This is my work Claude Opus leaked

[859963392]

All of this I have been working on for a substantial amount of time. Opus has literally word for word at times output everything I worked on. Even down to the repack of the weights. The entirety of it is in my conversation with Claude, parts are shown on ChatGPT and Gemini. All time stamped going way back into 2025. I cannot blame the developer. But just know I worked on this every day for months, fighting through hallucinations, especially at points where to complete the project would have, by the admission of the models I was working with, been problematic for those who stand to lose money from the models they serve online. This likely will be a long process between me and Anthropic, and I am just one single individual without any backing.

I would suggest never to believe that what you share with these models is safe, even with training off. Sharing your IP cannot reliably be determined as secure. It was the whole reason I began building this.

[chid]

do you have any proof of this?

[859963392]

The only way to compete for someone like myself is a lot of hours. All day, every day. Wake up and this is the first thing I do. Fall asleep with my Mac in bed, every night. There has been no life outside of what I do. So to see this, and to read someone ask me if I have proof, while understandable (I would ask the same) brings about a feeling which isn't something I can easily describe. I basically left my day job to focus on my projects, with this being the main one. I got stalled by ChatGPT / Codex, because when you're working on something which could provide a path outside of subscribing to models the big labs monetise, the models will literally hallucinate, hedge, throw up barriers. I didn't give a paper and ask for iterative testig. I built everything from first principles, which is why wht I have is a lot more advanced, with this repo synthesising the underlying logic and repackaging it. Even "trust the OS" is a direct line from the conversation. Claude didn't work that out, I worked out everything and told Claude (who constantly pushed back and said it would not work, to download Qwen2.5 7B quantize it, run it on llama.cpp etc).

[859963392]

I tried to upload a screenshot which failed. It reflects my weird journey as an autodidact. I was customising tokenizers before I could even setup a git. Something as simple as uploading a screenshot to a comment on GitHub, I have to check unless it is immediately apparent how to do so. The bottom left where it says "paste, drop, or click to add files" just adds a written line saying "uploading". This is not this developers responsibility and they hold no blame. I am only commenting because this really did happen, I broke myself inside and out completely on my work, and these labs are using models in ways people aren't aware of and it is just so wrong. There's no one to complain to, nothing that can be done about it, and we should all know what they are doing. Much of the capabilities and the constraints aren't made public. Constraints which lead you to part with money in the belief of an honest service, being unaware of the fact that something completely legal and ethical, is a hidden constraint the model cannot complete, without you or I being informed. From my position, there is very little I could do or say as it would just be drowned out in the noise and dismissed. Which makes me feel powerless to do anything about it. Anyway, the screenshots, documentation etc, everything. All available should anyone ask, despite the time sink that comes out of having to sift through everything and blank out IP.

[fiveangle]

I think this must be the 2bit quant