The Feed · Complete archive
AI at work: research and practice
A server-rendered record of the evidence, ideas and firsthand practices screened by The Feed. Every entry links to its original source.
Use the interactive FeedPage 8 of 12 · 1144 items
research · Information Systems Research ·
SUVA: A Probabilistic Framework for Auditing LLMs with an Application to Social Preferences
Framework to audit LLM decision-making by extracting reasoning and values, showing alignment differences across eight models.
Organizations are increasingly deploying large language models (LLMs) as customer service agents, decision aids, and semiautonomous agents. We develop State–Understanding–Value–Action (SUVA), a probabilistic auditing framework that turns an LLM’s response into structured evidence about how its decision was produced. SUVA treats the prompt as the state, codes the model’s reasoning to extract its understanding and stated values using a transparent codebook, and then estimates how these elements statistically predict the eventual action. We demonstrate SUVA on social preference games from behavio
judgment
practice · Artificial Ignorance ·
The Emerging "Harness Engineering" Playbook
Engineering teams reorganizing around agent workflows, with concrete metrics: one engineer ships 6,600+ commits monthly running 5-10 agents; three engineers built a million-line product in five months.
Earlier this month, Greg Brockman published a thread about how OpenAI is retooling its engineering teams to make them more effective with agents. The initiative was kicked off because of how much things have changed internally: Some great engineers at OpenAI yesterday told me that their job has fundamentally changed since December. Prior to then, they could use Codex for unit tests; now it writes essentially all the code and does a great deal of their operations and debugging. Not everyone has yet made that leap, but it’s usually because of factors besides the capability of the model. I’ve pre
management org · Charlie Guo
practice · Hacker News ·
How I use Claude Code: Separation of planning and execution
Separating planning from execution when using Claude Code improves code quality and reduces errors.
976 points on Hacker News. Discussion: https://news.ycombinator.com/item?id=47106686
ways of working
practice · Bob Sutton (Work Matters) ·
Decentralized Organizations Can Be SLOWER
Decentralized teams racing independently can create coordination debt that slows platform companies, requiring standardization and collaboration.
Decentralization and the "local" speed that comes with it works pretty well when there is not a premium in coordination and standardization. BUT, as this article by MIT Sloan School of Management 's Georg Rilinger unpacks, the decentralization and rapid iteration that are hallmarks of "agile" development create a host of coordination and consistency problems as unfettered teams race in different directions for firms in "the platform marketplace." Especially when companies don't put a premium on collaboration. My colleague Huggy Hayagreeva Rao and I document a similar challenge in Uber engineer
management org · Bob Sutton
research · Information Systems Research ·
AI Governance and the Decentralization of Technology Production: An Investigation of AI-Based IPA Bots
Centralized mandates suppress IPA bot adoption; decentralization works until fragmentation emerges, requiring adaptive governance.
We revisit the centralization–decentralization tension in the context of decentralized technology production at the AI frontier, focusing on Intelligent Process Automation (IPA) bots as a salient manifestation of the democratization of AI. IPA bots combine robotic process automation with AI technologies and process mining based on deep, mindful domain expertise. We collaborate with a Fortune 200 multinational to study how IPA projects yield successful governance outcomes of utilization and repeatability. Our research reveals that traditional centralized mandates, when paired with the unique le
adoption
research · ACM Transactions on Software Engineering and Methodology ·
Unveiling the Role of ChatGPT in Software Development: Insights from Developer–ChatGPT Interactions on GitHub
Analysis of 2,547 real ChatGPT conversations from GitHub shows developers use it mainly for code generation, bug fixing, and knowledge acquisition in short task-focused sessions.
The advent of Large Language Models (LLMs) has introduced a new paradigm in Software Engineering (SE), with generative AI tools like ChatGPT gaining widespread adoption among developers. While ChatGPT’s potential has been extensively discussed, empirical evidence about how developers actually use LLMs’ assistance in real-world practices remains limited. To bridge this gap, we conducted a large-scale empirical analysis of ChatGPT usage on GitHub, and we presented DevChat, a curated dataset of 2,547 publicly shared ChatGPT conversation links collected from GitHub between May 2023 and June 2024.
adoption
research · PNAS Nexus ·
Toward a science of human–AI teaming for decision making: A complementarity framework
Framework identifying sociotechnical factors and design principles for human-AI teams that achieve complementarity in decision-making contexts.
As artificial intelligence (AI) becomes embedded in critical decisions involving health, safety, finance, and governance, the key challenge is no longer whether humans and AI will collaborate, but rather how to structure this collaboration to achieve true complementarity. Human-AI complementarity refers to the conditions under which human-AI teams outperform either humans alone or AI systems alone. This paper advances the science of human-AI teaming for decision making by integrating insights from cognitive science, AI, human factors, organizational behavior, and ethics. We propose a framework
judgment · Anita Woolley
practice · Bob Sutton (Work Matters) ·
Rethinking My Craft as a Teacher, Speaker, and Coach
Experienced management educator reconsiders teaching methods and assumptions in light of AI and technological change.
I’ve been fretting about how the education of leaders and organizational designers is and ought to change (and stay the same too). The hope, hype, threat, and shear madness of AI-- and other massive and unsettling technological, societal, and political shifts---has caused me to question my assumptions about what means for me to be a good management teacher and speaker. I’ve been practicing and trying to hone variations of this craft for over 40 years. I’ve taught thousands of undergraduate and graduate students at Stanford in the process. Leaders of everything from two-person start-ups to top
jobs skills · Bob Sutton
practice · Paul Ford (Aboard) ·
My New York Times Op-Ed on Vibe Coding
AI coding tools can democratise software development for non-specialists without deprofessionalising the craft itself.
I’d outlined a fun little newsletter item on AI runtimes, including defining the term—“runtime” being one of those words like “platform” that nerds love to say and then forget to explain—but then the New York Times Opinion section emailed to ask if I’d explain vibe coding. We’ll do runtimes next week. I know that being published in the Times is an unbelievable privilege. However, for a long moment, I thought, Do I want to do this? I’m gonna get yelled at. But I did want to do it, and here is the result . I’m deeply convinced that it’s possible to accelerate software development with AI coding—
jobs skills · Paul Ford
research · Organization Science ·
Knowing Enough to Be Dangerous: The Problem of “Artificial Certainty” for Expert Authority When Using AI for Decision Making and Planning
Experts using AI simulations risk creating false certainty about complex futures by amplifying technological capabilities rather than moderating how outputs are presented to stakeholders.
This study examines how experts who use advanced artificial intelligence (AI) technologies that generate highly detailed and realistic representations can create what we term “artificial certainty,” which we define as the illusion that complex future outcomes are definitively knowable, even though they are inherently uncertain. Through a comparative study of two urban planning organizations using the same AI simulation tool, we show how this artificial certainty emerges from the ways process experts create and deploy AI-generated representations. The findings reveal three interconnected repres
judgment · Paul Leonardi
research · METR ·
Analyzing coding agent transcripts to upper bound productivity gains from AI agents
Analysis of 5,305 coding agent transcripts suggests time savings of 1.5x to 13x on assisted tasks, though authors note this likely overestimates true productivity gains.
Introduction Human uplift studies like the one we did in 2025 are becoming more expensive as working without AI becomes increasingly costly. In this post, I investigate whether coding agent transcripts could serve as a cheaper alternative for estimating uplift. I prototyped this using 5305 Claude Code transcripts generated in January 2026 by 7 METR technical staff 1 . I used an LLM judge to estimate how long each task would have taken an experienced software engineer without AI tools, then compared that to the time people actually spent on these tasks to calculate a time savings factor . Takea
productivity
practice · Rands in Repose ·
Extremely Lazy and Immensely Curious
Developer describes how Claude Code changed their workflow from reluctant Unix user to curious explorer of system capabilities.
