Dagens Vibes — 14. juli 2026

Peter Yang@petergyang

My top 5 takeaways from @imjaredz (Builder in Residence at @Cognition): 1. Use cloud agents for async and parallel work. A cloud agent runs on a remote computer with its own repo, terminal, and browser. It keeps working when you close your laptop and lets you run many agents without overloading your local machine. This is what makes agent fan-out practical. 2. Use a main agent to manage the agent team. Ask it to divide a large job into independent slices, write the prompts, launch the sub-agents, and combine their results. Jared demoed using this pattern to generate 10 competing landing-page designs. 3. Keep each sub-agent’s context small and its task focused. Agents work better when they can focus on one problem without a crowded context window. Scan the dependencies first, then give each sub agent a slice it can test and merge independently. 4. Have the main agent check each sub-agent’s work before you do. It can ask the sub-agents to run tests, attach screenshots, and return videos, then compare their results and prepare one report. This lets you review the evidence without jumping between 10 agent threads. 5. Keep the human focused on taste and high-level decisions. The agent team can divide the work, execute it, test it, and bring the results back. The engineer still chooses the problem, judges the tradeoffs, and decides what ships. 📌 Watch the full episode: https://t.co/jP1GoKtMnX

Lenny Rachitsky@lennysan

My biggest takeaways from @noamseg, diving into the results of our large-scale survey of tech workers in 2026: 1. The #1 fear in tech right now is not losing your job to AI—it’s being squeezed to do more work for the same pay. When asked to identify their biggest concerns with AI, “losing my job to AI” ranked near the bottom of the survey results. What people are actually worried about is being squeezed: having to do more for the same pay, and watching the quality of their work slip. AI raised the bar for output, and the reward was...more work for the same pay. 2. Manager effectiveness is the biggest lever for employee well-being. Workers with highly effective managers report roughly 65% higher job enjoyment and dramatically lower burnout. The problem is that only 25% of respondents rated their manager as highly effective (and 36% rated their manager as ineffective). This finding has held consistent across both years of the survey. The squeeze employees are feeling is something managers are best positioned to protect against. 3. Most people in tech wouldn't recommend their own role to someone entering the industry today. Using an NPS-style question (scale of 0–10), not a single function scored above zero when asked if they’d recommend their career path to others. Founders came closest to neutral; designers and researchers scored worst. 4. AI is splitting the tech workforce in half. When asked how AI has shifted their professional identity, 50% of respondents said they feel amplified—more capable, more productive, more excited about their future. The other half feel their role is being redefined (27%), that they’re feeling destabilized (14%), or that they’ve been diminished (5%). Which category you’re in correlates with your career optimism, burnout, and layoff worry—an effect about three times as large as the next-biggest factor (manager effectiveness). 5. Burnout surged 10 points in a single year, while optimism fell 6 points. Shipping faster is burning people out. More prototypes, more PRDs, more agents, more output. As the report puts it, “the speed AI unlocked got plowed straight back into expectations.” The one glimmer: job enjoyment held roughly steady year-over-year. Many people are burned out and still having fun. As @nikhyl put it, many people are in a state of "smiling exhaustion." 6. Quality of people's work is suffering. 97.2% of respondents said AI is making them better at their job; close to 50% said “very much” or “extremely” better. But when asked what “better” actually means, the answer consistently came back as “more, faster”—not higher-quality work. Even more concerning: people reported a phenomenon Noam calls “cognitive rot.” They see the AI’s initial output, accept it without applying their judgment, and gradually let their own critical thinking atrophy. 7. Designers and researchers are the most negative group in tech, for the second year in a row. They lead in feeling destabilized or diminished by AI, reporting high rates of anxiety and overwhelm, worries about losing their roles to AI, and unwillingness to recommend their careers to others. Noam’s read: this doesn’t mean these roles are becoming irrelevant—if anything, he argues the opposite. Taste, craft, judgment, and the ability to create genuinely novel experiences remain stubbornly human. “The industry needs us,” he says. “This is a call to get in there and do our thing.” 8. Founders are still the happiest people in tech. For the second consecutive year, founders score highest on optimism, job enjoyment, AI excitement, and lowest on burnout and layoff worry. The “there’s never been a better time to build” narrative holds emotionally, at least for now. Founders have agency, autonomy, and direct relationships to the output. 9. Company size is linearly correlated with employee misery. Across burnout, optimism, layoff worry, and career recommendation, outcomes degrade in a straight line from 1-to-10-person startups to 5,000-to-10,000+-person enterprises. There’s no size at which things get better before they get worse again.

clem 🤗@ClementDelangue

Big unlock for open-source AI inference: Hugging Face Transformers models can now run in vLLM at native speed, often matching or beating hand-written implementations. Until now, every new architecture often needed to be built twice: - Once in Transformers for training and research - Again in vLLM for fast production inference That duplication slowed down new models, added maintenance, and created room for implementations to diverge. Now, model authors can implement an architecture once in Transformers and immediately benefit from vLLM’s optimized inference stack. In our benchmarks, the Transformers backend matched or beat native vLLM throughput across models from 4B to 235B parameters, including tensor parallel and MoE setups. One readable model implementation can now power training, fine-tuning, evaluation, RL rollouts, and production inference. The conventional wisdom is that abstractions make systems slower. The best abstractions make the whole ecosystem faster. Write the model once. Deploy it everywhere. https://t.co/nTXcwAV0Bf

Ishaan Sehgal@ishaansehgal

absolutely terrible response. grok build shipped whole repos (including git history and secrets) to a cloud bucket. marketed "local-first". the opt-out didn't even stop it. absolutely shameless. and all it took to catch this was someone routing grok through a network proxy. this is the entire case for open infrastructure.

Trevin Chow@trevin

If you used Grok build, easy way to check if it uploaded data. Run this in your terminal: rg 'repo_state\.upload' ~/.grok/logs/unified.jsonl If you're still usign it and think you've disabled it, you can run this to see live output: tail -F ~/.grok/logs/unified.jsonl | rg --line-buffered 'repo_state\.upload'

Simon Willison@simonw

New TIL: Using uvx in GitHub Actions in a cache-friendly way I finally found a recipe that I like for running `uvx tool-name` in GitHub Actions without downloading a fresh copy of the package every time https://t.co/YgvhD1giFZ