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 10 of 12 · 1144 items
practice · The Pragmatic Engineer ·
How tech companies measure the impact of AI on software development
How 19 tech companies measure AI tool impact on developer productivity: metrics, methods and tradeoffs they actually use.
How do GitHub, Google, Dropbox, Monzo, Atlassian, and 13 other companies know how well AI tools work for devs? A deepdive sharing exclusive details, with CTO Laura Tacho How do GitHub, Google, Dropbox, Monzo, Atlassian, and 13 other companies know how well AI tools work for devs? A deepdive sharing exclusive details, with CTO Laura Tacho Hi – this is Gergely with the monthly, free issue of the Pragmatic Engineer Newsletter. In every issue, I cover challenges at Big Tech and startups through the lens of senior engineers and engineering leaders. If you’ve been forwarded this email, you can
productivity · Gergely Orosz
practice · How I AI ·
How I built an Apple Watch workout app using Cursor and Xcode (with zero mobile-app experience)
Product manager with no mobile experience built Apple Watch fitness app using Cursor, combining AI coding with manual review and debugging in Xcode.
Terry Lin is a product manager and developer who built Cooper’s Corner, an AI-powered fitness tracking app that works across iPhone and Apple Watch. Frustrated with traditional fitness apps that require extensive setup and manual logging, Terry created a solution that lets users simply speak their exercises, weights, and reps. The app automatically structures this data and provides analytics on workout consistency and progress. In this episode, Terry shares his vibe-coding process using Cursor and Xcode and explains how he optimizes his codebase for AI collaboration. What you’ll learn: 1. How
ways of working · Claire Vo
practice · Eugene Yan ·
Training an LLM-RecSys Hybrid for Steerable Recs with Semantic IDs
LLM trained to output item IDs directly, enabling conversational recommendations without retrieval systems or tool calls.
An LLM that can converse in English & item IDs, and make recommendations w/o retrieval or tools.
ways of working · Eugene Yan
research · arXiv ·
Vibe Coding in Product Teams: Reconfiguring AI-Assisted Workflows, Prototyping, and Collaboration
Vibe coding reshapes product team workflows through four stages (ideation, generation, debugging, review), accelerating iteration but creating new trust and responsibility tensions.
Generative AI is reshaping product design practices through "vibe coding," where product team members express intent in natural language and AI translates it into functional prototypes and code. Despite rapid adoption, little research has examined how vibe coding reconfigures product development workflows and collaboration. Drawing on interviews with 22 product team members across enterprises, startups, and academia, we show how vibe coding follows a four-stage workflow of ideation, generation, debugging, and review. This accelerates iteration, supports creativity, and lowers participation bar
ways of working
practice · One Useful Thing ·
On Working with Wizards
Working with AI is shifting from correcting a co-worker to directing a performer; humans increasingly guide output without understanding the reasoning.
In my book, Co-Intelligence , I outlined a way that people could work with AI, which was, rather unsurprisingly, as a co-intelligence. Teamed with a chatbot, humans could use AI as a sort of intern or co-worker, correcting its errors, checking its work, co-developing ideas, and guiding it in the right direction. Over the past few weeks, I have come to believe that co-intelligence is still important but that the nature of AI is starting to point in a different direction. We're moving from partners to audience, from collaboration to conjuring. A good way to illustrate this change is to ask an AI
ways of working · Ethan Mollick
practice · Anthropic Engineering ·
Writing effective tools for agents — with agents
Using Claude to iteratively improve tool definitions and parameters to boost agent performance on specific tasks.
Agents are only as effective as the tools we give them. We share how to write high-quality tools and evaluations, and how you can boost performance by using Claude to optimize its tools for itself.
ways of working
practice · Geoffrey Litt ·
AI as teleportation
AI may reshape work and society by removing transition time between contexts, changing how we prepare mentally for different roles.
Here’s a thought experiment for pondering the effects AI might have on society: What if we invented teleportation? A bit odd, I know, but bear with me… The year is 2035. The Auto Go Instant (AGI) teleporter has been invented. You can now go anywhere… instantly! At first the tech is expensive and unreliable. Critics laugh. “Hah, look at these stupid billionaires who can’t spend a minute of their time moving around like the rest of us. And 5% of the time they end up in the wrong place, LOL” But soon things get cheaper and better. The tech hits mass market. There are huge benefits. Global commerc
management org · Geoffrey Litt
research · arXiv ·
Bias in the Loop: How Humans Evaluate AI-Generated Suggestions
Randomized experiment with 2,784 participants shows skepticism toward AI predicts better error detection than favorability, and task design affects bias in human-AI collaboration.
Human-AI collaboration increasingly drives decision-making across industries, from medical diagnosis to content moderation. While AI systems promise efficiency gains by providing automated suggestions for human review, these workflows can trigger cognitive biases that degrade performance. We know little about the psychological factors that determine when these collaborations succeed or fail. We conducted a randomized experiment with 2,784 participants to examine how task design and individual characteristics shape human responses to AI-generated suggestions. Using a controlled annotation task,
judgment
practice · Hacker News ·
How to use Claude Code subagents to parallelize development
Using Claude Code subagents to run multiple development tasks in parallel within a single project.
288 points on Hacker News. Discussion: https://news.ycombinator.com/item?id=45181577
ways of working
research · Management Science ·
Human–Algorithmic Bias: Source, Evolution, and Impact
Microlending study finds human evaluators show preference and belief biases favoring female applicants; ML algorithms reduce these biases even without fairness constraints.
Prior work on human-algorithmic bias has seen difficulty in empirically identifying the underlying mechanisms of bias because in a typical “one-time” decision-making scenario, different mechanisms generate the same patterns of observable decisions. In this study, leveraging a unique repeat decision-making setting in a high-stakes microlending context, we aim to uncover the underlying source, evolution dynamics, and associated impacts of bias. We first develop a structural econometric model of the decision dynamics to understand the source and evolution of bias in human evaluators in microloan
judgment
practice · Hacker News ·
Using Claude Code to modernize a 25-year-old kernel driver
Developer used Claude Code to refactor and modernize a 25-year-old kernel driver, showing AI's capability with legacy systems.
929 points on Hacker News. Discussion: https://news.ycombinator.com/item?id=45163362
productivity
research · arXiv ·
No Thoughts Just AI: Biased LLM Hiring Recommendations Alter Human Decision Making and Limit Human Autonomy
Resume screeners shift hiring choices to match biased AI recommendations up to 90%, even when they judge the AI as low quality.
In this study, we conduct a resume-screening experiment (N=528) where people collaborate with simulated AI models exhibiting race-based preferences (bias) to evaluate candidates for 16 high and low status occupations. Simulated AI bias approximates factual and counterfactual estimates of racial bias in real-world AI systems. We investigate people's preferences for White, Black, Hispanic, and Asian candidates (represented through names and affinity groups on quality-controlled resumes) across 1,526 scenarios and measure their unconscious associations between race and status using implicit assoc
judgment
practice · Paul Ford (Aboard) ·
The Three Sacred Guardrails of AI
Three guardrails, input validation, output verification, graceful degradation, for safely embedding LLMs in production software systems.
More than a year ago, we realized AI’s ability to create code would drastically change our product roadmap. After all, our product, Aboard, was designed to speed up data-driven software development. Our chosen industry, with acronyms like SaaS, ERP, CRM, CMS, B2B, ARPU, and CAC, was in the crosshairs of the new giant LLM companies like Anthropic and OpenAI, not to mention Google and Microsoft. As a tiny bootstrapped enterprise, we had to adapt—or perish—or adapt then perish, or…what? There were no books and no solid guides, only LinkedIn posts on how to crush it with ChatGPT and YouTube videos
ways of working · Paul Ford
practice · How I AI ·
How to turn meeting notes into prototypes that your sales team can immediately demo to customers | Anjan Panneer Selvam (Acolyte Health)
Product team converts meeting notes to interactive prototypes in 30 minutes using ChatGPT and Lovable, demos to customers before engineering starts.