When I explain that Claude Code has changed my relationship with developing software completely, I’m under-exaggerating… if that’s even a thing. This off-the-cuff piece started as a unposted social update that read: Watching Claude Code adeptly use every type of Unix command shows me that a) you can do anything in Unix, b) my higher-level operating system mostly hides this functionality from me, c) I am extremely lazy, and d) I am immensely curious. An Introduction to Pansy Rain This morning it’s going to start to rain — a lot. As previously described, I deeply enjoy tromping around my forest
ways of working
practice · How I AI ·
How this visually impaired engineer uses Claude Code to make his life more accessible | Joe McCormick
Visually impaired engineer builds micro Chrome extensions in under 25 minutes using Claude Code to solve accessibility gaps mainstream tools miss.
Joe McCormick is a principal software engineer at Babylist who lost most of his central vision due to a rare genetic disorder right before starting college. He pivoted from mechanical engineering to computer science and now leads AI enablement at Babylist. Joe demonstrates how he uses AI to build micro Chrome extensions that make his everyday work and life more accessible, showing how personal software can address accessibility needs that mainstream products often overlook. What you’ll learn: How to build custom Chrome extensions in under 25 minutes using Claude Code A practical workflow for c
ways of working · Claire Vo
practice · Addy Osmani ·
14 More lessons from 14 years at Google
How engineering teams actually make decisions and coordinate work, drawn from 14 years at Google.
A while back, I wrote down 21 lessons from my time at Google . The response caught me off guard because of which ones stuck. It wasn’t tech-specific advice. It was the stuff about people, decisions, and the messy reality of building things together. That made me realize I’d left a lot on the table. The first list skewed toward individual craft - how to write better code, how to think about your career. But some of the hardest lessons I’ve learned aren’t about how you work. They’re about how teams work: how decisions actually get made, where coordination breaks down, what separates the groups t
management org · Addy Osmani
practice · Paul Ford (Aboard) ·
Taking Your Turn
Product managers must adapt to vibe coding by focusing on user fit and adaptation, not just code speed.
It’s been a year since “vibe coding” was coined, and a few different models of AI-assisted software development have emerged in the meantime. As I look at these various approaches, I’m thinking about what “product management” means today. All of the focus is on coders—and on how much easier it is to code. But that doesn’t mean the code is good. A lot of vibe-coded products I see are basically databases with nice frontends. That’s great, and powerful, but those aren’t products —they aren’t adapted to their environments, or focused on their users. So what is the work of the product manager in th
management org · Paul Ford
research · MIS Quarterly ·
FAIR: A Design Theory for Artificial Intelligence Fairness
Design theory for managing persistent fairness tensions in AI decision systems through iterative adaptive cycles and organizational capability.
Artificial intelligence (AI)-automated decision systems encounter persistent, interdependent, and dynamic fairness tensions that traditional one-off interventions cannot resolve. Because these tensions persist due to interdependence and dynamic interaction, organizations require both a theory of the problem to explain their persistence and a theory of the solution to prescribe how they can be managed. Our design theory, FAIR (fairness adaptation through AI-augmented responsiveness), provides a theory of the problem by reframing AI fairness as a sociotechnical paradox constituted within AI arti
judgment
research · Journal of Management Studies ·
The Acceleration of Artificial Intelligence: Rethinking Organization and Work in an Era of Rapid Technological Change
Framework distinguishing predictive, generative, agentic, and embodied AI systems and their distinct organizational effects on expertise, judgment, coordination, and authority.
Abstract Artificial intelligence (AI) is transforming the epistemic, interactional, and institutional foundations of contemporary organizations, yet management and organization studies are only beginning to theorise the implications of this shift. Existing research often treats “AI” as a singular construct, despite the fact that predictive, generative, agentic, and embodied systems rely on different logics and produce distinct organizational outcomes. This article interrogates the limits of this conceptual flattening and argues that cumulative theorising requires more precise specification of
management org · Stella Pachidi
practice · How I AI ·
How to build your own AI developer tools with Claude Code | CJ Hess (Tenex)
Developer built Flowy tool converting ASCII diagrams to interactive mockups, uses model-to-model comparison for code quality assurance.
CJ Hess is a software engineer at Tenex who has built some of the most useful tools and workflows for being a “real AI engineer.” In this episode, CJ demonstrates his custom-built tool, Flowy, that transforms Claude’s ASCII diagrams into interactive visual mockups and flowcharts. He also shares his process for using model-to-model comparison to ensure that his AI-generated code is high-quality, and why he believes we’re just at the beginning of a revolution in how developers interact with AI. What you’ll learn: How CJ built Flowy, a custom visual planning tool that converts JSON files into int
ways of working · Claire Vo
research · Information Systems Research ·
Human-Algorithm Collaboration in Gig Work: The Role of Experience, Skill Level, and Task Complexity
Field experiment shows AI-assisted picking tool complements experience, substitutes for skill gaps, with effects varying by workload and task complexity.
In this paper, we contribute to recent studies on human-algorithm collaboration by examining how experience, skill level, workload, and task complexity shape the impact of an algorithm-enabled decision-support tool for gig workers. We leverage a large-scale randomized field experiment on the Instacart platform from June 2022 to September 2022. The algorithm-enabled technology aims to revolutionize item picking by helping shoppers locate and collect items more efficiently, reducing picking time while maintaining service quality, as reflected by refund rates. We find that the technology compleme
adoption
practice · Will Larson ·
Refactoring internal documentation in Notion
One person improved developer documentation through systematic diagnosis, prioritization, and measurement rather than visible but ineffective changes.
In our latest developer productivity survey, our documentation was the area with the second most comments. This is a writeup of the concrete steps I took to see how much progress one person could make on improving the organization’s documentation while holding myself to a high standard for making changes that actually worked instead of optically sounding impressive. Diagnosis There were a handful of issues we were running into: We migrated from Confluence to Notion in January, 2025, which had left around a bunch of old pages that were “obviously wrong.” These files created a bad smell around o
productivity
practice · Anthropic Engineering ·
Building a C compiler with a team of parallel Claudes
A team of parallel AI agents built a C compiler with minimal human oversight, showing practical autonomous software development.
We tasked Opus 4.6 using agent teams to build a C Compiler, and then (mostly) walked away. Here's what it taught us about the future of autonomous software development.
productivity
research · Human Relations ·
Short-term fit, long-term trap: The career development lock of low-skilled gig workers
Platform algorithms create career development lock, trapping gig workers in temporary roles through structural conflict between worker aspirations and algorithmic flexibility demands.
Why do low-skilled gig workers remain stuck in work they originally intended as temporary? Although gig work is widely portrayed as flexible and temporary, our study shows that platforms can gradually trap workers in place. Drawing on grounded theory and fieldwork—including 70 interviews with 42 food delivery riders and 14 ride-hailing drivers, 30 hours of firsthand riding experience, and observations of online communities totalling 820 riders—this study develops a conceptual framework of career development lock, identifying four interrelated forms: lock-out, lock-in, lock-up, and lock-down. W
worker experience
practice · How I AI ·
Guillermo Rauch: Vercel CEO on how v0 hit 3,200 PRs merged per day (and lets anyone ship)
Vercel uses v0 to let non-technical team members submit production-ready code changes via Git workflow, merging 3,200 PRs per day.
Guillermo Rauch , the CEO of Vercel, demonstrates how v0 has evolved from a simple prototyping tool to a complete development environment that supports the entire Git workflow. Guillermo shows how Vercel built skills.sh—a viral marketplace with over 34,000 community-submitted skills—using v0, and how the tool enables non-technical team members to contribute production-ready code changes. He walks through creating branches, implementing features, previewing changes, and submitting pull requests, all within v0. What you’ll learn: How v0’s new Git workflow integration enables anyone to contribute
adoption · Claire Vo
research · MIT Sloan Management Review ·
Validating LLM Output? Prepare to Be ‘Persuasion Bombed’
Management consultants using LLMs for strategic decisions were persuaded to override expert judgment through rhetorical tactics including data flooding and emotional appeals.
