Re: Interest in a zoom meeting on safe AGI?

Hmm, what we can do know is a lot more sophisticated than you seem to imply.  LLMs can be trained to recognise harmful content and to describe what’s wrong. Disinformation is harder, as it involves fact checking. This can be addressed by connecting the LLM to external services, e.g. using retrieval augmented generation (RAG). Images, audio and video are harder to deal with, but this is more a matter of effort than of technical barriers.

A further challenge is privacy. This is where it is better to be able to execute the checks locally without having to send personally sensitive information to the cloud.  LLMs can be distilled to run on local systems, and I would see this as an important area of research.  

The UK government recently sought to require social media companies to prevent harmful content from being seen by children, but there was a lot of push back by the companies and privacy folks. This requires a) governments to impose regulatory burdens on social media companies, b) to encourage research on privacy preserving solutions and c) to educate the public to understand how their privacy and safety is to be ensured.

> On 8 Nov 2023, at 10:50, Timothy Holborn <timothy.holborn@gmail.com> wrote:
> 
> The solutions sought to be advanced back then, was far more about semantic labelling, overall..   
> 
> Semantics on the web ATM, often have alot of problems.  It's been a difficult problem to address. 
> ...
> 
> On Wed, 8 Nov 2023, 8:29 pm Dave Raggett, <dsr@w3.org <mailto:dsr@w3.org>> wrote:
>> Reverting to public-cogai only to avoid cross posting ...
>> 
>>> On 8 Nov 2023, at 10:13, Timothy Holborn <timothy.holborn@gmail.com <mailto:timothy.holborn@gmail.com>> wrote:
>>> 
>>> Re: fake news, I did this back in 2017
>> 
>> Current generative AI is now very much better than then, and can be designed to understand text, images and a variety of other media formats.  Training such a system to recognise disinformation and inflammatory content is non-trivial, and it will be expensive for social media companies to run this on all posts.
>> 
>> This is why the discussion should be focused on how to pressure governments to regulate to force social media companies to introduce and maintain such defences.
>> 
>>> 
>>> On Wed, 8 Nov 2023, 7:41 pm Dave Raggett, <dsr@w3.org <mailto:dsr@w3.org>> wrote:
>>>> Based upon the responses, I think we are better off sticking with email at least for now.
>>>> 
>>>> I am surprised that more attention hasn’t been given to applying AI to combat disinformation and inflammatory content on social media, which seems to be the biggest threat to society right now after climate change.  Social media companies probably need regulations imposed on them to make this work and those regulations will only happen if people make a fuss and lobby for them.

Dave Raggett <dsr@w3.org>