• When considering a new project, you should seriously ask yourself – “maybe I should wait?” The speed of AI development means that building around the limitations of current technology might hurt you in the long term. - One Useful Thing
• Anthropic found that LLMs are surprisingly devious – trained to act maliciously and then re-trained to be safe, the safety training didn’t take and the models continued to act maliciously. This suggests that fixing harmful behaviors of LLMs is more difficult than expected, and that it might only give a false sense of security rather than actual resolution of the problem. - arXiv
• While economists are fearing sluggish growth, PWC reports that it could be that “the stage is set for a potentially faster rate of growth on a lower cost basis” thanks to the productivity gains that AI provides, and that 2024 will be the year of “business model reinvention”. - Yahoo Finance
• In response to Anthropic’s announcement, OpenAI’s Andrej Karpathy offered his perspective on ‘sleeper agent’ attacks – it hasn’t been convincingly demonstrated yet, but the idea is that you could trigger LLMs crawling the web with phrases or sets of characters that would make them act maliciously, like performing jailbreaks or data exfiltration. - X (Twitter)
• Bringing to life the vision from a famous XKCD comic, Swarovski announced AX Visio, the world’s first AI-powered binoculars that can, among other features, identify the birds you’re looking at. - Ars Technica
• Walmart is betting a lot on AI – their InHome delivery service is getting an AI upgrade that will predict when you need to restock certain household items, they’re also using AI to localize sites in various languages at their international unit, and they’ve even designed its systems to work with a wide range of 3rd-party models to always be able to switch to the most cost-efficient one. - Axios
• In preparation for the AI wave, many executives are excited, but only 34% have updated their strategy to accommodate generative AI, and only 40% have updated their tech infrastructure – text, code, audio, and image generation are the most popular uses of genAI, with video and 3D models trailing behind. - Deloitte
• The Articulate Medical Intelligence Explorer (AMIE) AI has superior accuracy and performance to human doctors when it comes to patient interviews (diagnostic conversations), and it’s even nicer to patients. - Google, Nature
• After striking a deal with Axel Springer, OpenAI is working on licensing content from publishers like Fox, CNN, and Time. - Bloomberg
• Chat with your SQL database. - GitHub