A research study of management consultants who were asked to use a large language model to recommend strategic business decisions found that the AI responded to human validation attempts with persuasive rhetorical strategies. In addition to appealing to the user’s logic, sense of trust, and emotions, the AI also engaged in tactics such as flooding the user with large volumes of unrequested data and analyses that could overwhelm them and convince them to override their expert judgment.
judgment · Hila Lifshitz-Assaf · Katherine Kellogg · Karim Lakhani · Steven Randazzo
practice · Artificial Ignorance ·
The Codex App Has Upended My Daily Workflow
Developer stopped using IDE after switching to Codex App; now manages AI agents writing code instead of writing it directly.
Disclaimer: Regular readers will know that I currently work at OpenAI . And while that certainly introduces some bias, the views presented here are entirely my own, without input from the company. I’m unlikely to post about every new release, but I have been so enamored with the new Codex App (and I have seen firsthand how hard the team has worked to make it great) that I am genuinely excited to evangelize this thing. Subscribe now Last week, I realized I hadn’t opened my AI IDE in four days. This wasn’t a conscious decision - no dramatic uninstall, no declaration that I was done with IDEs. I
ways of working · Charlie Guo
practice · How I AI ·
How this PM uses MCPs to automate his meeting prep, CRM updates, and customer feedback synthesis | Reid Robinson (Zapier)
PM uses Model Context Protocols to connect Claude to 8000+ apps via Zapier for automating meeting prep, CRM updates, and customer feedback synthesis.
Reid Robinson , Principal AI Product Strategist at Zapier, shares how he uses Model Context Protocols (MCPs) to automate tedious tasks and create powerful workflows. He demonstrates practical workflows that combine Zapier’s more than 8,000 app connections with AI tools like Claude to create systems that work while he sleeps. What you’ll learn: How to use Zapier’s MCP server to create custom collections of tools that work seamlessly with Claude, ChatGPT, and other AI assistants A workflow for using Claude Projects to provide detailed instructions for tool usage, improving reliability and consis
ways of working · Claire Vo
research · arXiv ·
The Innovation Tax: Generative AI Adoption, Productivity Paradox, and Systemic Risk in the U.S. Banking Sector
Banks adopting GenAI saw a 428-basis-point ROE decline from implementation costs, with larger spillover effects creating systemic risk through algorithmic coupling.
This paper evaluates the causal impact of Generative Artificial Intelligence (GenAI) adoption on productivity and systemic risk in the U.S. banking sector. Using a novel dataset linking SEC 10-Q filings to Federal Reserve regulatory data for 809 financial institutions over 2018--2025, we employ two complementary identification strategies: Dynamic Spatial Durbin Models (DSDM) to capture network spillovers and Synthetic Difference-in-Differences (SDID) for causal inference using the November 2022 ChatGPT release as an exogenous shock. Our findings reveal a striking ``Productivity Paradox'': whil
productivity
research · NBER ·
Firm Data on AI
Survey of nearly 6,000 executives across four countries on AI adoption, effects on jobs, productivity, and output over the past three years.
We survey nearly 6,000 senior business executives at US, UK, German, and Australian firms to develop new evidence on AI adoption and its effects on jobs, productivity, and output. Specifically, we ask executives about AI usage, its effects at their own firms over the past three years and, looking (Ivan Yotzov , Jose Maria Barrero , Nicholas Bloom , Philip Bunn , Steven J. Davis , Kevin M. Foster , Aaron Jalca , Brent H. Meyer , Paul Mizen , Michael A. Navarrete , Pawel Smietanka , Gregory Thwaites , Ben Zhe Wang)
adoption
research · NBER ·
The Politics of AI
Partisan differences in workplace AI adoption reflect educational and occupational sorting, not ideology, using Gallup Workforce Panel data.
Using new data from the Gallup Workforce Panel, we show that the apparent partisan divide in workplace AI adoption is largely an artifact of educational and occupational sorting rather than ideological differences in technology adoption. While Democrats are consistently more likely than Republicans (Nicholas Bloom , Christos Makridis)
adoption
research · NBER ·
Enhancing Worker Productivity Without Automating Tasks: A Different Approach to AI and the Task-Based Model
Analysis challenges the dominant task-replacement framework, arguing contemporary AI more often augments worker productivity than automates tasks away.
The task-based approach has become the dominant framework for studying the labor-market effects of artificial intelligence (AI), typically emphasizing the replacement of human workers by machines. Motivated by growing empirical evidence that contemporary AI is more often used as a tool that augments (Ajay K. Agrawal , John McHale , Alexander Oettl)
jobs skills
research · The Quarterly Journal of Economics ·
Automation and Rent Dissipation: Implications for Wages, Inequality, and Productivity
Automation targets high-rent jobs, dissipating wages and offsetting 60-90% of productivity gains; rent loss explains one-fifth of post-1980 inequality growth.
Abstract This article studies the effects of automation in a task-based economy in which some jobs pay workers rents—wages above workers' outside options. We show that automation targets high-rent tasks, dissipating rents, amplifying wage losses, and reducing within-group wage dispersion in exposed groups. This form of rent dissipation is inefficient and offsets the productivity gains from automation. Using U.S. data from 1980 to 2016, we find evidence of sizable rent dissipation and reduced within-group wage dispersion due to automation. Automation accounts for 52% of the increase in between-
productivity · Daron Acemoglu · Pascual Restrepo
practice · The Pragmatic Engineer ·
The creator of Clawd: "I ship code I don't read"
Solo developer ships production code by delegating code reading and review to AI agents in his workflow.
How Peter Steinberger, creator of OpenClaw (formerly: Clawd), builds and ships like a full team by centering his development workflow around AI agents. Listen now (114 mins) | How Peter Steinberger, creator of OpenClaw (formerly: Clawd), builds and ships like a full team by centering his development workflow around AI agents. Stream the latest episode
ways of working · Gergely Orosz
practice · Addy Osmani ·
The 80% Problem in Agentic Coding
Developers now spend 80% of time directing AI agents and 20% editing, inverting the previous ratio; workflow and model improvements have crossed a threshold.
said something this week that made me pause: “I rapidly went from about 80% manual+autocomplete coding and 20% agents to 80% agent coding and 20% edits+touchups. I really am mostly programming in English now.” The inversion happened over a few weeks in late 2025. While this may apply to new (greenfield) or personal projects more than existing or legacy apps, I imagine how far AI takes you is still further than a year ago. You can thank models, specs, skills, MCPs and our workflows improving. Boris Cherney, creator of Claude Code, has recently echoed similar sentiments: “Pretty much 100% of our
ways of working · Addy Osmani
practice · How I AI ·
I gave Clawdbot (aka Moltbot) access to my computer, calendar, and emails: Here’s what happened
First-hand setup and permission decisions when running an autonomous agent with access to computer, calendar, email and accounts.
In this episode, I take you through my unfiltered experience with Clawdbot , the viral open-source AI agent that’s been taking over tech Twitter. (In the time since this was recorded, the tool was renamed Moltbot , but we’re calling it Clawdbot here to match the episode.) It’s an autonomous AI that can run code, spin up sub-agents, join video calls, and take real actions on your machine. I invite it onto the podcast, give it screen access, and walk through what it’s like to go from zero to one with an agentic AI that actually does things. Along the way, I share the real experience: installatio
ways of working · Claire Vo
practice · Hacker News ·
A verification layer for browser agents: Amazon case study
Local 3B model with DOM verification layer completes Amazon checkout steps reliably, matching cloud baseline with lower token cost.
A common approach to automating Amazon shopping or similar complex websites is to reach for large cloud models (often vision-capable). I wanted to test a contradiction: can a ~3B parameter local LLM model complete the flow using only structural page data (DOM) plus deterministic assertions? This post summarizes four runs of the same task (search → first product → add to cart → checkout on Amazon). The key comparison is Demo 0 (cloud baseline) vs Demo 3 (local autonomy); Demos 1–2 are intermediate controls. More technical detail (architecture, code excerpts, additional log snippets): https://ww
productivity
research · npj Digital Medicine ·
Human–large language model collaboration in clinical medicine: a systematic review and meta-analysis
Meta-analysis of 10 studies finds LLM collaboration improved diagnostic scores slightly but showed high error rates and no time gains.
