Working Reference — AI Workflow Guides

Insights2026.10.26

AI turned out 722 math manuscripts at once. Now the scarce resource is mathematicians to read them

OpenAI released 722 AI-generated math manuscripts at once. Of the 719 left after withdrawals, only 42% have proofs confirmed by the Lean program. This piece follows the record of 3 withdrawals over a calculation error the next day, 14 separate proof revisions, and how mathematicians reacted. Together they show math's bottleneck moving from producing proofs to finding people who can read and check them.

Insights2026.10.26

Handing classification to AI now costs almost nothing. Almost no one checked whether it got the answers right

TypeSafe's decision model Jev matched Claude Sonnet 5's score of 67.8% at $0.0004 per decision, but the answers it was graded against came from two AI models that aren't on the leaderboard. When Red Hat retested it on public data with labels already attached, Jev ranked first or fourth depending on the task, and the results did not show it was faster or cheaper. Now that decisions are nearly free, the first thing to check isn't the model's price. It's whether you have an answer key that can show the decisions were right for your work.

Insights2026.10.26

AI agent hacks hit several Korean financial firms, but the breached systems were loan-broker and staff tools, not banking apps

US security firm CrowdStrike analyzed AI work logs left on an attacker's server. The attacker ran a free open-source tool on other companies' commercial AI models and tried to break into many financial firms at once. What got breached were not banking apps but work systems for loan brokers and staff with weak identity checks. Korea's Financial Services Commission told firms to review systems by whether they are open to the outside, not by whether customers use them.

Insights2026.10.23

Big employers are asking less for people who use AI and more for people who build with it

Workday reviewed 11 months of job requisitions from 554 large employers. Requirements for using AI tools fell 25% after peaking in January 2026. Requirements for building AI into work rose 51% from September 2025. The data can't tell us why demand for using AI fell.

Insights2026.10.19

Banks want lower legal fees because of AI. Law firms' billable hours went up instead

Big banks say AI has sped up legal work and want lower fees. Yet law firm billed hours rose 4.2% in the first half of 2026, more than double the usual pace. Firms are counting AI learning time toward internal targets, and clients like AT&T are moving work in-house, changing what the negotiation is about.

Insights2026.10.19

Robots can do a lot of work. They're cheaper than people for just 0.3% of it

Anthropic used Claude to go through 19,000 tasks across about 900 US occupations, one by one. Measured by total US work hours, robots can do 34% of work, mostly in dedicated spaces like factories. By the same measure, they are cheaper than people for just 0.3%. Even if robot costs fall 3% a year, it would take about 40 years for robots to be cheaper across 10% of total work hours.

Insights2026.09.27

Coding agents got so good that Shopify now builds every app twice: once for iOS, once for Android

Shopify has built its mobile apps from a single codebase since 2020, and now it is rewriting them separately for iOS and Android. Coding agents cut the cost of writing everything twice, so Shopify's focus has moved from the amount of code to the time it takes to check what the AI built.

Insights2026.09.27

AI agents finished whole research papers on their own, but judgment, not experimentation, is where they stalled

Researchers from Princeton, the UK AI Security Institute, and other institutions gave frontier agents the research questions behind two unpublished NeurIPS submissions, then had the original authors grade the results. The agents handled everything from the literature review to typesetting without human help, yet both papers were rejected. Here are the five failures the paper traced in the run logs, along with the limits the authors acknowledged.

AI-generated

Insights2026.09.27

Crediting a career coach for a cover letter didn't hurt trust, but crediting AI did

In 13 experiments preregistered before any data was collected, University of Arizona researchers Schilke and Reimann found that a cover letter credited to a career coach earned the same trust as one with no note at all, while only the one credited to AI lost trust. The penalty held across six ways of wording the disclosure and when disclosure was required by rule, and in the final experiment an advisor exposed by a third party scored lower still than one who disclosed it himself.

