Hacker News: hedgehog

New comment by hedgehog in "Qwen 3.8 27B is excellent, but it defaults to overthinking things"

hedgehog — Mon, 17 Aug 2026 03:17:25 +0000

To be fair a lot of models have quirks, I've never found a model swap that was transparent.

New comment by hedgehog in "Qwen 3.8 27B is excellent, but it defaults to overthinking things"

hedgehog — Mon, 17 Aug 2026 03:12:57 +0000

Look for improved templates by "froggeric" on Hugging Face. I use Qwen 3.6 a fair amount using his template and it fixes some issues I saw with the upstream versions.

New comment by hedgehog in "Qwen 3.8 27B"

hedgehog — Fri, 14 Aug 2026 18:57:30 +0000

This is all a pretty active area of research, both adaptive drafting, different draft approaches (DFlash etc), I expect to see really big improvements land in the popular runtimes this year.

New comment by hedgehog in "Qwen 3.8 27B"

hedgehog — Fri, 14 Aug 2026 18:31:49 +0000

Yes, I was fixing issues piecemeal until I found the froggeric template, I've had to fix I think one issues with that one but it's better.

New comment by hedgehog in "Qwen 3.8 27B"

hedgehog — Fri, 14 Aug 2026 18:22:58 +0000

Same, I have one workload where on 3.6 drafting 6 tokens is the fastest setting.

New comment by hedgehog in "[dead]"

hedgehog — Thu, 13 Aug 2026 19:45:39 +0000

I think the idea is you do visual editing on your page as rendered, and then Pixy generates a log of what you did annotated in a way a coding agent can use to generate code changes. Seems convenient.

New comment by hedgehog in "DeepSeek Harness developer preview"

hedgehog — Thu, 13 Aug 2026 18:45:42 +0000

I have some self-improving harness tooling, I run it within Claude Code because of the cost advantage but that does mean restarting Claude Code periodically so it can pick up all changes. It sounds like the DeepSeek system won't need that, which is useful.

New comment by hedgehog in "Tracking down the 16-year-old WAL-reset SQLite bug"

hedgehog — Wed, 12 Aug 2026 18:36:24 +0000

The fun part is unless your app is pretty high traffic you might see the bug only once, or twice, and never have a reasonable way to even get close to a fix or know what your real exposure is.

New comment by hedgehog in "Nvidia Nemotron 3.5 Lightning and NeMo Switchyard"

hedgehog — Tue, 11 Aug 2026 20:16:45 +0000

See sibling answer but essentially the effectiveness of cache is not diminished by having a separate one per model (relative to the win of doing more turns and generation with a cheaper model).

New comment by hedgehog in "Nvidia Nemotron 3.5 Lightning and NeMo Switchyard"

hedgehog — Tue, 11 Aug 2026 20:13:48 +0000

To elaborate, because I don't think some of the people reading this understand the reason, typically a lot or most of the cost in "agentic" API usage is cached read + generation. Cached read costs scale with turn count, which multi-model switching doesn't increase, and of course generation gets cheaper if you do some of it with a cheaper model. When you switch models the "catching up" batch of messages is just a single prefill and then that goes into cache. You don't even need to have the same chat history across models so long as the view from each model's perspective looks like a series of appends.

The main problem with model routing in my experience is that to work well the router needs to be pretty strong, maybe even moreso than any of the actual models in service. There are probably clever solutions to this but I haven't seen any that look better than just using sub-agents.

New comment by hedgehog in "The Alpha 21264 CPU: NT's Greatest RISC (1998)"

hedgehog — Mon, 10 Aug 2026 17:26:12 +0000

As pjmlp points out we don't know what would have happened if Intel had spent another few billion dollars re-implementing IA64 in a lower cost / lower power form. To me one of the big lessons of the last few decades is that pragmatic microarch has convergent evolution due to the same underlying constraints of expensive memory access, clock speed scaling limits, general non-determinism of everything surrounding the core, etc, and within reason the ISA will end up conformed to those limits in a way that makes the initial ideology (RISC, CISC, VLIW) not so important. For example RISC is simple, except once you mix OoO, compressed instructions, macro op fusion, etc, those all have externally visible effects so c'est la vie simplicity.

One can imagine an evolution of IA64 that conceded microarchitectural complexity to allow adoption of a more compact instruction format with dependency tags rather than fixed bundle width, resulting in something more like an evolution of P6 with a different face, better x86 compatibility, and still ISA uniformity with the big iron. Or an evolution that involved buying Transmeta and adding IA64 support to that.

