AI Roundup 050: Synthetic Geometry
AI Roundup 050: Synthetic Geometry
January 19, 2024.
Here’s that survey again - I’m going to be making some newsletter changes in February, and want to do it based on reader feedback.
Synthetic Geometry
A new Google Deepmind model, named AlphaGeometry, can solve International Math Olympiad problems at a near-gold medalist level.
Between the lines:
There are a few caveats: the model only handles geometry problems (there are multiple types), and the IMO is meant for high school math students.
But it is another impressive achievement for generative AI - previous geometry problem solvers came nowhere close to AlphaGeometry.
The model takes a "neuro-symbolic" approach and was trained by generating purely synthetic geometry problems - no human examples were used!
Elsewhere in the FAANG free-for-all:
Mark Zuckerberg's new goal is creating AGI - and he's amassing 600,000 GPUs to make it happen.
Microsoft releases Copilot Pro with GPT-4 Turbo and Office integrations (for $20/month), and Reading Coach (for free).
Amazon is reportedly working on an AI-powered version of Alexa, but needs to work through internal politics and technical issues first.
Galaxy Unpacked
Samsung's Galaxy Unpacked event took place this week, with a bevy of new AI features for Samsung phones (and other hardware).
What to watch:
The Galaxy S24 lineup will get Gemini-powered AI features, as well as a new “circle to search” feature.
The phones can do live translation for phone calls or texts, and use in-house models to help rewrite messages and organize notes.
And Samsung teased the Galaxy Ring, a new health tracker that will use AI in some capacity.
Elsewhere in AI hardware:
The Rabbit R1 will use Perplexity's AI to answer questions using live data.
You can now generate AI images on your Fire TV for… some reason?
And a new report details how China's military and government are able to buy Nvidia GPUs (despite a US ban).
People are worried about AI elections
OpenAI has published its approach to the 2024 worldwide elections. Its goals are to prevent abuse, create transparency around AI content, and improve access to proper voting information.
Why it matters:
Just one year after world leaders embraced AI, many now fear it. Australia is the latest example of governments pushing AI regulations.
OpenAI, for its part, is trying to calm tensions: it is publishing findings from AI governance research and downplaying the impacts of AGI.
Even so, expect AI to become the bogeyman and/or the scapegoat as misinformation and propaganda inevitably play a role in this year's elections.
Elsewhere in AI anxiety:
Google News is boosting AI-generated articles, many of which were ripped from original sources.
Fairly Trained is a new nonprofit that aims to certify AI companies that use training data collected with consent.
And Anthropic researchers find that AI models can easily be trained to deceive users when given trigger words - AI sleeper agents.
Things happen
Is AI the death of IP? A new GPU vulnerability could expose AI data. Sam Altman's full Davos interview. Vodafone and Microsoft sign 10-year, $1.5B deal. Artisans bring AI tools to the workbench. Taking AI costs from $100/day to less than $1/day. AI girlfriend bots are already flooding the GPT Store. What about AI boyfriends? Stable Code 3B. "I'm sorry but I cannot fulfill this request it goes against OpenAI use policy." An open-source AI model for Earth. Racing team fires its AI influencer after fans balk. Politico launches AI-generated Bill summaries. TikTok can generate AI songs, but it probably shouldn’t. "I literally spoke with AI-powered NPCs." A shocking amount of the web is already AI-translated trash. OpenAI removes ban on ChatGPT for "military and warfare" purposes.