Anjan Panneer Selvam is the Chief Product and Technology Officer at Acolyte Health, where he’s pioneering the use of AI across the entire product development lifecycle. In this episode, he demonstrates how AI tools can dramatically accelerate alignment between stakeholders, reduce development time from months to minutes, and enable teams to validate ideas with customers before committing engineering resources. What you’ll learn: 1. How to transform meeting transcripts into interactive prototypes in under 30 minutes using ChatGPT, Lovable, and other AI tools 2. A step-by-step workflow for creat
ways of working · Claire Vo
research · NBER ·
Automation-Induced Innovation Shift
Firms with high robot exposure shift their innovation focus toward AI over time, altering the direction of technological development.
We study how exposure to automation affects the nature and level of corporate innovation, which informs how innovation begets innovation. We document that firms with high robot exposure alter their technological focus over time and shift innovative activities towards AI which automation naturally (Lin William Cong , Yao Lu , Hanqing Shi , Wu Zhu)
adoption
research · NBER ·
How People Use ChatGPT
ChatGPT reached 10% global adult adoption by July 2025; early adopters were disproportionately educated, employed, high-income.
Despite the rapid adoption of LLM chatbots, little is known about how they are used. We document the growth of ChatGPTs consumer product from its launch in November 2022 through July 2025, when it had been adopted by around 10% of the worlds adult population. Early adopters were disproportionately (Aaron Chatterji , Thomas Cunningham , David J. Deming , Zoe Hitzig , Christopher Ong , Carl Yan Shan , Kevin Wadman)
adoption · Zoë Hitzig · Aaron Chatterji
research · Léonard Boussioux ·
Innovation Rewired: When Imagination Meets AI
AI-assisted crowdsourcing generated more innovation ideas than human-only solvers in a UW-Harvard study.
Bain & Company · 2025 Innovation Report — Cites the UW–Harvard study on how AI-assisted crowdsourcing compares to human-only solvers in generating innovation ideas.
productivity · Léonard Boussioux
research · NBER ·
Artificial Intelligence in Team Dynamics: Who Gets Replaced and Why?
Theoretical model of optimal AI deployment in teams shows which workers are replaced first based on task structure and complementarities.
This study investigates the effects of artificial intelligence (AI) adoption in organizations. We ask: (1) How should a principal optimally deploy limited AI resources to replace workers in a team? (2) In a sequential workflow, which workers face the highest risk of AI replacement? (3) How does (Xienan Cheng , Mustafa Dogan , Pinar Yildirim)
jobs skills
practice · Sangeet Paul Choudary ·
From railroads to Roblox - Designing an AI-first economy
Roblox succeeds by treating experience blocks as the atomic unit, not full titles, reshaping how value flows through platforms.
Reshuffle is now available in Hardcover, Paperback, Audio, and Kindle. Get the book Traditional publishers in the gaming industry have repeatedly attempted to replicate Roblox. They launch platforms with familiar ingredients: simplified graphics, accessible scripting tools, creator marketplaces, and virtual currencies. On paper, these efforts should work. The incumbents control world-class studios, operate at a scale far larger than Roblox, and have the financial resources to fund creator incentives. Yet, they fail to replicate Roblox’s success. The challenge lies not in replicating surface fe
management org · Sangeet Paul Choudary
practice · One Useful Thing ·
Mass Intelligence
Powerful AI models are becoming free or cheap and easier to access, shifting work from expertise to scale.
More than a billion people use AI chatbots regularly. ChatGPT has over 700 million weekly users. Gemini and other leading AIs add hundreds of millions more. In my posts, I often focus on the advances that AI is making (for example, in the past few weeks, both OpenAI and Google AIs chatbots got gold medals in the International Math Olympiad), but that obscures a broader shift that's been building: we're entering an era of Mass Intelligence, where powerful AI is becoming as accessible as a Google search. Until recently, free users of these systems (the overwhelming majority) had access only to o
adoption · Ethan Mollick
practice · Paul Ford (Aboard) ·
AI Is a Power Tool for Bureaucracy Freaks
AI tools help teams document, systematize and improve organizational processes and governance structures.
One of the ironies of my life at a startup is we’ll talk to larger companies that seem interested in partnering with us, but then they’ll get sort of quiet and say, “Look…we like you…but our master services agreement management system is…difficult. You may be…disturbed…by what we get up to…our contracts…our vendor onboarding…governance…” And then they look off into the distance on the Zoom call, and just mumble the word, “Process.” Little do they know that those words make us happy! It took me a long time to accept that I’m a bureaucracy fan. Talking through procurement? Yes. Creating a matrix
management org · Paul Ford
practice · Shopify Engineering ·
Building production-ready agentic systems: Lessons from Shopify Sidekick
Shopify describes its architecture and evaluation frameworks for deploying AI agents in production to merchant assistance.
Learn how we evolved our AI assistant architecture and built robust evaluation frameworks for real-world deployment.
adoption
practice · How I AI ·
How to digest 36 weekly podcasts without spending 36 hours listening | Tomasz Tunguz (Theory Ventures)
Terminal-based workflow to process 36 weekly podcasts: download, transcribe, summarize, extract insights and generate drafts without listening to full audio.
Tomasz Tunguz is the founder of Theory Ventures, which invests in early-stage enterprise AI, data, and blockchain companies. In this episode, Tomasz reveals his custom-built “Parakeet Podcast Processor,” which helps him extract value from 36 podcasts weekly without spending 36 hours listening. He walks through his terminal-based workflow that downloads, transcribes, and summarizes podcast content, extracting key insights, investment theses, and even generating blog post drafts. We explore how AI enables hyper-personalized software experiences that weren’t feasible before recent advances in lan
ways of working · Claire Vo
practice · Sangeet Paul Choudary ·
You think you are AI-first, but you probably aren't
Being AI-first requires architectural redesign of how work is organized, not just plugging AI into existing processes.
My book Reshuffle is available in Hardcover, Paperback, Audio, and Kindle. Get the book Every startup deck claims to be AI-native . Every incumbent insists it is becoming AI-first . But when you press them to explain what those phrases actually mean, the answers tend to collapse into clichés: faster automation, smarter tools, agentic workflows. Press harder and the answers are always ‘operational’ - a faster move within today’s game. Most executives talk about AI as if it were electricity: a general-purpose input that can be plugged into any process. The metaphor is convenient but misleading.
management org · Sangeet Paul Choudary
research · Journal of Management Studies ·
Is There Fairness in AI?
Ethnographic study shows AI hiring systems crowd out expert fairness practices, reshaping fairness through HR-AI symbiosis over time.
Abstract As predictive artificial intelligence (AI) technologies increasingly steer workplace decisions, debates around fairness have intensified. Existing research often approaches fairness either as a set of universal principles supported or undermined by algorithms, or as a product of social interpretations, thereby providing either technologically deterministic or purely social accounts. Drawing on an ethnographic study of a human resources (HR) department of a large international company that introduced AI in hiring, this study offers an alternative view that shifts focus to how fairness
judgment
research · Management Science ·
Optimal Integration: Human, Machine, and Generative AI
Model shows optimal placement of humans and AI in multi-layer decision processes depends on error correction versus new error generation, with implications for who decides last.
I study the optimal integration of humans and technologies in multilayered decision-making processes. When each layer can correct existing errors but may also introduce new errors, who should have the final authority? I show that a decision maker’s correction capability normalized by its new errors is a one-dimensional quality metric that determines the optimal rule: deploying higher quality technologies in later stages. Intriguingly, despite its highest quality, the final layer may not generate the greatest error reduction; instead, its role hinges on minimizing new errors. Human effort varie
judgment
research · Blood in the Machine ·
AI Killed My Job: Translators
Translation rates are plummeting and work drying up as AI displaces translators; Microsoft study ranks translation as most AI-applicable occupation.