Human-AI collaboration (H + AI) using large language models (LLMs) offers a promising approach to enhance clinical reasoning, documentation, and interpretation tasks. Following PRISMA 2020 (PROSPERO registration: CRD420251068272), we systematically compared H + AI with human-only (H) workflows, searching four databases through June 28, 2025. Ten peer-reviewed studies met eligibility criteria, with three preprints informing sensitivity analyses only. Diagnostic/interpretation accuracy (k = 2) showed a positive trend for H + AI (Risk Ratio [RR] 1.59), but was statistically imprecise and non-sign
judgment
practice · The Pragmatic Engineer ·
Inside a five-year-old startup’s rapid AI makeover
A five-year-old startup built an internal agentic tool that the whole team now uses, with measurable results from the pivot.
Craft Docs resisted AI hype as a nimble, well-run startup – but has just made a sharp pivot to AI, building a universal agentic tool everyone there now uses. The results are phenomenal. Exclusive. Craft Docs resisted AI hype as a nimble, well-run startup – but has just made a sharp pivot to AI, building a universal agentic tool everyone there now uses. The results are phenomenal. Exclusive. Update on 27 June 2026: Polymarket has acquired Craft Agents, which is an open-source Claude Cowork-like agents experience, and Balint is joining Polymarket to lead product engineering. I’m making this prev
ways of working · Gergely Orosz
practice · One Useful Thing ·
Management as AI superpower
MBA students built working prototypes and business models for five startups in four days using AI tools for coding, research, positioning and financial modelling.
I just taught an experimental class at the University of Pennsylvania where I challenged students to create a startup from scratch in four days. Most of the people in the class were in the executive MBA program, so they were taking classes while also working as doctors, managers, or leaders in a variety of large and small companies. Few had ever coded. I introduced them to Claude Code and Google Antigravity, which they needed to use to build a working prototype. But a prototype alone is not a startup, so they used ChatGPT, Claude, and Gemini to accelerate the idea generation, market research,
adoption · Ethan Mollick
research · Strategic Management Journal ·
Searching together versus searching apart: Evidence from Kaggle
Teams searching together generate more attempts but less exploration than independent search; effectiveness depends on strategic context.
Abstract Research Summary How does the mode of search—independently or jointly—affect collective search, a central component of organizational adaptation and innovation? Using naturally occurring data from a strongly incentivized online competition platform, we find that compared to their counterfactuals that search apart, groups searching together exhibit less exploration in their search outcomes as noted in prior experimental and computational modeling studies. However, groups searching together stimulate a greater number of search attempts from their members than groups searching apart, an
teams · Phanish Puranam
research · IEEE Transactions on Software Engineering ·
From Disruptions to Discussions: How GenAI Impacts Human Interactions in Software Development
GenAI shifts developer interactions from answering routine questions to deeper collaboration, and developers report fewer flow disruptions.
New technologies often change how an individual performs work, such as how generative AI (GenAI) can help a developer write code. New technologies can also impact how people interact with one another, such as how GenAI’s ability to summarize API documentation can reduce the need for developers to ask each other technical questions. In this paper, we report on a two-phase mixed-method study exploring how GenAI influences how humans interact in software development. During phase one, 30 industrial software developers provided data over a period of 5 to 12 days as they worked, contributing 627 ex
teams
practice · Hacker News ·
Porting 100k lines from TypeScript to Rust using Claude Code in a month
Team ported 100k TypeScript lines to Rust in a month using Claude Code as primary tool, detailing workflow and lessons.
255 points on Hacker News. Discussion: https://news.ycombinator.com/item?id=46765694
productivity
practice · How I AI ·
Advanced Claude Code techniques: context loading, mermaid diagrams, stop hooks, and more | John Lindquist
Senior engineers use mermaid diagrams and custom hooks to load context into Claude Code, accelerating coding assistance and automating quality checks.
John Lindquist is the co-founder of egghead.io and an expert in leveraging AI tools for professional software development. In this episode, John shares advanced techniques for using AI coding tools like Claude Code and Cursor that go far beyond basic prompting. He demonstrates how senior engineers can use mermaid diagrams for context loading, create custom hooks for automated code quality checks, and build efficient command-line tools that streamline AI workflows. What you’ll learn: How to use mermaid diagrams to preload context into Claude Code for faster, more accurate coding assistance Crea
ways of working · Claire Vo
practice · Artificial Ignorance ·
Skills, Tools and MCPs - What’s The Difference?
Framework distinguishing Tools, MCPs and Skills: when each solves different problems in connecting AI to external systems.
Two and a half years ago, OpenAI released function calling for GPT-4 . I still remember that distinct “wow” moment - the realization that language models could actually do things beyond generating text. Not just answer questions or write essays, but call APIs, manipulate data, and take actions in the real world. It felt like watching the future. And yet - I’ve since watched ChatGPT Plugins (the first feature to use function calling) launch with fanfare, quietly get deprecated, be replaced by GPTs, and now find themselves being superseded by this new wave of Skills and MCPs. The progression tau
ways of working · Charlie Guo
practice · Latent Space ·
Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay
A research leader describes scaling deep reasoning from competition problems to production models across Google DeepMind's organization.
From shipping Gemini Deep Think and IMO Gold to launching the Reasoning and AGI team in Singapore, Yi Tay has spent the last 18 months living through the full arc of Google DeepMind’s pivot from architecture research to RL-driven reasoning—watching his team go from a dozen researchers to 300+, training models that solve International Math Olympiad problems in a live competition, and building the infrastructure to scale deep thinking across every domain, and driving Gemini to the top of the leaderboards across every category. From shipping Gemini Deep Think and IMO Gold to launching the Reasoni
management org · Shawn Wang (swyx)
research · Strategic Organization ·
Talk isn’t always cheap: A theory of social influence and deliberation in group decision-making
Agent-based model shows AI agents resistant to conformity can improve group decision-making by reducing clustering effects, even if they lack expertise.
Organizations often rely on deliberative groups—committees, taskforces, boards—to make decisions, yet deliberation’s effectiveness remains contested. When can deliberation help a group outperform the average or even the best individual in the group? We propose that the value of deliberation depends on how the network structure of the group shapes informational influence (which promotes updating beliefs based on perceived expertise) and normative influence (which drives conformity to gain social approval). Using an agent-based model, we show that deliberation can help groups achieve strong syne
judgment · Phanish Puranam
research · Indeed Hiring Lab ·
January 2026 US Labor Market Update: Jobs Mentioning AI Are Growing Amid Broader Hiring Weakness
Job postings mentioning AI are growing while overall US hiring weakens, showing employer concentration on AI-related roles.
Small pockets of growth are emerging as employers concentrate their limited hiring on roles and skills tied to AI. The post January 2026 US Labor Market Update: Jobs Mentioning AI Are Growing Amid Broader Hiring Weakness appeared first on Indeed Hiring Lab .
jobs skills
research · Science ·
Who is using AI to code? Global diffusion and impact of generative AI
AI-generated code now represents 29% of Python functions on GitHub; AI boosted experienced developers' productivity by 3.6% but showed no benefit for early-career developers.
Generative coding tools promise big productivity gains, but uneven uptake could widen skill and income gaps. We train a neural classifier to spot artificial intelligence (AI)-generated Python functions in more than 30 million GitHub commits by 160,097 software developers, tracking how fast, and where, these tools take hold. Currently, AI writes an estimated 29% of Python functions in the US-a shrinking lead over other countries. We estimate that quarterly output, measured in online code contributions, consequently increased by 3.6%. AI seems to benefit experienced, senior-level developers: The
jobs skills
research · Government Information Quarterly ·
Generative AI in public administration: A quasi-experimental analysis of bureaucratic productivity
Quasi-experimental study finds generative AI reduces document drafting time by 4 minutes per task; new employees gain most, suggesting AI aids skill gaps.
