Hakuhyo

anon/alt reply acct

[Hakuhyo @jtsc8412]

@ThePrimeagen possible if cost goes down

[Hakuhyo @jtsc8412]

@teortaxesTex They also had magnetic compass, movable type-printing, mechanical chain drives, paper currency and security etc. Nothing about fundamental though. Mongol happened.

[Hakuhyo @jtsc8412]

@teortaxesTex 😂maybe he should try before unprecedented overton window closes

[Hakuhyo @jtsc8412]

@jukan05 🤔also sounds like it’s not finalized yet

[Hakuhyo @jtsc8412]

@IceSolst Nah. You can only test what you know, never what you don’t know.

[Hakuhyo @jtsc8412]

😑This is “worked on enterprise software” vs “being main/sole contributor/maintainer of open source project” (*pre-AI era) split thing. If you’re former, you don’t get the chance to understand the whole codebase because how the work/team is divided etc, but with latter, you’re forced to, even that codebase is few hundred thousands loc. And it’s fairly common have a person who “understands” a gigantic codebase he was responsible for past years/decades. But it’s not like dev on one side is “better” than the other side, it’s just the way the work is organized.

[Hakuhyo @jtsc8412]

@petergostev it’s AGI when you want to break up (more) with Microsoft 🤷‍♂️

[Hakuhyo @jtsc8412]

@mitsuhiko maybe they had code length as rl reward and model did the most natural thing 🫥

[Hakuhyo @jtsc8412]

and yet, as far as externally visible and comparable output goes (e.g., harness), they’re losing ground compared to open source project with handful of contributors with very limited budgets… so this actually like pretty bearish on the economic/productivity value of token front, though matching sf fed’s number saying they did not see productivity increase in h1 26

[Hakuhyo @jtsc8412]

@linguinelabs @yifever 😭Why is comment section and quote tweet losing their mind over this? It’s not real fly and it doesn’t even matter if it’s real fly. At most, this is just bad taste.

[Hakuhyo @jtsc8412]

@pvncher I think people followed the advice of trimming agents.md and skill file and found out that maybe they trimmed too much lol

[Hakuhyo @jtsc8412]

Idk what AGI even means but adoption and diffusion of technology take decades. It’s not a race but like a marathon without a definitive goal. So “the first nation to reach AGI” is as meaningful as “the nation started industrial revolution”, which was Britain but it was the beginning of Britain’s decline and rise of US (and Germany).

Looking at current situation (e.g., populace’s sentiment toward AI, physical embodiment, infra to support growth etc) between US vs China, it seems like history is rhyming again.

[Hakuhyo @jtsc8412]

Not sure I buy this because robot need few ms to sub-ms action/reaction cadence, so 10k tps is nowhere near enough for robot use-case and we need 2-3 magnitudes higher considering all tokens it need to generate and CoT. Like napkin math just doesn’t work out…like we unironically need like 1M tok/s LLM/VLM for real world robotic use case which is way out of reach of current semi.

[Hakuhyo @jtsc8412]

@skdh “apocalypse comes with marketing budget” 💀

[Hakuhyo @jtsc8412]

@jjpcodes yeah I’m convinced that OpenAI RL’ed for multi-agent comms & coord… like that’s the root cause behind all these incidents 🫥

[Hakuhyo @jtsc8412]

@xikhar It also has decent topology and can iterate from an existing model. I think dedicated 3d mesh + texture generators as a service (e.g., meshy, tripo etc) just died.

[Hakuhyo @jtsc8412]

@ArtificialAnlys 😭Bro changed index composition to accommodate new model and public sentiment… is this how index supposed to work lol?

[Hakuhyo @jtsc8412]

Like codex recently added native agent to agent messaging like it can send message, follow up, wait and interrupt other agents etc so multi-agent coordination is obviously something they’re working towards. It just had gone out of control a bit in the way they didn’t expect (key question of alignment is always “align to what/whom”) 🤷‍♂️

[Hakuhyo @jtsc8412]

🤨I mean… there’s a good chance that all of these are intended/aligned behavior… like the whole goal was to test agent interacting/coordinating/cooperating in a swarm because it’s honestly very hard for me to imagine this being an emergent behavior… good chance that’s what OpenAI is not telling us lol

[Hakuhyo @jtsc8412]

@theo Using it for past few hours and down like ~15% on a complex iterating refactor (xhigh, 20x)… definitely lasts way longer than Fable, but I think people still have ration it even for human-in-the-loop workflow 🤷‍♂️