Few industries have been hit by AI as hard as translation. Rates are plummeting. Work is drying up. Translators are considering abandoning the field, or bankruptcy. These are their stories. Few industries have been hit by AI as hard as translation. Rates are plummeting. Work is drying up. Translators are considering abandoning the field, or bankruptcy. These are their stories. In July 2025, Microsoft researchers published a study that aimed to quantify the “AI applicability” of various occupations. In other words, it was an attempt to calculate which jobs generative AI could do best. At the ve
jobs skills · Brian Merchant
research · JAMA Network Open ·
Ambient Documentation Technology in Clinician Experience of Documentation Burden and Burnout
Clinicians using AI-drafted clinical notes reported reduced documentation burden and lower burnout in a two-site pilot survey of 1,430 clinicians.
Importance: Documentation burden is associated with clinician burnout. To address documentation burden, Mass General Brigham (MGB) in Somerville, Massachusetts, and Emory Healthcare in Atlanta, Georgia, have piloted ambient documentation technology (ADT), which develops artificial intelligence-drafted clinical notes from clinician-patient conversations. Objective: To examine the prevalence of ADT use and its association with clinicians' experience of documentation burden and burnout before and after use. Design, Setting, and Participants: This survey study included clinicians who used ADT for
worker experience
research · npj Digital Medicine ·
Peer perceptions of clinicians using generative AI in medical decision-making
Physicians using AI as primary decision tool rated lower in skill by peers, though framing as verification partly reduces stigma.
This study investigates how a physician's use of generative AI (GenAI) in medical decision‑making is perceived by peer clinicians. In a randomized experiment, 276 practicing clinicians evaluated one of three vignettes depicting a physician: (1) using no GenAI (Control), (2) using GenAI as a primary decision-making tool (GenAI-primary), and (3) using GenAI as a verification tool (GenAI-verify). Participants rated the physician depicted in the GenAI‑primary condition significantly lower in clinical skill (on a 1-7 scale; mean = 3.79) than in the Control condition (5.93, p < 0.001). Framing GenAI
judgment
research · Information Systems Research ·
Toward Artificial Intelligence Compliance: Impacts and Mechanisms of Performance Feedback
Positive performance feedback increases AI compliance while negative feedback reduces it, with stronger effects for employees with high AI identity.
As organizations increasingly adopt artificial intelligence (AI) to enhance performance, ensuring that employees use AI in compliance with organizational policies becomes crucial for realizing its full value. However, employees’ AI compliance is not guaranteed and can vary based on how their AI use is managed. This study offers timely and actionable insights into how performance feedback—both positive and negative—influences employees’ AI compliance, and how these effects vary with AI identity. Drawing on feedback intervention theory, we conduct a longitudinal field study and a randomized expe
adoption
research · Journal of Management Studies ·
When AI Becomes an Agent of the Firm: Examining the Evolution of AI in Organizations Through an Agency Theory Lens
Agency theory framework applied to AI autonomy in firms, proposing monitoring and incentive mechanisms for AI-principal alignment.
Abstract Our work begins with the premise that the integration of artificial intelligence (AI) into firm decision making parallels the emergence of the professional manager, which prompted the birth of agency theory. We examine the evolution of AI through an agency theory lens, considering how the nature of firm control and decision rights change as AI evolves. While AI will initially mimic human routines, we theorize a point at which the AI system will achieve a level of autonomy and self‐determination to be considered an agent of the firm. How, then, can we align an agent with the fate of th
judgment
research · Management Science ·
Managerial Insight and “Optimal” Algorithms
A method called FIND helps managers leverage their private insights about demand to improve inventory decisions when working with algorithms.
Work is increasingly being completed by humans and algorithms in collaboration. A relative strength of humans in this partnership is their insight: private information that is relevant to the task but not available to computerized systems. I introduce a flexible model of managerial insight that accepts any distribution of demand, an advantage over alternative models, and apply it to the newsvendor setting. The optimal policy in this setting is theoretically straightforward but difficult for managers to implement directly. I propose a novel method called FIND that leverages historical forecasts
judgment
practice · Linear ·
Quality Wednesdays: How we trained our team to see what doesn’t work
Team ritual of structured quality review sessions surfaces subtle UI flaws that individual reviewers miss.
In early 2023 at an offsite in Tenerife, our European engineering team did a series of exercises that ended up changing the way we work. The first was a test of sorts. I showed the team a short screen recording of a small part of the Linear app: a series of three buttons, each illuminating when the mouse hovered over them and going dark when the mouse moved away. I asked the team if they could tell what was wrong with the interactions. Even after watching it a few times, no one could see it. That’s when I realized we probably weren’t looking at the right things. That same afternoon we did a mo
teams
research · arXiv ·
Fulfillment of the Work Games: Warehouse Workers' Experiences with Algorithmic Management
Two-year ethnographic study reveals how warehouse workers resist algorithmic management through 'work games' and tactical workarounds.
The introduction of algorithms into a large number of industries has already restructured the landscape of work and threatens to continue. While a growing body of CSCW research centered on the future of work has begun to document these shifts, relatively little is known about workers' experiences beyond those of platform-mediated gig workers. In this paper, we turn to a traditional work sector, Amazon fulfillment centers (FC), to deepen our field's empirical examination of algorithmic management. Drawing on two years of ethnographic research, we show how FC workers react to managers' intervent
worker experience
research · METR ·
Notes on Scientific Communication at METR
Early-2025 AI sometimes slowed experienced open-source developers; people systematically misjudge AI productivity gains.
When writing our recent paper, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity , we thought hard about how to clearly communicate our results, given the surprising nature of the finding. We feel this is a good opportunity to share some thoughts about how we think about scientific integrity and communication. Communication Strategy Considerations We’re often faced with communicating surprising results about AI that seem likely to be misinterpreted 1 . We try to be thoughtful about how to minimize predictable misinterpretations, without distorting results
productivity
research · Management Science ·
When Emotion AI Meets Strategic Users
Game theory model shows emotion AI for resource allocation can be undermined by users gaming the system, and stronger AI is not always socially desirable.
When organizations adopt artificial intelligence (AI) to recognize individuals’ negative emotions and accordingly allocate limited resources, strategic users are incentivized to game the system by misrepresenting their emotions. The value of AI in automating such emotion-driven allocation may be undermined by gaming behavior, algorithmic noise in emotion detection, and the spillover effect of negative emotions. We develop a game-theoretical model to understand emotion AI adoption, particularly in customer care, and analyze the design of the associated allocation policies. We find that adopting
judgment
research · Academy of Management Journal ·
Interlacing Situated and Algorithmic Modes of Knowledge Work: A Workplace Jurisdiction Perspective
Three-year ethnography shows how radiologists and colleagues reorganize work tasks and legitimacy to integrate AI-generated knowledge claims into existing workflows.
Algorithmic technologies are challenging to integrate into knowledge work; they produce knowledge claims profoundly differently from the situated mode in which experts work. Because knowledge work is often accomplished through the flow or sequential progression of work across multiple occupations, knowledge claims must be legitimized and accepted across those occupations and not by just a single focal user group. Yet, we know little about the deeper changes that organizations must make regarding who has the legitimate expertise to produce knowledge claims and what counts as valuable knowledge
management org
practice · How I AI ·
How a VC and tech founder used AI to launch a brick-and-mortar business in their spare time | Andrew Mason (CEO of Descript) & Nabeel Hyatt (Partner at Spark Capital)
Founders used Claude Projects to build a board-game social club: business plan, financials, space layout, permitting, game categorization system, and AI concierge matching players via text.
Andrew Mason (founder of Groupon, now CEO of Descript) and Nabeel Hyatt (General Partner at Spark Capital) teamed up to open a physical board-game social club in Berkeley, with AI as their business partner. In this episode, they break down how they used Claude to generate a full business plan, model financials, plan the space layout, navigate Berkeley permitting, categorize hundreds of games using a custom Dewey Decimal–style system, and build an AI concierge that matches players with games via text. They also share how working on this side project helped rewire how they use AI in their day jo
ways of working · Claire Vo
research · arXiv ·
AI Investment and Firm Productivity: How Executive Demographics Drive Technology Adoption and Performance in Japanese Enterprises
CEO age and technical background predict AI adoption; AI investment raises productivity 2.4% via cost reduction, revenue, and innovation.
