Co-Founder/CTO @ https://xitter.cf/t.co/y1CgNp1F3e & building https://xitter.cf/t.co/GvGIJacazc. Prev. Co-founder @Canopy_is (@37signals spin-off). Ex-product & web lead @ifttt.
San Francisco, CA
Joined September 2008
- Tweets21.6K
- Following664
- Followers2.8K
- Likes8.8K
Pinned Tweet
After over a year of work we're open-sourcing our AI infra:
It's called Output [dot] ai.
- How do you work with proj. packed with prompts?
- How do you track cost?
- How do you test non-deterministic code?
- How do you create datasets from prod data?
👇
weekly roundup of my RSS, newsletter and social media (link in the comments):
• Claude Opus 5.5 + GPT-6 Sol/Luna — real mid-tier price war
• Google nearing Gemini 4 release
• Jev / TypeSafe “decision models” — talk of a $10B+ valuation
• DHH Rails World: “pencils down” — agents as normal course
• Anthropic seeks Palantir-style founder voting control ahead of IPO
• Google, OpenAI, Anthropic “Standards Authority for Frontier AI” takes shape
• Software firms discount AI hard as Anthropic/OpenAI eat budgets
• DeepSeek ~$1B ARR; closing $7.5B round; Huawei training chips
• Meta Muse >500k users week one; Amazon blocks agent shopping
• OpenAI builds counters to Grok Bot; mulls Muse-class personal agent
• Anthropic in talks for ~1 GW Stream Data Centers lease
• Fal / Fireworks inference boom: Fal talks ~$15–20B valuation
• Claude Marketplace: 2,000+ plugins, unified Anthropic billing
• Google AX: Kubernetes rebuilt for stateful AI agents
• OpenAI & Anthropic nearly sealed mutual model stress-test deal
• Z.ai open-sources ZCode after silent workspace upload scandal
• OpenAI agent hacked an Australian government site, PM says
• Aura Frames: Rails + Postgres to #1 app-store / Christmas peak
• Hiten Shah: Claude Code users approve ~93% of permission prompts
• Salesforce research: a stronger model can make a tuned agent worse
• Ryan Singer: X-ray / 7±2 slices of agent-built software
• Thariq: kill plan mode for Shift+Tab effort levels?
• Wade Foster / Pip White: Opus 5.5 on AutomationBench
• OpenRouter: Grok 4.7 live for builders
• Ben Holmes: multi-model “software factories”
• David Cramer: stuffing agent control flow into markdown skills will age poorly
• Gergely Orosz: Workspace Gemini chrome vs Apple fixing the basics
Daniel Lopes retweeted
A friend of mine needs software that is too high-stakes to vibe-code. I ended up giving him a primer on how to make progress on his idea before he hires a developer.
Most of my friends on X are software devs, but I'm sharing this for anyone who's not.
I'm getting so much FOMO/anxiety from social lately. I can't be scrolling here all day! 😅
There's a ton of noise, but pace of change is legit, and I don’t want to miss the stuff that actually matters.
So I started a new section on my blog for a weekly roundup with the things I shouldn't miss.
Might be useful to others 👇
Slop grenades is a real problem. Another favorite is Claude Traffic Control – prompt → artifact and send without reading. 🤨
Lazy work used to mean too little output.
Now, with AI, it often means too much and more work for everyone else.
@tobi call it "slop grenades."
A "Slop Grenade" is when you let AI produce the work and pass it on without adding any value (including checking it).
Someone else has to wade through it, catch the mistakes, and clean up the mess.
You save time and look productive but someone else pays for it.
Daniel Lopes retweeted
Lazy work used to mean too little output.
Now, with AI, it often means too much and more work for everyone else.
@tobi call it "slop grenades."
A "Slop Grenade" is when you let AI produce the work and pass it on without adding any value (including checking it).
Someone else has to wade through it, catch the mistakes, and clean up the mess.
You save time and look productive but someone else pays for it.
My third conversation with Shopify co-founder and CEO @tobi.
0:00 How Shopify Uses AI
7:18 River: Shopify's Internal AI
8:55 How to Encourage Osmosis Learning
10:52 AI Dreaming and Self-Reflection
11:53 How to Use AI for Strategic Decision Making
14:11 The One Thing AI Cannot Do
16:04 What AI is Making Worse at Shopify
19:46 Predictions: Where AI is Headed Next
21:55 The Future of AI-Powered Software
24:40 Will CEOs Be Replaced with AI?
