Kurt Manninen

@nsxdavid @tibor_tee 100% of my PRs come from AI. All 1 one of them.

@tibor_tee Weak sauce. Way more than 40% of PRs come from our dark factory. Somewhere around 90%. That's because most of these PRs are cleanups, documentation alignment, test coverage and correctness, etc. Real feature work requires human-in-the-loop (even still!) so they come slower.

40% of Cursor's internal PRs now come from cloud agents. Not demos. Not examples. Actual production PRs from agents running for hours or days in isolation. Temporal handles 50 million agent actions daily across 7 million workflows making this possible. What percentage of your team's PRs could realistically come from agents in the next year?

@natmiletic 🤦‍♂️🤣

LeBron fans would never

Dire Straits 🎶 Sultans of Swing 🎸

South Region @RTX_News UAV Touchdown Competition just happened and it was EPIC 🔥  Autonomous UAV. Moving UGV. One shot to land then navigate to target together. Every team delivered 🤯 West Coast is next. THIS SATURDAY. Ready? 👀 #RTX #BrainChip #EdgeAI

@nsxdavid @omarsar0 Totally agree. I wonder who’s doing this with git. A repo (or sub repo) just for agent context (markdown, html, etc) for the project.

@k_manninen @omarsar0 Version history tells you what changed. Comment history used to tell you why. With agents, I think the missing object is a review trail: prompt/context, proposed delta, evidence, human acceptance, and open questions. Not a full transcript.

Increasingly, HTML Artifacts are becoming a core part of how I work with AI agents. Long-horizon agent sessions need a better way to surface insights about what work it has done. This may not be obvious right now, but as you start to let your agent work on dynamic workflows, large codebases, long-running loops (e.g., using /goal), and deep research tasks, you need a good way to present results. Chat window is not it. You also don't want to just trust everything the agents do. Artifacts help provide an important verification layer, which in turn enables important decision-making. I like HTML artifacts because I can just ask the agent to produce as many of them (and in whatever form) as I need to verify the work and make sense out of everything. I even built a nice tab system for my artifacts. They are great for continual learning and research. I use HTML artifacts for logging, tracking experiments, brainstorming, managing my inbox, code reviews, agent session management, deep research, writing, reading, and so much more. I believe @karpathy wrote about this somewhere: As we move on to more advanced applications of AI agents and outputs get more complex, we will start to find the need for even more advanced forms of interactions with AI, including interactive neural videos/simulations.

@omarsar0 @nsxdavid But it’s the version history and comment history that is starting to tie us up. In the past we would use Word comments in a shared doc to collaborate on a design. Now we are using agents to help with our collaboration, but how do we attribute the thoughts behind the changes?

@omarsar0 @nsxdavid I’m finding that with globally dispersed teams (different time zones and first languages) it’s been helpful to share agent created markdown files and each engineer uses their own agent to help them interpret the md and build upon the design.

@nsxdavid @omarsar0 @nsxdavid this is my mental model as well. I’m having a hard time keeping track of review comments and suggestions when sharing these artifacts across my global team. What is your experience with the human to human collaboration side of artifact reviews?

@omarsar0 There are many places where HTML artifacts can be very useful, even fun. But I argue it is not the right default artifact file type for a few very good reasons, here: aicrier.com/notes/markdown…

@svpino Is Framework sponsoring all of your posts about their laptop?

Battery life is a solved problem with Framework 13 Pro running on Intel Panther Lake and Omarchy. I took the laptop on my second trip and here are the numbers: • I charged it before leaving home. • I used it for 1.5 hours at the airport. • Then, used it for about 5 hours during the flight. • Then, used it for about 1 hour the next morning. I plugged the laptop with 36%. My usage was mostly browsing the web and social media, about 1 hour of Claude Code and Visual Studio Code, and I had a docker container running a database all the time. Honestly, this is more battery life than what I need. I can leave my house early on without carrying a charger and know I'll be back in the afternoon with enough juice.

@jun_song @mervenoyann @huggingface I agree.

@mervenoyann @huggingface 네 가장 큰 문제는 대부분의 사람들은 safetensor 원본버전이 아닌, 본인 하드웨어에 맞는 양자화 버전을 다운로드 받는다는겁니다. 당장 trending 페이지만 봐도 safetensor가 아닌 모델이 수없이 많습니다. 그래서 저는 “Random model”을 다운로드 받지 말라고 한거에요. 그것에 문제가 있나요?

저는 @huggingface 를 굉장히 높게 평가하고 가장 좋아하는 기업중 하나이지만, 보안 취약점 발생에 대한 이 직원분의 대응은 믿을수가 없습니다. 공격자들의 저장소로 사용이 되었고, 그것을 지적하는 사람에 대해 멍청이 취급을 하고있어요. ”우리는 보안스캐너가 있습니다“. 이것은 해킹사건에 대한 적합한 대응이 아닙니다.

@JimZarkadas I’d enjoy hearing about what that transition looked like. Do you still use figma for UI/UX mockups before jumping into Claude to prototype?

We hired an AI mentor recently to help us become a truly AI-native design team. To transition from designers living in Figma and basic prototypes to designers shipping with Claude and prototyping real interactive tools. It's not that we stopped using Figma or that suddenly we only code but we are now even more focused on building *interactive and realistic* prototypes that would take dozens of hours in Figma and on getting things SHIPPED to production. This week we opened our first Pull Request at ZenMaid and most of the code got accepted with only a few comments 😍 This is a big milestone and it's super exciting for us! As product and growth designers, we can have more impact, get closer to the dev team and get unblocked and ship more growth experiments and more UX improvements. We are using AI to create MORE VALUE and NOT to 'speed up' or 'cut costs'. AI is not eating our job. And AI is NOT doing design for us. AI is a tool we use to double down on craft, growth and impact and it's just unblocking us and giving us new skills we didn't have before. And it's freaking amazing! Will share more in the coming months :)

@natmiletic @WebsiteSquadron Yes!!

