AI Top-of-Mind for Dec 13
Top-of-mind is yesterday’s announcement by Microsoft of their new ‘Phi-2’ ‘small language model.’ Since it requires less computing power, and therefore cost, it is more flexible in deployment. However, it is only for non-commercial use. The Microsoft blog is great reading and some of the published benchmarks vs other open-source models are below.
With all the innovation in the AI image generation space, it is sometimes confusing as to what models underly the various applications. For example, we sometimes use Bing Image Creator, but it is not a model in itself, leveraging DALL-E3 which is not a surprise given Microsoft’s partnership with OpenAI. Or Leonardo.ai, that is considered a model itself with Leonardo Diffusion XL, but that also supports a variety of other models, some open-source, including Stable Diffusion 2.1. Confused yet?
A few months, back, Zapier published a great list of AI art generators, many probably familiar to you. Though the list is probably somewhat out-of-date, most notably omitting Leonardo.ai, it is a good landscape of the many options. The almost 30 listed all leverage a short set of fundamental models, with the majority tied to Stable Diffusion, DALL-E3, Midjourney, and Adobe Firefly. But there are others including open source or in the case of Getty, a co-development with NVIDIA. Zapier also published their own take on the best currently available models, listed here.
Over the next week, I’ll try to dig up the best papers and presentations on the most popular models and re-publish them here, starting with Stable Diffusion. There have also been several comparisons published, some in the eye of the beholder, and I’ll cover these as well.
Speaking of Getty Images, they and Runway ML are partnering to create legally compliant AI-generated videos for advertising, media, and other creative endeavors.
Moving to the public space, we’ve all heard of Smart Cities, but how are governments using AI as part of this? It is not only the infrastructure, but leveraging Gen AI to summarize large datasets, as Alison Brooks from IDC notes:
“Think about [city] council notes and complaint records, and policing, and the generation of outreach to citizens. All of those things are fairly low risk and are fairly high value.”
And on policy, Stanford has created a new initiative to help people better understand many new technologies, AI included. The new ‘Stanford Emerging Technology Review,’ or SETR, has already published a baseline document on ten major technologies. Key takeaways from the AI focus area:
· AI is a foundational technology that is advancing other scientific fields and, like electricity and the internet, has the potential to transform how society operates.
· Even the most advanced AI has many failure modes that are unpredictable, not widely appreciated, not easily fixed, not explainable, and capable of leading to unintended consequences.
· There is substantial debate among AI experts about whether AI poses a long-term existential risk to humans, and whether the most important risks are also current weaknesses of AI.
On the random front, Meta is rolling out some limited AI capabilities within its 2nd generation Ray-Ban glasses. Not real-time since it works by taking a picture and then analyzing the results, and it also takes a voice prompt to get going, which could be embarrassing at times. But still, a view of the future. ‘CNET’ describes the demo.
Finally, how AI can help with personalizing employee communications. The author of the ‘Ragan’ article used ChatGPT to develop some of the recommendations, and the bot produced this last bit of guidance:
Remember, while ChatGPT can assist in generating ideas and content, it’s crucial to review and adapt these suggestions to ensure they align with your brand voice, goals, and ethical considerations. Always keep your communications authentic and genuine, even as you leverage AI to personalize your messages.
