Thanks Bastian. Couple of questions - if i use groqcloud , will it amount to violating data privacy ? I can not afford to deploy vicuna or llama in private cloud . But i want to ensure data privacy concerns of users . How to fix this problem ? - when are you going to conduct the seminar . I am looking forward to it . Very informative session. Thanks Dharani
That's an interesting question indeed. I discussed this with my LinkedIn network. See the discussion thread here: www.linkedin.com/posts/bastianbest_question-in-my-youtube-comments-do-you-activity-7233699713638662144-jHhN?
Hi Bastian. Many thanks for the video - you inspired me to see what I could squeeze out of GTP4 with minimum prompting - only using its internal knowledge of the patent process - but allowing it to fill in some details from the internet. It generated what looked like a patent - but of course, far short of what we would expect from a professional. However, I will push it further with some specific prompting and exposure to examples. It is interesting to ask it about the weaknesses of its own draft - and what would be needed to improve it. It can then ask you questions to fill in the gaps..... Of course, we can't use these models due to disclosure. Is there a reason to choose the Vicuna 13B model? Do you know if any other models may be more aligned to the task - or even pre-trained on patent law and practices? (I too use LM Studio - and haven't seen anything compatible)
Actually, I picked Vicuna-13B only because I read a blog post saying it's almost as good as GPT-4. I've just tested some other open-source LLMs here: ruclips.net/video/yMTbJdxwjUw/видео.html. Currently, I'm not aware of any LLM that would be specifically pre-trained on patent knowledge. That's why I'm going to build one myself in 2024. Stay tuned :) I'll answer the other questions in your comment in an upcoming video.
@@beyondbestpractice Hi Bastian. I was playing today with unaligned models like Dolphin 2 1 Mistral 7B - The results were much better, even with simple prompts (as it is not constrained by the normal guardrails). I am not a patent attorney, so would find it fascinating to see a critique of the output of these models by someone trained in the arts!
It would be nice to draft a PPA on USPTO. Some minor adjustment would be required. Not too much for a real patent, but a diy PPA, I will give it a go for sure.
My prompt library is currently offline for an overhaul. But I've just launched the first one in a series of online courses on gen AI fpr patent drafting. Maybe that's helpful: powerclaim.io/course-overview/
@@beyondbestpractice concerning a beautiful piece of fine art the US government has ruled it can't be copyrighted because it's too much “machine” and not enough “human.” An award-winning piece of AI art cannot be copyrighted, the US Copyright Office has ruled. Following this line of reasoning if an ai program writes a patent a trademark a novel etc would it not be considered in the same light, the machine did it not the human, I’d be afraid every patent trademark and copyright created by ai might be challenged.
But patent attorneys say it sound very 'patenty' but contains a lot of nonsense that would not make them pass as a patent. and the claims are weak to protect the patent or help them pass.
Well done! This was very helpful and I can’t thank you enough for putting this together!
Thanks so much!
Thanks Bastian. This i some grerat resource. May I ask what HW are you using? My i7 laptop is struggling despite 16GB of ram and GPU offload of 20
Yes, these models are a bit resource-intensive. I'm using a MacBook Pro 16", 2021 Version with M1 Max chip and 32 GB of RAM.
Thanks Bastian. Couple of questions
- if i use groqcloud , will it amount to violating data privacy ? I can not afford to deploy vicuna or llama in private cloud . But i want to ensure data privacy concerns of users . How to fix this problem ?
- when are you going to conduct the seminar . I am looking forward to it . Very informative session. Thanks
Dharani
Thanks for the nice words! I haven't tried groqcloud, so can't comment on that. Depends on their privacy policy, I guess...
I'm currently revising the seminar contents. Expect an update soon!
To the USPTO do you have to disclose to the patent office the patent application was assisted by AI?
I was wondering this exact question
That's an interesting question indeed. I discussed this with my LinkedIn network. See the discussion thread here: www.linkedin.com/posts/bastianbest_question-in-my-youtube-comments-do-you-activity-7233699713638662144-jHhN?
Thanks for demonstrative insights.😮
You're welcome!
Hi Bastian. Many thanks for the video - you inspired me to see what I could squeeze out of GTP4 with minimum prompting - only using its internal knowledge of the patent process - but allowing it to fill in some details from the internet. It generated what looked like a patent - but of course, far short of what we would expect from a professional. However, I will push it further with some specific prompting and exposure to examples. It is interesting to ask it about the weaknesses of its own draft - and what would be needed to improve it. It can then ask you questions to fill in the gaps.....
Of course, we can't use these models due to disclosure. Is there a reason to choose the Vicuna 13B model? Do you know if any other models may be more aligned to the task - or even pre-trained on patent law and practices? (I too use LM Studio - and haven't seen anything compatible)
Actually, I picked Vicuna-13B only because I read a blog post saying it's almost as good as GPT-4. I've just tested some other open-source LLMs here: ruclips.net/video/yMTbJdxwjUw/видео.html. Currently, I'm not aware of any LLM that would be specifically pre-trained on patent knowledge. That's why I'm going to build one myself in 2024. Stay tuned :) I'll answer the other questions in your comment in an upcoming video.
@@beyondbestpractice Hi Bastian. I was playing today with unaligned models like Dolphin 2 1 Mistral 7B - The results were much better, even with simple prompts (as it is not constrained by the normal guardrails). I am not a patent attorney, so would find it fascinating to see a critique of the output of these models by someone trained in the arts!
Very impressive and educational 👍
Thank you!
It would be nice to draft a PPA on USPTO. Some minor adjustment would be required.
Not too much for a real patent, but a diy PPA, I will give it a go for sure.
Good luck!
link to the prompts?
My prompt library is currently offline for an overhaul. But I've just launched the first one in a series of online courses on gen AI fpr patent drafting. Maybe that's helpful: powerclaim.io/course-overview/
Would i create a challage for patent agents job in the future?
Absolutely. GenAI will fundamentally change any job that involves understanding language and producing text, and patent work is no exception.
If ai created content isn’t protectable under patent trademark and copyright law wouldn’t this pose the same problem?
What do you mean exactly?
@@beyondbestpractice concerning a beautiful piece of fine art the US government has ruled it can't be copyrighted because it's too much “machine” and not enough “human.” An award-winning piece of AI art cannot be copyrighted, the US Copyright Office has ruled. Following this line of reasoning if an ai program writes a patent a trademark a novel etc would it not be considered in the same light, the machine did it not the human, I’d be afraid every patent trademark and copyright created by ai might be challenged.
But patent attorneys say it sound very 'patenty' but contains a lot of nonsense that would not make them pass as a patent. and the claims are weak to protect the patent or help them pass.
I’d say that too if I was concerned it was going to replace me 😅
I guess they either used a poor model, or poor prompting (more likely)