Open source vs closed LLM: Regulate only Open Source LLM, AI?

[@sebastiansosnowski3859]

In my view, the recent tour by Sam Altman where he was speaking of AI regulations with governments is an attempt at regulatory capture to widen the moat they have on open source, If there is a limit on max compute that can be owned or even operated by an organization or an individual It'd be only to take away people's ability to compete with government and corporate entities. I'm very much against it. The most scary scenario for me is when the AI models are black boxes that only government and corporations can train and deploy and the same thing done by an individual is penalized or restricted to the point where no real research and development can be done by individuals in the open source space. This is the most dystopian outcome IMO

[@sebastiansosnowski3859]

@dulles.gehlen It pretty much would destroy our ability to do so in the open. When you throw in a requirement of having good OPSEC to do any real work on AI, by definition you'll be unable to reap the benefits of collaborative open source. More people are afraid of the consequences of going against the law or even consider it just than people who would disregard it on principle.

[@Sydra.]

The best way to minimize harm globally is to make AI open and unregulated as possible. Regulated AI just makes YOU and only YOU defenceless against big tech and governments who will use their AI against you.

[@Pure_Science_and_Technology]

Who will pay the bill of training? A public company that gets its funding from the public? I’d donate an A100 80GB if there was such a initiative.

[@zackbreckenridge3213]

This video was not at all what I expected and was very good. It’s nice to hear your thought process as someone sitting in Europe pondering these issues. As an American developer, I can now more easily envision the calculations the rest of the world will make should our government choose to “regulate” AI. Also, I do not currently have experience with h2oai but you’ve inspired me to look.

[@jmirodg7094]

You might want to have a look at the Laion petition initiative to support opensource models

[@spoonikle]

the internet needs a redesign. LLM’s make big data science a breeze - just add compute. I bet the feds have trained models to track everything now and may even be able to generate possible password/secret data or even capture passwords and credentials via sms/email and automatically pair that with profiles. With how much data they have collected, I think they can make targeted hacks trivial.

[@Leon-uz1rz]

Can you make a video about quantization/gunk and hosting LLMs as efficiently as possible?

[@jamesjonnes]

A good way to prevent bad use of LLMs (and other crimes) would be to focus on creating and facilitating LLMs such as GPT-4 for therapy to people with problems in their lives. A bad approach is to do the opposite and nerf GPT-4 while large companies continue to train ever more powerful LLMs and preventing the creation of new companies due to lack of access to open-source LLMs. Once again in human history, centralization of power creates more problems than solves them.

[@joser100]

just in case... Bard can be easily accessed from anywhere by just using a VPN (Opera browser has VPN as a native feature very easy to use)

[@eraznafre]

Thanks for the content ❤

[@jayhu6075]

I personally believe what you all mention Opensource is good for the customer and for developer that will create new things and build specific models for the people.

[@SirajFlorida]

I'm just so surprised that lmsys and others are still training Llama traditional when OpenLlama has trained a Llama model without the constraints of Meta Llama.

[@blablabic2024]

Meta is financing all these offshoots of its model and that's why we've got inundation of LLaMA models... they are essentially trying to corner the AI market. As for Apple M1 Ultra ... too expensive for what's it giving. You will essentially be locked in for everything. No expandable RAM, VERY expensive NVME, pathetic GPU compute (for the price), pathetic CPU compute (for the price). Only plus is the very nice looking interior... Worst part is that everyone will go via integrated systems where RAM will be baked into APU combo, and pretty much entire PC board will be baked into one SoC. I mean, I like the idea and I don't like the idea at the same time. As for limiting LLMs ... how?! Models are already in the wild.

[@NeuroScientician]

Bard is being worked on and it keeps getting better. How about the Apple machine? Isn't that the most cost-efficient kit for a solo dev/enthusiast?

[@PatrickDunca]

How healthy is open source if the only way to run it is via the AWS/Meta/Microsoft gang, or the smaller startups now claiming billion dollar valuations? Whenever I consider these options for exporting my data to for operation I can’t help but expect that, even if I’m paying them, they are using small operators like me to cherry pick the best ideas for them to capitalise on with their greater resources. For this reason I think open source needs computing at the edges, not in the data centres, to be healthy. I’m currently considering building a 4x 4090 workstation for my house. Power and heat and noise are all major concerns. Models keep getting bigger so I figure that 96GB of VRAM gives me enough room for the Falcons et al for the next wee while, without everything being quantised to 4 bit. As much as it is berated, the home friendly version of this is that 192GB M2 Mac Studio. But no CUDA. And it’s much slower. And no upgrades. And even with the amazing work by GGML, it’s probably still some ways behind operationally. E.G. I am yet to see someone show a model that uses both the GPU and ANE to its fullest simultaneously. So despite finally being ready to use an open source OS exclusively, I’m back looking at Apple. Open source is the only future I want to invest in long term. I think computing at the edge is vital for its health. I do wonder if a SETI style project for training open source models is possible? Could training that is usually done across massive connected GPUs be chunked up up to fit on 8GB+ VRAM home machines? I’d happily donate 16 hours of compute time per day to training the next version of open source model I’m using.

[@PatrickDunca]

Actually if someone has a useful perspective on running locally based home friendly LLMs then I would really appreciate this.

[@andrewdang-u5h]

can i get the sources? im making a presentation for work ... love your vids!

[@andrewdang-u5h]

also contact email ?

[@henkhbit5748]

Ofcourse the closed source LLM has money to lobby and fear that open source model become better and better. Especially open source should not be regulated. If u can run your LLM locally its the best to protect your personal data. Can we imagine that the progress of Linux is regulated? Btw: mtb30b has also a 4 bit version