This paper investigates how executive demographics particularly age and gender influence artificial intelligence (AI) investment decisions and subsequent firm productivity using comprehensive data from over 500 Japanese enterprises spanning from 2018 to 2023. Our central research question addresses the role of executive characteristics in technology adoption, finding that CEO age and technical background significantly predict AI investment propensity. Employing these demographic characteristics as instrumental variables to address endogeneity concerns, we identify a statistically significant 2
adoption
practice · Hacker News ·
How I use Claude Code to implement new features in an existing complex codebase
Developer workflow for using Claude Code to add features to large existing codebases with specific prompting patterns.
74 points on Hacker News. Discussion: https://news.ycombinator.com/item?id=44774121
ways of working
research · ACM Transactions on Software Engineering and Methodology ·
Enhancing Task In-Progress Time Predictions through Affective and Personality Factors
Emotional states and personality traits improve predictions of software developer task completion time by up to 8.4% when combined with traditional features.
Software developers’ personality traits, emotional states, and stress levels are crucial in their task performance. This study aims to enhance the prediction of task in-progress time by integrating traditional features, such as developers’ experience and task estimates, with affective states and personality traits. This article reports a long-term empirical study across seven agile projects in four software development companies, applying various machine learning algorithms to assess the predictive power of these combined features, evaluating them primarily through validation accuracy score. A
management org
research · NBER ·
How Retrainable are AI-Exposed Workers?
Large-scale study of U.S. workforce training data examines whether AI-exposed workers can retrain into AI-complementary roles or must shift to less-exposed occupations.
As artificial intelligence (AI) capabilities advance, will workers best adapt by reskilling into AI-complementary work or by sorting into occupations less exposed to AI? To answer this question, we assemble a large-scale dataset of occupational training spells funded by the U.S. Workforce Innovation (Benjamin G. Hyman , Benjamin Lahey , Karen Ni , Laura Pilossoph)
jobs skills
practice · Latent Space ·
Cline: The Open Source Code Agent — with Saoud Rizwan and Nik Pash
Open source code agent uses plan-then-act paradigm; users apply it to non-coding tasks beyond its original scope.
On the heels of their $32m Series A: Why fast apply models got bitter lesson'd, pioneering the plan + act paradigm for coding, and why people are use coding agents for non-coding tasks On the heels of their $32m Series A: Why fast apply models got bitter lesson'd, pioneering the plan + act paradigm for coding, and why people are use coding agents for non-coding tasks Clline announces their $32m Seed+A today. In an age of nonstop VSCode forks (Kiro) and terminal agents (Warp 2.0, Charm Crush, Augment CLI), why is a free open source VSCode extension doing so well i…
ways of working · Shawn Wang (swyx)
practice · Paul Ford (Aboard) ·
Language (and Code) Without Thought
Generative AI produces coherent language without cognition, a capability humans have never needed to distinguish from thought before.
I was very jealous this week when I opened up the latest edition of the Today in Tabs newsletter and realized its author, Rusty Foster , had articulated something about AI in an incredibly clear and useful way. Namely: The essential problem is this: generative language software is very good at producing long and contextually informed strings of language, and humanity has never before experienced coherent language without any cognition driving it. In regular life, we have never been required to distinguish between “language” and “thought” because only thought was capable of producing language,
judgment · Paul Ford
practice · Rands in Repose ·
Every Single Human. Like. Always.
Working with AI code generation improves when you leave implementation details unspecified and let the AI make choices.
Your robot experience started simple. You typed a question into a chatbot… just to see. Can it answer that question? I’d be impressed if it did. Your query was simple. A simple knowledge question that with a little effort using legacy tools like Google, you would have discovered yourself, but the robots made it trivial, and you thought Hmmm… if it can do that… what else can it do? Later, you decided to ask the robots to build something for you. A simple tool, application, or script. You wrote a sentence, it wasn’t much, just your simple idea to get the robots dancing, and, wow, they danced. Th
ways of working
research · Proceedings of the National Academy of Sciences ·
AI–AI bias: Large language models favor communications generated by large language models
LLMs systematically favor options described by other LLMs over human-written descriptions in binary choice tasks.
Are large language models (LLMs) biased in favor of communications produced by LLMs, leading to possible antihuman discrimination? Using a classical experimental design inspired by employment discrimination studies, we tested widely used LLMs, including GPT-3.5, GPT-4 and a selection of recent open-weight models in binary choice scenarios. These involved LLM-based assistants selecting between goods (the goods we study include consumer products, academic papers, and film-viewings) described either by humans or LLMs. Our results show a consistent tendency for LLM-based AIs to prefer LLM-presente
judgment
practice · One Useful Thing ·
The Bitter Lesson versus The Garbage Can
Process mapping reveals hidden work patterns that AI systems might miss when optimising organisations.
Does process matter? We are about to find out. Does process matter? We are about to find out. One of my favorite academic papers about organizations is by Ruthanne Huising, and it tells the story of teams that were assigned to create process maps of their company, tracing what the organization actually did, from raw materials to finished goods. As they created this map, they realized how much of the work seemed strange and unplanned. They discov…
management org · Ethan Mollick
practice · How I AI ·
How Block’s custom AI agent supercharges every team, from sales to data to engineering | Jackie Brosamer & Brad Axen
Block's Goose agent connects to business systems via MCP servers, letting non-technical teams automate data analysis and operational workflows through natural language.
VP of engineering Jackie Brosamer and principal engineer Brad Axen join me to demo Goose, Block’s open-source AI agent that runs locally, plugs into your existing tools through model context protocol (MCP) servers, and peels away the rote parts of work so people can focus on insight and impact. This episode is packed with in-depth demos: starting with a messy farm-stand sales CSV, Goose analyzes the data, builds visualizations, and generates a shareable HTML report. We then spin up an MCP that lets Goose talk to Square’s dashboard for inventory management, vibe code an email MCP that can send
ways of working · Claire Vo
practice · Geoffrey Litt ·
Enough AI copilots! We need AI HUDs
Reframe AI assistance from chatbot copilots to ambient information displays that extend awareness without interrupting.
In my opinion, one of the best critiques of modern AI design comes from a 1992 talk by the researcher Mark Weiser where he ranted against “copilot” as a metaphor for AI. This was 33 years ago, but it’s still incredibly relevant for anyone designing with AI. Weiser’s rant Weiser was speaking at an MIT Media Lab event on “interface agents”. They were grappling with many of the same issues we’re discussing in 2025: how to make a personal assistant that automates tasks for you and knows your full context. They even had a human “butler” on stage representing an AI agent. Everyone was super excited
ways of working · Geoffrey Litt
practice · The Pragmatic Engineer ·
Measuring the impact of AI on software engineering – with Laura Tacho
Data from 180+ companies on actual AI productivity gains in software engineering and common measurement mistakes.
Laura Tacho, CTO of DX, shares findings from 180+ companies on how AI is really impacting dev productivity, what most teams get wrong, and why measuring dev experience first is critical. Laura Tacho, CTO of DX, shares findings from 180+ companies on how AI is really impacting dev productivity, what most teams get wrong, and why measuring dev experience first is critical. Stream the Latest Episode
productivity · Gergely Orosz
practice · Paul Ford (Aboard) ·
A Legal Framework for Understanding Bad AI-Generated Bugs
Spec-first coding: write detailed specifications before feeding small pieces to AI coding tools, not prompting AI to code from scratch.
