What AI Users Want: Anthropic Releases 81,000-User Survey
March 19, 2026
D.A.D. today covers 12 stories from 7 sources. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
D.A.D. Joke of the Day: My AI gave me three different answers to the same question. Finally, a coworker who's just as confused as I am but way more confident about it.
What's New
AI developments from the last 24 hours
OpenAI Declares 'Code Red,' Tells Staff to Drop Side Projects Before IPO
OpenAI leadership told staff to drop 'side quests' and focus on core priorities, with COO Simo declaring 'code red' urgency and calling Anthropic a 'wake-up call,' according to a Wall Street Journal transcript review. The article, published earlier this week, is generating intense discussion on the chatboards. The internal push comes amid what analysts see as a narrow IPO window: The Economist notes that if OpenAI, Anthropic, and xAI each offered 15% of shares, the combined sum would roughly equal every dollar raised across all American IPOs in the past decade. Reuters reports OpenAI is in talks with TPG, Bain Capital, and others for a $10 billion enterprise distribution joint venture.
Why it matters: The three leading AI labs may be racing to go public before market appetite—or sovereign wealth fund backing—runs thin, which could reshape how aggressively each prioritizes revenue-generating products over research.
Discuss on Hacker News · Source: om.co
Sci-Fi Story Imagines Future Where Domain Expertise Beats Coding Skills
A science fiction short story imagines a near-future where AI-generated software has inverted the value hierarchy in tech work. In the story, traditional programming skills become obsolete as software is 'regenerated' from plain-language specifications rather than repaired. The protagonist, a former agricultural equipment technician, finds his domain expertise in farming more valuable than coding ability—software mechanics now need to understand what systems do, not how they're built.
Why it matters: The story crystallizes a question already circulating in boardrooms: as AI handles more code generation, does competitive advantage shift from technical talent to deep industry knowledge?
Discuss on Hacker News · Source: nearzero.software
Rob Pike's 1989 Programming Rules Resurface as AI Prompting Guide
A 1989 document resurfaced this week outlining computer scientist Rob Pike's five rules of programming—principles that emphasize measuring before optimizing, keeping algorithms simple until data proves otherwise, and prioritizing data structure design over clever algorithms. Pike's core argument: get your data structures right and the code practically writes itself. The rules predate modern AI coding assistants but speak directly to how developers should think about prompting them. Community discussion noted the 'premature optimization' maxim may originate with Donald Knuth, not Tony Hoare as often cited.
Why it matters: As AI tools generate more code, these 35-year-old principles offer a useful filter: focus your prompts on data organization rather than algorithmic cleverness—the same wisdom that guided human programmers applies when directing AI ones.
Discuss on Hacker News · Source: cs.unc.edu
New Workplace Frustration: Colleagues Who Blindly Trust AI Outputs Without Critical Thinking
A Hacker News thread surfaced a growing workplace friction: colleagues who treat LLM outputs as authoritative sources rather than starting points for research. The discussion revealed no consensus solution, but commenters offered practical tactics—like demonstrating how easily an LLM will reverse its position when challenged, which can quickly illustrate the technology's limitations. Others pushed back, noting that blind trust in 'reputable' search results or social media isn't meaningfully different. The thread suggests AI literacy gaps are becoming a real collaboration challenge.
Why it matters: As AI tools spread beyond early adopters, teams will increasingly need shared norms around when to verify LLM outputs—a management problem, not just a technical one.
Discuss on Hacker News · Source: news.ycombinator.com
What's in the Lab
New announcements from major AI labs
Anthropic's 81,000-User Survey Reveals Top Desire: Handle Mundane Work
Anthropic surveyed 80,508 Claude users across 159 countries in 70 languages—what the company calls the largest qualitative study of AI users ever conducted. The top desire: professional excellence, with 19% wanting AI to handle mundane tasks so they can focus on meaningful work. Another 9% saw AI as a potential entrepreneurial partner. The interviews, conducted by an AI system over one week in December, aimed to capture what real users actually want from these tools rather than theoretical debates about AI's future.
Why it matters: This is Anthropic positioning itself as the user-focused AI lab—but it's also rare large-scale data on what people actually do with chatbots, useful context as AI companies compete to define what these tools should become.
Measuring Progress Toward AGI: Google Releases Cognitive Framework
Google DeepMind published a paper proposing a cognitive science-based framework for measuring progress toward AGI. Rather than testing narrow task performance, the approach would evaluate AI systems across 10 cognitive abilities—including reasoning, memory, learning, and social cognition—benchmarked against human performance. The paper outlines a three-stage evaluation protocol but doesn't include results for current AI systems. DeepMind partnered with Kaggle to launch a $200,000 hackathon (March 17–April 16) inviting researchers to build evaluations based on the framework.
Why it matters: How you measure AGI shapes what gets built—DeepMind is staking out a position that cognitive science, not just benchmark leaderboards, should define the goalposts, which could influence where the field directs research investment.
