Diarmid Mackenzie

If the work stood a good chance of leading to AGI, I suspect you wouldn't have to pay the researchers anything like these amounts (cf most scientists doing fundamental research in academia)

The prospect of being a part of that kind of groundbreaking research would be incentive enough.

My guess is the reason they have to pay top researchers such vast amounts of work on LLMs is that they all know the path to AGI lies elsewhere.

Top researchers would prefer to be working in areas that might lead towards AGI, not the dead-end that is LLMs.

So they have to paid extra to compensate.

Something I've been working on over the last couple of months.

Driven by the frustration that I couldn't find a decent, simple, free online drawing tool to help with our building project.

Given time, I think it has the potential to be really great!

"I like how [Claude 3.7 Sonnet] solved the problem of pelicans not fitting on bicycles by adding a second smaller bicycle to the stack."

This is a great metaphor for the sort of thing Claude will often do when attempting to fix problems in code it has written...

Yes, the plug/socket components can be used completely independently of the brick geometry.

diarmidmackenzie.github.io/aframe-compo...

So they could be used with any 3D model you like.

I don't have any sloping/curved bricks either yet, but that might not be too hard to add.

Snappable flags and flagpoles might also be do-able.

Did I miss anything else from the picture?

Things that would be hard:
- minifigures
- bricks with holes in
- physics integration

Things that are easy:
- different sizes of brick / plate, that already exists.
- placement of bricks with controllers, including snapping into place.

Happy to help if you want some help building something here.

Stack overflow is almost dead Today, Stack overflow has almost as few questions asked per month, as when it launched back in 2009. A recap of its slow, then rapid, downfall.

Sad to see Stack Overflow decline so fast.

I never found asking or answering questions on SO to be a particularly pleasant experience, but as a resource for passive consumption, it was (and still is) invaluable.

blog.pragmaticengineer.com/stack-overfl...

Just discovered this

An extremely useful, thoughtfully developed set of resources on the many challenges and issues presented by the proliferation of LLMs.

GitHub - johnousterhout/aposd-vs-clean-code: A discussion between John Ousterhout and Robert Martin about differences between John's book "A Philosophy of Software Design" and Bob's book "Clean Code". A discussion between John Ousterhout and Robert Martin about differences between John's book "A Philosophy of Software Design" and Bob's book "Clean Code". - johnousterh...

A very helpful set of links!

Particularly enjoyed this discussion with Bob Martin about the issues that arise when applying "Clean Code" principles without some level of restraint.
github.com/johnousterho...

Thoughtful post about the issues that arise with agentic coding.

An interesting question IMO is to what extent multi-agent flows with tailored prompts per-role can mitigate these effects.

Due to its openness, RooCode seems to be the best currently available tool for this kind of experimentation.

For clarity, it's not that Cline/Roo fill up the LLM context window any faster than other agentic coding tools like Cursor or Windsurf.

The difference is that Cline and Roo are transparent about how much of the context window they are using, while other tools are opaque.

I asked ChatGPT to estimate the resulting CO2 emissions from these 1T tokens.

It came up with a figure of 57,000 tonnes of CO2, which is the same as the emissions from 253 round-trip transatlantic flights of a full Boeing 747.

@annecurrie.bsky.social

Open Router shows 1.35T tokens from AI coding tools Cline and Roo this month.

Easily 75% of this is waste, because client-side tooling is ineffective at concisely providing LLMs with relevant context.

That's 1T tokens/month of pure waste - and that's just the traffic passing through Open Router.

When I switched to using Cline (and then Roo) for AI coding, I realized how rapidly my LLM's context window was filling up.

An LLM chewing through an entire module just to see what's happening in a few lines of code is supremely wasteful.

github.com/buger/probe is going to help change that.

Asked GPT-4o-mini to analyse a business idea.

It's response was reasonable, coherent, but also utterly generic and bland, and contained nothing that was in any way thought provoking.

Then I tried asking GPT-4.5-preview...

Just one question cost me $0.50, but its response was absolute 🔥

If you ever wanted to listen to an upbeat podcast about an obscure github project that you've personally put many hours of work into....
www.gitpodcast.com

It's very obviously artificial, and makes some odd choices about what to focus on.

Nevertheless, it's surprisingly uplifting and validating!

Very nice.

I tried to do something similar on a pancake screen a while back, using head tracking...

diarmidmackenzie.github.io/aframe-compo...