AI News Roundup – OpenAI’s new AI-powered internet search engine, Google DeepMind’s AI model is solving complex math problems, Meta unveils new AI model, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• • OpenAI has announced a prototype of SearchGPT, its AI-powered internet search engine. SearchGPT will “quickly and directly respond to your questions with up-to-date information from the web while giving you clear links to relevant sources,” according to the company, and allows for follow-up conversational questions in the style of OpenAI’s ChatGPT. SearchGPT will provide prominent links to further information about search queries and will partner with content publishers to allow for greater control in how they appear in the search engine. SearchGPT is not yet publicly available, but the company stated that many features are intended to make their way into ChatGPT after further testing.

• • The New York Times reports that researchers at Google’s DeepMind have developed an AI model capable of solving complex math problems. The new model, known as AlphaProof, working alongside an improved version of DeepMind’s AlphaGeometry math-solving model, achieved a silver medal-level performance at the 2024 International Mathematical Olympiad (IMO), solving four out of six problems for a total of 28 points. This marks the first time an AI system has achieved a medal-worthy performance at the IMO, traditionally a competition for high school students. While the AI model took longer than human competitors to solve some problems, researchers emphasized that this breakthrough represents a significant step towards computers being able to tackle increasingly complex mathematical challenges.

• • A new study has found that AI models could “collapse” within just a few generations if trained on data they generate themselves, according to Semafor. The study, published in Nature, claims that “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models,” causing the models to generate gibberish. The effect is compounded as the models continue to be trained on synthetic data, as the model finds it more and more difficult to find “real” data to source things from. The study concludes that in order to “sustain learning over a long period of time, we need to make sure that access to the original data source is preserved and that further data not generated by LLMs remain available over time.”

• • Meta has unveiled Llama 3.1 405B, its newest AI model. According to the company, the new version of Llama “is the first openly available model that rivals the top AI models when it comes to state-of-the-art capabilities in general knowledge, steerability, math, tool use, and multilingual translation.” The new model was made available in Meta’s WhatsApp messaging app as part of its Meta AI platform. In a statement released along with the new model, Meta CEO Mark Zuckerberg declared that “open source AI is the path forward,” distinguishing the open-source Llama from its proprietary competitors. Zuckerberg argues that “open source is necessary for a positive AI future… ensur[ing] that more people around the world have access to the benefits and opportunities of AI, that power isn’t concentrated in the hands of a small number of companies, and that the technology can be deployed more evenly and safely across society.”