• MemGPT teaches LLMs to manage their own memory, like an operating system, to achieve unlimited context windows – in testing, it successfully analyzes large documents and enables multi-session chats where the LLM remembers you and evolves from chat to chat. - MemGPT
• AI helped decode letters on an ancient, unopened scroll in a revolutionary breakthrough for papyrology – the AI was built for the Vesuvius Challenge, which aims to decode scrolls carbonized by the Mount Vesuvius eruption in 79AD. - Scrollprize
• Report time – if you’re looking for a comprehensive overview of the current state of AI, check out these reports from Kaggle and Air Street Capital. - Kaggle, Air Street Capital
• Prompt injection is still an unsolved problem for LLMs, and now vision capabilities offer a fun new way of doing it – you can break models quite easily by putting prompts into images, and it’s even possible to insert them in a way that humans can’t see, only GPT. - Simon Willison
• Tech analysts from CCS Insights predict that genAI hype will get a “cold shower” in 2024 due to its costs, risks, and complexities – also, AI is moving too fast for legislators, and it will enable 50% of European companies to try 4-day workweeks by 2030. - CCS Insight
• Well, that was too easy – you can overcome the guardrails (i.e. censorship) of LLMs by simply fine-tuning them with as little as 10 harmful examples at a cost of less than $0.20 via OpenAIs APIs. - The Register
• Optimizing traffic lights with AI to reduce emissions – Google’s Green Light initiative (live in 12 cities, affecting 30m car rides each month, set to expand) is driving 30% fewer stops and 10% less emissions. - Google
• “The ultimate signal of transparency in digital content” – the new icon from Adobe’s Content Credentials effort is gaining steam with supporters like Bing, Nikon, Leica, or Publicis, and it’s also integrated into Adobe’s Firefly to distinguish generated images. - Adobe
