. .AI: Google AI generates rabbits in hats (1.17.24)

Author

Marshall Kirkpatrick
January 17, 2024

Friends, just three big stories in the AI world today (not including the consumer news of Google Gemini Pro coming to Samsung Galaxy devices) , but the first one here is especially remarkable. You might think "I don't use AI for geometric proofs, what does this have to do with me?" But it's really about creativity in problem solving, and there's a reason the New York Times wrote it up today.

The second story is technically valuable and the third one will update your understanding of the State of the Art on autonomous agents.

As always, these are the stories the AI community is talking about most. Thanks for joining us; today a bunch of people from the writing & editing world have joined! Welcome, folks. And thanks to everyone sharing this newsletter.

Now here’s today’s news,

First impacted: Mathematicians, AI researchers
Time to impact: Short

Want to feel inspired about AI and creativity? Google's DeepMind and Google Research say their new AI system, AlphaGeometry, nearly equaled a human gold medalist's performance in the International Mathematical Olympiad (IMO) by solving 25 out of 30 geometry problems within the standard time. Trained on generated synthetic data, the system identifies & is then trained on millions of "magic construction points" that act like rabbits pulled out of a hat to solve hard geometry problems. At least I think that's what I took from this fascinating YouTube video the team published about it.

Key takeaway: Sometimes hard problems can be solved by substituting a simple factor for a complex equivalent – and good candidates for meaningful, applicable substitute factors can be found in the following way. By thinking through known scenarios where you identify two factors, an outcome of those factors, then think backwards from the outcome to discover interstitial dependencies along the way. Those newly discovered dependencies are good candidates for being swapped out in other scenarios for more complex factors. Google did it 100M times and found 9M really good ones. Wow! [AlphaGeometry: An Olympiad-level AI system for geometry] Explore more of our coverage of: Google DeepMind, AlphaGeometry, AI in Mathematics. Share this story by email

First impacted: AI researchers, Software developers
Time to impact: Short

"Large Language Models are increasingly utilized for complex tasks that require multiple chained generation calls, advanced prompting techniques, control flow, and interaction with external environments. However, there is a notable deficiency in efficient systems for programming and executing these applications. To address this gap, we introduce SGLang, a Structured Generation Language for LLMs. SGLang enhances interactions with LLMs, making them faster and more controllable by co-designing the backend runtime system and the frontend languages." [Fast and Expressive LLM Inference with RadixAttention and SGLang | LMSYS Org] Explore more of our coverage of: Large Language Models, Structured Generation Language, AI Programming Efficiency. Share this story by email

First impacted: Business professionals, Social Media Managers
Time to impact: Medium

MultiON AI has launched an upgrade that reportedly enables its agents to execute over 500 steps for a single task without losing context. Unfortunately, the examples shared in the announcement ("not a dream anymore!!!") are sending LinkedIn spam and automated ecommerce discount hunting. A good reminder to check on your dreams now and again. [via @DivGarg9] Explore more of our coverage of: AI Upgrade, Automated Tasks, Ecommerce Discount Hunting. Share this story by email

Ok, that’s all for today! More AI news tomorrow!