Lance R. Vick

Anthropic just legally threatened Opencode to make them drop support: https://web.archive.org/web/20260221041617/https://github.com/anomalyco/opencode-anthropic-auth/pull/15#issuecomment-3930558874

Archive link because they deleted the repo after to comply with demands.

In short, Anthropic only wants you using their official walled-garden clients to access the models trained on our open source code.

They are not a lesser evil. They are just as evil as OpenAI.

Stop giving these assholes money. Rent or buy hardware to self-host with privacy and freedom. It is not that hard, I promise.

fix: Align Anthropic OAuth requests with Claude Code by deveworld · Pull Request #15 · anomalyco/opencode-anthropic-auth

Summary Normalize Anthropic OAuth requests to match Claude Code's headers, betas, metadata, tool casing, and model IDs. Remove tool_choice and inject metadata.user_id from ~/.claude.json to sa...

GitHub
Mar 26 at 6:09pmWeb
Show threadxyhhx 2d ago
@lrvick 👏 dont 👏 use 👏 generative 👏 ai 👏
Show threaddegenerating degenerate2d ago

@lrvick

> Rent or buy hardware to self-host with privacy and freedom. It is not that hard, I promise.

I would really like to do this instead of using Gemini 3.1 / Antigravity. And I have a 5090 to do it with. But what "self-hostable" solution (because of the models, none are FOSS) is actually competitive with those?

Show threadLance R. Vick2d ago

@hopeless a 128G Framework AI Max 395+ gives you ~110GB of VRAM and I am getting ~35t/s w/ qwen3.5 122b for reference once switching to 7.0.0 kernel series which is a big perf boost.

https://frame.work/products/desktop-diy-amd-aimax300/configuration/new

Configure Framework Desktop DIY Edition (AMD Ryzen™ AI Max 300 Series)

Choose from AMD and Intel system options, select your preferred memory and storage, operating system, and more customizations. Available in DIY and pre-built configurations.

Framework
Show threadLance R. Vick2d ago
@hopeless #notsponsored
Show threadnabeards1d ago
@lrvick @hopeless I’m honestly impressed by that pricing! I thought it would be much higher
Show threadLance R. Vick1d ago

@nabeards @hopeless Everyone only wants to buy Nvidia, but the AMD stuff is much cheaper if you are willing to use bleeding edge kernel+drivers as support is very recent.

But if you wait to buy hardware when it is plug and play, prepare to pay double.

Show threaddegenerating degenerate1d ago

@lrvick @nabeards

... sure. But clearly, you don't need a H200 to run the inference for SOTA coding assist. And if it works but is slower on a 5090, I will definitely be motivated to spend money to make it work faster.

The actual inference horsepower seems to be becoming less of a blocker

https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/

One big missing part is the tight loop eg Antigravity has around the LLM such that, eg if it produces a bad diff, it retries it from scratch at the cost of a few seconds.

TurboQuant: Redefining AI efficiency with extreme compression

Show threadnabeards1d ago
@lrvick @hopeless is there a Confidential Compute equivalent on AMD GPUs? The EPYC SEP system is awesome.
Show threadLance R. Vick1d ago

@nabeards @hopeless I would argue Nvidia does not actually support useful confidential compute given it requires you giving them root access to your host operating system which breaks the trust boundary for no reason.

That said, with Linux 7.0.0+ and SEV-SNP, you get to take advantage of PCI Link Encryption where every PCIe device gets a dedicated encrypted communications channel to the VM it is bound to, which could also be an SEV-SNP VM.

I think this setup is best effort today.

Show threaddegenerating degenerate1d ago

@lrvick @nabeards Let's say I did those contortions. And I forgot this thread started with you promising how easy selfhosting it was.

It still won't compete with the effectiveness of paid-for coding assistance in March 2026. It's going to be very compelling if it closes the gap effectiveness-wise.

Show threadLance R. Vick1d ago

@hopeless @nabeards I mean I went from having almost no idea how to use this tech at all, to having a self hosted setup in one evening. I would not worry about enclaves or SEV-SNP encryption or any of that for a personal home setup. That stuff is most relevant if you want to be able to trust a third party managed system.

Feel free to join #!ai:matrix.org if you want help at any point!

I am very motivated to help more people learn how to have digital sovereignty and privacy with this tech.

Show threadLance R. Vick1d ago
@hopeless @nabeards I work on some of the most demanding projects that exist, including entire operating systems, with self hosted LLMs, in March 2026, and am convinced anyone else modestly technical can too.
Show threadnabeards22h ago
@hopeless @lrvick not everyone is using AI for coding assistance. There are other features of AI to be used and shared with multiple users, hence the need for SEP/Confidential Compute.
Show threaddegenerating degenerate1d ago

@lrvick ... which means what for code quality competitiveness? It's like claiming a faster CPU will solve every problem in a non cpu-bound domain.

Using the model directly means it typically can't produce a usable diff one out of two times, and lacks a framework to coordinate its investigations into bigger problems, which are solved problems in paid-for land.

Show threadLance R. Vick1d ago

@hopeless Well sure, I tend to use tools like opencode in an untrusted virtual machine, and playing with making my own tooling because I hate NPM the dependency hell of opencode, but I do really like the functionality and UX.

I have found it to be incredibly useful for research and debugging tasks, or writing test suites for existing code, or writing code to pass a test suite. Just gotta give it verifiable definitions of done and it will crank on problems all night while I sleep.

Show threadTuskun2d ago
@lrvick z.ai ?
Show threadLance R. Vick1d ago
@tuskun their free public anonymous tokens are great for experimenting, but I would not rely on any third party to do my job personally.
Show threadEli1d ago
@lrvick Qwen3-Coder:30b works on my laptop until I need to use the RAM for anything else.