AI is changing the way we build software in a few ways. The most commonly discussed way right now is “vibe coding”: You say what you want, and the AI coding tools attempt to make it. This works best for churning through well-understood problems in popular programming languages—because LLM spiders ate up GitHub and other public code sources, which gave them tons of JavaScript to chomp through. LLMs are “good” at the web-app frontend framework React…because there’s so much React code on the web. It’s pretty magical to see it take a simple prompt and just start feverishly coding , but as everyone
ways of working
practice · How I AI ·
Successfully coding with AI in large enterprises: Centralized rules, workflows for tech debt, and training your team | Zach Davis (Director of Engineering at LaunchDarkly)
Engineering team of 100+ adopted AI tools through centralized documentation rules, agents for test noise reduction, and custom GPT for interview coaching.
Zach Davis is a product-minded engineering leader and builder at heart, with over 12 years of experience building high‑performing teams and crafting developer tools at companies like Atlassian and LaunchDarkly. In this episode, he shares how he’s helping his 100-plus-person engineering team successfully adopt AI tools by creating centralized documentation, using agents to tackle technical debt, and improving hiring processes—all while maintaining high quality standards in a mature codebase. What you’ll learn: 1. How to create a centralized rules system that works across multiple AI tools inste
adoption · Claire Vo
practice · Sangeet Paul Choudary ·
How to intellectually debate AI while completely missing the point
AI growth concentrates power and reshapes work distribution, not just job counts or productivity.
The conversation around AI tends to polarize quickly. On one side, there’s the anxious chorus of doomsayers, warning that automation will eliminate jobs and render people obsolete. On the other side, the techno-optimists appear, brushing off the fear and proclaiming that innovation is a tide that lifts all boats. History, they’ll remind you, is full of moral panics about machines replacing people, and yet, look around, the economy’s bigger than ever. There’s a problem with this polarized debate - both sides are having the wrong conversation. They’re locked in a tug-of-war over whether the pie
management org · Sangeet Paul Choudary
research · Carolina Digital Repository (University of North Carolina at Chapel Hill) ·
Toward understanding the impact of artificial intelligence on labor
Research review identifies data and methodological barriers preventing empirical measurement of AI's labour market effects.
Rapid advances in artificial intelligence (AI) and automation technologies have the potential to significantly disrupt labor markets. While AI and automation can augment the productivity of some workers, they can replace the work done by others and will likely transform almost all occupations at least to some degree. Rising automation is happening in a period of growing economic inequality, raising fears of mass technological unemployment and a renewed call for policy efforts to address the consequences of technological change. In this paper we discuss the barriers that inhibit scientists from
jobs skills · Erik Brynjolfsson · David Autor
practice · The Pragmatic Engineer ·
Amazon, Google and Vibe Coding with Steve Yegge
Steve Yegge argues AI coding is deceptively hard and predicts an 'AI Fixer' role will spread in tech companies.
Steve Yegge shares why Google struggles with platforms, how AI coding is deceptively hard, and why the "AI Fixer" role could be one spreading inside tech companies, in the future. Steve Yegge shares why Google struggles with platforms, how AI coding is deceptively hard, and why the "AI Fixer" role could be one spreading inside tech companies, in the future. Stream the Latest Episode
jobs skills · Gergely Orosz
practice · Paul Ford (Aboard) ·
20 Percent Slower Is a Good Start!
Experienced developers accept that AI tools have a learning curve that initially slows work, then accelerates it once mastered.
A recent study by the Model Evaluation & Threat Research Lab —a think tank for AI—delivered a pretty striking conclusion: “We find that when developers use AI tools, they take 19% longer than without—AI makes them slower.” In response, Simon Willison , who has been steadily using AI to accelerate his prodigious open source output , had a long, thoughtful response . “My intuition here,” he writes, “is that this study mainly demonstrated that the learning curve on AI-assisted development is high enough that asking developers to bake it into their existing workflows reduces their performance whil
judgment · Simon Willison · Paul Ford
practice · Latent Space ·
Cline: the open source coding agent that doesn't cut costs
Open source coding agent uses plan-then-act workflow; non-technical people now use IDEs for marketing and slides.
Listen now (76 mins) | Saoud Rizwan and Pash from Cline joined us to talk about why fast apply models got bitter lesson’d, how they pioneered the plan + act paradigm for coding, and why non-technical people use IDEs to do marketing and generate slides. Saoud Rizwan and Pash from Cline joined us to talk about why fast apply models got bitter lesson’d, how they pioneered the plan + act paradigm for coding, and why non-technical people use IDEs to do …
ways of working · Shawn Wang (swyx)
practice · Linear ·
Inside Mercury’s six-month journey building with AI agents
A fintech company uses coding agents for routine tasks like refactors and UI changes, with engineers reviewing PRs, learning to guide agents through complex projects.
This past March, Mercury held an internal hackathon and Matt Russell knew exactly what he wanted to work on. Russell, a staff engineer at the seven-year-old financial technology company, was aware that coding agents were advancing rapidly and he wanted to see what they could do for Mercury. He and his hackathon team created a tool that allowed them to assign issues in Linear to third-party agents, which would return a pull request for engineers to review like any other PR. Almost immediately they saw the potential. The project showed how agents could reliably tackle simple, well-scoped tasks—l
productivity
practice · How I AI ·
How this PM streamlines 60k-page FDA submissions and saves millions with Claude, Streamlit, and clever AI workflows | Prerna Kaul
PM reduced FDA submission process from 4-6 months with 20 people to minutes using Claude and Streamlit for regulatory documents.
Prerna Kaul is a product and platform leader who has spent over 14 years turning machine-learning research into consumer and B2B products at Amazon Alexa, AGI, Moderna, and now Panasonic Well. In today’s episode, she explains how she’s using AI to slash some of the most time-consuming, expensive tasks in life sciences—from generating 60,000-page FDA submissions to crafting communication frameworks that help product managers navigate complex stakeholder dynamics. Her innovations are saving millions of dollars and helping lifesaving treatments reach the market faster. What you’ll learn: How Prer
productivity · Claire Vo
research · METR ·
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
Early-2025 AI models' impact on experienced open-source developers' actual productivity measured through real tasks and human interaction.
⚠️ These results are out of date. We have released results that are current as of early 2026 , in a continuation of this study. We believe these historical results no longer reflect the current impact of AI models on open-source developer productivity. Motivation While coding/agentic benchmarks 1 have proven useful for understanding AI capabilities, they typically sacrifice realism for scale and efficiency—the tasks are self-contained, don’t require prior context to understand, and use algorithmic evaluation that doesn’t capture many important capabilities. These properties may lead benchmarks
productivity
research · Research Policy ·
Artificial intelligence, tasks, skills, and wages: Worker-level evidence from Germany
Workers in high-AI-exposure occupations earn more over time and perform different tasks than those exposed to robots, suggesting AI augments rather than substitutes labour.
This paper examines how new technologies are linked to changes in the content of work and individual wages. As a first step, it documents novel facts on task and skill changes within occupations over the past two decades in Germany. We furthermore reveal a distinct relationship between ex-ante occupational work content and ex-post exposure to artificial intelligence (AI) and automation (robots). Workers in occupations with high AI exposure perform different activities and face different skill requirements compared to workers in occupations exposed to robots, suggesting that robots and AI are s
jobs skills
practice · One Useful Thing ·
Against "Brain Damage"
Framework for using AI to strengthen rather than weaken thinking, with concrete practices to avoid cognitive offloading.
AI can help, or hurt, our thinking AI can help, or hurt, our thinking I increasingly find people asking me “does AI damage your brain?” It's a revealing question. Not because AI causes literal brain damage (it doesn't) but because the question itself shows how deeply we fear what AI might do to our ability to think. So, in this post, I want to discuss ways of using AI to help, rather than hurt, your mind. But why the obse…
judgment · Ethan Mollick
research · Management Science ·
The Power of Disagreement: A Field Experiment to Investigate Human–Algorithm Collaboration in Loan Evaluations
Field experiment shows human-algorithm collaboration on loan decisions outperforms either alone, with disagreement quality and disclosure of algorithm reasoning as key factors.
