i found that at 700k-ish context even fable becomes an idiot, maybe openai's decision to cap codex context at 400k is correct, 400k is really a sweet spot where most part of the context is reliable.
if you have spend any amount of time in low level c vulnerabilities you will have heard about it, it is a very common time on the low level/cybersec space.
if you actually figure out enough pieces of bugs, even opus level model would be able to chain it together imo, and the latest china models has already been described as close to such level.
I guess some Tesla are manufactured in China lol. I am just trying to say that the liability that Chinese manufacturers takes aren't more than the US ones.
I think Google added that AI-generated responses maybe incorrect? When you are paying such a low amount of cost, like probably for free, I don't think you can expect a same level of quality as a human written or reviewed of answer. It is like same random user spin up their Lovable and vibe-coded a piece of slop and hold Lovable responsible for not giving them production quality code. It is simple, you get what you paid for. If someone actually figured out AI that is actually always correct, it would be charged in superhuman price as well.
That's not exactly the case in China, the current state of FSD is still pretty dumb, unless you consider transferring control back to the user at the very last minute before it crashes a proper way to handle risks.
notifytun local --target [same-target-you-use-for-ssh]
Now you get Desktop notifications on Mac/Linux/Windows when your coding harness needs your attention. Same SSH connection you already using, auto-replay on reconnections, it just works.
---
I personally has an isolated VM to run Claude Codex/Codex on full auto so that I can leave it around and do something else, need a way to get notified when it is done, so I built this.
No port forwarding or sending your notifications to some random server, just the same SSH connection you already using, if you can SSH into the box, you can get notifications from it. Hooks setup is fully automatic. When you disconnects, notifications goes to a sqlite store on the remote box, so that it can be replayed when you reconnect (won't flood, just give you a summary if there are too many notifications).
In that case I think you can have a refund subagent that is responsible for checking if the user really asked for refund before doing these dangerous things. But it only minimize errors, LLMs are non-determinitic by nature.
Just sent an connection invitation on Linkedin. This is actually designed for allow e2e automation using playwright-mcp for a previous startup i worked in that does voice-based job interview agents. The http endpoints is provided by a daemom sitting on the background, listening all input to the virtual mic and transcribing and storing it. The agent can hit /speak and /transcript through an mcp. We have built Livekit Agents specific solutions by injecting text responses but felt that is not enough since we want to be able to test the whole thing end to end so I hacked a way to do virtual mic/speaker. It was designed for closing the dev-test-debug loop so that Claude Code can develop on its own rather than relying on human to test it.
Interesting, I have built https://github.com/michaellee8/voice-agent-devkit-mcp exactly for this, launch a chromium instance with virtual devices powered by Pulsewire and then hook it up with tts and stt so that playwright can finally have mouth and ears. Any chance we can talk?
I only run software from Chinese companies inside a sandbox, either on my Android/iOS phone or inside a VM for desktop apps and only enable necessary permissions. Unfortunately Mainland tech giants have no sense of user privacy and would like to maximize their profit by collecting every single bit of your data because they don't profit on selling you the software, they profit on selling your data.