@WebsiteSquadron Agreed

If you use cPanel you should update your server right away. 🚨 There is a critical vulnerability being exploited and sites are getting wiped out.

@karpathy @jenzhuscott nealstephenson.com/fall,-or-dodge…

Yes it's the tractable form of brain upload. There's a ton of scifi on brain uploads that requires way too exotic tech (scanning and simulating brains etc), when we're about to get a lossy and approximate version of that *a lot* sooner via LLM simulators. You can easily imagine a "brain upload" startup - you show up for a few days to carry out detailed video interviews, then they use all that data with an LLM finetuning process to "upload" you and give you an API endpoint of your simulation that you can talk to. Look at what's already possible with HeyGen as an example, but combine it with an LLM model that has deep knowledge and personality. Trippy and admittedly kind of dystopian but in principle quite possible around now.

As I build my own 2nd brain 🧠 on Obsidian using @karpathy ‘s wiki idea, it suddenly dawned on me - one day when we r gone, our kids could inherit an interactive map to your mind, passion, obsessions, work, fascinations… It’s kind of beautiful way to think abt your 2nd 🧠.

@dvassallo @BrianMSkinner Really?

@BrianMSkinner Funny I don’t have it because my claw signed up for it on its own with its browser. I didn’t even see it. But I think you have to choose DIY and then opt for the Premium plan under DIY. You don’t get to see this option if you say you have a business through their onboarding quiz.

OpenClaw just saved me $500 by unburying the $129 TurboTax plan I actually needed. Intuit was trying to sell me a $639 package via a recommendation quiz that said there's no other option for my situation. There was!

@natmiletic @natmiletic this is great. My first thought, however, would be how does my domain rank against other companies in the same (competitive industry and region).

We are super excited to add a domain rating score in our free website assessment tool Check it out: cliowebsites.com/free-website-a…

@levelsio @StevieZollo My idea for a business is called fast food. You ask for a hamburger and fries and you receive a paper bag with fresh food in under a few minutes. The idea is to make it a business.

✨ To prove my friend @StevieZollo (who's visiting me in Brazil) you don't need an idea, or even a lot of time these days to ship a little app that might make money I took the top idea from IdeasAI.com: "A startup that uses AI to generate personalized bedtime stories for kids based on their interests, family photos, and daily activities, delivered via a voice app. (❤️ 110 likes, 3 days ago)" So I copy pasted it into Claude Code and asked it to build it The first version of course didn't work, and I had to tell it some endpoints didn't work properly but then it fixed it The bedtime stories are generated by @xAI Grok 4.1, then sent to TTS with @GoogleAI Gemini and payment with @Stripe Checkout Total time from start to live: 24 minutes

@aakashgupta CLAUDE DM

JUST IN: Claude Code hit $2.5B ARR in 12 months. That makes it the fastest growing software ever. Anthropic buried the Claude Code ARR news in its announcement today of raising $30B at a $380B valuation, but I think it might just be the most important news of the day. Here's your guide to getting started: 1. Understand what you get It's not just for coding. It's an LLM with the best agentic harness: file system access, the ability to connect to MCPs like your analytics and docs tools, and autonomous planning. Karpathy called it "a little spirit/ghost that lives on your computer." 2. Setup in 5 minutes Install with npm install -g @anthropic-ai/claude-code, run "claude," log in with Claude Pro ($20/mo), navigate to your project. You can also download the native app if you prefer skipping the terminal install. 3. Start by asking questions, not giving orders Ask "what does this codebase do?" before requesting changes. Let Claude map the architecture first. Then try "find the bug in UserAuth.js" and watch it pinpoint the exact line and fix. 4. Update your CLAUDE.md This is your secret weapon. It gives Claude your project context, coding standards, common commands, team conventions, and architecture notes. Claude reads it first every session. Run /init to auto-generate one, then keep customizing it for your workflow. 5. Level up with advanced features - Hit Shift+Tab to make Claude think before coding. Review the approach, then execute. Use message queuing to stack multiple requests. - Use /clear to reset context without losing your CLAUDE.md. - Spin up subagents for specialized tasks like testing, security, and docs. 6. Master the use cases for you For PMs, these include: • Draft PRDs from templates • Parse CSVs for feature requests • Prep interview questions by company and level • Run competitive research across 5 pricing pages • Build working React prototypes from wireframes When you engineer the context properly, you'll find the output is 10x regular Claude or ChatGPT. Here's everything you need to go from 0 to 1: 1. Claude Code basics: youtube.com/watch?v=4nthc7… 2. Claude Code advanced: youtube.com/watch?v=59gy_2… 3. Claude Code operating sys: news.aakashg.com/p/pm-os 4. Claude Code workflows: news.aakashg.com/p/how-to-use-c… 5. Claude skills: news.aakashg.com/p/steal-6-of-m… 💬 Comment "CLAUDE" + DM for my Claude.md. ➕ Follow @aakashgupta for daily AI PM insights. Claude Code is the fastest growing product of all time. You're not too late to learn it.

@mervenoyann @huggingface MobileNets?

I'm keeping a track of real-time vision models (mostly detectors) on @huggingface we have RT-DETR, YOLO, RF-DETR and D-FINE for now what other models should we add?

@trq212 Nice! Looks amazing.

@k_manninen Yeah it just runs android

my DC-1 is quickly becoming one my favorite things