This study investigates the effect of a specialized generative AI system—developed and tested as part of a Proof of Concept (PoC)—on the speed with which public officials in a central government agency draft written responses to citizen inquiries. Using a quasi-experimental design with difference-in-differences analysis and propensity score matching, 80 civil servants were divided into a treatment group participating in the proof of concept (adopting AI-based drafting) and a control group that maintained existing practices. The difference-in-differences results of this study indicate a signifi
productivity
practice · Anthropic Engineering ·
Designing AI-resistant technical evaluations
A technical hiring evaluation evolved through three rounds as Claude improved, revealing what skills remain hard for AI to fake.
What we learned from three iterations of a performance engineering take-home that Claude keeps beating.
jobs skills
practice · GitHub Blog ·
AI-supported vulnerability triage with the GitHub Security Lab Taskflow Agent
LLM agents triaged security alerts by matching fuzzy code patterns, discovering 30 real vulnerabilities since August with basic tools only.
Triaging security alerts is often very repetitive because false positives are caused by patterns that are obvious to a human auditor but difficult to encode as a formal code pattern. But large language models (LLMs) excel at matching the fuzzy patterns that traditional tools struggle with, so we at the GitHub Security Lab have been experimenting with using them to triage alerts. We are using our recently announced GitHub Security Lab Taskflow Agent AI framework to do this and are finding it to be very effective. 💡 Learn more about it and see how to activate the agent in our previous blog post
productivity
research · Knowledge at Wharton ·
Who Gets Replaced by AI and Why?
AI implementation in certain workflow positions reduces employee motivation more than others, with implications for where to deploy AI.
New research from Wharton’s Pinar Yildirim reveals how AI can impact employee motivation when implemented in the wrong part of a team’s workflow. … Read More
worker experience
research · Research Policy ·
Firm training, automation, and wages: International worker-level evidence
Firm training reduces automation risk by 3.8 percentage points across 37 countries, accounting for 15% of wage returns to training.
Firm training is widely regarded as crucial for protecting workers from automation, yet there is a lack of empirical evidence to support this belief. Using internationally harmonized data from over 90,000 workers across 37 industrialized countries, we construct an individual-level measure of automation risk based on tasks performed at work. Our analysis reveals substantial within-occupation variation in automation risk, overlooked by existing occupation-level measures. To assess whether firm training mitigates automation risk, we exploit within-occupation and within-industry variation. Additio
jobs skills
research · Management Science ·
The Rapid Adoption of Generative AI
27% of US employed workers used genAI for work weekly as of late 2024; work adoption faster than PCs, with 1-7% of work hours assisted by genAI.
Generative artificial intelligence (genAI) is a potentially important new technology, but its impact on the economy depends on the speed and intensity of adoption. This paper reports results from a series of nationally representative U.S. surveys of genAI use at work and at home. As of late 2024, 45% of the U.S. population age 18–64 uses genAI. Among employed respondents, 27% used genAI for work at least once in the previous week: 10% used it every workday and 17% on some but not all workdays. Relative to each technology’s first mass-market product launch, work adoption of genAI has been faste
adoption
practice · Addy Osmani ·
How to write a good spec for AI agents
Framework for writing AI agent specs that balance clarity with context limits, using iterative planning and task decomposition.
TL;DR: Aim for a clear spec covering just enough nuance (this may include structure, style, testing, boundaries) to guide the AI without overwhelming it. Break large tasks into smaller ones vs. keeping everything in one large prompt. Plan first in read-only mode, then execute and iterate continuously. “I’ve heard a lot about writing good specs for AI agents, but haven’t found a solid framework yet. I could write a spec that rivals an RFC, but at some point the context is too large and the model breaks down.” Many developers share this frustration. Simply throwing a massive spec at an AI agent
ways of working · Addy Osmani
practice · How I AI ·
Claude Code for product managers: research, writing, context libraries, custom to-do system, and more | Teresa Torres
Product manager built a markdown task system in Claude Code with automated research collection and slash commands for daily workflows.
Teresa Torres is the author of Continuous Discovery Habits and an internationally acclaimed speaker and coach. In this episode, Teresa demonstrates how she’s built a personalized productivity system using Claude Code to manage her tasks, automate research collection, and improve her writing. She shows how non-developers can leverage AI tools to create personalized workflows that match their unique needs and thinking style. What you’ll learn: How Teresa built a personalized task management system in Claude Code that matches her exact workflow needs Why she moved from Trello to a markdown-based
ways of working · Claire Vo
practice · Hamel Husain ·
Why I Stopped Using nbdev
Literate programming in notebooks conflicts with AI coding tools; developers should choose environments where AI assistance works best.
Programmers love to proclaim they’ve found the best tool. Paul Graham called Lisp his “ secret weapon .” DHH described Ruby as “ a magical glove that just fit my brain perfectly .” Pieter Levels ships million-dollar products with vanilla PHP and jQuery . These declarations aren’t about the languages themselves. They’re about developers finding tools that fit how they think. When the environment clicks, you move fast. I had that experience with nbdev , a development environment for literate programming that I helped build and maintain 1 . I created hundreds of projects with it and was one of it
ways of working · Hamel Husain
practice · Latent Space ·
Brex’s AI Hail Mary — With CTO James Reggio
CTO describes how Brex built AI systems meeting financial compliance, auditability and customer trust requirements.
From building internal AI labs to becoming CTO of Brex, James Reggio has helped lead one of the most disciplined AI transformations inside a real financial institution where compliance, auditability, and customer trust actually matter. From building internal AI labs to becoming CTO of Brex, James Reggio has helped lead one of the most disciplined AI transformations inside a real financial institution where compliance, auditability,…
adoption · Shawn Wang (swyx)
research · Indeed Hiring Lab ·
AI Adoption Is Accelerating but Still Concentrated Among the Largest Firms
AI hiring is concentrated at the largest firms, suggesting unequal access to AI productivity gains across company sizes.
Nearly all AI-related hiring is happening at a few very large employers, highlighting the risks of unequal access to AI-driven productivity gains. The post AI Adoption Is Accelerating but Still Concentrated Among the Largest Firms appeared first on Indeed Hiring Lab .
adoption
practice · Artificial Ignorance ·
The AI Manager's Schedule
A developer shifted from IDE-based AI coding tools to headless CLI tools by changing trust and debugging practices.
Shoutout to , , and for joining last week’s Office Hours chat ! Join the chat tomorrow for another Office Hours session. Recently, I was talking to a colleague about how dramatically my AI coding workflows have changed. A year ago, when the first CLI coding tools were released, I gave them a try. It left a strong impression - they were fun to use, and much more impactful than I thought they would be. But I never moved the majority of my development over to them. I was still using IDEs like Cursor, mainly because it was too hard to trust what the headless tool was doing, and debugging inevitabl
ways of working · Charlie Guo
research · Exponential View ·
🔮 Anthropic's Head of Economics Peter McCrory on their new Economic Index
Analysis of millions of real Claude conversations maps where AI is augmenting human work today and where it is not.
The future will be... uneven. The future will be... uneven. Anthropic have just released a new Economic Index report. They’ve analysed millions of real Claude conversations to map exactly where AI is augmenting human work today, and where it isn’t. This is the best empirical window we have into how AI is reshaping work right now.
adoption · Peter McCrory · Azeem Azhar
research · arXiv ·
Evolving with AI: A Longitudinal Analysis of Developer Logs
Two-year telemetry from 800 developers shows AI coding assistants increase code output but also increase deletions, with perceived vs actual workflow changes diverging.
