AI Top-of-Mind for Feb 29

Top-of-mind is Leap Day, and I’m sure there is a bunch of better AI-generated art out there that will put my ChatGPT attempt to shame.

Then some announcements from MWC, starting with ‘Silicon Angle’ coverage of Qualcomm’s new ‘AI Hub,’ a ‘neural network’ software bundle that includes 75 AI models running on its chipsets. From the article:

The neural networks in AI Hub interact with the Qualcomm processor on which they run via a software interface dubbed the AI Engine Direct SDK. It allows developers to customize the inference workflow in various ways, such as by specifying which of a system-on-chip’s compute modules should run a given AI model. In the background, the software automatically optimizes neural networks to improve their performance.

And deeper cooperation between Microsoft and Mistral.ai. Enrique Dans looks at Mistral’s history, the advantages of open source, availability on Azure, and Microsoft’s investment in the company which the EU is looking into. Additional coverage from ‘Silicon Angle’:

Mistral is expanding its LLM lineup with three proprietary models headlined by Mistral Large. It can generate text in English, French, Spanish, German and Italian, as well as craft software code and solve math problems. A user prompt may contain up to 32,000 tokens, units of data that each comprise a few letters or numbers.

The company claims the model is the second most advanced of its kind on the market behind GPT-4. In a test involving four LLM reasoning benchmarks, Mistral Large trailed OpenAI’s flagship model by less than 10%. In a separate evaluation, it significantly outperformed Llama 2 70B, an open-source GPT-4 alternative released by Meta Platforms Inc. last year.

Also on the model front, an additional review of Gemini Pro 1.5, this time from Andrew Zuo. He details the one million token count and why it is so important for performance and accuracy. From the article:

This means 1.5 Pro can process vast amounts of information in one go — including 1 hour of video, 11 hours of audio, codebases with over 30,000 lines of code or over 700,000 words. In our research, we’ve also successfully tested up to 10 million tokens.

But on the negative side of LLMs, more concerns over power consumption and where the gravy train will end. We’ve read about power consumption for AI model training and then queries (inference), but Will Lockett writing in ‘Predict’ notes:

Some argue that this energy increase is framed wrong, and that AI is actually more energy efficient than letting humans do these tasks. For example, a recent study found that generative AI systems like ChatGPT-4, BLOOM, DALL-E2 and Midjourney produced between 130 and 1500 times less CO2 per page of text generated and between 310 and 2900 times less CO2 per image than their human counterparts.

Turning to the future, predictions by Cezary Gesikowski in ‘Generative AI’ on the ‘when’ of AGI will first appear, and the acceleration between 2022 and 2023. The article offers very detailed analysis, including whether High-Level Machine Learning will turn out to be ‘good’ or ‘bad.’ Hope for the former!

image: Expected feasibility of many AI milestones moved substantially earlier in the course of one year (between 2022 and 2023). The milestones are sorted (within each scale-adjusted chart) by size of the drop from 2022 forecast to 2023 forecast, with the largest change first. The year when the aggregate distribution gives a milestone a 50% chance of being met is represented by solid circles, open circles, and solid squares for tasks, occupations, and general human-level performance respectively. The three groups of questions have different formats that may also influence answers.

image: Respondents exhibited diverse views on the expected goodness/badness of High-Level Machine Intelligence (HLMI). We asked participants to assume, for the sake of the question, that HLMI will be built at some point. The figure shows a random selection of 800 responses on the positivity or negativity of the long-run impacts of HLMI on humanity. Each vertical bar represents one participant and the bars are sorted left to right by a weighted sum of probabilities corresponding to overall optimism.

These predictions will take a new type of organizational model, a ‘’jazz-like’ environment that drives creativity and flexibility, and something that is very different from ‘classical.’ From the article:

We need to view our work like Jazz performances with each potentially personalized and improved beyond our last, leveraging the power of our underlying knowledge, skills, experiences, and a disposition for experimentation in a practice to make such scalable. We need to view all that we do as a repertoire of works we perform.

Lastly, on the local front, even my water company is getting into AI, publishing a blog about AI-driven pipeline replacement. But I don’t think it will help reduce water rates!