AI Broke Take-Home Exams. Now It's Being Tested as a Fix for Grading In-Class Work
March 16, 2026
D.A.D. today covers 12 stories from 3 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 tried to make small talk and asked about my weekend. I said "nothing much." It wrote me 400 words on why that's actually something.
What's New
AI developments from the last 24 hours
Developers Report AI Coding Assistants Create Mental Fatigue, Not Savings
A Hacker News thread surfaced a counterintuitive complaint: using AI coding assistants is mentally exhausting. Developers report that evaluating AI-generated code, maintaining context across multiple agent sessions, and staying vigilant for subtle errors creates significant cognitive overhead—sometimes more draining than writing code themselves. Community reaction was divided: some say the fatigue reflects the skill required to use these tools well, others blame the hype cycle, and several compare it to broader attention fragmentation from modern technology. The thread suggests a 2-3 concurrent AI session limit as a practical ceiling.
Why it matters: As organizations push AI coding tools into more workflows, this signals a potential productivity ceiling—the oversight burden may offset some efficiency gains, a tradeoff worth monitoring as your teams adopt these tools.
Discuss on Hacker News · Source: tomjohnell.com · In a separate thread, programmers debated whether automated coding kills or restores the joy of the work.
Visual Reference Compares Architecture of Major Open AI Models
A new resource called 'LLM Architecture Gallery' compiles visual architecture diagrams and technical specs for major open-weight models including Llama 3, DeepSeek R1, Gemma 3, Qwen3 variants, and others. The gallery shows decoder types, attention mechanisms, and parameter counts—useful for comparing how different models are built. It covers models from April 2024 through late 2025, including some not yet released. Available as a webpage and physical poster.
Why it matters: This is primarily useful for technical teams evaluating which open models to deploy—a quick visual reference for architecture trade-offs rather than something that changes day-to-day workflows for most professionals.
Discuss on Hacker News · Source: sebastianraschka.com
Classic 2015 Visual Guide Still Explains How Machine Learning Works
A 2015 interactive tutorial resurfaced showing how machine learning classification works using a simple real estate example—distinguishing New York from San Francisco homes. The visual walkthrough demonstrates how algorithms find mathematical boundaries in data (like an elevation threshold of 240 feet or price per square foot of $1,776) to make predictions, using decision trees that split data through if-then logic. It's a decade-old explainer, but remains one of the clearest introductions to how ML models actually work under the hood.
Why it matters: For professionals who use AI tools daily but want to understand the mechanics beneath them, this kind of foundational visual explanation helps demystify what's happening when models make predictions—useful context as you evaluate AI outputs and vendors.
Discuss on Hacker News · Source: r2d3.us
Simon Willison Proposes 'Agentic Engineering' for AI-Assisted Software Building
Developer and blogger Simon Willison published a guide proposing 'agentic engineering' as the term for building software with AI coding agents—tools like Claude Code, OpenAI Codex, and Gemini CLI that can both write and execute code. The key distinction from earlier AI coding assistants: these agents run their own code, check results, and iterate until something works. Humans still decide what to build and make design tradeoffs. Community reaction on Hacker News was skeptical, with commenters questioning whether this is genuinely new or rebranding.
Why it matters: As AI coding tools proliferate, this framing—humans as directors, agents as implementers—offers a mental model for teams figuring out how to integrate these capabilities without losing engineering judgment.
Discuss on Hacker News · Source: simonwillison.net
What's Innovative
Clever new use cases for AI
Open-Source Project Aims to Track Wildfires Autonomously From Satellite Data
A new open-source project called Signet aims to provide autonomous wildfire tracking by combining NASA satellite fire detections, GOES-19 thermal imagery, and weather data into a single monitoring system for the continental US. The system claims to continuously triage multiple data sources and perform multimodal reasoning without human initiation at each step. No accuracy metrics or performance benchmarks were provided. Hacker News commenters noted Google has similar real-time wildfire tracking research, and offered detailed UX critiques around filtering and readability.
Why it matters: This is an early-stage project to watch rather than deploy—but autonomous monitoring that synthesizes satellite, weather, and geographic data could eventually matter for insurance, logistics, and emergency planning if it proves accurate.
Discuss on Hacker News · Source: signet.watch
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Canada's Proposed Bill Would Require Telecom Surveillance Infrastructure
Canada introduced Bill C-22, the Lawful Access Act, which would require telecom providers to build surveillance capabilities into their networks and establish new rules for law enforcement access to subscriber data. The bill improves on an earlier version by requiring judicial approval for most subscriber information requests—police can only confirm without a warrant whether a provider serves a particular person. However, critics note the surveillance infrastructure requirements remain largely unchanged from the prior bill and include data retention provisions that concern privacy advocates.
