Hello, Starters!

How we look up answers online is expected to change radically in the AI era. There have been attempts by several companies, and now, unsurprisingly, OpenAI has joined the trend.

Here’s what you’ll find today:

• DeepMind’s models win a silver medal in maths

• OpenAI announces SearchGPT

• Google upgrades its Gemini chatbot

• Apple’s AI may get delayed

• Udio releases a new version of its AI model

• And more.

🥈 DeepMind’s models win a silver medal in maths (6 min)

AGI can be achieved by developing models that solve advanced maths problems. Google DeepMind is aware of this, and that's why most of its efforts have been directed towards the creation of models like AlphaProof and AlphaGeometry 2, which have recently showcased amazing results by solving four of six problems from the International Mathematical Olympiad (IMO), one of the main benchmarks for these capabilities.

As has been discussed several times, most AI models are still in the very early stages of development, struggling to catch up with general maths problems due to reduced reasoning skills and limitations on training data. Yet, the transformations we expect from AI don't seem as far off as we witness these advances.

🔍 OpenAI announces SearchGPT (2 min)

After months of speculation, OpenAI has finally released SearchGPT, their take on an AI-powered search experience. The feature is a prototype, and it's currently available to selected users, although there's a possibility to join the waitlist. The company expects to integrate these capabilities into ChatGPT once the testing process ends.

Everything points to the next aspect AI will change: how we search the web, as we've seen with offerings like Perplexity, Google, and Bing. Now, here's OpenAI's approach, which is based on quick and direct responses, with links to relevant sources as part of their ongoing partnerships with publishers. SearchGPT is not related to the training of OpenAI's foundation models.

🤖 Google upgrades its Gemini chatbot (2 min)

Gemini users can now enjoy an improved experience as Google has decided to make 1.5 Flash its new base model. With this integration, the multimodal model will now be available for Gemini's free tier and will broaden its availability to over 40 languages and 230 countries.

1.5 Flash will provide faster and more insightful responses, as Google points out that many of the main use cases for Gemini focus on productivity and streamlining tasks. Also, the company has expanded the context window for the chatbot, which now goes up to 32K tokens.

🍎It seems that Apple Intelligence won't be rolled out at the same time as iOS 18 as initially expected. According to reports, the company's widely anticipated foray into AI will be released for customers in October. However, beta testers could be getting a glimpse of it this week.

🎶Udio has introduced a new version of its novel AI music generator, with tools to enhance music production and catch the attention of creatives and music experts. The latest upgrades include improved audio quality, more human-like voices, and a revamped user experience.

⚡️Quick links

Microsoft unveils serverless fine-tuning for its Phi-3 small language model

Study reveals major weaknesses in AI's ability to understand diagrams and abstract visuals

Generative AI requires massive amounts of power and water, and the aging U.S. grid can’t handle the load

Google Deepmind Researchers Introduce Jumprelu Sparse Autoencoders

Paris Olympics broadcasters diverge on AI approach

Thank you for reading!