AI News Roundup – Researchers unveil memory-focused AI architecture, AI-developed antivenoms, new translator 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.

• • Researchers at Google have developed a new machine learning architecture, dubbed “Titans,” that are positioned as successors to the Transformer framework that underlies many modern AI systems. In a recent paper, the researchers discuss a new neural long-term memory module that, when combined with a more conventional short-term module, can be used to build machine learning systems that greatly outperform current models. In tests, the performance was found to remain consistently high even as the input sequence length increases, at which point current models often experience steep drop-offs in accuracy. This architecture more closely mimics the way in which human brains retrieve memories and thus could bring future models a step closer to more advanced cognition.

• • AI-designed proteins may have the potential to combat toxins found in the venom of deadly snakes, according to a new study in Nature. In 2022, researchers at the University of Washington in Seattle developed RFdiffusion, an AI-based program to generate proteins that have the ability to bind to target proteins, unlocking a series of medical applications from cancer treatments to autoimmune diseases. The same team recently used RFdiffusion to design binding proteins for toxins found in the venoms of several snakes, including cobras, mambas and adders. These proteins were found to neutralize the effects of several venom components and protected lab mice against simulated snakebites. The researchers admitted that, in order to protect fully against such venoms, further proteins would need to be developed, but that the results were incredibly promising for the future of this type of research.

• • Meta has released a new AI model that can translate speech from over 100 different languages, according to MIT Technology Review. In contrast to existing machine translation models, which generally covert speech into text, translate the text into text in the target language, and then convert the translated text into speech, “SeamlessM4T” directly translates speech from one language into speech in another. The company claims that the new model can translate text 23% more accurately than existing models. Meta researchers, writing in a paper published in Nature this past week, claim that they were able to train the model on parallel data, which matches instances of audio that match a subtitle in another language, and from that the model learns how to match audio fragments together. The system is open-source, and the researchers are optimistic that it will encourage others to build on the model’s capabilities.

• • Apple has temporarily suspended a news-summarization feature in its flagship Apple Intelligence AI software, according to The Washington Post. The feature, intended to allow users to quickly catch up on notifications, produced widely publicized “hallucinations,” including a false alert that Donald Trump had endorsed Democratic vice-presidential nominee Tim Walz for President during the 2024 election. In a beta update that will be included fully in iOS 18.3, the software will provide a notice that the summaries can produce unexpected results and display them in an italic font in order to distinguish them from standard notifications. The rollback, however temporary, demonstrates the difficulty companies have had in combatting the “hallucinations” that appear to plague all major AI systems.

• • OpenAI’s ChatGPT now has the ability to set reminders and schedule tasks for users, according to The Verge. The feature, rolled out to paying OpenAI customers this past week, is an add-on to OpenAI’s 4o model, and can be accessed directly through a chat or through a dedicated Tasks UI. The new features highlight the company’s shifts towards AI agents that work autonomously on behalf of users.