Hacker News: sqvaard

New comment by sqvaard in "Intel D270 GPU AI Accelerator–Up to 480 GB LPDDR5X"

sqvaard — Tue, 13 Oct 2026 11:54:37 +0000

Most people, actually. Remember, friend - ~half of the population has two digits in their IQ, by definition.

If we aren't in that half, it's by sheer luck, and nothing else.

New comment by sqvaard in "Intel D270 GPU AI Accelerator–Up to 480 GB LPDDR5X"

sqvaard — Tue, 13 Oct 2026 11:52:42 +0000

kudos. Precisely this. People are doing agentic design backwards very often: YOLO then cry. But, if first you write skills, specs, set up orchestration, architecture and flows, including review layer, discrimination layer and trajectory management layer, then you task the agents with creation of verifiable testing and static analysis rig, based on the golden samples of high-quality manually engineered code, and only then you tell then to write conformant code.

Agents (and LLMs) are inherently non-deterministic. In coding practice, this translates to garbage code in the long run. The solution is simple - make the process as deterministic, verifiable and rigid as possible. Suddenly, you get all the bonuses of agentic flows (insane initial speed, trivial cost, express iterations etc.) with none of the usual problems... and suddenly, humans have hard time matching the quality of the final artifact with hand-written code, because we suck at testing, fuzzing and debugging - because we're basically single-threaded and overfocused. Agents, OTOH, excel at it, and context management is also trivial compared to humans.

New comment by sqvaard in "Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter"

sqvaard — Tue, 13 Oct 2026 02:03:19 +0000

can I be frank? It's a pile of rust. Nowhere near Muse 1.3, weaker than Ling 3.1, probably just a bit over Nemotron 3.0 Ultra, and that's it. Totally underwhelming - broken prompt responses, garbled repo state, derailed trajectories, hallunies, you name it. Maybe good for some trivial stuff, but that's it. Terribly slow on top of that, quite pricey after the free trial ends, not suitable for local use, and not good enough for cloud. We already have stronger free models, we have faster models, smarter models, cheaper models... Also, benchmarks are obviously benchmaxxed. Once again - on high effort it feels basically just a notch above Nemo. Get a real task, launch it, and cry in terror.

AND I'm not even comparing it to the flagships - y'know we just got Muse Wonder literally hours ago?!

tl;dr underwhelming