A little late this week again, but here are this week's stand-out links for me:
- 2024: The State of Generative AI in the Enterprise (thank you for the hat-tips I received in sharing this one !). This Menlo Ventures survey of 600+ Enterprise software leaders doesn't have anything super surprising in it, but it does clearly show an uptick in real spending for AI adoption. The chart from the review below shows the top use-cases making it into production. Code is unsurprisingly high up the list along with support chatbots and search. I'm a little surprised that meeting summarization is not higher, but perhaps this is revenue adjusted. What is striking about the adoption chart, though, is that the top use cases are all very general and things used by / relevant to human knowledge workers. I'd be willing to bet that deep automation of processes (without humans) will be a much bigger part of the picture once people figure out how to do that properly.
- AI found a new way to create quantum entanglement [Paywall - free coverage is here]. More groundbreaking science with AI. In this case, the team involved aimed to use an AI to iteratively improve existing methods of entangling photons, but instead, it kept converging to a simpler method. This new method initially looked right but was then verified. Another step towards your passwords all being busted!
- From Code to Paper: Using GPT Models and Python to Generate Scientific LaTeX Documents,. As if scientific conferences had already received too many AI submissions, one of the last barriers was removed: generating LaTeX markup directly from GPT-4o. LaTex is a markup language used to capture scientific formulas and text, so this could be a boon for researchers.
- Latent Space Podcast: Compound AI with Lin Qiao, CEO of Fireworks.AI. If you want to understand where the technical architecture of AI systems will end up in the mid-term, there's probably no better short conversation than this. The moderators and Lin talk about how serving models for inference isn't just about speed; it's about optimizing which models to use for what and even that it's possible to create overlays that do things like use an LLM iteratively to get iterative reasoning like that available in OpenAI o1 Strawberry models. Fireworks seems to be ahead in thinking about such architectures holistically, but you can bet that most major inference providers will migrate to similar approaches.
- X’s Grok AI chatbot is now available to all users. Elon Musk's X.ai bot Grok is now available to all twitter/x.com for free. It will be interesting to see how this evolves but clearly X.ai is now also in the race for consumer mindshare in AI. OpenAI is building its own user base, Meta and Google are adding AI to their user touchpoints, and now Twitter/X.com is doing the same. Grok seems to be alive and kicking already: "Generate an image of a steampunk Elon Musk admiring a rocket" (see below). It looks like they are using Flux for image generation.
Part of the reason I was late this week was that I was at Amazon's annual AWS Re:Invent conference in Las Vegas. It was interesting to be back at the event after a a year of two out. More on impressions in a future post, but in the meantime, here is a news round-up.
Wishing you a wonderful week!
