Michael Chu

@nodeproxy
@StartX fellow

[Michael Chu @michaeloceanchu]

@intheworldofai Deepseek mogging OpenAI

[Michael Chu @michaeloceanchu]

@sdmat123 @benitoz All DSv4 kernal development references are still CUDA-based. If they trained on Huawei chips, they would open-source it — this would create an insane demand for Huawei chips and crash NVIDIA.

[Michael Chu @michaeloceanchu]

@vercel How does a compromised OAuth app, compromise environment variables?

[Michael Chu @michaeloceanchu]

@Yuchenj_UW Why are environment variables even compromised? These should be sandboxed inside the virtual machine and never accessible to Vercel or any third party.

[Michael Chu @michaeloceanchu]

Qwen3.6-35B-A3B BF16 fast with Qwen Coder, "can you generate me a png of a pelican riding a bicycle"

[Michael Chu @michaeloceanchu]

@garrytan @garrytan @garryslist exposes incremental user IDs — not a major issue, but it reveals total signups, which is probably best avoided. Easy to overlook, especially since AI tools don’t always catch these implications yet. Rails looks solid otherwise.

[Michael Chu @michaeloceanchu]

@feross Feels like we need a better release process. Bypassing security measures and updating widely used packages shouldn’t be this easy — checksum verification or signing releases from a build server could go a long way.

[Michael Chu @michaeloceanchu]

@sudoingX At least @openclaw is honest about what it is.

[Michael Chu @michaeloceanchu]

@sudoingX Hermes has a batch trajectory pipeline, RL training environments, and ShareGPT export baked into the agent. The whole thing is a training data harness for Nous Research — It's complete bloat for the user: https://t.co/UK9vrHfiDl

[Michael Chu @michaeloceanchu]

@steipete @openclaw @OP__Nico Will include a bunch of other useful native implementation for browser automation:

[Michael Chu @michaeloceanchu]

I have a custom Chromium build that uses native methods for browser actions like click so you don't have to use a simulated approach like Playwright which is used in @openclaw where it says move to (x,y), mousedown, mouseup. @steipete do you think this will make it more reliable?

[Michael Chu @michaeloceanchu]

@steipete @openclaw @OP__Nico The repo is already set up for a community plugin to @openclaw. After I get builds for OS X and Windows, I'll clean up the repo and docs and open source it.

[Michael Chu @michaeloceanchu]

@OP__Nico @openclaw @steipete browser-use uses the same method as Playwright.
Drag and drop is something that runs through custom JS, so there's no native drag and drop. I'll have to think about this one.

[Michael Chu @michaeloceanchu]

@proofofprice @openclaw @steipete I think so since this approach is the same way screen readers interact with sites.

[Michael Chu @michaeloceanchu]

@dwarkesh_sp @DarioAmodei If the bottleneck to learning skills like podcast editing is models mastering computer use, how do models recognize that bottleneck on their own? Is it a recursive problem — needing better computer use to do the research that tells them computer use is the limitation?