I'm quite fascinated by the potentials of AI as you probably noticed and have been exploring the potential for AI to simulate or acquire conscious experience.
oilprice.com
www.newstalk.com
Matthew Berman just released a yt video explaining how Claude Sonnet 'thinks'. It's much more complicated than most of us imagined how it does 'mental math' or speak a dozen languages or reasoning. Anthropic released a couple of papers recently on the subject. Their own researchers build a kind of digital microscope to detect patterns in their models. They reckon that we will need AI to understand how AIs actually work.
yes - I am interested like you in the parallels between the way we learn and the way AI 'learns'. I put that word in quotes because the two phenomena are different, but have similarities too. The debate about copyright seems on the writer's side to assume that their novels / articles / papers have been plagiarised - copied, stored and poured out by AI. It is nothing like that at all. The training of models involved the text being scraped into the machine, but the text is not stored and NEVER will it be re-produced by the machine. Only the relationships between words and concepts is absorbed and combined with huge amounts of other data. The AI may have 'read' a book, but it only read it and its content became part of the 'knowledge' of the machine, pretty much like your own reading of a book. I don't know about you, but I would have a hard time reading even a paragraph and reciting it back to you word for word. I could certainly tell you what it was about, but I could not recite it without a lot of re-reading and rote learning. When I attempt to tell you what the piece was about, I am actually combining and integrating the new knowledge with my past learning and experience. In the same way, AI's combine their training content and produce output word by word by word, on the basis of the giant vector database inside them which represents how words and concepts fit together. You can see it happening word by word if you try and run local AI models (ie your own ) on your phone. Because the hardware is really over stretched when you try that on a phone, everything slows right down and you see the stutter between words. It is actually computing what the next word should be, with no idea until it finds the next most likely word to print.
When people don't like you, they tend to say bad things about you. The side that owns the press and the mainstream media can write hit pieces on the other side all day long, which is a tactic that's been used for years. These stories are then laundered around using their own biased fact checkers in an attempt to get them established as some sort of truth. Having established in lore that somebody is a bad person, they can then use any authority that they have control of to go through that person's dealings and affairs with a fine tooth comb until they find something to target them with. Most politicians have done something that can be targeted, like fiddled their expenses, shouted at their staff, treated someone badly, flirted with a pretty girl, etc.
AI evolves so quickly that it is hard to keep abreast with its progress. In the earlier generations up to about 6 months ago, models use the temperature setting as random seed number for their transformer. Now with reasoning, the same temperature setting may not reproduce the earlier reply. What Claude Sonnet does is to combine the knowledge that it gets from various languages into its own world where languages are replaced by 'concepts' like large, small, etc. The papers I mentioned gave the example of how Claude Sonnet writes poetry like Haiku. The model run several processes concurrently and some combinations come out on top, similarly to the way we play chess for example, you play a lot of what ifs but your minds eye can see which resulting position is best. When you ask it to show reasoning, it has then to translate its thinking into words.
www.bbc.co.uk
All of these transformer models do that - run many processes in parallel and understand meaning by working out which processes come out with the highest probability of matching with 'learned' material. The same happens when the model is working out the output to send to the user. Parallel processing is one of the key enablers of AI. The big leaps in progress flowed from a paper written in 2017 by Ashish Vaswamy at Google and others.
www.bbc.co.uk
I misssed this remark about 'temperature' earlier Tony. The temperature setting can be varied. On the LLM models that I have downloaded and am running on a pc here in the house, you can adjust that parameter and others.
www.bbc.co.uk
I wonder how many countries would ring Trump up and offer him concessions.oh well 25% import fees for all uk cars to usa from today
You can't assume because a company has a D&B number, it's for profit. D&B pre-register these companies before they ask, if they ask at all. You don't need to ask them. It's a pretty useless registration anyway, like many free trade directories.
