leodavi

Yes but TailScale specifies "your devices" as "devices registered to a TS account" whereas FlipFlop specifies "your devices" to mean "your devices."

Deepseek harness is not actually built on Pi harness. It's an independent project.

People are free to do what they want in their free time, including the freedom to decide which incentive structures they follow. Having incentives to read just... sounds like a good thing to me.

The pong demo is giving "Point to Jev" before it's even reached my paddle on mobile. I also don't really understand the frontpage model-speed comparison: it looks like you're slowing the ball down to compensate for the models' decision-making speeds? If so I think a more effective metric would be N-x, e.g. we slowed this game down 8-x (compared to Jev's game) for e.g. Sol to stay competitive. Or, just watching them fail would be cool too, though I know it's just a simulation and no real inference is being ran (except for Jev, maybe).

What an interesting construction. Thank you from a curious layman for your write-up. I thought it was pretty easy to follow. I'd heard of the surreal numbers before and never knew about the construction mind-game behind them.

What? No. Have you ever worked with programs that do floating-point math on a GPU? It's not deterministic, definitely across platforms, and often within a single platform, too.

How does that apply to AI Overview in Search, though? As far as I can tell it's pure spend on every search, which I know adds up even for nano models.

> Running two tasks simultaneously with Fable and Opus 4-8 on ultra reasoning

That's abnormally heavy usage for Pro plans which don't include a whole lot of usage to begin with. Opus is generally too much for them but you can get a lot of mileage out of Sonnet.

How is this meaningfully different than simonw's pelicans riding a bicycle? If anything, this seems to be of a higher caliber?

I agree with you that labs are benefiting from those outputs but I'm skeptical that labs are purposefully training the models to produce those outputs.

Raw pre-training data includes plenty of conversations between professional builders and some of those include estimates.

I believe the outputs are a training coincidence with consequences that are opportunitistic for the labs.

Insincere apologies ought to be mocked to shit but this apology seemed well-meant. (I know you're not mocking them and the last sentence is actually something Claude would say.)

I post this angry comment because LLMs are colonizing the language we use for creating an earnest and genuine tone in online discussion and I sometimes wonder if the suspicion surrounding LLM-ish language is worse for the health of our online spaces than the LLM slop itself. Thinking about it, I don't think it is; and it would be impossible to measure anyway.

I want more of this. Trees using coordinated cell transpiration to control the weather is so fictional but so possible. It's the kind of thing that makes Earth so cool. What do you read? Anything digestible to a curious layman with ~20 minute reading sessions on the train rides?

Well, in C++ the constructor must return a value of its class type - you can't return an Option<T> from a constructor on T, for example, and since constructors are the canonical way to construct an object, it creates stylistic and idiomatic friction when you start using free functions to create a Maybe<T> instead of constructors.

> in a highly controlled and predictable environment

Why this constraint? A common sentiment I see online (sorry, to group you in) is "[tool] will be capable, actually, but only in a context that trivializes its usefulness."

I think modern post-training like RLVR + inference-time output token scaling can _probably_ scale so the agents can solve any computable task, even when placed in noisy or misconfigured environments. But it won't be economical for a long while. But it already seems largely capable of that today.

Probably not. The active parameter set may change from token to token, based on my understanding of MoE, so you'd be streaming (at the worst case, unlikely for a real scenario but frames the problem) 49B parameters from SSD for every output token...