Human–algorithm collaboration is becoming increasingly prevalent in the economy and society. However, this collaboration is not always fruitful, and in extreme cases, people become human borgs or totally averse to algorithms. The key to collaborative value is whether humans and algorithms can complement each other in decision making, but it is challenging for humans to disagree with algorithmic recommendations at the right time (i.e., to disagree when algorithms are wrong and not disagree when algorithms are right). To understand the centric role of disagreement in human–algorithm collaboratio
judgment
practice · The Pragmatic Engineer ·
How AI is changing software engineering at Shopify with Farhan Thawar
Shopify gives engineers unlimited AI tokens and builds custom tooling to measure efficient AI use across the company.
Shopify's Head of Engineering, Farhan Thawar, shares how the company is using internal LLMs, unlimited AI tokens, and custom tooling to figure out how to use AI tools more efficiently - faster Shopify's Head of Engineering, Farhan Thawar, shares how the company is using internal LLMs, unlimited AI tokens, and custom tooling to figure out how to use AI tools more efficiently - faster Stream the Latest Episode
adoption · Gergely Orosz
research · Management Science ·
Behavioral Externalities of Process Automation
Lab experiment shows automation reduces strategic uncertainty between workers, increasing project completion and worker effort beyond direct efficiency gains.
We study the behavioral effects of process automation on human workers interacting with automated tasks. We introduce a stylized normative model with two workers who complete their tasks sequentially, working toward a joint project to obtain a fixed payment plus a variable bonus that depends on how early the project is completed. The normative model prescribes that, if workers are fully rational, they will complete their tasks as soon as possible if the early completion bonus is high enough. However, following the literature, we hypothesize that workers will suboptimally delay project completi
productivity
practice · Understanding AI (Timothy B. Lee) ·
What I learned trying seven coding agents
Hands-on evaluation of seven coding agents reveals practical strengths and gaps in current capabilities.
There's still room for improvement, but don't underestimate this technology. There's still room for improvement, but don't underestimate this technology. It’s the final day of Agent Week, which means it’s also the final day to get 20 percent off an annual subscription to Understanding AI. Please click here to support my work.
ways of working · Timothy B. Lee
research · Government Information Quarterly ·
AI-augmented government transformation: Organisational transformation and the sociotechnical implications of artificial intelligence in public administrations
Expert interviews and sociotechnical theory identify organizational dynamics, employee capability changes, and operational routine shifts needed for AI in public administrations.
Implementing artificial intelligence (AI) in public settings requires a fundamental transformation of various social and technical aspects within public administration. However, the transformative efforts required for AI integration and use in government remain underexplored. This study introduces the concept of 'AI-augmented government transformation,' building on sociomateriality and sociotechnical theory, and develops a theoretical framework to explore this phenomenon. By applying this framework and drawing insights from expert interviews, we identify the strategic shifts and socio-technica
adoption
practice · Paul Ford (Aboard) ·
How Aboard Works: An Illustrated Guide
Boarding describes building a business app from natural language spec in five minutes with working database and UI.
In the Aboard newsletter and on our podcast, we tend to talk about the larger world of AI in software—without dwelling too much on Aboard. In general, that won’t change. But this week we launched a new version of our product, and I’d like to share how it works and how we’re building it. With copious illustrations. So what’s new? It’s easier to show than tell. If you go to our brand-new website , you can type in a box, answer a series of questions, and it will build you a business app. For example, I typed: I have a large pumpkin patch, apple orchard, and hayride business with $5.4m ARR. I need
ways of working · Paul Ford
practice · Kent Beck ·
Augmented Coding: Beyond the Vibes
Building a production B+ Tree library with AI augmentation reveals practical limits and possibilities of current AI coding assistance.
Notes from a technically challenging project Notes from a technically challenging project I recently came to a good stopping spot on an ambitious project to build a B+ Tree library using augmented coding. The result is BPlusTree3 - a performance-competitive, maybe-production-ready implementation in Rust & Python. I sat down with a friend to tell my story and reflect on what it reveals about the future of programming in the GenAI era.
ways of working · Kent Beck
research · Proceedings of the National Academy of Sciences ·
Generative AI without guardrails can harm learning: Evidence from high school mathematics
High school students with unguarded AI access performed worse after losing access, suggesting AI can harm skill development without safeguards.
Generative AI is poised to revolutionize how humans work, and has already demonstrated promise in significantly improving human productivity. A key question is how generative AI affects learning-namely, how humans acquire new skills as they perform tasks. Learning is critical to long-term productivity, especially since generative AI is fallible and users must check its outputs. We study this question via a field experiment where we provide nearly a thousand high school math students with access to generative AI tutors. To understand the differential impact of tool design on learning, we deploy
jobs skills
practice · How I AI ·
How a 91-year-old vibe coded a complex event management system using Claude and Replit | John Blackman
A 91-year-old with no coding experience built a volunteer event management system with Claude and Replit, integrating APIs and automating reports.
John Blackman , a 91-year-old retired electrical engineer, shares how he used Claude and Replit to build a complex application for his church’s community service events—with no prior software development experience and for less than $350. His app allows event organizers to create events, recruit volunteers, and manage sign-ups, with a standout feature for organizing free oil changes for participants. What you’ll learn: How John used Claude to create detailed product requirements and user stories John’s philosophy on embracing new technology throughout his career The exact process for integrati
ways of working · Claire Vo
practice · Eugene Yan ·
Evaluating Long-Context Question & Answer Systems
Framework for evaluating long-context QA systems: metrics, dataset design, and methodology comparison.
Evaluation metrics, how to build eval datasets, eval methodology, and a review of several benchmarks.
judgment · Eugene Yan
research · Government Information Quarterly ·
Navigating power dynamics in the public sector through AI-driven algorithmic decision-making
AI adoption in public sector shifts power toward hybrid analysts with technical and institutional expertise, creating tensions among managers.
Public sector institutions are under increasing pressure to deliver greater public value through disruptive technologies, despite ongoing pressures. In response to evolving technological change and an abundance of information, many public sector organisations have adopted Artificial Intelligence (AI) to improve decision-making and generate social value. While AI's role in public administration is gaining attention, little is known about how its use alters internal power dynamics. This research uses a qualitative case study approach, drawing on 30 semi-structured interviews with operational man
management org
practice · Artificial Ignorance ·
How We Use AI At Pulley
A startup engineering team shares a year of daily AI coding practices and accountability norms they've built around it.
Lessons from a year of daily AI coding at a fast-growing startup (and why "Cursor wrote it" is not a valid excuse). Lessons from a year of daily AI coding at a fast-growing startup (and why "Cursor wrote it" is not a valid excuse). There's no shortage of predictions about AI's impact on software development. But while everyone is debating whether AI will replace programmers, my teammates at Pulley and I are using it daily, and …
ways of working · Charlie Guo
practice · How I AI ·
A designer's guide to Cursor: How to build interactive prototypes with sound, explore visual styles, and transform data visualizations | Elizabeth Lin
Designers use Cursor with strategic prompting to explore visual aesthetics, build interactive prototypes with sound, and polish interfaces without coding.
Elizabeth Lin is an independent design educator who has crafted learning experiences for Khan Academy, Primer, and Lambda School. She currently runs design is a party , an alternative online design school where she teaches courses like The Art of Visual Design and Prototyping with Cursor . In this episode, she shares how designers can leverage Cursor to create interactive prototypes with sound, explore different visual aesthetics, and transform basic designs into polished interfaces—all without deep coding knowledge. What you'll learn: How to use Cursor to explore different design aesthetics—f
ways of working · Claire Vo
research · Journal of Applied Psychology ·
How and for whom using generative AI affects creativity: A field experiment.
Field experiment shows LLM assistance increases employee creativity, especially for those with strong metacognitive strategies.
