Hacker News: edg5000

New comment by edg5000 in "A new spam policy for “back button hijacking”"

edg5000 — Tue, 14 Apr 2026 12:06:11 +0000

Power is taken but also given. It's a dynamic and I agree it's gotten way, way out of hand. It may eventually supress progress and become a real parasitic presence, but we've not reached that point yet (in net terms). Google has been relatively responsible with the power, but cracks have been starting to show. It will get a whole lot worse before it gets better. That is why I embrace vertical integration despite the tremendous cost. Call it the cockroach approach; it allows being partially decoupled from outside fluctuations.

Addition: People underestimate Google's influence. It's easy to forget they de-facto control Firefox, leaving only Apple and Google in control of the Web. Scary, but looking away won't help either. The Americans have been consistently competent with technology since the advent of the transistor right after WW2. They're reaping the benefits of that still to this day. I say that as a European.

New comment by edg5000 in "Someone bought 30 WordPress plugins and planted a backdoor in all of them"

edg5000 — Tue, 14 Apr 2026 03:57:02 +0000

If the plugins were bought for six figures, then it must be incredibly lucrative. How on earth could they be making it back? Is injecting spam into Google results THAT lucrative?

New comment by edg5000 in "Someone bought 30 WordPress plugins and planted a backdoor in all of them"

edg5000 — Tue, 14 Apr 2026 03:55:03 +0000

> In 2017, a buyer using the alias “Daley Tias” purchased the Display Widgets plugin (200,000 installs) for $15,000 and injected payday loan spam.

Is that it? Going through all that trouble just for some spam? Surely more lucrative criminal actions can be imagined with a compromised WP plugin?

New comment by edg5000 in "Ask HN: What Are You Working On? (April 2026)"

edg5000 — Mon, 13 Apr 2026 16:02:18 +0000

Cool use case. I have it PDFs for my electronic load and lab PSU, and it was able to write drivers in C++. Quite a powerful use case. I did battery discharge tests this way.

New comment by edg5000 in "Ask HN: What Are You Working On? (April 2026)"

edg5000 — Mon, 13 Apr 2026 16:01:15 +0000

For large georeferences textured meshes, generally 3D Tiles is used which uses a hierarchy of oriented bounding boxes to support various LODs. If you split up your model into chunks as well as diferent LOD levels, the viewer can then request chunks based on what is in the view port as well as the zoom level. The Cesium implementation leaves a lot to be desired; it't pretty tricky to get right. This will become commonplace since 3D scans are getting more common.

New comment by edg5000 in "Sequential Optimal Packing for PCB Placement"

edg5000 — Sun, 12 Apr 2026 04:50:53 +0000

They do a have a vision decoder like many other LLMs, so in theory it should be able to write the positions textually, then call a render command, then look at the rendered bitmap. I's all very opaque though; I'd love a visualisation of the latent space data that it's converting the image to. I found that very long vertical images throw Opus off completetely for example. It's very interesting to experiment with this. Let it play with placement and let it call a render command. Then let is describe in detail what it sees. I'll be looking into this a lot this year. Maybe there will be niche models that will be smaller but have better vision capabilities then Opus. A world where one model rules would be incredibly depressing (kinda like what we saw with some software companies since the 90s).

New comment by edg5000 in "France to ditch Windows for Linux to reduce reliance on US tech"

edg5000 — Sun, 12 Apr 2026 04:42:57 +0000

They all have Intel or AMD chips. My surprise is that there are no good, non-US ARM options. I'd love to be proven wrong though.

New comment by edg5000 in "France Launches Government Linux Desktop Plan as Windows Exit Begins"

edg5000 — Fri, 10 Apr 2026 11:15:28 +0000

Interestingly, there are zero non-US powerful laptops. The closest option is the Moore Threads MTT AI Book (12-core 2.65Ghz, 32GB DDR5, 1TB SSD, 14 inch). It cannot reach a modern Ryzen in performance though. It's fascinating that only the US can make good computers. I'm not from/in the US so I'm not saying that from a patriotic point of view. How hard can it be to pop a good ARM chip in a laptop and compete with HP, Apple and the likes?

