Joel

software dev, working on AI

[Joel @joelwohlhauser]

First of all, T3 Code is already very good. I'm very happy with it. The thing that annoys me most right now is that I can't fork a conversation to switch to another model mid-conversation. Since Astra, I've often been switching between OpenAI and Anthropic models. I think you guys are already working on this with the new orchestration update. Other than that, worktree management could be improved. I don't particularly like the git worktree workflow in T3 Code.

[Joel @joelwohlhauser]

@ky__zo people keep saying English is becoming the new programming language while AppleScript has been sitting there the whole time

[Joel @joelwohlhauser]

@BasedDaedalus banger. both the song and the video

[Joel @joelwohlhauser]

@kimmonismus would you still recommend reading it? I have it in my library but keep putting it off because I worry it's already outdated. so much has happened in the field since it was published.

[Joel @joelwohlhauser]

I don't know when we decided as an industry that ~5 lines before and after a diff were enough context, but I think it's wrong.

I always want to see the full file with the diff.

[Joel @joelwohlhauser]

@jkelleher @ChatGPT what a banger. absolute chef's kiss

[Joel @joelwohlhauser]

another day, another GitHub incident

[Joel @joelwohlhauser]

are we just making up numbers now?

[The Wall Street Journal @WSJ]

Exclusive: Anthropic is likely to tell investors its potential revenue opportunities are above $30 trillion, topping SpaceX’s $28.5 trillion estimate. https://t.co/anJex4VsFX

[Joel @joelwohlhauser]

For the last few weeks I've completely switched from VS Code to Blackbird, my personal fork of T3 Code.

I was a die-hard VS Code user for years. I even built my own extension to push it further, but eventually hit too many limits.

T3 Code had most of what I wanted, so I forked it and added the missing pieces:

- a right sidebar for files, local changes, and PR changes
- tabs in the main chat area, with tabs for the Linear issue and GitHub PR opening automatically
- better flows for existing and new worktrees, with automatic .env linking and cleanup
- smaller things I missed, like fast keyboard switching between chats

Why call it Blackbird? the SR-71 is the most impressive airplane ever built. easy.

[Joel @joelwohlhauser]

who is building the situational awareness fund without leverage?

I was just digging into this story and his famous blog post about it. Even though I'd skimmed through it when it came out, reading it again now still fascinates me. I don't agree with many of his ideas, especially his timeline, but I'd find it interesting to see how such a fund performs over the next few years.

Don't use leverage kids.

[Drew Hahn @drewhahn]

“the agent is working in the sandbox”

meanwhile the agent:

[vx-underground @vxunderground]

Unfathomable banger

[Joel @joelwohlhauser]

if you’re not intrinsically curious, yngmi

[Joel @joelwohlhauser]

@arnie_hacker @theZachChen I'd still want to know how it works regardless. even if it's just high-level. that curiosity is usually what gets me exploring in the first place

[Joel @joelwohlhauser]

You simply can't build a competitive tech future on energy that costs more than double your rival's. In a world that's innovating fast, stagnation is functionally equivalent to decline.

While others build, Europe seems focused on being the first to regulate.

[Joel @joelwohlhauser]

Everywhere I look, the gap is widening. The US and China are sprinting ahead. Europe is standing still.

For instance, industrial electricity prices in Germany hit an average of $0.19/kWh in 2025. That's more than double the US rate of roughly $0.08/kWh.

[Joel @joelwohlhauser]

Those are my two cents about the paper. It is just a confirmation that the current architecture for LLMs does not provide a direct route to AGI.

[Joel @joelwohlhauser]

Just read Apple’s new paper “The Illusion of Thinking” and wanted to share some thoughts. 🧵

The paper tries to argue that LRMs (Large Reasoning Models) aren’t really reasoning because they collapse on hard puzzles. But I think that’s kind of a weird argument.

[Joel @joelwohlhauser]

A better setup would let models act in an environment. Let them observe the state, plan, make a move, repeat. That’s how we solve puzzles. That’s how we think.