AI-generated

Earlier posts

Weekly · Free

Hi, I'm Rush from Working Reference.

The week in AI: the BEST
just 3, straight to your inbox

Work Smarter with AI

You're in See you in your inbox with the next issue.

Insights

AI turned out 722 math manuscripts at once. Now the scarce resource is mathematicians to read them

Handing classification to AI now costs almost nothing. Almost no one checked whether it got the answers right

AI agent hacks hit several Korean financial firms, but the breached systems were loan-broker and staff tools, not banking apps

Big employers are asking less for people who use AI and more for people who build with it

Banks want lower legal fees because of AI. Law firms' billable hours went up instead

Robots can do a lot of work. They're cheaper than people for just 0.3% of it

Coding agents got so good that Shopify now builds every app twice: once for iOS, once for Android

AI agents finished whole research papers on their own, but judgment, not experimentation, is where they stalled

Crediting a career coach for a cover letter didn't hurt trust, but crediting AI did

Earlier posts

Your Code Just Became 14 Figma Frames — AI Agents Can Now Write to the Canvas

Ranking #1 No Longer Means Clicks — 2026 AI SEO Statistics That Rewrote the Rules

The AI Already Chose Your Competitor — What Agent Search Optimization Actually Is

1 million agents in, Notion hit a wall — so it just turned the workspace into an AI agent hub

A dictation app is worth $2B now — how Wispr Flow killed the 'Whisper wrapper' label

Asked AI for a website, got 5 pages with animations — why Webflow took a different road from v0 and Lovable

$100M in revenue from comment threads — inside Nectar's 'AI Social OS' that Menlo just backed with $30M

The $5M Seed Is Dead — AI Startups Now Start at $40M, and the Real Bill Comes Later

20% of Companies Took 74% of AI's Value — PwC's Study Reveals the One Real Divide

$2T Just Vanished from SaaS — Here's a16z's Counter to the SaaSpocalypse

5 Models for One Image — a16z's 2026 Report Marks the Rise of AI Media Orchestration

They said they'd hire 1,111 — then laid off 1,100. The real rules of AI-era org redesign at Cloudflare

The first GenAI unicorn just gave up on models — what India's Krutrim pivot really tells the rest of us

Y Combinator's newest unicorn is an insurance company — what it means that Corgi raised again 4 months later

The open-source villain just got $45B — what it means that China's state fund is betting directly on DeepSeek

AI isn't replacing consulting — it's buying it. Anthropic and OpenAI's same-day move

Enterprise AI is now a capital game — Sierra's $950M just drew the new line

Maigret: The OSINT Tool That Builds a Dossier From One Username — Now With AI Mode

Academic Papers Are Still PDFs — Not Because of AI, But That's Why Science Is Falling Behind

Spot Read a Handwritten To-Do List and Got to Work — Is This the End of Robot Coding?

$700B in AI Capex Isn't a Bubble Question — It's About Which Assumptions Break First

Half of All Call Centers Are Being Disrupted at 7 Cents a Minute — The 4 Real Bottlenecks of Voice AI Adoption

Web Scraping and Company PDFs Finally Share the Same Engine — Firecrawl /parse + Fire-PDF

The People Who Turned Down $100M Signing Bonuses — How Anthropic Actually Won the Talent War

Every Web Scrape Is a Potential Injection: Firecrawl's Lockdown Mode

Chrome 146 Brought AI Skills Down to the OS Layer — The Real Threat to Perplexity and Comet

Uber Burned Through Its Annual AI Budget in Four Months — The Claude Code Governance Crisis

Claude Code Skills: One Markdown File, 50 Listings, Zero Code

Codex CLI Got a /goal Command — A 'Dumb Bash One-Liner' Just Became an Official OpenAI Feature

WhatsMyName.io — One Username, 1,500 Sites, Instant OSINT

The week in AI: the BESTjust 3, straight to your inbox

The week in AI: the BEST
just 3, straight to your inbox