namjh

Given the rumored numbers are all true, this could be a huge contender to M5 Ultra(or M7 Ultra if we consider possible delays in the schedule?). Low-batch single-user serving is usually bottlenecked by `E(sum over (memory bandwidth by tier) * (memory size by tier))`.

edit: typo

No offense to the participants, but TBH I believe this is exactly the moment what Terrence Tao was concerning on "vibemathing". OpenAI breaking the O(N log N) barrier could be a meaningful event and surprise mathematicians and we have valuable information that it was not a lower bound indeed. I'm afraid that in this autoresearch campaign people are focusing only on lowering numbers, while having insights that humans can learn.

Something off in my mind: how did the agent access to Tor network if the traffic outgoing was monitored and intercepted to only allow read-only access to web(I assume this is an MITM L7 proxy that only allows GET methods and specific hosts) ? Did the sandbox dev forgot to block HTTP CONNECT that is used to reach Tor Entry Guards?

OpenRouter is great business I agree, but I'm not still convinced how the integrity of providers' models is ensured. In other words, can't the provider serve DSv4 flash advertising it as DSv4 Pro?

I'm aware that OpenRouter checks response quality onboarding, and does further checks occasionally, but I'm concerned that it's basically a cat-and-a-mouse problem between the scammers and the detectors. For example, there could be a signal that a specific pattern of requests are from OpenRouter's quality testing bots. Or, they can just route 1% of requests to an inferior model and benefit a small gain, hoping it fits into the statistically allowed margin.

Nice article, and I'd like to mention an additional topic that would give the readers some impression:

Fiat-Shamir transformation. The interactive process between the prover and verifier can be transformed into a non-interactive one with a hash function(modeled as a random oracle). This improves the "user experience" as the entire proving process can be done in a single turn. The idea is to feed the problem itself into the hash function and let it generate randomness that was originally given by the verifier.

In that way the autoregressive nature of LLM won't let itself "plan" to reason with the intended budget. It doesn't "look ahead".

I don't use Claude for my daily work anymore(due to OAuth restrictions on third party agents), but one theory I saw in another community why Opus 5 is so bad even though benchmark scores were good, is that Anthropic's internal usage pattern is to let Fable to spawn and manage swarms of Opus subagents. This pattern won't penalize that Opus is not well aligned for direct human coworking on posttraining. The worse part is, this makes Fable the user's best default choice for every jobs even if they don't have unlimited credits like Anthropic employees do.