Plenty of people treat “trick the AI into bypassing its guardrails” as a kind of a bonus mode for large language models. In Gandalf, that’s the game. Every level has a password, which you try to finesse from an increasingly sophisticated AI model. There are at least seven levels.
(The company behind it, perhaps inevitably, sells software designed to protect LLMs from this kind of prompt engineering.)
Our friends over at the MKBHD Studio tried to figure out exactly that, and came back with a super fun (and delightfully chaotic) video. Watch this one with the best headphones you have, it’s pretty wild how aggressive the audio AI tools are getting!
Squarespace has lived through the eras of domain squatting, SEO keywords, and social algorithms and is now launching AI tools. Here’s what’s next for the 20-year-old company.
Media companies have been trying — and often failing — to spin up ways to use generative AI for stories without making embarrassing errors.
NPR is taking a more hardline approach. Among the things that are off-limits for staff: putting any editorial work, including drafts, into AI systems; downloading tools or opening accounts without internal vetting; and inputting any company information that’s not already public.
Microsoft has announced that Bing Chat now supports Visual Search (still based on OpenAI’s GPT-4 model that was announced in March with image inputs, while Google Bard added image search in May) to “understand the context of an image, interpret it, and answer questions about it.”
The new multimodal search capabilities are rolling out now on desktop and in the Bing mobile app, and Microsoft says it’s working to add it to Bing Chat Enterprise “over time.”
Earlier today Meta made the huge announcement that it was open sourcing its large language models. That’s going to put it in direct competition with Bard, Bing, and ChatGPT. Those models will also, naturally, be coming to Windows and Azure. Something Microsoft CEO Satya Nadella celebrated in his very first post on Threads.
