Saturday Links: ChatGPT Atlas Jailbroken, AI Music, and Continual Learning

After a busy week with new AI browsers, lawsuits, and a relentless march of tech advances, here are the most interesting links:

• Introducing ChatGPT Atlas. Continuing the theme of attacking almost every adjacent business to their own, OpenAI launches the Atlas Browser. Much like Perplexity's Comet browser, Atlas is a fully fledged web browser that includes deep AI integration. AI agents within the browser can access browsed content and can take actions on behalf of the user. An example would be summarizing the content of certain types or populating an e-commerce shopping basket. Unfortunately, the AI browser is also vulnerable to attacks from hidden prompts on pages. I'm not sure I've ever seen a product launch announcement with an inclusion like this "ChatGPT's agent capabilities still carry risk. Besides simply making mistakes when acting on your behalf, agents are susceptible to hidden malicious instructions, which may be hidden in places such as a webpage or email with the intention that the instructions override ChatGPT agent’s intended behavior. This could lead to stealing data from sites you're logged into or taking actions you didn't intend." Probably something to stay away from for now!
• Andrej Karpathy on the Dwarkesh Podcast. This podcast episode is a great (and fairly technical) deep dive into the state of LLM tech. Andrzej comes up with some useful analogies and highlights a number of key points on which we see progress in some areas and not in others. The podcast headline of "AGI is 10 years away" is a bit of an over-rotation of what he says. However, I agree that we need to go significantly beyond current LLM tech to get to really flexible reasoning and agents that don't need supervision.
• Rainbow of Yours – Psychedelic Trip-Hop - Portishead & Morcheeba Inspired AI Mix [2000s Album]. What happens when people figure out how to use advanced AI tools? Something like this – really impressive and immersive music combining the styles of multiple bands. It's amusing to see people in the comments asking where the tracks are on Spotify... these tracks didn't exist long before they were posted. The comment block encourages listeners to engage with the original artists. Maybe they will, but this does point to a future where AI remixes and Fusions become potentially more popular in aggregate than all the source tracks.
• Meta will ban rival AI chatbots from WhatsApp. From January 2026, third-party chatbots will no longer be able to interact with customers via WhatsApp. This is Meta shutting a distribution door for OpenAI, especially. By favouring its own Meta AI, the company likely wants to stop the leakage of usage minutes and engagement to other platforms. There is a logic to this move, but one could imagine that A) there might be anti-trust suits filed about this, B) it's a little late since ChatGPT already has 800M+ users. It seems likely that it's too late for ChatGPT to be the "AI app" users focus on. Unless Meta manages to make Meta AI compelling, there's a risk that ChatGPT ultimately flips the script and adds messaging ("have your agent talk to my agent").
• Continual Learning via Sparse Memory Finetuning. A new paper from FAIR (Meta) and the University of Berkley outlines techniques for adjusting model weights over time to learn continually. The technique showcases methods that lead to retention and use of new knowledge, but that cause less damage to existing linkages than more brute force methods, such as Lora. These are early days, but it is interesting to see progress towards AI that could adapt even after it is deployed. Of course, it likely poses a fundamental problem for AI models served to many users: one probably wants to learn for every user individually. How do you do this without exploding the number of models you need to serve?

Wishing you a good weekend!