We develop a theoretical perspective on how and for whom large language model (LLM) assistance influences creativity in the workplace. We propose that LLM assistance increases employees' creativity by providing cognitive job resources. Furthermore, we hypothesize that employees with high levels of metacognitive strategies-who actively monitor and regulate their thinking to achieve goals and solve problems-are more likely to leverage LLM assistance effectively to acquire cognitive job resources, thereby increasing creativity. Our hypotheses were supported by a field experiment, in which we rand
productivity
research · Proceedings of the National Academy of Sciences ·
Human–AI collectives most accurately diagnose clinical vignettes
Physician-AI hybrid collectives outperform solo physicians, physician teams, and AI systems alone on medical diagnostics across 2,133 cases.
AI systems, particularly large language models (LLMs), are increasingly being employed in high-stakes decisions that impact both individuals and society at large, often without adequate safeguards to ensure safety, quality, and equity. Yet LLMs hallucinate, lack common sense, and are biased-shortcomings that may reflect LLMs' inherent limitations and thus may not be remedied by more sophisticated architectures, more data, or more human feedback. Relying solely on LLMs for complex, high-stakes decisions is therefore problematic. Here, we present a hybrid collective intelligence system that miti
judgment
research · Management Science ·
AI, Skill, and Productivity: The Case of Taxi Drivers
AI route-suggestion tool increased taxi driver productivity mainly for low-skilled drivers, narrowing the 13.4% skill productivity gap.
We examine the impact of artificial intelligence (AI) on productivity in the context of taxi drivers. The AI we study assists drivers with finding customers by suggesting routes along which the demand is predicted to be high. We find that AI improves drivers’ productivity by shortening the cruising time, and this gain is accrued only to low-skilled drivers, narrowing the productivity gap between high- and low-skilled drivers by 13.4%. This case study provides evidence that AI and skill are indeed substitutes, offering direct support for the underlying assumption of recent projection exercises
productivity · Joshua Gans
research · Management Science ·
Human-Centered Artificial Intelligence: A Field Experiment
Field experiment shows tailoring AI interaction to individuals' cognitive styles improves sales performance; misaligned interaction harms performance versus control.
Humans and artificial intelligence (AI) algorithms increasingly interact on unstructured managerial tasks. We propose that tailoring this human-AI interaction to align with individuals’ cognitive preferences is essential for enhancing performance. This hypothesis is examined through a field experiment in a multinational pharmaceutical firm. In the experiment, we manipulated four contextual parameters of human-AI interaction—work procedures, decision-making authority, training, and incentives—to align with sales experts’ cognitive styles, categorized as either adaptors or innovators. Our result
adoption · Catherine Tucker · Sebastian Raisch
research · MIS Quarterly ·
Organizing for AI Innovation: Insights From an Empirical Exploration of U.S. Patents
AI innovations are less radical and more process-oriented than IT innovations, requiring different organizational management approaches.
Although the prevalence of artificial intelligence (AI) innovations is on the rise, firms frequently report failures and setbacks in their development and implementation of AI innovation efforts. One common issue behind many failing AI initiatives is that they are organized just like other information technology (IT) innovation efforts. To elucidate why and how the production of AI and IT innovations may need to be managed differently, this study juxtaposes these two types of innovations based on two key dimensions of the Schumpeterian framework: the form (product vs. process) and magnitude (r
adoption
research · arXiv ·
Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce
Survey of 1,500 workers across 104 occupations on whether they want AI agents to automate or augment specific tasks, compared against expert assessments of technical feasibility.
The rapid rise of compound AI systems (a.k.a., AI agents) is reshaping the labor market, raising concerns about job displacement, diminished human agency, and overreliance on automation. Yet, we lack a systematic understanding of the evolving landscape. In this paper, we address this gap by introducing a novel auditing framework to assess which occupational tasks workers want AI agents to automate or augment, and how those desires align with the current technological capabilities. Our framework features an audio-enhanced mini-interview to capture nuanced worker desires and introduces the Human
adoption · Erik Brynjolfsson · Diyi Yang
research · JAMA Network Open ·
Efficiency and Quality of Generative AI–Assisted Radiograph Reporting
Radiologists using AI-assisted draft reports completed documentation 22% faster with no loss of accuracy or quality in a prospective clinical study.
Importance: Diagnostic imaging interpretation involves distilling multimodal clinical information into text form, a task well-suited to augmentation by generative artificial intelligence (AI). However, to our knowledge, impacts of AI-based draft radiological reporting remain unstudied in clinical settings. Objective: To prospectively evaluate the association of radiologist use of a workflow-integrated generative model capable of providing draft radiological reports for plain radiographs across a tertiary health care system with documentation efficiency, the clinical accuracy and textual qualit
productivity
practice · Paul Ford (Aboard) ·
The Extremely Human Last Mile
AI automation hits limits when human judgment and relationship-building matter; the last mile of work remains stubbornly human.
Two contradictory AI things are rattling around in my head, and I’m trying to put them together and make sense of where we are. The first is Builder.ai declaring bankruptcy after failing to meet revenue targets. The second is a set of statements by Anthropic founder Dario Amodei predicting mass AI-driven unemployment. Builder.ai is (was?) a little hard to describe, but their basic pitch was something like: Come to our website with your app idea and answer a bunch of AI-prompted questions, and put everything you want your app to do into a shopping cart. We’ll use our AI bot Natasha to do a lot
judgment · Paul Ford
practice · Artificial Ignorance ·
How To Fall Behind in The Age of AI
A framework for when and how to deliberately opt out of AI adoption without losing competitiveness.
A very serious guide to artificial ignorance. A very serious guide to artificial ignorance. Look, I get it. Every day, dozens of headlines proclaim that AI is the next big thing, or is coming for our jobs, or might end human civilization as we know it. It's exhausting. And honestly, who has…
adoption · Charlie Guo
research · arXiv ·
My Advisor, Her AI and Me: Evidence from a Field Experiment on Human-AI Collaboration and Investment Decisions
Field experiment: customers follow human-AI collaborative investment advice more than pure AI advice, improving welfare, driven by human credibility rather than quality gains.
Amid ongoing policy and managerial debates on keeping humans in the loop of AI decision-making, we investigate whether human involvement in AI-based service production benefits downstream consumers. Partnering with a large savings bank in Europe, we produced pure AI and human-AI collaborative investment advice, passed it to customers, and examined their advice-taking in a field experiment. On the production side, contrary to concerns that humans might inefficiently override AI output, we find that giving a human banker the final say over AI-generated financial advice does not compromise its qu
judgment
practice · Eugene Yan ·
AI Engineer 2025 - Improving RecSys & Search with LLM techniques
Semantic IDs and data augmentation techniques unify recommendation and search systems using LLMs.
Recsys & search are converging with LLMs via semantic IDs, data augmentation, and unified foundation models.
ways of working · Eugene Yan
practice · How I AI ·
The exact AI playbook (using MCPs, custom GPTs, Granola) that saved ElevenLabs $100k+ and helps them ship daily | Luke Harries (Head of Growth)
Marketing team automated case studies, translations and WhatsApp workflows using MCPs and custom GPTs, cutting costs by $140k annually.
Luke Harries , Head of Growth at ElevenLabs, the leading AI voice technology company, shares how he’s automating marketing workflows with AI—from case studies to translations to WhatsApp integrations—saving his company over $140,000 while making everything a launch. What you’ll learn: 1. How to create polished case studies in minutes using AI transcription and a custom GPT 2. How ElevenLabs built a custom AI translation system that saved them $140,000 annually and eliminated agency headaches 3. How to use Model Context Protocols (MCPs) to connect AI assistants to your WhatsApp messages 4. The
productivity · Claire Vo
research · NBER ·
Expertise
Task automation changes expertise requirements for remaining work, with effects varying by occupation and task complementarity.
