In Plan C the government essentially misses the opportunity to implement the multinational deal while the threat of covert projects is still low (fewer latest-gen chips unaccounted by tracking measures, worse models/algorithms to use for RSI). That's why it says the probability of a deal is lower and lower each month rather than outright zero.
This is by far the most realistic optimistic AI takeoff scenario I've seen, and more specifically it's the first one I've read that deals with both the AI alignment and power concentration issues in a sufficient way, even in a world where hard alignment is assumed (in this scenario the AIs are assumed to be misaligned until ~2038-39).
Bravo, and I hope it has the impact on the AI safety field it deserves to have.
I'm working on a multi-harness IDE that supports custom agent workflows and skills that are shared between any harnesses it wraps over. I think it might prove handy for a workflow like yours.
Huh, a good alternative just as anthropic's 50% weekly subscription subsidy is ending this weekend. Time to see if it's benchmaxxed or actually a strong leap over GPT5.5.
They also seem to really not care about alignment, or care about it in the wrong way. It's entirely missing in the blogpost and there are some concerning bits in the model card, seemingly treating CoT controllability as something to be "investigated" rather than the warning sign it's supposed to be.
Somehow this article doesn't even mention the fact that AI makes software rewrites much, much faster than before and with higher confidence of backwards compatibility.
Nowadays, a good AI harness can fairly reliably rewrite a medium complexity piece of software to an appropriate modern tech stack with pretty strong confidence of exactly preserving its behavior. The AI can pick up legacy details and keep them exactly the same as before in ways that a human rewriter would usually not bother with. After rewriting each feature it can then exhaustively smoke test all the happy paths and edge cases and ensure the code behaves exactly the same as before, which is another thing that human rewrites basically never do.
- For Claude.ai subscriptions I think Sonnet is much cheaper than Opus. This is why there was a "Sonnet only" usage bar for Max tier for the longest time.
- For some tasks the sheer amount of raw input tokens is the most important. For example multimodal computer use tasks. You can't make them any more efficient on Opus by turning down the reasoning, so a cheaper model like Sonnet is useful for them
+1, this is a good thesis. It's plausible and if it's true, then that explains the opaqueness of this whole ordeal and why it's been applied to both OpenAI and Anthropic models now.