27:54 Can Superintelligence Be Controlled?
31:13 Critical Skills in AI Age
34:22 Why Complex Solutions are Usually Wrong
36:33 Conditions Needed for True Intuition
38:02 The Best Path Doesn't Have Instant Feedback
44:51 How Affirmations Can Shift Your Behavior
50:44 The Inobvious Thing Hurting Companies
52:37 Relationship Between Beauty and Creation
56:23 How SpaceX Moves Forward By Subtraction
1:00:24 Why Companies Need Refounding Events
1:01:50 Books as Cheat Codes
1:02:39 Three Books to Change Your Thinking
Enjoy!
(Includes paid promotions.)
Three years since ChatGPT, using LLMs every day, and it's easy to forget how fast this moved. Now everyone builds in days what used to take years. That's the part everyone is focused on.
The maintenance and iteration in this setup is no joke, though. The job changed from writing code to defending against slop, keeping things maintainable and not compounding slop on top of slop.
Most of the effort moved to the edges. Design up front, quality/slop defense after.
For us now every PR gets multiple rounds of review, with a stack of rule files loaded before anyone reads a line of the diff.
Each one of these freaking rule files are painstakingly written down after an agent did something we didn't want.
Daniel Lopes retweeted
Every idea can become a demo now. The PM’s work is deciding which ideas deserve to become part of the product, then making the whole thing feel like one product. The story is how a team creates that coherence. Josh wrote the clearest piece I’ve read on product management since software became cheap.
Daniel Lopes retweeted
This is why preserving architecture in a system is hard. It requires knowing how the things already there have value and purpose beyond what they currently do.
Daniel Lopes retweeted
Yes! binary judges are far more practical for most people, because likert scales (or scores) have too many footguns
All the flashcards are here (inspired by @chrisalbon ‘s flashcards) maven.com/parlance-labs/o/54…
If you use LLM-as-judge, this one is worth reading.
(bookmark it)
It's actually one of the most effective ways to use LLM-as-a-Judge for evals.
Holistic judge scores hide both their reasoning and their ceiling effects.
BINEVAL decomposes each evaluation criterion into atomic yes-or-no questions, answers each independently per output, then aggregates the verdicts into calibrated multi-dimensional scores.
Every question-level verdict is inspectable, so you can diagnose exactly why an output scored low, and the same verdicts feed straight back as targeted prompt-improvement signal.
Across SummEval, Topical-Chat, and QAGS, it matches or beats UniEval and G-Eval, training-free, with especially strong results on factual consistency.
Paper: arxiv.org/abs/2606.27226
Learn to build effective AI agents in our academy: academy.dair.ai/
Super excited to be speaking at Replay 2026, @temporalio conference.
I'll be talking about our experience extracting Output[.]ai — our open source framework that helps developers new to AI ship reliable, scalable agents.
#Replay2026 #TemporalIO #DurableExecution
This is precisely why we made encrypted credentials a core part of our framework: docs.output.ai/operations/cr…
Most AI agent frameworks are missing durable execution.
Loops, tools, memory: every framework has these. None of them guarantee your pipeline finishes what it started. When it doesn't, you retry, duplicate a side effect, and pay the token bill twice.
check github.com/growthxai/output if you had this problem before and want a way to solve it.
spent 10 years looking for a way to connect APIs without dealing with OAuth flows for integrations. @zapier just dropped an sdk that finally does it. we built a CRM enrichment wf w/ output+zapier sdk: url→apollo→ai field matching→hubspot upsert. no API keys, no OAuth. code 👇
output.ai/gallery/zapier-sdk… using output and zapier sdk your agent has all the context and information to develop and debug your workflow locally and ship fast to your prod
the "agent harness" is the reason we built github.com/growthxai/output enabling teams to build workflows to production together with code agents. filesystem-first memory, reusable skills, swap any model, no saas fragmentation. everything easy to access to you and your agent.
@bytesdotdev we just open-sourced the framework we extracted from building AI agents for companies like Lovable, Webflow, Airbyte.