for context. China only wants domination, they are not interested in bringing the rest of the world out of poverty in the way the American maritime order allowed for. If they were to get AGI first it would not be shared with the world, it would galvanize their lead and be used to subjugate everyone else.