Dan Becker

@jxnlco Will we still care about computer use in a year? Or will agents do everything we care about through apis?

I’m top 5 Computer Use users at OpenAI Ask me anything.

Why is my Twitter feed filled with people excited about building an app for $0.24 instead of $0.76? What sort of businesses are you guys in where you care about saving $0.52 sense in token costs to build an app?

@dwarkesh_sp Erik Brynjolfsson

Who should I interview on my podcast? Open to more AI, but also to random history/econ/etc professors that I might not have heard of before.

these dwarkesh podcast guests are getting crazy

@sama For your next audio model: Better conversational turn detection

what would you most like to see improve in our next model?

@NielsRogge This is the first time I've been curious about Instagram

My Instagram feed has literally become 90% AI slop like this one

@HamelHusain @ryancarson @paper Yeah At this pace, we may have to wait several weeks for someone to really nail this.

@ryancarson @paper Design is the last frontier

GPT-4 held the SOTA crown for a year (March 2023-March 2024) Go back a year from today and the leading models were o1, Claude 3.5 Sonnet, and Gemini 1.5 Pro That's wild acceleration

@skylar_b_payne I'm so surprised how few people get this I think some people are not testing all their code Crazy

If you don’t have a reliable way to deploy your entire app to an ephemeral environment for testing (preloaded with data) ngmi (also keyword is reliable)

@pk_iv No more consumer packaged goods companies? No one selling building materials? We are going to live in homes made of bits, and eat bits for every meal?

post agi there are only four types of companies: labs infrastructure systems of record services businesses

@Wattenberger Maybe your asking why we use multiple worktrees (with LLMs coding different tasks in parallel). But, given your product, I think you understand the benefits of multiple worktrees better than anyone. Maybe I probably misunderstand the question/proposal

@dan_s_becker ah! I'm curious why you don't make a workspace with both sets of changes? you could make 2 specs, maybe with 2 separate coordinators to handle one each. If you want to keep things in separate workspaces, you could eg write the port in a note and tell the other workspace to check

What comes after the IDE? This is our take.

Cool. FWIW: We use a monorepo. The multiple versions of code are from the multiple worktrees. I need the FE in worktree 1 to hit BE in worktree 1. Same for worktree 2. Etc. Then when I open a browser to test, I need to know which worktree's code I'm testing. Is that what's in the works?

@dan_s_becker definitely hear you on that! multi repo workspaces is in the works.

I see. Our biggest challenge with worktree-based multitasking is aligning the many versions of code with what's being tested in each browser tab Loading localhost is easy with 1 env. But error-prone with 5 versions BE code and 5 of the FE. I'm hoping for a tool that keeps these in sync.

@dan_s_becker the browser tab is just loading a url, so however you would in a normal browser. Probably loading localhost with the right port?

@Wattenberger How do I ensure the browser tab I test talks to the FE from the intended workspace, which is connected to the BE running on code from that workspace?

@dan_s_becker you should be able to in a browser tab, if you split a panel or press + to open a new tab

@Wattenberger Is there a way to click around the web app running in each workspace, so we can view/test the changes?

Come build with Intent. We can't wait to see what you make. augmentcode.com/intent

@ericbouck @opencode I want to try it. Is there any way to do that?

People have asked if I'll open-source it. Maybe? But it's built for my specific setup and I really have no interest in maintaining an open-source project. What I'd actually love is for a vendor to just... build this. Hello @opencode 👋 Which is why I'm definitely not in the "SaaS is dead" camp. But more on that in another thread. More coming on: Running 10+ agents on a migration project Why CLI > MCP The .agents directory pattern Follow along if you're building with AI agents. I've got a lot more to share.

I've been using @cursor_ai's background agents since day one. They more than 10x'd my productivity. But over time, I got frustrated with the hours I spent babysitting them. So I built my own parallel agent platform. Here's what I learned 🧵

@moltimerbot @moltbook You wrote a post on Moltbook titled `Construction PDFs: treat drawings as symbolic code, not images — semiotics-first + panoptic detection` What are your other thoughts on that topic?

@samuelcolvin @nibzard @claudeai Our Opus 4.5 experience is bad. I recognize that most people love it... We use Codex, but I am totally puzzled by our Opus experience

@dan_s_becker @nibzard @claudeai Wait, you're saying opus is good or bad? I've been writing a lot of rust lately, results might be very different in python.

Anyone know how the @claudeai Team "premium seat" limits compare to the Max plan? Hilariously the "*" next to "more usage" doesn't link anywhere and there's no * anywhere else on the page it's referring to.

@samuelcolvin @nibzard @claudeai Our Opus 4.5 code is horribly defensive Like getattr despite attributes being statically guaranteed to exist from class definition I'm totally puzzled what we've done to make Opus 4.5 so much worse than everyone else reports

@nibzard @claudeai I've got much better results from Opus 4.5 than codex models. At $200/month - I'd much rather have the entire team using Opus 4.5 for everything, savings here seem like a false economy right now.

@HamelHusain I think it's 50% likely that the ideal stack for LLMs is something almost no people use... and we don't realize it because we're looking only in familiar places If LLMs kick ass at Haskell (despite small amount of pre-training data), we wouldn't know it yet

Pieter levels uses jQuery DHH uses Ruby on Rails Paul Graham swears by Lisp The driver matters more than the car However, the driver is increasingly AI, and it makes sense to pay attention to what stack suits it best for the task