dirceu

Qwen is irreplaceable. Has been going from strength to strength in recent times. Things will always be different, I'm hopeful we can find groups of other models to fill the void. RIP

an increasingly large part of the job of an engineer is deciding how much compute to spend on a problem

this is my favourite prompt of all time: “how could this be better?” reply with yours and why it rocks!

trying this now

1) What

1. Install Readwise CLI 2. OpenClaw, Claude Code, Codex, or any coding agent can now: * Search your entire readwise library * Read full content of anything you've saved * Tag, create highlights, organize on your behalf * Pipe reading data into any workflow Your agent now knows everything you've read :)

if your skill depends on dynamic content, you can embed !`command` in your SKILL.md to inject shell output directly into the prompt Claude Code runs it when the skill is invoked and swaps the placeholder inline, the model only sees the result!

@GenAiAlien @WisprFlow I’ve tried WisprFlow and SuperWhisper (which I use all the time on macOS) but there’s something about using a custom keyboard… I don’t like it. On Android the default dictation works decently well but on iOS I just end up using ChatGPT 😅

@dirceu Oof you guys gotta check out @WisprFlow

looool that’s exactly what I do

@viemccoy thanks, this tweet made me give it another try!

I think Atlas browser has become a surprisingly load-bearing tool for me. There is an effort threshold for which I'll see a word/phrase/concept I don't understand and actually take the time to look it up, and the sidechat does seem to have meaningfully lowered that threshold. This is obviously a good thing because it means I am engaging more actively with what I am reading, and more critically, allowing me to both curate my information diet further and expanding the range of my knowledge.

New chapter for Agentic Engineering Patterns: I tried to distill key details of how coding agents work under the hood that are most useful to understand in order to use them effectively simonwillison.net/guides/agentic…

Btw you can add `context: fork` to run a skill in an isolated subagent. The main context only sees the final result, not the intermediate tool calls It gets a fresh context window with CLAUDE.md + your skill as the prompt. The `agent` field even lets you set the subagent type!

I see people at Anthropic who didn't necessarily start that way getting better at it. Part of it is being surrounded by others who are AGI-pilled + watching how they push the models. But ultimately... 1. Ask yourself: what if the exponential actually continues 2. Take a task and handhold the AI less, be more ambitious, try to do more of it end-to-end with AI 3. Do #2 enough until you reach the limits of current AI and it fails 4. Wait until the models get better and can successfully complete that task 5. Learn from this. Update your strategy. Rethink what the future looks like. And practice that over & over

This will help to decide who's turn it is.

@douglasandrade @_st0012 Not sure about Stan's case, but a bunch of us are using github.com/davebcn87/pi-a…

@_st0012 Hey @_st0012! How y'all are using it? I'm really curious to try it out in a big monolith we have at work

There’s some hype around autoresearch inside Shopify, so I asked it to speed up Ruby LSP’s tests. And the result is a 33% reduction on CI (~30min -> ~18min) 🤯 github.com/Shopify/ruby-l…

A few end of week ships: You can now set effort to 'max' which reasons for longer and uses as many tokens as needed. This will spend your usage limits more quickly so you have to activate it per session. Hit /effort to try it.

Flash Attention's biggest strengths become bottlenecks during LLM decoding. Q-tiling, causal masking, 2D combine grid- all overhead when Q is 1 token. So I built a decode-specific kernel that beats it by 21% (-7.2 μs) using KV-cache parallelism and temp buffers in ScratchPad Memory. Here's what I learned building it- and why Flash Attention was never really designed for this in the first place.

Well I usually have multiple working if I can help it. The key I found is to limit context switching. Its better to run multiple agents in parallel with xhigh thinking that take 30+ minutes so I can get into some flow with another task then go back to babysit them when I'm done with my won't ask. The agents don't interrupt me, I interrupt the agents.

I met Nick Land a few weeks ago. He mentioned that many people in his circles were anti-LLMs. Someone asked why he thought so many people were. His answer was better than anything so short I thought of: “People like to exist critically with respect to something.” This I think accurately characterizes a lot of people whose outputs and inputs primarily consist of “discourse” about rather than direct contact with the reality at hand. Existing critically with respect to something makes it easy to seem cool, sophisticated, above something, hard-to-impress and therefore worth trying to impress, especially to others who also don’t have contact with the phenomena itself. And for that reason I think it’s cheap. And to someone who has an inside view of what is being discussed, it’s always so transparent and boring and compressible. I’m far more impressed by someone who is capable of loving something and showing others why it’s beautiful or good. Doesn’t have to be LLMs, but anything at all.

love this analogy