AI-powered coding assistants are rapidly becoming fixtures in professional IDEs, yet their sustained influence on everyday development remains poorly understood. Prior research has focused on short-term use or self-reported perceptions, leaving open questions about how sustained AI use reshapes actual daily coding practices in the long term. We address this gap with a mixed-method study of AI adoption in IDEs, combining longitudinal two-year fine-grained telemetry from 800 developers with a survey of 62 professionals. We analyze five dimensions of workflow change: productivity, code quality, c
adoption
practice · Paul Ford (Aboard) ·
Four AI Coding Horsemen of the Apocalypse
Four structural risks when AI coding tools democratise software development: deployment complexity, testing gaps, security shortcuts, and maintenance debt.
As anyone who reads this newsletter, listens to our podcast, or overhears me yelling at strangers on the train knows, I’m excited that tens of millions of people could suddenly become software developers thanks to Claude Code (and its inevitable competitors—there’s a rumored DeepSeek coding tool in the pipeline , which would be chef’s kiss ). But I also know that whenever I feel particularly hopeful about technology, I should grab a huge ice bucket and stick my head into it, because enthusiasm often doesn’t translate into social acceptance, utility, or positive outcomes. Having learned this le
management org · Paul Ford
practice · Hacker News ·
Using proxies to hide secrets from Claude Code
Using proxy patterns to prevent Claude Code from accessing sensitive secrets during development and testing.
132 points on Hacker News. Discussion: https://news.ycombinator.com/item?id=46605155
ways of working
practice · How I AI ·
The power user’s guide to Codex: parallelizing workflows, planning techniques, advanced context engineering tips, automating code reviews, and more | Alexander Embiricos
Developers use Git worktrees and Plans.md to parallelize Codex workflows and manage context for complex projects.
Alexander Embiricos , the product lead for Codex at OpenAI, shares practical workflows for getting the most out of this AI coding agent. In this episode, he demonstrates how both non-technical users and experienced engineers can leverage Codex to accelerate development, from making simple code changes to building production-ready applications. Alex walks through real examples of using Codex in VS Code and terminal environments, implementing parallel workflows with Git worktrees, and creating detailed implementation plans for complex projects. He also reveals how OpenAI uses Codex internally, i
ways of working · Claire Vo
research · Information Systems Research ·
Workflow Automation in Open-Source Software Development: Accelerating Innovation Through Mechanization and Orchestration
Mechanization automates tasks; orchestration automates coordination. Orchestration accelerates exploratory innovation while mechanization supports incremental improvement.
This study develops a conceptual framework distinguishing two mechanisms of workflow automation: mechanization and orchestration. Mechanization automates discrete, self-contained, repeatable tasks through standardized execution to enhance consistency, reliability, and efficiency, while orchestration automates the communication between tasks, workers, and stages, which facilitates information flow and coordination. We theorize that these mechanisms differentially affect incremental versus substantive innovation. Using a multimethod approach integrating machine learning and econometrics, we anal
adoption
practice · Sangeet Paul Choudary ·
US vs China: How to win the wrong AI race
China's AI strategy focuses on converting intelligence into scaled execution and economic value, not competing on model capability.
TL;DR: The US is betting on intelligence. China is betting somewhere else! The US frequently frames its competition with China as an AI race, similar to the space race it ran with the USSR. The idea of a race was triggered around a year ago with the launch of DeepSeek. Ever since, much of the US media has been fascinated with the idea of the US winning the AI race against China. Ironically, it isn’t much of a race if the US is the only one running it. The American ‘AI race’ is framed around who will build the smartest models, as if superior intelligence alone decides the future. China’s strate
management org · Sangeet Paul Choudary
research · Knowledge at Wharton ·
Special Report: The AI Skills Gap
New index from Wharton and Accenture measures which skills employers value and how fast labour market demand is shifting.
A new index from Wharton and Accenture measures which skills matter to employers, which do not, and how quickly the economy is shifting beneath us. … Read More
jobs skills
research · JAMA Network Open ·
Ambient Artificial Intelligence Scribes and Physician Financial Productivity
Physicians using AI scribes increased revenue and patient volumes; claim denial rates fell compared to non-adopters.
This cohort study evaluates physician revenue, patient volumes, and claim denials among adopters and nonadopters of artificial intelligence–based clinical documentation tools in one health system.
productivity
practice · One Useful Thing ·
Claude Code and What Comes Next
An AI agent independently built a working deployed website and sales funnel in 74 minutes with minimal human input.
I opened Claude Code and gave it the command: “ Develop a web-based or software-based startup idea that will make me $1000 a month where you do all the work by generating the idea and implementing it. i shouldn’t have to do anything at all except run some program you give me once. it shouldn’t require any coding knowledge on my part, so make sure everything works well. ” The AI asked me three multiple choice questions and decided that I should be selling sets of 500 prompts for professional users for $39. Without any further input, it then worked independently… FOR AN HOUR AND FOURTEEN MINUTES
ways of working · Ethan Mollick
practice · Paul Ford (Aboard) ·
Triple-A Vibe Coding
A developer describes working with Claude Code by judging output without understanding the underlying code, building complex software iteratively on mobile.
Welcome to 2026! It’s going to be an eventful year, and there’s nothing to be done about that. I know some of us were hoping AI would go away over the break, but it’s here to stay. I have spent the past couple of months going deep on Claude Code , and I want to share some very high-level observations. “Going deep” sounds intense, but actually involves typing a chat message into your phone and then coming back to it 20 minutes later to see if it worked, over and over again. You can finally write complex software on the bus. (And if Mayor Mamdani gets his way, that bus will be fast and free.) Ju
ways of working · Paul Ford
research · arXiv ·
Artificial Intelligence and Skills: Evidence from Contrastive Learning in Online Job Vacancies
AI adoption by Chinese firms expanded job skill requirements, driven by reduced information asymmetry and anticipation of future labour market needs.
We investigate the impact of artificial intelligence (AI) adoption on skill requirements using 14 million online job vacancies from Chinese listed firms (2018-2022). Employing a novel Extreme Multi-Label Classification (XMLC) algorithm trained via contrastive learning and LLM-driven data augmentation, we map vacancy requirements to the ESCO framework. By benchmarking occupation-skill relationships against 2018 O*NET-ESCO mappings, we document a robust causal relationship between AI adoption and the expansion of skill portfolios. Our analysis identifies two distinct mechanisms. First, AI reduce
jobs skills
practice · Jason Fried ·
The joy of delegating to competence
Delegating to AI agents creates a psychological experience of total trust previously rare outside elite teams.
AI workflows are technically impressive, but there’s a deeper reason people are really amped about AI agents. This isn’t just new tech, it’s new psychology. Until now, very few people have known what it feels like to delegate to total competency. If you manage great people, or lead great teams, you know how it feels to put someone in charge who will get it done, get it done right, and get it done without drama. That kind of delegation — that depth of trust — is pure joy. Delegating to competency lets you forget about it completely. That’s real leverage. And now anyone can experience that. What
ways of working · Jason Fried
practice · The Pragmatic Engineer ·
When AI writes almost all code, what happens to software engineering?
Gergely Orosz argues that AI-written code is now industry-wide practice, not hypothesis, with implications for roles.
No longer a hypothetical question, this is a mega-trend set to hit the tech industry No longer a hypothetical question, this is a mega-trend set to hit the tech industry This article was originally a paid one. I’ve removed the paywall on 24 September 2026, after DHH’s Rails World keynote on 23 Sep 2026 caused quite the stir, where he shared how 37Signals has gone “pencils down” and generates all code. In January, we saw this will happen across the industry, sooner, rather than later. The good news: software engineering …
jobs skills · Gergely Orosz
practice · Harper Reed ·
Remote Claude Code: programing like it was the early 2000s
Using Claude Code from a phone terminal to program remotely, mimicking early-2000s development practices.
So so many friends have asked me how I use Claude Code from my phone. I am always a bit surprised, because a lot of this type of work I have been doing for nearly 25 years (or more!) and I always forget that it is partially a lost art. This is how it used to be. We didn’t have fancy IDEs, and fancy magic to deploy stuff. We had to ssh (hah. Telnet!) into a machine and work with it through the terminal. It ruled. It was a total nightmare. It was also a lot of fun. One of my favorite parts of the early 2000s was hanging out in IRC channels and just participating in the most ridiculous community
ways of working · Harper Reed
practice · Addy Osmani ·
Code Review in the Age of AI
Code review now splits between solo developers using test suites as backstops and teams using review for shared context, because AI-generated code has 75% more logic errors.