OpenAI Tailors Teen Safety Rules for Japan's Regulatory Environment
OpenAI Japan announced a "Japan Teen Safety Blueprint," a framework of safety measures designed specifically for teenage users in the country. The plan includes stronger age verification, parental controls, and well-being safeguards for teens using generative AI tools. OpenAI did not release specific technical details about implementation. The Japan-specific approach suggests OpenAI is tailoring its safety policies to different regulatory environments rather than applying uniform global standards.
Why it matters: This signals OpenAI is pursuing region-specific safety frameworks—potentially a template for how AI companies navigate varying international expectations around youth protection.
Facebook Reels Now Highlights Videos Your Friends Liked
Meta published a technical breakdown of Friend Bubbles, a Facebook Reels feature that surfaces videos your friends have liked or reacted to. The system uses machine learning to estimate relationship strength between users and rank friend-interacted content in your feed. Meta says the feature creates a feedback loop: the more you engage with friend-recommended content, the better the system gets at predicting what you'll want to see. The company frames this as blending social signals with interest-based recommendations.
Why it matters: This is Meta showing its algorithmic hand—useful context if you're thinking about how social platforms are evolving beyond pure engagement metrics toward relationship-weighted discovery, which affects both content strategy and ad targeting dynamics.
What's in Academe
New papers on AI and its effects from researchers
Researchers Release Tool to Test Whether AI Can Invent New Algorithms
Researchers released DiscoGen, an open-source tool that automatically generates millions of algorithm discovery tasks—challenges where AI systems must invent new optimizers, loss functions, or other ML components. The project aims to solve a growing problem in AI research: existing benchmarks for testing whether AI can discover novel algorithms are running out of fresh problems and may be contaminated by training data. DiscoGen can create configurable tasks across difficulty levels and ML subfields, accompanied by DiscoBench for standardized evaluation.
Why it matters: This is research infrastructure for studying whether AI can automate AI research itself—a capability that would accelerate the entire field if achieved, making it a closely watched area among labs and investors.
Smaller AI Models Match Larger Rivals on Code Search, Study Finds
Researchers developed CodeScout, a reinforcement learning method that trains AI coding agents to locate relevant code in large repositories using only basic Unix terminal commands—no specialized analysis tools required. On standard benchmarks, CodeScout models matched or beat models 2-18x their size, and in some cases approached Claude Sonnet's performance despite Sonnet having access to more sophisticated tooling. The approach suggests smaller, cheaper models can handle complex code search tasks when trained with the right techniques.
Why it matters: For teams building AI-assisted development workflows, this signals that effective code navigation may not require the largest (most expensive) models—a potential cost advantage as coding assistants become standard infrastructure.
Bug-Fixing AI Learns From Its Own Failed Repairs
Researchers introduced FailureMem, a framework for automated software repair that learns from its own mistakes. The system combines visual analysis (reading screenshots and UI elements) with a "Failure Memory Bank" that stores unsuccessful repair attempts and converts them into guidance for future fixes. On SWE-bench Multimodal, a benchmark for GUI-based bug fixing, FailureMem improved resolution rates by 3.7% over the previous best approach. The core innovation: rather than starting fresh each time, the system builds institutional memory of what doesn't work.
Why it matters: For teams evaluating AI coding assistants, this signals a shift toward tools that improve through accumulated experience—potentially reducing the repetitive trial-and-error that frustrates current automated debugging.
Compression Method Cuts AI Memory Use While Preserving Accuracy
Researchers introduced CARE, a method for converting standard AI attention mechanisms into a more memory-efficient architecture called multi-head latent attention (MLA). The technique compresses how models store context during inference—the "KV-cache" that grows with conversation length—without the accuracy losses typical of compression approaches. Tested on Qwen and Llama models, CARE dramatically outperformed simpler compression baselines: up to 215x lower perplexity degradation and 1.70x better accuracy retention at equivalent memory budgets. With brief additional training, models fully recovered their original performance.
Why it matters: This is infrastructure research, but the practical goal is running powerful AI models with less memory—which could eventually mean faster responses, longer context windows, or cheaper deployment for the tools you use.
Game Theory Approach Aims to Improve Multilingual AI Training
Researchers developed ShapleyLaw, a method that treats multilingual AI training as a cooperative game where each language is a "player" contributing to the model's overall capabilities. The approach borrows from game theory to measure how learning one language helps (or hurts) performance in others—cross-lingual transfer effects that current training methods don't capture well. The team claims ShapleyLaw outperforms existing approaches for predicting model performance and optimizing the mix of languages in training data, though specific benchmark numbers weren't released.
Why it matters: For companies building or fine-tuning multilingual AI products, better language mixture optimization could mean stronger performance in underrepresented languages without sacrificing quality in high-resource ones like English.
What's On The Pod
Some new podcast episodes
AI in Business — Why Financial AI Can't Scale Without Unified Governance with James Dean of Google and Mark Crean of Securiti
AI in Business — How Digital Workers Are Changing Industrial Performance - with Somya Kapoor of IFS Loops