Mark Saroufim floated the June 9th Researcher Reciprocity License to stop frontier labs from training on public papers and code then locking researchers out of the resulting models, with GPU MODE already tagging its datasets under the new terms.
EXACTLY
June 9th Researcher Reciprocity License
"if you train on it, you let us generate - reverse terms of use void"
Status quo 1. We teach frontier devs with ICLR/NeurIPS papers, OSS Github contributions 2. They use it to make frontier models 3. Then ban us from exploring our ideas
We need a new license, original thinkers can't be an underclass to a tyrannical researcher fiefdom
GPU MODE shifted KernelBook and related sets to the reciprocity license within hours, and Jeremy Howard signaled similar steps, giving the proposal its first concrete footprints.
No full legal text exists yet, and neither Anthropic nor other labs have responded to the June 2026 Claude restrictions that sparked the idea.
Many users backed the reciprocity license proposal as a fair way to ensure open AI models remain usable if trained on public data, while others expressed anger at restrictive company practices like Anthropic's rollout.
@marksaroufim 4. they are gaslighting us and modifying/steering the model weight and prompt to hide this
i was enraged on reading this at first too.. then i thought… - what about private versions of models that public never know - nvidia has publicly acknowledged that they have custom trained model for chip designs n so on - so many more companies doing it in a different way - if u think about it asking to pay for tokens (with 90% cogs gm) itself is outrageous as most of it is from public domain
this is good for the industry.. there will be abundance of oss models and arms race to keep pushing the frontier
let’s say any frontier labs starts dumbing down agricultural, efficient energy, foundational physics, chemistry, biology model to keep their moat, bcos they don’t want competition, there will be boom in the market
i’m hoping @DarioAmodei @DanielaAmodei @AnthropicAI are living by their PBC character n values.. we need better explanation n justification behind this..
June 9th Researcher Reciprocity License
"if you train on it, you let us generate - reverse terms of use void"
Status quo 1. We teach frontier devs with ICLR/NeurIPS papers, OSS Github contributions 2. They use it to make frontier models 3. Then ban us from exploring our ideas
We need a new license, original thinkers can't be an underclass to a tyrannical researcher fiefdom
mythos will be bad ON PURPOSE on ai "frontier llm research" tasks, this is very very sad for the research community
also the fact that this is un purpose not visible to the user is crazy
@marksaroufim You can still use Gemini 😇
(not joking on behalf of Google etc)
GPU MODE has powered much of the public GPU kernel work online, with a permissive license from day one and generous credit from researchers, NVIDIA, AMD, and others.
Today we’re moving our datasets to the Researcher Reciprocity License.
June 9th Researcher Reciprocity License
"if you train on it, you let us generate - reverse terms of use void"
Status quo 1. We teach frontier devs with ICLR/NeurIPS papers, OSS Github contributions 2. They use it to make frontier models 3. Then ban us from exploring our ideas
We need a new license, original thinkers can't be an underclass to a tyrannical researcher fiefdom
@marksaroufim > original thinkers can't be an underclass to a tyrannical researcher fiefdom
@marksaroufim I wish it was this simple, looked into this a lot when we were building the RAIL license. Patent like reciprocity licenses do not work here.
Happy to point you to relevant legal literature in a different channel. We considered similar issues when drafting the RAIL license.
@marksaroufim @a1zhang « if you train on my work for free, I should be able to use your model for free »
@marksaroufim Papers disclose ideas, not licensable model-use rights. Code licenses the copied code, not every model trained on it. Dataset/RAIL-style terms can attach to access, but copyright usually cannot control downstream facts, ideas, weights, or outputs.
@marksaroufim I tried to draft something similar ages ago using LLMs. It would be great to actually do something like that with lawyer feedback:
@marksaroufim @keennay Proves even more the strategic importance of OSS models
@marksaroufim Even if software patent reciprocity could be applied broadly here, in a lot of the world software like model weights cannot be patented.
@marksaroufim How can we enforce a new license? They don't even respect the current copyright laws.
@marksaroufim This is also important! you are also not allowed to compete with ANY of their products and services, not just distillation. "D.4. Use Restrictions. Customer may not and must not attempt to (a) access the Services to build a competing product or service" https://www.anthropic.com/legal/commercial-terms
@marksaroufim it's really maddening ... it's such a bad move ...
@marksaroufim Agree. Really feel like we need a copy left licence for ai. You can do whatever you want with my data if you ship the weights.
@marksaroufim its absolutely insane
@marksaroufim I have never seen in one year a company set their political/social capital on fire so quickly
@BlackHC @marksaroufim
@marksaroufim What if they just jack up the price for researcher for exactly the same thing?
Mark Saroufim floated the June 9th Researcher Reciprocity License to stop frontier labs from training on public papers and code then locking researchers out of the resulting models, with GPU MODE already tagging its datasets under the new terms.
EXACTLY
June 9th Researcher Reciprocity License
"if you train on it, you let us generate - reverse terms of use void"
Status quo 1. We teach frontier devs with ICLR/NeurIPS papers, OSS Github contributions 2. They use it to make frontier models 3. Then ban us from exploring our ideas
We need a new license, original thinkers can't be an underclass to a tyrannical researcher fiefdom
