bellowsgulch

Don't give away our job to Indians? I mean what do you want here, it's not a post-grad course. If H1Bs are for "highly specialized knowledge," why the fuck are we taking in Indians at below par cost for American "highly specialized knowledge?"

Why is there not a highly significant fee or compensation threshold to prevent abuse?

Track record of the Democratic party vs the Republican party

I vote for Democrats. Sad they didn't step up for American tech workers though, and the guy shitting himself did.

It seems to me that Qwen3.6-35B-A3B is still the local LLM leader, because 27B in either of the latest releases is just too slow to be usable compared to OpenCode Zen free models, or OpenRouter free models.

Sad a Qwen3.8-35B-A3B model wasn't released.

Also, I think this is just an ad.

Uh...

    brew install llama.cpp
    llama serve -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_M
    open http://127.0.0.1:8080

Does anyone remember when you posted on a place like Something Awful or Facepunch getting banned for a day for saying something so fucking stupid you needed to learn a lesson? No? Anyone?

Honestly, I wish other language developers would chime in, too. It's insightful.

Who cares what you think?

I agree. It's actually quite insulting to other countries whose actual atrocities were measured in lives lost. I voted for Kamala but Trump Derangement Syndrome is real. He's just a loser, he's not even remotely brave enough to be Hitler Jr.

The guy doesn't have the stomach for real totalitarianism. Just populism, corruption, and weakening the country.

Growing up with PageRank search felt so, so much different than whatever the hell we get with Google now, which is almost certainly overtuned to some crap metrics that some software engineers are getting paid good for and getting promoted over because they measured some crap that doesn't matter. Such a shame.

GPT‑Realtime‑2 is really weird. Perhaps just because it's bidirectional and now has the failure mode as a possibility, it responds too soon with filler at awkward times, and it's generally overeager. I feel like there was plenty of opportunity to just work on latency engineering like this effort.

poob has it for you

> Pity you had to ruin it by telling us you're exceptionally thoughtful and profound, and that those who disagree, suck!

It's just a comment on the Internet, you don't have to take it so seriously.

I've been using OpenCode Zen and OpenCode Go for a while now, and I've been a bit disappointed with how they communicate to customers. There's no status page that I'm aware of, and at least with my private harness, connections frequently drop. Maybe OpenCode itself does this, too?

Anyway, I suspect it was just pricing. They seemed to suggest they were running it on their own hardware? But I'm not sure that was actually the case. As soon as DeepSeek's pricing changed, they basically fell back to higher prices as well.

I was very confused why I was getting approximately $30 of spend when I thought I was paying for $60, but I guess I shouldn't complain when I have a DeepSeek account and I was paying for far more token usage on OpenCode Go.

I have since defaulted my private harness to muse-spark-1.2-contributor-free, because that seems to be the prevailing model out of the current Zen list.

Futurama did it first!

The older I get, the less I want to use new browser features that have already been somehow possible with existing ones that typically have been around for literal decades.

Adopting them makes a future with multiple browser engines more labor intensive and unlikely as the number of standards targets one is required to support grows and it becomes a game of asking yourself what sites you want your browser to support.

If you don't have a lot of experience in web software, you might think, that's a silly thing to think about, but I think it's that level of maybe exceptional thoughtfulness that has really profound ramifications. If you do have a lot of experience in web software and you think it's silly to think about, you suck.

They successfully did it, basically, if you read about Xbox's architecture. It's a really interesting look into how we basically wouldn't own our Windows PCs if they used that tech in their mainstream software and hardware.

OpenCode Go is too good to be true, and I suspect it won't last, like GitHub Copilot.

I have no idea why anyone would spend the money Anthropic asks for Claude models these days. There was a brief window in time when it was worth the money to use them a bit through GitHub Copilot, but there's no way in hell I'm spending that for Claude Sonnet when DeepSeek is basically doing all of my work right now.

I want them to improve the product meaningfully beyond lane keeping, which besides lane changing (partially), it's the only thing it's capable of because they willingly choose not to extend the system to use more visibility.

I thought I would add a comment asking the maintainers to consider not shipping major version zero, since it basically makes semantic versioning worthless and using SemVer and ignoring this advice basically suggests one wasn't reading it to begin with, but then I asked myself, "Why bother?"

No one seems to take it seriously, and why would my voice matter anyway? I think https://0ver.org shows us the industry doesn't really care about software versioning.

Hell, just drop the pretense and tell people to stick to commit hashes and dates.

This is definitely such a cool feature that I wish cloud providers would expose.

Two things: 1. they have a cheap fuse (cents on the dollar) that’s not automotive-grade so your windshield mounted device fails in hot weather and doesn’t last past the hardware warranty. Known issue, doubt they ever fixed it.

2. 1B parameter model to stay between two lines. Why. Waste of time. Their product hasn’t actually changed in several years because they will never advance past lane keeping SAE Level 2.

I suspect for what they’re selling for $1,000+ these days, could you find a way to do for half? Less than that?

I mean after all, it still runs off a Snapdragon 845.

Nice tech, disappointing limitations and too expensive. The world moved on, Tesla is bad but a Waymo competitor would be more serious than this.

I've tried this, too, and for whatever reason, simple things like having a subagent with vision capabilities hallucinates responses to the blind models. I don't know why, and it's not consistent, but I don't see this happening when the primary agent is a vision model.

The existing ones definitely are, so you know they're bullshitting.

Counterpoint: something has happened in frontier models, and yes they now get discouraged and will sometimes prefer to not continue working on a problem unless you tell them to anyway.

I don't know how or why this would be trained on behavior, but no, it isn't true anymore that models don't say things like, "Ugh," or "this is going to take hours and maybe we should stop here."

Framework, Bevy is definitely not an engine.

Does it do one thing? It's probably a library.

Is it a collection of libraries that work together focusing on primitives with callbacks? It's probably a framework.

Does it load levels? Have its own schema for an entity? Provide a multiplayer server protocol, and build on primitives? It's probably an engine.

Libraries -> frameworks -> engines -> games.

Fuck Apple.

Yeah, it's really weird watching everyone complain about Claude Code and continue to give Anthropic what amounts, in most cases, to 20 years of my current spend in 1 year, while I'm using a private harness and Pareto frontier models.

But for good or even exceptional engineers to exist, by definition, bad ones have to, too.

Yep. Exact same take here. I have a private harness I actually integrated GitHub Copilot Chat support with, but you lose an "entire months" credits with any nominal task.

They knew what they were doing and wanted to kick people off.

Thank you for sharing your experience, too. I suspect a lot of us all took this path.