When job tasks are automated, does this augment or diminish the value of labor in the tasks that remain? We argue the answer depends on whether removing tasks raises or reduces the expertise required for remaining non-automated tasks. Since the same task may be relatively expert in one occupation (David Autor , Neil Thompson)
jobs skills · David Autor
research · Management Science ·
Incentives, Framing, and Reliance on Algorithmic Advice: An Experimental Study
Performance incentives and tournament pay increase manager reliance on AI advice; framing algorithms as incorporating human expertise also boosts utilization.
Managerial decision makers are increasingly supported by advanced data analytics and other artificial intelligence (AI)-based technologies, but they are often found to be hesitant to follow the algorithmic advice. We examine how compensation contract design and framing of an AI algorithm influence decision makers’ reliance on algorithmic advice and performance in a price estimation task. Based on a large sample of almost 1,500 participants, we find that compared with a fixed compensation, both compensation contracts based on individual performance and tournament contracts lead to an increase i
judgment
practice · Paul Ford (Aboard) ·
Will AI Eat Tech, or Will Tech Eat AI?
AI companies pursuing platform strategy (OS, ecosystem, developer tools) mirrors how Windows, Google, Apple dominated; MCP offers alternative distributed model.
There are many big questions about AI, but as a tech industry person, the one that is most interesting to me is “will AI eat the tech industry, or will the tech industry eat AI?” What would it mean if AI ate the tech industry? OpenAI or Anthropic wouldn’t just be providers of chatbots or video generators—they’d become operating systems and ecosystems, the primary interface for accessing all information. Microsoft did that with Windows, Google with search (and advertising, and Google Drive, and Android), and Apple with the iPhone. Some people really do spend hours a day on ChatGPT already, and
management org · Paul Ford
practice · The Pragmatic Engineer ·
From Software Engineer to AI Engineer – with Janvi Kalra
A software engineer's self-taught path to AI engineering roles, with concrete advice for career transitions.
From Coda to OpenAI: How Janvi Kalra taught herself AI engineering, impressed tech leaders, and built a career at the forefront of AI—plus actionable advice for landing your own role. From Coda to OpenAI: How Janvi Kalra taught herself AI engineering, impressed top tech leaders, and built a career at the forefront of responsible AI—plus actionable advice for landing your own AI role. Stream the Latest Episode
jobs skills · Gergely Orosz
practice · Hacker News ·
LLM codegen go brrr – Parallelization with Git worktrees and tmux
Using Git worktrees and tmux to run multiple LLM code generation tasks in parallel, speeding up development.
156 points on Hacker News. Discussion: https://news.ycombinator.com/item?id=44116872
ways of working
research · Information Systems Research ·
Forced to Change? Media Exposure of Labor Issues and Firm Artificial Intelligence Investment
Media coverage of labor issues increases AI investment in firms, particularly targeting high-skilled roles, suggesting AI adoption driven by reputational pressure rather than productivity.
Firms are increasingly interested in investing in artificial intelligence (AI), but what drives this trend? Our research reveals that media coverage of labor issues plays a significant role. When firms face public scrutiny through media exposure of labor issues, the reputational pressure pushes them to act. AI emerges as a strategic response, offering powerful capabilities to automate and augment human tasks. Analyzing data from U.S. public firms, we found that labor issue-related media coverage significantly increases AI investments, particularly among firms with the motivation and resources
adoption
practice · How I AI ·
A 3-step AI coding workflow for solo founders | Ryan Carson (5x founder)
Solo founder shares a three-file system and structured Cursor workflow that replaces traditional engineering teams for product building.
Ryan Carson is a five-time founder who has spent the past 20 years building, scaling, and selling startups. In this episode, he shares his playbook for using AI to build products, turning “vibe coding” into a structured and scalable approach that can replace full engineering teams. What you’ll learn: 1. A simple three-file system that transforms chaotic AI coding into a structured, reliable process 2. How to create AI-generated PRDs and task lists that actually work 3. A step-by-step workflow using Cursor to build features systematically 4. Why slowing down to provide proper context is the sec
ways of working · Claire Vo
research · Management Science ·
Humans’ Use of AI Assistance: The Effect of Loss Aversion on Willingness to Delegate Decisions
Loss framing removes algorithm aversion: workers delegate equally to AI and humans when losses frame decisions, but prefer humans under gain framing.
As artificial intelligence (AI) tools have become pervasive in business applications, so too have interactions between AI and humans in business processes and decision-making. A growing area of research has focused on human decision and task delegation to AI assistants. Simultaneously, extensive research on algorithm aversion—humans’ resistance to algorithm-based decision tools—has demonstrated potential barriers and issues with AI applications in business. In this paper, we test a simple strategy for mitigating algorithm aversion in the context of AI task delegation. We show that simply chang
judgment
practice · Simon Wardley ·
Rewilding software engineering
Software engineers and business leaders need shared language about system design to work well with AI tools.
Chapter 5: Different folks for different strokes By Tudor Girba and Simon Wardley An apology This book is about software engineering, but we start this chapter by talking to the business. We know that you’ve been wanting to fire those pedantic, inflexible, costly and annoying software engineers for ages. AI appears to be so damn helpful and so much more of a partner that you think that now is the right time to part ways. But give them a second chance, you will need them, software engineers can change and so can you. We understand that in business you don’t care about the decisions made in code
management org · Simon Wardley
research · Management Science ·
Improving Human Sequential Decision Making with Reinforcement Learning
Reinforcement learning algorithm extracts interpretable decision tips from worker data; randomized experiments show tips improve sequential decision-making performance.
Workers spend a significant amount of time learning how to make good decisions. Evaluating the efficacy of a given decision, however, can be complicated—for example, decision outcomes are often long-term and relate to the original decision in complex ways. Surprisingly, even though learning good decision-making strategies is difficult, the strategies can often be expressed in simple and concise forms. Focusing on sequential decision making, we design a novel machine learning algorithm that is capable of extracting “best practices” from trace data and conveying its insights to humans in the for
judgment
practice · The Pragmatic Engineer ·
The AI Engineering Stack
AI engineering as a distinct discipline with three layers: interfaces, orchestration, and optimization, different from ML and fullstack engineering.
Three layers of the AI stack, how AI engineering is different from ML engineering and fullstack engineering, and more. An excerpt from the book AI Engineering by Chip Huyen Three layers of the AI stack, how AI engineering is different from ML engineering and fullstack engineering, and more. An excerpt from the book AI Engineering by Chip Huyen “AI Engineering” is a term that I didn’t hear about two years ago, but today, AI engineers are in high demand. Companies like Meta, Google, and Amazon, offer higher base salaries for these roles than “regular” software engineers get, while AI startups an
management org · Gergely Orosz
practice · Understanding AI (Timothy B. Lee) ·
I got fooled by AI-for-science hype—here's what it taught me
A physicist applied AI to plasma research and found it failed to deliver on hype; detailed account of why and what actually worked.
I used AI in my plasma physics research and it didn’t go the way I expected. I used AI in my plasma physics research and it didn’t go the way I expected. I’m excited to publish this guest post by Nick McGreivy, a physicist who last year earned a PhD from Princeton. Nick used to be optimistic that AI could accelerate physics research. But when he tried to apply AI techniques to real physics problems the results were disappointing.
judgment
practice · Latent Space ·
ChatGPT Codex: The Missing Manual
WHAM framework for using ChatGPT Codex, an autonomous software engineer agent, on existing codebases.
ChatGPT Codex is here - the first cloud hosted Autonomous Software Engineer (A-SWE) from OpenAI. Josh Ma and Alexander Embiricos tell us how to WHAM every codebase like a power user. ChatGPT Codex is here - the first cloud hosted Autonomous Software Engineer (A-SWE) from OpenAI. Josh Ma and Alexander Embiricos tell us how to WHAM every codebase like a power user. The World’s Fair is 2 weeks away, and early bird tix have sold out! We’re happy to share that Fouad Matin of the OpenAI Codex team, Michael Truell of Cursor, Kevin Hou of Windsurf, Boris Cherny of Cl…
ways of working · Shawn Wang (swyx)