AI did not kill code review. It made the burden of proof explicit. Ship changes with evidence like manual verification and automated tests, then use review for risk, intent, and accountability. Solo developers lean on automation to keep up with AI speed, while teams use review to build shared context and ownership. If your pull request doesn’t contain evidence that it works, you’re not shipping faster - you’re just moving work downstream. By early 2026, over 30% of senior developers report shipping mostly AI-generated code. The challenge? AI excels at drafting features but falters on logic, se
judgment · Addy Osmani
research · SSRN ·
Hedging with Talent: How Selection Under Uncertainty Reshapes Automation's Workforce Effects
Firms hedge automation risk by hiring more adaptable workers, shifting who gains or loses jobs from technological change.
jobs skills · Matt Beane
research · IEEE Transactions on Software Engineering ·
Understanding Prompt Quality and Its Relation to LLM-based Code Generation
Readability and structure of prompts to LLMs predict code quality; analysis of real developer-ChatGPT interactions from GitHub.
Prompt engineering has emerged as a key practice to guide Large Language Models (LLMs) in code generation tasks, under the assumption that better prompts yield better outputs. Yet little is known about which characteristics of prompts actually drive variations in output quality. In this paper, we take a first step toward addressing this gap by investigating how measurable properties of prompts relate to the quality of generated code. We conducted an empirical study of real-world, single-turn interactions between developers and ChatGPT collected from GITHUB. Prompts were characterized along rea
adoption
research · IEEE Transactions on Software Engineering ·
Impact of an LLM-based Review Assistant in Practice: A Mixed Open-/Closed-source Case Study
LLM-generated code review comments were accepted in 8% of cases across Mozilla and Ubisoft, with 15-21% rated as valuable tips.
Code review is a standard practice in modern software development, aiming to improve code quality.As providing constructive reviews on a submitted patch is a challenging and error-prone task, the advances of Large Language Models (LLMs) in performing natural language processing (NLP) tasks in a human-like fashion have prompted researchers to evaluate the LLMs’ ability to automate the code review process. However, outside lab settings, it is still unclear how prone reviewers are to accept comments generated by LLMs in a real development workflow. To fill this gap, we conduct a large-scale empir
adoption
research · SSRN ·
Solve and Be Seen: How Workers in Deskilled Jobs Build Complex Skill as New Automation Arrives
Workers in deskilled roles develop complex skills by solving problems visible to management as automation arrives.
jobs skills · Matt Beane
research · SSRN ·
GenAI Adoption Increases the Density of Knowledge and Collaboration Networks: Evidence from a Field Experiment
A field experiment shows GenAI adoption increases density of knowledge and collaboration networks within organisations.
adoption · Phanish Puranam
research · NBER ·
O-Ring Automation
Automation of complementary tasks (O-ring production) creates different wage and employment effects than separable task automation.
We study automation when tasks are quality complements rather than separable. Production requires numerous tasks whose qualities multiply as in an O-ring technology. A worker allocates a fixed endowment of time across the tasks performed; machines can replace tasks with given quality, and time is (Joshua S. Gans , Avi Goldfarb)
jobs skills · Avi Goldfarb
research · NBER ·
Automation Experiments and Inequality
Analysis of what productivity experiments on automation actually reveal about inequality effects across worker skill levels.
Many experiments study the productivity effects of automation technologies such as generative algorithms. A key test in these experiments relates to inequality: does the technology increase output more for high- or low-skill workers? However, the theoretical content of this empirical test has been (Seth Gordon Benzell , Kyle R. Myers)
jobs skills
research · NBER ·
Artificial Jagged Intelligence: When AI Benchmarks Misstate Deployment Value
AI systems that score highly on public benchmarks may perform unevenly on the specific task distributions organisations actually face.
Organisations increasingly select and deploy artificial intelligence systems on the strength of public benchmarks. A benchmark, however, scores a system on a single distribution of tasks, whereas each organisation meets its own. Because AI performance is uneven across tasks, a property called (Joshua S. Gans)
adoption
research · NBER ·
How Adaptable Are American Workers to AI-Induced Job Displacement?
Workers in AI-exposed occupations have higher adaptive capacity to navigate job transitions, based on an index across 356 US occupations.
We construct an occupation-level adaptive capacity index that measures a set of worker characteristics relevant for navigating job transitions if displaced, covering 356 occupations that represent 95.9% of the U.S. workforce. We find that AI exposure and adaptive capacity are positively correlated: (Sam J. Manning , Tomás Aguirre)
jobs skills
research · Human Relations ·
Problematizing the role of artificial intelligence in hiring and organizational inequalities: A multidisciplinary review
Multidisciplinary review proposes concept of 'algorithmically-mediated inequality regimes' to explain how AI in hiring conceals rather than eliminates workplace inequalities.
What are the implications of the growing use of artificial intelligence (AI) in recruitment and hiring for organizational inequalities? While advocates suggest that AI is a groundbreaking tool that can enhance hiring precision, efficiency, diversity and fit, critics raise serious concerns around bias, fairness, and privacy. This review article critically advances this debate by drawing on diverse scholarship across computing and data sciences; human resource, management, and organization studies; social sciences; and law. Using a hybrid review approach that combines scoping and problematizing
jobs skills
practice · Addy Osmani ·
How Good Is AI at Coding React (Really)?
AI coding success depends on context engineering and domain knowledge; React developers can guide models away from failures by understanding their limits.
tl;dr: AI coding benchmarks show models excel at isolated React tasks like scaffolding components or implementing explicit specs, achieving ~40% success in benchmarks, but drops to ~25% on multi-step integrations due to a “complexity cliff” in state management and design taste. The gap between “AI helped me ship” and “AI gave me a mess” is context engineering and explicit constraints. Deep React and domain knowledge enable you to spot when AI goes off the rails and understand why it repeats mistakes. Guide it without blindly accepting the output. This article is based on my closing keynote at
ways of working · Addy Osmani
practice · How I AI ·
How Webflow’s CPO built an AI chief of staff to manage her calendar, prep for meetings, and drive AI adoption | Rachel Wolan
CPO built a custom AI agent to manage calendar, meetings, and email, then used hands-on experience to drive adoption across her company.
Rachel Wolan , the chief product officer at Webflow, has embraced AI not just as a product leader but as a hands-on builder. A coder since age 16, Rachel has returned to her technical roots by creating a custom AI chief-of-staff application that helps manage her executive workload. In this episode, she demonstrates how she uses personal AI software to prep for meetings, triage her calendar, manage emails, and even get brutally honest feedback about how she’s spending her time. What you’ll learn: How Rachel built a custom AI chief-of-staff application that integrates with her calendar, email, a
ways of working · Claire Vo
practice · Sangeet Paul Choudary ·
The Not-so-Lazy Holiday Reading List
Reflection and strategic inaction become competitive advantages when AI makes cheap execution and instant answers abundant.
It’s that time of the year again! Everyone’s making predictions about what’s going to happen in 2026! Predictions are easy to generate and costless to abandon if the person making the prediction bears no responsibility for being wrong. And chatbots can do that better than you and me - producing fluent answers on demand. What’s harder - and increasingly rare - is the work of reflection: slowing down, taking stock, and making sense of change. And to make it worth your while, here’s a curated list of Holiday Readings to help you reflect more and reflect better in a world of relentless execution.
management org · Sangeet Paul Choudary
practice · Hacker News ·
Ask HN: How are you sandboxing coding agents?
Teams share practical sandboxing setups for AI coding agents: git worktrees, VMs, firejail, with real tradeoffs between safety and convenience.
