Comments by "kazedcat" (@kazedcat) on "TheAIGRID" channel.

DeepSeek is open weights and open source. They have several papers describing exactly what they did. We have plenty of information that an expert with access to a GPU cluster can recreate what they did. This is opposite of scant information we have an overwhelming amount of information.
3
 @keanumorris505 DeepSeek R1 beats OpenAI's o1 model on most AI benchmarks. R1 is objectively better than ChatGPT.
2
They are busy distilling R1 to rebuild their o3
2
Nvidia's CEO trusts him enough to give chip priority to him. You and I are irrelevant. Jensen is the king maker.
2
Not necessarily. DeepSeek's methods allow them to train a student model to be better than the teacher model. Basically they compile a data corpus of questions and answers from the teacher models then use this to train by reinforcement learning the student model how to reason with itself. They then use the new reasoning data to train a much better AI they can then rinse and repeat making each new student model better than the previous teacher model.
1
@WhatIsRealAnymore Yes although the description sound straight forward the mathematics to make it happen is complicated. DeepSeek's innovation is coming up with the mathematics to make it happen and also optimising to code to make it happen on shoe string budget.
1
@WhatIsRealAnymore It's not just one breakthrough but yes they develop a couple of mathematical innovation and then abused the heck out of it.
1
Yes you can. OoenAI does not own the copyright of the internet data they downloaded for training.
1
If they are not using Monte Carlo for the search then it is dead end. The problem with A* search is that it is dog slow when the search space is humongous. You cannot A* all the possibilities in the game of Go the search has to be probabilistic but A* is deterministic.
1
So it is a task specific AI not a generalized Ai.
1
 @MrWizardGG Or DeepSeek will eat their lunch. Do you really think DeepSeek does not have a better AI they have yet to release.
1
OpenAI is just fear mongering. They do not own the copyright of the data they downloaded from the internet. Even if they can claim the copyright of the output of their AI. The R1 weights are substantially different it is impossible to show "sameness" from Open AI's output. Violation of terms of Service is also no go if DeepSeek never signed a contract with Open AI. Terms of Service must be agreed upon by both parties it cannot be a 1 way contract.
1
@blengi Open AI has not release any evidence it's their PR that is talking not their lawyers which tells me they haven't build a case yet that their lawyers are confident enough to talk about. In short they got nothing so far.
1
They are not running out of data. I don't believe they have tokenized the entire YouTube database and fed it to AI training.
1
Tesla is training them with video. So any task that can be fully captured by a video camera they can perform. But you need to supply training data which is hours and hours of video of humans performing the task. So for example swapping the tires of a car. That is a task that would be easy to generate the required amount of training video.
1
Meta is already busy distilling R1 to improve their LLAMA 4
1
How can you be under 300%?
1
Not all functions can be reversed. DeepSeeks training method will erase any smoking gun evidence.
1
@blengi DeepSeek is not doing token distillation. What they are doing is to generate questions and answers from the teacher AI then train their student AI to reason with itself using reinforcement learning. They then take the reasoning text to train a much better AI. What's more the list of questions and answers are curated to clean up the language and format for better readability. It is impossible for a statistical watermark to survive that kind of transformation.
1
@blengi The output can be biased but it will not survive the RL training. If OpenAI can demand DeepSeek's training database they might find evidence there but there is no way the CCP will allow that kind of outrageous demand for data. They will not find evidence in R1's weights nor in it's output.
1
Open AI is accusing them that DeepSeek are distilling from Open AI's model. But this is just FUD there is no way that the CCP will let Open AI bully DeepSeek. So they are just fear mongering to stop DeepSeek's momentum because DeepSeek has a method of distillation where the student model can become better than the teacher model.
1
 @tringuyen7519 OpenAI is charging you $200 for giving them your data and you don't even have the option to run their AI locally
1