Pixie is an AI assistant from Google that could launch with the Pixel 9. Powered by Gemini, it will use data from Google products on devices to perform complex and multimodal tasks. Google has discussed creating glasses with the ability to recognize objects and advise wearers on how to use tools. The last bit of news on Google AR was how the hardware offer effort was being sidelined for a Samsung partnership.
DeepMind AI outdoes human mathematicians on unsolved problem (3 minute read)
FunSearch is an AI system that can interact with a specially trained large language model to create short computer programs that generate solutions to math problems. The system checks to see whether these solutions are better than known ones, and if not, it provides feedback to the LLM so that it can improve in the next round. It was able to discover new solutions to an old problem. One important feature of FunSearch is that people can see the code created by the model and learn from it.
Moderna-Merck vaccine cuts odds of skin cancer recurrence in half, study finds (3 minute read)
An ongoing randomized vaccine trial involving 157 patients with high-risk stage III/IV melanoma is showing promising results. Those who received the vaccine were 49% less likely to die or have their cancer come back after three years than those who only received treatment with a cancer drug. Phase 3 trials of the vaccine have begun.
Meta's Ray-Ban Glasses Added AI That Can See What You're Seeing (5 minute read)
Meta's second-generation Ray-Ban glasses can interpret scenes and make judgments using generative AI. A new update is rolling out a feature that uses the on-glasses cameras to look at images and interpret them. The update also brings Bing-powered search to the glasses. This article contains some examples of how the feature can be used.
Prompt engineering (35 minute read)
This guide shares strategies and tactics for getting better results from large language models. The methods can sometimes be combined for greater effect. Some examples only work with the most capable models. Strategies for getting better results include writing clearer instructions, providing reference texts, splitting complex tasks into simpler subtasks, and using external tools.
Significantly Enhancing Adult Intelligence With Gene Editing May Be Possible (72 minute read)
There are no known fundamental barriers preventing humans from significantly enhancing adult intelligence with gene editing - many pieces of a viable editing protocol have already been demonstrated independently.
OpenAI buffs safety team and gives board veto power on risky AI (4 minute read)
OpenAI has announced a new safety advisory group that will sit above the technical teams and make recommendations to leadership as part of an expansion of its internal safety processes to fend off the threat of harmful AI. The company has updated its Preparedness Framework to show a clear path for identifying, analyzing, and deciding what to do about catastrophic risks inherent to models they are developing. Leadership will be able to make decisions on whether to ship products, but the board has the power to reverse those decisions.
Beeper vs Apple battle intensifies: Lawmakers demand DOJ investigation (2 minute read)
US senators have written to the Department of Justice over Apple's alleged anti-competitive conduct towards Beeper.
All the changes coming to Google Play and sideloading following $700M settlement (3 minute read)
Google will pay $700 million as part of a settlement with the US attorney general for a lawsuit over the Google Play Store. It has also agreed to implement third-party app stores, side-loading, and alternative billing on Android. Google will not be able to enter deals with phone makers to have Google Play as an exclusive app store on devices for at least five years. More details about the settlement agreement are available in the article.
Electric scooter company Bird files for bankruptcy (1 minute read)
Bird was once valued at $2.5 billion by investors.
SSH protects the world’s most sensitive networks. It just got a lot weaker (4 minute read)
Researchers have devised an attack with the potential to undermine cryptographic SSH protections. Terrapin is a man-in-the-middle attack that affects SSH connections secured by ChaCha20-Poly1305 or CBC with Encrypt-then-MAC. Around 57% of servers list these encryption modes as their preferred choice. Anyone using SSH should check whether they are vulnerable to the attack and make the appropriate updates. The attack invalidates proofs published in 2016 concluding that such attacks weren't possible. Details about the attack and how it works are available in the article.