🥈AI Takes IMO Silver!

DeepMind’s AI wins silver medal at International Mathematical Olympiad

Key Points 

• AlphaProof and an improved AlphaGeometry were tested together, solving four out of six IMO questions and earning a silver medal.

• AlphaProof uses a trial-and-error approach called reinforcement learning, similar to Google DeepMind's AIs that excel at chess and Go.

☕News - An AI developed by Google DeepMind has won a silver medal at this year's International Mathematical Olympiad (IMO). This is the first time an AI has ever reached the podium in the competition, which is renowned for its challenging questions that usually require a level of mathematical skill beyond what AI systems can typically handle.

🤓For context -  In January, Google DeepMind showcased AlphaGeometry, an AI that could solve some IMO geometry questions as well as humans, but it wasn't tested live and couldn't handle other math areas like number theory, algebra, and combinatorics, which are essential for winning an IMO medal. Recently, Google DeepMind introduced AlphaProof, a new AI that can solve a wider range of math problems, and an improved AlphaGeometry that can tackle more geometry questions.

When both systems were tested together on this year's IMO questions, they correctly answered four out of six, scoring 28 out of 42 points. This was enough to win a silver medal, just one point shy of the gold medal threshold.

🧮What makes them good at math? AlphaProof uses a trial-and-error approach called reinforcement learning, where the system finds its own way to solve a problem over many attempts. This is similar to Google DeepMind's AIs, which have achieved remarkable success in games like chess and Go. However, this method requires a large set of problems written in a language the AI can understand and verify. Since most IMO problems are written in English, the team, led by Thomas Hubert, used Google's Gemini AI chatbot to translate these problems into a programming language called Lean.

A point worth noting here is that despite its impressive performance, AlphaProof works slowly, taking up to three days for some solutions, compared to the 4.5 hours competitors have per three questions. It also couldn't solve both the combinatorics questions. Nonetheless, this achievement is a significant milestone.

OpenAI announces SearchGPT, its AI-powered search engine

Key Points 

• SearchGPT’s interface is similar to ChatGPT, featuring a large text box that asks, “What are you looking for?”

• Instead of simply listing links, SearchGPT organizes and summarizes the information.

• It also uses your location for certain searches, such as finding nearby restaurants or checking the weather.

• SearchGPT is currently just a prototype. The service, powered by GPT-4, will only be available to 10,000 test users for now. The plan is to eventually integrate these search features directly into ChatGPT.

♨️News - OpenAI has announced its highly anticipated search engine, SearchGPT. It’s an AI-powered tool that gives real-time access to information from the web.

🧐What's it like? SearchGPT’s interface is similar to ChatGPT, featuring a large text box that asks, “What are you looking for?” Instead of simply listing links, SearchGPT organizes and summarizes the information. For instance, it can give an overview of music festivals with brief descriptions and links, or provide planting tips and details on different tomato varieties.

It also uses your location for certain searches, such as finding nearby restaurants or checking the weather. You can adjust this in the settings if you prefer to share more precise location information.

🔍What's more? SearchGPT is currently just a prototype. The service, powered by GPT-4, will only be available to 10,000 test users for now. The plan is to eventually integrate these search features directly into ChatGPT.

OpenAI is going for a more responsible and careful rollout with SearchGPT. In a blog post, the company explained that SearchGPT was developed with input from major news partners like The Wall Street Journal, The Associated Press, and Vox Media. It noted that publishers have control over how their content appears in SearchGPT and can opt out of having their content used for training the models while still being part of search results.

👨🏻‍💻In conclusion - Launching SearchGPT as a prototype has several advantages for OpenAI. If the results aren’t perfect, they can easily attribute any issues to it still being a prototype. Nevertheless, this move puts OpenAI in direct competition with Google, which is rapidly adding AI features to its search engine to stay ahead. Plus, it positions OpenAI directly against the startup Perplexity, which promotes itself as an AI “answer” engine.