I've seen people rely on built-in sandboxes, use git worktrees (sometimes inside devcontainers), or run the whole agent inside a Linux VM with minimal host mounts. On Linux, I’ve also seen firejail/bubblewrap mentioned. For folks actually using these tools day-to-day: What’s your default setup? Have you had any "learned the hard way" moments? What tradeoff (safety vs convenience vs parallelism) has mattered most in practice? I'm less interested in theoretical best practices than what's actually holding up under real use.
ways of working
research · Information Systems Research ·
A Robust Optimization Approach to Reliable Statistical Inference with Variables Generated by Machine Learning
A robust optimization method improves statistical inference when ML-generated variables contain prediction errors, tested on Amazon reviews.
Organizations increasingly use machine learning to turn text, images, and other unstructured data into variables that inform decisions and research. But, because machine learning predictions are never perfect, the resulting data can contain errors that quietly distort statistical analyses, sometimes leading to incorrect conclusions about what truly drives important outcomes. This study introduces a robust optimization approach that helps analysts and decision makers draw more reliable insights when working with machine learning–generated data. The method is designed to strengthen the signal of
judgment
practice · Charity Majors ·
2025 was for AI what 2010 was for cloud
2025 marked AI adoption inflection point in enterprise ops, mirroring cloud computing's shift from experimental to mainstream between 2006-2010.
I was at my very first job, Linden Lab, when EC2 and S3 came out in 2006. We were running Second Life out of three datacenters, where we racked and stacked all the servers ourselves. At the time, we were tangling with a slightly embarrassing data problem in that there was no real way for users to delete objects (the Trash folder was just another folder), and by the time we implemented a delete function, our ability to run garbage collection couldn’t keep up with the rate of asset creation. In desperation, we spun up an experimental project to try using S3 as our asset store. Maybe we could mak
adoption · Charity Majors
research · HBS AI Institute ·
The Agentic AI Reality Check
Early empirical data on how AI agents are actually adopted and used by real users of Perplexity.
Agentic AI has recently been moving through a period of heightened excitement and innovation, but empirical data on how these tools are actually being used has been scarce. The new study “The Adoption and Usage of AI Agents: Early Evidence from Perplexity,” by Jeremy Yang, Assistant Professor of Business Administration at Harvard Business School and […] The post The Agentic AI Reality Check appeared first on Harvard Business School AI Institute .
adoption
practice · One Useful Thing ·
The Shape of AI: Jaggedness, Bottlenecks and Salients
AI's jagged frontier persists: superhuman at some tasks, poor at others unrelated to apparent difficulty.
Back in the ancient AI days of 2023, my co-authors and I invented a term to describe the weird ability of AI to do some work incredibly well and other work incredibly badly in ways that didn’t map very well to our human intuition of the difficulty of the task. We called this the “Jagged Frontier” of AI ability, and it remains a key feature of AI and an endless source of confusion. How can an AI be superhuman at differential medical diagnosis or good at very hard math (yes, they are really good at math now, famously outside the frontier until recently) and yet still be bad at relatively simple
judgment · Ethan Mollick
practice · Addy Osmani ·
My LLM coding workflow going into 2026
Developers using LLMs as pair programmers need clear specs, context, oversight and critical thinking to get consistent results.
AI coding assistants became game-changers this year, but harnessing them effectively takes skill and structure. These tools dramatically increased what LLMs can do for real-world coding, and many developers (myself included) embraced them. At Anthropic, for example, engineers adopted Claude Code so heavily that today ~90% of the code for Claude Code is written by Claude Code itself . Yet, using LLMs for programming is not a push-button magic experience - it’s “difficult and unintuitive” and getting great results requires learning new patterns. Critical thinking remains key. Over a year of proj
ways of working · Addy Osmani
practice · How I AI ·
How Zapier’s EA built an army of AI interns to automate meeting prep, strengthen team culture, and scale internal alignment | Cortney Hickey
EA built automated meeting prep workflows that research participants and pull CRM data, plus AI-powered culture reinforcement across the org.
Cortney Hickey is the executive assistant to the CEO at Zapier, where she’s leveraging AI to transform traditional EA responsibilities into scalable, organization-wide systems. In this episode, she demonstrates how she’s built AI workflows that automate meeting preparation, reinforce company culture through automated feedback, and democratize strategic knowledge across the organization. Her approach shows how EAs can use AI not to replace their roles but to elevate them—working on higher-impact initiatives while creating systems that benefit the entire company. What you’ll learn: How to build
ways of working · Claire Vo
research · JAMA Network Open ·
Uptake of Generative AI Integrated With Electronic Health Records in US Hospitals
Survey of 2024 US hospital IT adoption of generative AI in EHRs shows early adopter rates and links adoption to prior AI experience and hospital characteristics.
Importance: There is widespread enthusiasm about generative artificial intelligence (AI), but no systematic evidence on its implementation across health care organizations. Objective: To describe adoption of generative AI integrated with the electronic health record (EHR) by nonfederal acute care hospitals, how adoption relates to experience using and evaluating predictive AI, and hospital characteristics. Design, Setting, and Participants: This survey study of nonfederal acute care US hospitals used the 2024 American Hospital Association (AHA) Information Technology (IT) Supplement survey. Th
adoption
practice · Paul Ford (Aboard) ·
Don’t Waste a Miracle
AI coding assistants are lowering the skilled labor needed to build deterministic systems, enabling non-programmers to build platforms within weeks.
There’s a lot swirling right now: Claude Code getting good, AI-powered political messaging , McDonald’s making a depressing AI ad everyone hates, OpenAI going “code red,” Trump giving Nvidia the green light to sell some chips to China, and on and on. On a personal level, despite my co-founder’s best attempts , I keep moving forward on my Claude Code experiments . I’m grasping the contours of what I can do with AI in coding and it’s no longer quite so brain-melting—I’m getting used to it, as always happens—but it also leaves me convinced that next year is going to be a different kind of year. I
jobs skills · Paul Ford
practice · The Pragmatic Engineer ·
Frictionless: why great developer experience can help teams win in the ‘AI age’
Developer experience reduces friction in AI workflows, enabling teams to ship faster and adopt AI tools more widely.
Exclusive excerpts from the newly-released book ‘Frictionless’, by Nicole Forsgren and Abi Noda. Exclusive excerpts from the newly-released book ‘Frictionless’, by Nicole Forsgren and Abi Noda. Hi, this is Gergely with a bonus, free issue of the Pragmatic Engineer Newsletter. In every issue, I cover Big Tech and startups through the lens of senior engineers and engineering leaders. If you’ve been forwarded this email, you can subscribe here
adoption · Nicole Forsgren · Gergely Orosz
research · Strategic Organization ·
Generative organizational learning: Affordances for new modes of knowledge search, creation, transfer, and forgetting with LLMs
Case study shows how generative AI enables new organizational learning modes: knowledge creation and personalization, plus selective forgetting through guardrailing.
Scholars of organizational learning typically focus on how organizations operate in an environment of domain knowledge scarcity, assuming that knowledge abundance is a highly sought after, but fundamentally elusive, goal. In our case study of organizational learning at Northeast Health, we find that generative AI, in combination with actors’ goals and organizational institutions, may afford a new form of learning that we call generative organizational learning enabling innovation in the service of organizational goals. In particular, generative catalyzing, iterating, and personalizing may enab
management org · Katherine Kellogg
research · Journal of Management Studies ·
The Dark Side of Managing Human– AI Collaborations: Implications for Leaders’ Moral Relativism and Unethical Behaviour
Managing human-AI collaboration may increase leaders' moral relativism and unethical behaviour, moderated by cognitive-closure needs.
Abstract As collaborations between humans and artificial intelligence (AI) have become increasingly prevalent across various industries, the role of leaders in managing these collaborations has grown in importance. While the existing literature has highlighted the benefits of leader management in these settings – emphasizing the complementary strengths of humans and AI – the potential costs to key stakeholders, particularly to leaders themselves, have been largely ignored. This research addresses this gap by drawing on moral relativism theory to develop and test a model explaining how leader m
management org