Open sourcing the AI ecosystem ft. Arthur Mensch of Mistral AI and Matt Miller
HTML-код
- Опубликовано: 11 июл 2024
- Arthur Mensch, founder of Mistral AI, speaks with Matt Miller at Sequoia Capital's AI Ascent about his mission to bring AI to all developers, pushing for more open platforms and spreading the adoption of AI, as well as the balancing open source efforts while pursuing commercial opportunities.
Amazing talk :) Excited to see Mistral in action in Taskade's next Multi-Agent update! 🌈
Bro compare avg salary vs avg home cost in USA 20 years ago with current situation. What do you see?
So much work being done right now. Jump in and swim or be left behind.
Great interview
Wow, Arthur Mensch is brilliant 💡
Releasing the weights and the necessary code for inference is not open source in the traditional sense, not sure why they are using that term.
Because they get the benefits of saying that it is open source with out being open source
very interesting and great humble personality
Open weights isn’t open source unless they provide full access to their training set and source code. In all respect to the capabilities of Mistral’s models, it is an extreme stretch to call company that’s dropping torrents of weight binary, an OPEN SOURCE. What’s the benefit of this abuse of semantics? Marketing tactic?
It's even worse, the illusion that they're open source may give them leniency against future regulation of AI, and even tax breaks. I consider this a dangerous fraudulent precedent.
Open weights is by far the most important part.
It allows people to fine tune the model for their own applications (commercial or not), and allows anyone to run it locally rather than have to pay for a handful of private models.
Open source would only benefit major researchers that can afford to train similarly sized LLMs
@@AhmedKachkach
I suppose that in the ideal case, open source would include open weights. Furthermore, fine-tuning with the weights only isn't always the best methodology. The model can 'forget' (I use that term loosely), ideally what you want to do is have a mix of old data that the model was initially trained on + new data to ensure the model doesn't lose anything previously learnt etc.
@@quantum_encrypt Correct, you can mess up Mistral with just ~10 epochs of fine-tuning on a single example. It will "transfer" into examples not in the training set in an undesired manner. As a concrete case, consider training on "Q: What is your favorite fruit? A: Tomato." After a few rounds of training it will also start replying that 1. The favorite ANIMAL/PERSON/whatever is tomato and 2. The least liked fruit is tomato. It will also kill the ability to provide longer answers to similar questions. Etc. This highlights how having an unbalanced training set destroys the model. Rather than just incorporate the new information, as some may naively think, it also overwrites and erases the old information along.
@@clray123 Thank you! Great example 😊
Wow, Grok is too large for them to deploy. Interesting how large is Au Large then.
that was cool 🎶
Moderator need to remember the best questions can come from.the audience. Open up at half point
he is high? ) Mistral is best .
No, he is French.
@@VJ-lt9uk that's cool. I'm EU the same )) I love Mistral 4x70
Google deep mind should investigate on this guy.
Has that microphone been dropped on its head?
This French dude is hot... now that Elon is old, decrepit and looks more and more like his dad.... this Arthur is the new young musk.