New comment by edg5000 in "GitHub has DMCA'd nearly all forks of the official Claude-code repo"

edg5000 — Sat, 04 Apr 2026 11:08:13 +0000

:)))

New comment by edg5000 in "Sequential Optimal Packing for PCB Placement"

edg5000 — Sat, 04 Apr 2026 10:08:45 +0000

I think using the vision decoder baked into modern LLMs is the way to go. Have the LLM iterate; make sure it can assert placement qualities and understands the hard requirements. I think it can be done.

New comment by edg5000 in "Qwen3.6-Plus: Towards real world agents"

edg5000 — Fri, 03 Apr 2026 04:30:11 +0000

Hahaha

New comment by edg5000 in "Qwen3.6-Plus: Towards real world agents"

edg5000 — Fri, 03 Apr 2026 03:42:44 +0000

I build custom harnesses (like many of us) and I genuinely think Anthropic will eventually sue their customers if they detect they are selling competing harnesses (competes with their vertically integrated offerrings).

I feel Alibaba and DeepSeek see themselves more as infra. No urge to control the stack and litigate competition out of existence.

New comment by edg5000 in "Qwen3.6-Plus: Towards real world agents"

edg5000 — Fri, 03 Apr 2026 03:37:21 +0000

What are the reasons for switching? Personally I got into the habit of doing a bit of a round robin with Codex/Claude (CLI) and then DeepSeek and Qwen web chat. And Claude in web chat. I like to switch just to learn the differences, otherwise I'd never know what the other models can do. But I still feel attached to Opus, but this can be fammillarity. If I only had Qwen maybe it would be effectively identical at the end of the day. Hard to say.

New comment by edg5000 in "Qwen3.6-Plus: Towards real world agents"

edg5000 — Fri, 03 Apr 2026 03:35:17 +0000

Do they have an API where you can control the chat template or at least just put everything in the system prompt? This way you can control everything including the tool calling syntax. Even if you use the trained tool syntax, it allows you to control the tool system prompt which you may want to tweak. With DeepSeek this is all possible. An undocumented feature, great for harness builders. Anybody got info on Qwen regarding this?

New comment by edg5000 in "Qwen3.6-Plus: Towards real world agents"

edg5000 — Fri, 03 Apr 2026 03:29:15 +0000

How can it be free? What do you mean?

EDIT: Ah, I see. Some kind of promotion. Pretty cool.

New comment by edg5000 in "Qwen3.6-Plus: Towards real world agents"

edg5000 — Fri, 03 Apr 2026 03:22:01 +0000

Has anybody done serious agentic work (e.g. using a CLI harness or simmilar) with 3.5 Plus/3.0 Max and such? How does it compare against Opus with Claude Code? I've used the chat quite a bit and I can't say at this point.

New comment by edg5000 in "Trinity Large Thinking"

edg5000 — Thu, 02 Apr 2026 04:03:01 +0000

They are repeating a million times on their huggingface page that the thinking output should be included in the conversation history for multiturn use. That makes me wonder, is this generally needed for LLMs? Because that implies that they only really function well on typicial multiturn flows; I'm experimenting with a completely different approach: there is still the main message stream in the context, but the agent can use structured means to exchange messages and interact with terminals and the file system in a statefull manner. The state is rendered to the context on every cycle, with the message history just being a "panel". I'm still in the middle of trying this out so I can't say yet if it will work. But I hope the models are flexible enough for this.

New comment by edg5000 in "ChatGPT won't let you type until Cloudflare reads your React state"

edg5000 — Mon, 30 Mar 2026 16:36:49 +0000

The chat client has serious performance issues on lower end systems. Now I see why!

New comment by edg5000 in "Iran-linked hackers breach FBI director's personal email"

edg5000 — Sat, 28 Mar 2026 10:25:33 +0000

We'd have to look at the longest-running democracies and observe how they handled periodic refactorings

New comment by edg5000 in "TurboQuant: Redefining AI efficiency with extreme compression"

edg5000 — Wed, 25 Mar 2026 08:35:38 +0000

So limiting max context length also reduces VRAM needs a bit? If cache is 20% of total, 1/10th of context as a limit would mean 18% total memory reduction.