AI Top-of-Mind for Jan 3
Happy New Year, and second to last long one due to skipping days… AI never sleeps
Diving into edge AI and LLMs, Ignacio de Gregorio reports on what could be the most important AI development in 2024 that could impact the average consumer. He details Apple’s innovations in Flash-memory LMMs. As a review, running models at the edge requires memory and processing. Thus, running LMMs in the cloud has been the path to-date. To-date, we’ve seen SLM announcements by the likes of Google due to smartphone RAM limitations, but the ability to run part of the model in flash totally changes the game. It paves the way for edge LMMs and all the benefits thereof. Read the article… fascinating.
On a more technical front, not that the above wasn’t ;), a bit on text embeddings as the basis for NLP (Natural Language Processing), and therefore as a foundation for AI text queries. The basis of text embedding is the translations of words into a set of numbers, ‘vectors,’ that may be arbitrarily long and that capture essential characteristics of the word and their relative sameness to others. The diagram below is an example of this mapping.
The same concept extends to sentences but this also requires awareness of the positions of the different words. As you’d expect, the larger the corpus of training data, as well as the quality and diversity, the better the potential model, and the IP of the different AI companies is to best create the associations between the training data via embedding scoring. Both Cohere and Deepset.ai provide some great background on the above concepts. The Cohere site also dives into multi-lingual support and Retrieval Augmented Generation (RAG), the basis of integrating external data into the LMM. If you want to learn more about RAGs, ‘Towards Data Science’ offers a great primer.
As a more practical case, I leveraged the Cohere dashboard to enter four sentences, with the initial output below. Note the ‘distance’ between Los Gatos and Saratoga, both cities in California, but not including this as part of the prompt. Within the system, they obviously have differences, but I can’t tell where. A Google search will bring up Saratoga, CA as the first result.
I then added CA to both, and the distances are as expected.
On the corporate front, ‘The Information’ reports that Elon Musk’s xAI has incorporated as a benefit corporation, equivalent to Anthropic and analogous to OpenAI (but not the same structure which is a non-profit with a subservient for-profit side) The concept of a benefit corporation is that the company may take initiatives that don’t automatically prioritize profit, as is the case with your typical for-profit corporation where the board could get sued or replaced by activist investors.
In recent weeks I’ve covered a host of 2023 year-in-reviews, so time to look forward to 2024. There are multiple reports on 2024 being a crucial year for the future of what we consider democracies, across the globe. We’re in a place and time where it becomes increasingly impossible to tell the difference between falsehoods and truth, and ‘deepfakes’ will be common as we enter our election year in the US. From a thoughtful (and downright scary) article at ‘The Hill’:
· “2024 will be an AI election, much the way that 2016 or 2020 was a social media election,” said Ethan Bueno de Mesquita, interim dean at the University of Chicago Harris School of Public Policy. “We will all be learning as a society about the ways in which this is changing our politics.”
· Experts are sounding alarms that AI chatbots could generate misleading information for voters if they use it to get info on ballots, calendars or polling places — and also that AI could be used more nefariously, to create and disseminate misinformation and disinformation against certain candidates or issues.
· “I think it could get pretty dark,” said Lisa Bryant, chair of the Department of Political Science at California State University, Fresno and an expert with MIT’s Election lab.
The increasing use of AI is even top-of-mind within the US Supreme Court, where Justice Roberts in his year-end report warned against the indiscriminate use. From the ‘CNN’ article:
· “AI obviously has great potential to dramatically increase access to key information for lawyers and non-lawyers alike. But just as obviously it risks invading privacy interests and dehumanizing the law,” Roberts wrote.
· But he appeared wary toward a wholesale adoption of the technology in courts, noting that “studies show a persistent public perception of a ‘human-AI fairness gap,’ reflecting the view that human adjudications, for all of their flaws, are fairer than whatever the machine spits out.”
· Roberts also pointed to a flaw in one prominent AI tool that he said caused “lawyers using the application to submit briefs with citations to non-existent cases. (Always a bad idea.).”
· The chief justice said he welcomed coming efforts by the Judicial Conference — the policy-making body for the federal courts — to examine AI’s “proper uses in litigation.”
Where the Wild Things Are…. A new take on bedtime stories, as I remember sitting in bed with my daughters trying to come up with new takes on their favorite themes, and think back to one of my favorite childhood books. Gen AI to the rescue, but do we lose something in translation? Is there less of a creative bond. The article by Enrique Dans looks at options including prompting asking for a specific style, or uploading a full work and whether it is currently under copyright. These issues will also come to the forefront as some of our most treasured cartoon characters, including Poo and Mickey, have either lost copyright protection or will lose it in the coming year. If you have access to ‘The New York Times,’ Brookes Barnes offers his perspective.