Obviously Intel didn't end up doing any of these things, probably because distribution strength and process superiority allowed them to do less and still post strong quarterly numbers. I'm not even sure if it was a blunder given the local incentives of the people who worked there at the time.

New comment by hedgehog in "Auto mode is now the default in Claude Code"

hedgehog — Mon, 10 Aug 2026 07:40:38 +0000

You can set it in the config file so it's the default every time.

New comment by hedgehog in "The Alpha 21264 CPU: NT's Greatest RISC (1998)"

hedgehog — Sun, 09 Aug 2026 18:55:47 +0000

It's unclear to me what would have likely happened had AMD stayed on IA32 and Netburst been uncontested. Itanium was far too expensive for consumer machines and it seems very possible that Intel would have fumbled IA64 in a way similar to more recent AVX-512 and their big.little efforts. Easy enough to port XP to Itanium, hard to get those machines into Best Buy or the Dell catalog in any way that makes sense. Had AMD continued Athlon performance improvements and Intel still responded with a P6-derived Pentium M (which even in our timeline was 32-bit) it still seems like there would have been no volume market for IA64 and we still would have eventually ended up with an evolution of IA32 even if it came from Intel instead of AMD.

New comment by hedgehog in "Europe EV Sales BEVs Jump 50% & Reach 26% Market Share"

hedgehog — Sun, 02 Aug 2026 22:02:08 +0000

They Toyota hybrids are still gasoline, what's the objection to the eCVT? So long as we're still using combustion engines it seems like the best design anyone's come up with.

New comment by hedgehog in "Linux desktop market share has hit over 10% in North America"

hedgehog — Sun, 02 Aug 2026 04:43:01 +0000

https://archive.is/kaef0

New comment by hedgehog in "Agent Skill to Force Docs in ASD-STE100 Simplified Technical English"

hedgehog — Thu, 30 Jul 2026 23:29:20 +0000

They are useful for packaging all the related bits together. Some instructions on the line to add to the main prompt, detailed policy doc that spells out details, troubleshooting docs, the code itself, etc.

New comment by hedgehog in "Stacked PRs are now live on GitHub"

hedgehog — Thu, 30 Jul 2026 22:02:31 +0000

I like it based on initial testing today. I already use my own local UI for managing stacked PRs so I can see the dependencies as a tree view and the review + CI status for each, it would be good to also have those in the GH web UI. Maybe I missed it but it looks like merging just the bottom of the stack in the web UI might not be supported? Happy to share the workflow / code, it would be nice if it was supported within GitHub's native tools.

New comment by hedgehog in "Kimi K3-256k"

hedgehog — Wed, 29 Jul 2026 21:31:58 +0000

I do essentially all of my work with auto compact set to 250k and it's fine. It may be due to the way the project tooling is set up and the use of sub agents?

New comment by hedgehog in "Handbook.md shows that long policy documents do not reliably govern agents"

hedgehog — Wed, 29 Jul 2026 14:41:10 +0000

What models have you found to work for which tasks? I find the local models very useful but not in a way that's replaced cloud models (yet, I remain hopeful).

New comment by hedgehog in "Half-Life ported to Mac OS 9"

hedgehog — Wed, 29 Jul 2026 04:20:52 +0000

I don't know the details but I would imagine that the comparison would shift quite a lot over those years due to the rollout of SIMD features on both sides along with the other microarchitectural improvements. I'm not sure I would characterize PPC as being mediocre, I think the designers hit more or less the target they wanted even if the top line performance numbers weren't hands down higher than the contemporary x86 parts and it didn't have the shine of IA64. There's a reason we still fly PPC 750 variants in some critical systems. If you think back to that period Apple was moving from 68k (where they were both one of the biggest customers for the chips but also not enough volume to field a part competitive on desktop) to PPC where they were no bigger (maybe declining unit volume) but at least part of a stronger coalition of users. The group's interest diverged (no laptop roadmap) and I'm guessing Apple wanted to switch to x86 as early as 2002 and it just took a few years to get the company stabilized enough to make the jump.

Edit: It would be interesting to look at the software rasterizer and game engine problems in a modern context of big compute and superoptimizers and see what is possible. I'm guessing not much above what the hand-tuned code of the era got, but maybe significant in the context of doing these kinds of retro-ports on a small budget.