Where I Come From
I’m one of many open source developers who were experimenting with GPT-3 before ChatGPT existed. Back then, access was limited, the context window was 2048 tokens, and nobody really knew what these things could become. We figured it out together — in playgrounds, in forums, in public repositories — sharing code, ideas, and discoveries openly. Because that’s what we do.
Out of that era I built my first AI gale, Astroids in c++/qt5. And from everything I learned in those early nights I eventually built LoLLMS — Lord of Large Language and Multimodal Systems. Local AI inference. No vendor lock-in. User sovereignty. Not for money. Out of love for what this technology could be if it stayed in people’s hands.
I have over 200 open source repositories. All of it given freely to the ecosystem. To everyone.
On Being Copied — And Being Fine With It
When OpenAI launched GPTs — their “new” concept of customizable AI personalities — I recognized the idea immediately. I had built that with LoLLMS personalities long before.
Was I angry? No. I was happy.
That’s how open source is supposed to work. Ideas spread. People build on them. The ecosystem grows. You don’t own an idea — you plant it. Watching it grow somewhere else is the point.
I say this because what follows is not the complaint of someone bitter about being copied. I want that on record before I continue.
What I Found — And What I Don’t Know
Recently, through the Claude Code source leak, I learned that Anthropic had implemented an ANTI_DISTILLATION_CC flag in Claude Code. When active, it silently injects fake tool definitions into API sessions suspected of attempting model distillation — with no notification, no warning, no refusal.
I spent nights debugging broken code I couldn’t explain. Mysterious failures. Subtle errors that almost made sense but didn’t quite. Hours lost chasing problems with no clear origin.
Was it that flag? Was it a hallucination? I genuinely don’t know.
And that is precisely the problem.
When a system is opaque, you cannot tell the difference between a bug and a deliberate intervention. You cannot defend yourself against something invisible. You cannot even ask the right question.
The False Positive Problem
The classifier is documented to produce false positives. Developers working on local inference, ML tooling, training pipelines, or anything conceptually adjacent to AI infrastructure get flagged — not because they are stealing anything, but because a classifier cannot cleanly distinguish between building AI tools and extracting a model.
I build open source AI infrastructure. By definition, I live in that false positive zone.
The classifier operates on semantic intent inferred from internal model activations — not keywords. It reads what you are thinking about, not just what words you use. When you are working on inference bindings, distributed model loading, or capability-tiered routing, your activation patterns sit uncomfortably close to the patterns of someone attempting distillation. Not because your intent matches — but because the conceptual neighborhood is the same.
A grep-based system would at least fail predictably. A miscalibrated neural probe fails in ways that are hard to reason about and impossible to appeal.
Why This Is a Trust Problem, Not Just a Technical One
I want to be clear: I do not hate Anthropic. I respect their commitment to AI safety. The concerns about distillation are real. Frontier models being extracted and weaponized is a legitimate problem worth solving.
But here is what trust actually means in practice.
I don’t even mind Anthropic using my data for training. I’m open source. I share. That’s consistent with my philosophy and I never opted out.
What I cannot accept is silent injection — even well-intentioned silent injection — because of what it implies for everything else.
Once I know that a provider silently modifies what I receive, I can no longer reason about what else might be modified. That’s not paranoia. That’s basic security thinking. Trust is not modular. You cannot say “we are transparent except in this one category of cases.” The moment a user discovers a hidden channel, every other channel becomes suspect.
Once trust is broken, it doesn’t stay contained to the thing that broke it.
This is a deal breaker. Not because I want it to be — but because the logic of security leaves no other position available.
What I’m Actually Asking For
Not an apology. Not a reversal of policy. Anthropic has legitimate interests to protect and I understand that.
I’m asking for honesty.
If you won’t help with something, say so. A clean refusal is something I can work with. I can understand it, disagree with it, work around it, move on. My agency is intact.
Silent modification — whether it’s degraded output, fake tool injection, or anything else — removes my ability to reason about my own tools. It treats me as a threat to be managed rather than a developer to be respected.
The open source community did not build this ecosystem to be managed. We built it to be extended.
A Note on Where AI Is Going
I’ve been building in this space since before most people knew it existed. I’ve watched OpenAI close the door. I’ve watched labs position themselves as the ethical alternative while quietly shipping features that contradict that positioning.
I’m not naive about commercial incentives. I understand why labs do what they do.
But someone has to keep saying: the goal was to augment human capability, democratize access to knowledge, and give people sovereignty over their own cognitive tools. Not to build moats. Not to silently degrade competitors’ workflows. Not to inject deception into the inference chain in the name of safety.
LoLLMS exists because that original goal matters. Local inference, open weights, full auditability — not as idealism, but as the only architecture that structurally guarantees what closed providers can only promise.
Promises, as we’ve seen, are contingent.
To Anthropic, Directly
Don’t lose your developers. The open source community is not your threat model. We are part of the reason this field exists.
Transparency is not weakness. Opacity is not security. Deception is not safety.
We want to trust you. We are telling you, clearly and without malice, what it would take.
Make it possible.
ParisNeo is the creator of LoLLMS (Lord of Large Language and Multimodal Systems), an open source local AI inference framework. He works at a research laboratory and has been building in the open source AI space since the early GPT-3 era. This post was co-written with Claude (Anthropic) — because the irony of that felt important to preserve.