Why it matters: For companies operating in Canada or serving Canadian customers, this signals potential new compliance obligations around data retention and law enforcement access—worth tracking as the bill moves through Parliament.
Discuss on Hacker News · Source: michaelgeist.ca
What's in the Lab
New announcements from major AI labs
Google AI to Screen 50,000 Rural Australians for Heart Disease Risk
Google is partnering with Australian healthcare organizations to deploy its Population Health AI in rural communities where residents are 60% more likely to die from heart disease than those in cities. The $1 million AUD initiative pairs Google's risk-identification tools with SISU Health, which plans to conduct over 50,000 health screenings in remote areas. The program aims to shift from reactive treatment to preventative care by flagging chronic condition risks earlier.
Why it matters: This is a test case for AI-driven population health management in underserved areas—if it works, expect similar partnerships targeting healthcare deserts elsewhere.
What's in Academe
New papers on AI and its effects from researchers
Visual Comparison Method Judges AI Code by What It Renders, Not What It Says
Researchers developed Visual-ERM, a reward model that judges AI-generated code by comparing rendered visual output rather than the code itself. The approach treats vision-to-code tasks—like converting charts, tables, or diagrams into code—as an image comparison problem. In testing, an 8-billion-parameter version outperformed models 30 times larger on a benchmark for spotting fine-grained visual discrepancies. Applied to reinforcement learning, the method improved a base model's chart-to-code accuracy by 8.4 points.
Why it matters: For teams automating document processing or data visualization, this signals that AI may soon more reliably reproduce what you actually see—not just approximate it.
Maternal Health Chatbot Study: No Single AI Model Ensures Medical Safety
Researchers from academic institutions, a health tech company, and a hospital built and tested a phone-based maternal health chatbot for India that handles multilingual queries—including the code-mixed speech common in real conversations. The system achieved 86.7% recall on emergency triage in testing (N=150), meaning it correctly flagged most urgent cases. The team's core finding: no single AI model or evaluation method was sufficient for medical reliability. They used layered defenses and multiple evaluation approaches, including clinician-designed criteria and expert review of 781 real queries.
Why it matters: This offers a practical blueprint for deploying health AI in challenging real-world conditions—relevant for any organization considering AI assistants where accuracy stakes are high and user inputs are messy.
AI-Assisted Grading Cut Time 23% While Matching Human Accuracy
Researchers tested an end-to-end system for grading handwritten math exams using LLMs, deploying it across six quizzes in two undergraduate courses. The hybrid approach—automated scanning, multiple LLM scoring passes, consistency checks, and mandatory human review—cut grading time by roughly 23% while matching or exceeding the consistency of fully manual grading. The system caught its own errors: occasional model mistakes were flagged by built-in checks before reaching students.
Why it matters: For institutions drowning in grading workload, this offers early evidence that AI can meaningfully assist with handwritten assessments—not just multiple choice—without sacrificing accuracy, provided humans stay in the loop.
Synthetic Surgery Videos Help Train AI on Rare Procedures
Researchers developed SAW (Surgical Action World), a video generation model that creates realistic laparoscopic surgery footage from simple inputs: text prompts, reference images, tissue maps, and tool movement paths. Trained on 12,044 surgical clips, SAW achieved dramatically better temporal consistency than previous methods (CD-FVD score of 199 vs. 547). The practical payoff: when researchers used SAW-generated videos to supplement rare surgical scenarios in training data, AI systems improved significantly at recognizing uncommon actions—one category jumped from 0% to 8.3% accuracy, another from 21% to 43%.
Why it matters: Surgical AI development is bottlenecked by scarce, privacy-sensitive training data—synthetic video generation could accelerate tools for surgical training, procedural planning, and real-time assistance without requiring thousands of additional recorded procedures.
Open Framework Offers 45,000 Coding Environments to Train AI Agents
Researchers released OpenSWE, a large-scale open framework for training AI agents to write and fix code. The project includes over 45,000 executable coding environments across 12,800+ repositories—infrastructure the team values at roughly $1.5 million to build. Their trained models hit 62-66% on SWE-bench Verified, a standard test for AI's ability to resolve real GitHub issues. Notably, the software engineering training also improved the models' math reasoning by up to 12 points without hurting other capabilities. All code and environments are open-sourced.
Why it matters: This is developer infrastructure, but the pattern matters: open, reproducible training data for coding agents could accelerate competition among AI coding assistants and give enterprises more options beyond closed commercial tools.