Sunday Links: ChatGPT Agent, No Code, and Sneaky Text Prompts

Greetings from London, here are this week's links.

The Big LLM Architecture Comparison. This is a nice, compact technical overview of today's different LLM architectures. Great post by Sebastian Raschka. The latest LLMs do all borrow from each other, but it's interesting to see the rise and fall of different architectural approaches.OpenAI unveils ‘ChatGPT agent’. OpenAI is pushing hard into automating web and tool use (also a detailed blog post here). The new "Agent mode" takes previous work on operator and starts long-running tasks (like deep research), which is a very similar trajectory to what Anthropic is taking with its coding agents work. There will be a lot of challenges with long-running tasks running into errors, but OpenAI is clearly pushing hard to try to change web and office tool use from today's user-centric experience to an AI-centric one.Substack just raised $100M. Substack isn't an AI company directly, but I do think a "subscription" era of web content is going to be inevitable as human web browsing gives way to AI aggregation. Substack and other subscription platforms might be one of the engines of content distribution. I hope there will be lots of platforms that do this, not just a few, but they'll probably be needed. At some level, the fundraise seems small to me for something that could end up being a new distribution hub in an AI world.No Code Is Dead. Clickbait title, but a nice article from The New Stack. No-code has been "almost here" for the last 10-15 years of enterprise development. There has been a slow, painful advance to give more configuration power to business users, and hence unburden development teams from building custom code. The result has been an uneasy alliance between development teams delivering APIs and components, SaaS platforms building increasingly sophisticated UIs, and a range of drag-and-drop integration tools that try to span the two worlds. Mostly, this has resulted in a fragile ecosystem of interdependencies. Now that Vibe coding promises executives and business users the promised land of building apps entirely from scratch by prompt, it's unclear what will happen. For now, even the best vibe coding tools like Lovable don't quite provide enough precision and definitely not enough security or manageability to be durable in an enterprise environment, but that future is surely coming. It seems likely that the best players will integrate with existing enterprise ecosystems rather than try to rewrite them from scratch. Ingesting existing no-code integrations and making them controllable and adjustable via Vibe coding prompts seems like a way to leapfrog to utility quickly.Scientists reportedly hiding AI text prompts in academic papers to receive positive peer reviews. Scientific conference reviewers are often volunteers, and they are generally supposed to use their own judgment to determine the grading of submitted papers. Unfortunately, many increasingly use AI to generate reviews, and it's a frequent occurrence that "AI reviews" are returned to authors. By encoding messages directly to AI in the submitted papers, authors are fighting back. This seems natural, though perhaps going as far as asking for a favourable review breaks some ethical boundaries!

In other news, there was a cool mathematical theorem proof revealed this week. Wishing you a Happy Sunday.