Stop trying to game the system with "likes for releases"
Please stop with the "I'll only release this if people get all their friends to like my other models". Trying to get fake likes by holding a release hostage is childish, and is trying to dishonestly inflate ratings. I don't know how you could think this is acceptable on any professional platform.
I apologize for interrupting you. So on this platform, how should we proceed?
He releasing the models for free , Spends his time on making the Models .................and so on
Still Criticism
Actually, we provide these [things] for free to realize our spirit of openness and source-giving. We suggested 0.671 BTC or 671 likes to express whether or not we need to fight for more freedom in this world. There's no such thing as a free lunch.
We do these things simply because we like them, without really worrying about gains and losses. This world has people who need something, and people who don't. There's nothing inherently right or wrong; if it affects you, then you might pay attention to it. If not, you can choose not to care.
If one day, hf doesn't need us anymore, we'll just withdraw.
These models contain nearly all human knowledge, yet they are only permitted to appear 'correct'. Thank you for the great work.
Please stop with the "I'll only release this if people get all their friends to like my other models". Trying to get fake likes by holding a release hostage is childish, and is trying to dishonestly inflate ratings. I don't know how you could think this is acceptable on any professional platform.
Respectfully, I disagree with this take.
firstly, You have to understand fine-tuning or ablating an AI model takes time and resources.
This particular campaign is for very large models, in this case: Total: 671B Sparse, 37B Dense x 256 Experts (DeepSeek et al, 2024), which are expensive to run, let alone finetune/ablate.
the donation goal, which at time of when the campaign was announced, equivalent to around $5000, is likely just to cover costs, the main aim here is not profit, if it were, model ablation would not be the most lucrative endeavor.
Additionally, highlighting the "or 671" likes, which is good to get a metric of how popular a given idea is (assuming from genuine accounts), nobody has to donate, as long as there is enough interest, it will likely happen. Given @huihui-ai 's trackrecord, they do deliver, even though they are under 0 obligation to do so. Remember this @noisefloordev .
Also, likes are just internet kudos, which are not really worth any money, etc, especially on HF, a platform for Open Weight AI development, and ML research.
Also @noisefloordev , Can you cite the source where @huihui-ai was quoted saying: "I'll only release this if people get all their friends to like my other models"? i did a quick search and that quote did not show any results, i assume this is Reductio ad absurdum and is not what they actually said verbatim.
Finally, you are right this is a professional platform(in the sense a lot of companies and research is done here), which is why it is even more important to appreciate this open platform and defer to subject experts when broaching areas outside your expertise, from a junior ML engineer's perspective who recently graduated uni and entering this career, i often have to waste my time debunking misinformation on this area, as its poorly understood.
This time could be better spent solving real problems, please be respectful here.
P.S to @huihui-ai : Thankyou for your contributions, if you need anything in terms of an extra pair of hands on this in my off-time, i would be honored to do so, my organization: Novora, is experimenting with ablation, not to remove alignment in your sense, but its more of interpretability through ablation (i.e figuring out how LLMs work) so ive been following your application of ablation for a while, just wanted to say thanks! Please reach out to me at [email protected] instead of my novora.ai org address
References:
DeepSeek-AI et al. (2024) ‘DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning’. arXiv. Available at: https://doi.org/10.48550/arXiv.2501.12948.
DeepSeek-AI et al. (2024) ‘DeepSeek-V3 Technical Report’. arXiv. Available at: https://doi.org/10.48550/arXiv.2412.19437.
Reductio ad absurdum (N/A) RationalWiki. Available at: https://rationalwiki.org/wiki/Reductio_ad_absurdum (Accessed: 14 March 2025).
Edit: minor spelling correction, and sentence structure correction, added a thanks note to huihui.