Author: Jason

For the past decade of scaling, we’ve been spoiled by the enormous amount of internet data that was freely available for us to use. This was enough for cracking natural language processing, but not for getting models to become reliable, competent agents. Imagine trying to train GPT-4 on all the text data available in 1980—the data would be nowhere near enough, even if we had the necessary compute. In 2025, our situation when it comes to automating software engineering is no different

— Read on www.mechanize.work/blog/how-to-fully-automate-software-engineering/

But the fundamental problem is that LLMs don’t get better over time the way a human would. The lack of continual learning is a huge huge problem. The LLM baseline at many tasks might be higher than an average human’s. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box. You can keep messing around with the system prompt. In practice this just doesn’t produce anything even close to the kind of learning and improvement that human employees experience.

The reason humans are so useful is not mainly their raw intelligence. It’s their ability to build up context, interrogate their own failures, and pick up small improvements and efficiencies as they practice a task.

— Read on www.dwarkesh.com/p/timelines-june-2025

Instead of the meticulously reported summer entertainment coverage the Sun-Times staff has published for years, these pages were filled with innocuous general content: hammock instructions, summer recipes, smartphone advice … and a list of 15 books to read this summer.

Of those 15 recommended books by 15 authors, 10 titles and descriptions were false, or invented out of whole cloth. Sixteen hours later, the journalist behind the piece ‘fessed up: rather than a reported recommendation list, this one had been generated by an AI agent.

Lessons (and an apology) from the Sun-Times CEO on that AI-generated book list – Chicago Sun-Times

It’s not that LLMs randomly mess up. It’s that they’re built to sound right, not be right. They generate what looks plausible based on the patterns they’ve seen, without fact-checking.

That’s fine for brainstorming. Dangerous when you hit publish without looking twice.

Marin Smiljanic – LinkedIn

Star Trek: The Next Generation contains evidence that the actor Patrick Stewart exists in the Star Trek universe!

— ironicsans.ghost.io/proof-that-patrick-stewart-exists-in-the-star-trek-universe/

The point is, you can read all the Wikipedia summaries you want, but they won’t give you a holistic understanding of an idea. That only happens once you have a layered, three-dimensional perspective, which writing helps you achieve. 

Charlie Munger calls this the difference between “real knowledge” and “chauffeur knowledge.” He tells an apocryphal story about Max Planck, who went around the world giving the same knowledge about quantum mechanics after he won the Nobel Prize. After hearing the speech multiple times, the chauffeur asked Planck if he could give the next lecture. Planck said, “Sure.” At first, the lecture went well. But afterwards, a physics professor in the audience asked a follow-up question that stumped the chauffeur. Only Max Planck, who had the background knowledge to support the ideas in the talk, could answer it.

— Read on perell.com/essay/how-philosophers-think/

It’s both sad and not at all surprising. After letting it languish for years, Mozilla has now put Pocket out to pasture. And it will be formally slaughtered in October. Along with all your saved items. All those glorious saved items that you never got to but were totally just about to.

— Read on spyglass.org/rip-pocket/

In a safety test, the AI company asked Claude Opus 4 to act as an assistant to a fictional company, but then gave it access to (also fictional) emails saying that it would be replaced, and also that the engineer behind the decision was cheating on his wife. Anthropic said the model “[threatened] to reveal the affair” if the replacement went ahead.

— Read on www.semafor.com/article/05/23/2025/anthropics-ai-resorts-to-blackmail-in-simulations