nickbec.co.uk RSS

Stream

Stream Archive

They're Made Out of Weights

Amazing, also make sure to read the original before you read this (at the link at the top of the post) if you don't know it.

https://maxleiter.com/blog/weights

I'm partway through Anthony Shaw's NDC talk, "Are LLMs good software engineers?"... | Hacker News

This is one of the things that needs to be really understood if you're going to make progress with LLMs. Every time you call one, you need to give it a version of this kind of briefing so they can go from nothing to having these lessons embedded. I'd love to see a way for agents to truly learn rather than just spending more tokens on context.

https://news.ycombinator.com/item?id=48278419

Leaving the Physical World | Electronic Frontier Foundation

"In the Silicon Valley, I never know whether I'm in Mountain View or Sunnyvale or Campbell. They all look the same, and, in fact, they all look more or less like what one finds outside Austin or Boston. Most Americans can't be said to come from anywhere at all." This desire to be involved in something physical, to have place and presence, is powerful stuff. Could it be why I've gone quite deep on 35mm film photography?

https://www.eff.org/pages/leaving-physical-world

MacBook Neo Review: The Laptop For The Rest Of Us — fireborn

This says everything I’d want to say about the MacBook Neo. I’d not buy one for myself, but I’d recommend it to anyone who fits in the segment described here.

https://fireborn.mataroa.blog/blog/macbook-neo-review-the-laptop-for-the-rest-of-us/

agents need control flow, not more prompts | brian’s thoughts

This is the heart of what I'm working on at the moment, trying to build reliable systems that combine the benefits of LLMs for language processing, but don't require them to define their own workflow.

https://bsuh.bearblog.dev/agents-need-control-flow/

Adding a feature to a closed-source app - Stavros’ Stuff

Interesting to see someone try and modify a closed source app with Claude. I’m far more interested in the implications of LLMs for Open Source, but this shows that closed source as a protection mechanism for your app isn’t quite what it was.

https://www.stavros.io/posts/adding-a-feature-to-a-closed-source-app/

Spectre

This is truly the best of the internet. Far too much detail, abstract mathematical concepts and lots of diagrams.

https://www.revk.uk/2026/04/spectre.html?m=1

If America’s So Rich, How’d It Get So Sad?

So much to unpack here, it’s probably worth a proper blog post. Well worth the read (lots of good charts too).

https://www.derekthompson.org/p/if-americas-so-rich-howd-it-get-so

nial.se/blog/less-human-ai-agents-please/

How do we train the AI agents that we want when the training data makes it easiest for them to imitate humans?

https://nial.se/blog/less-human-ai-agents-please/

solar pi

I’ve always wanted to do a solar based machine and love the idea of this just running in memory. My one slight issue is that you are, of course, dependent on power for your internet connection. I might see if I can set up something similar.

https://hackaday.io/project/205403-solar-pi

@adlrocha - How the "AI Loser" may end up winning

I think this is really perceptive - my dream is to run something fully locally that can match the performance of the current Claude Sonnet. I have a box that can run very small models on it, and would happily invest in something powerful to run more capable models. I suppose I'm waiting for the models and hardware to align. I'd prefer to run something like a Framework Desktop than a Mac Mini, and then deliberately give it context instead of just letting context be my Apple / iPhone / iCloud account.

https://adlrocha.substack.com/p/adlrocha-how-the-ai-loser-may-end

Protect Your Shed · dbut2

I’ve really enjoyed getting more into side projects, mostly enabled by LLMs. Either in fixing problems or helping me triage tech choices. My other thought would be why we seem unable or unwilling to make similar spaces to try and fail in professional environments.

https://dylanbutler.dev/blog/protect-your-shed/

Multi-agentic Software Development is a Distributed Systems Problem (AGI can’t save you from it)

I’m not sure my brain was really in the right place to process this at 8am, but I do like some formal rigour as opposed to lots of LLM hot takes. I don’t think I’m qualified to comment as to whether this is true or not but it seems reasonable as a real concern for agentic swarm development. Maybe a single superintelligence wouldn’t have such problems, but that feels like running away from reality to a safe, warm delusion that just feeds more of the economy into the AGI dream.

https://kirancodes.me/posts/log-distributed-llms.html

Bad vibes

Interesting as always, especially the part about Labour losing their traditional base because of their continued stance on Brexit.

https://chrisgreybrexitblog.blogspot.com/2026/04/bad-vibes.html?m=1

The Subprime AI Crisis Is Here

"The death of the AI industry would be cataclysmic to venture capitalists, bring about the end of the hypergrowth era for the Magnificent Seven, and may very well kill Oracle, but — seriously — that is nothing in comparison to the scale of the Great Financial Crisis. This isn’t me minimizing the chaos to follow, but trying to express how thoroughly fucked everything was in 2008." Not sure this is a bad thing.

https://www.wheresyoured.at/the-subprime-ai-crisis-is-here/

AI Agents Could Make Free Software Matter Again | George London

I am finding being able to quickly build and modify my own software massively liberating, but hadn't thought of how this impacts 'free software' more philosophically.

https://www.gjlondon.com/blog/ai-agents-could-make-free-software-matter-again/

Do LLMs Break the Sapir-Whorf Hypothesis?

This guy's content is amazing. Such depth and really understandable. Is there an underlying logic for how we structure concepts? Who knows, but it looks like there might be for LLMs.

https://dnhkng.github.io/posts/sapir-whorf/

Zack Shapiro on X: "The Input Layer" / X

If something is good enough that I am prepared to cross post something from X, it's worth a read. The blending of technical and domain knowledge is key to any successful 'digital transformation', and getting it right is almost impossible. That's why I don't think I've ever seen one.

https://x.com/zackbshapiro/status/2036791156915290271

← Back to stream