Developer Builds Quiz Pitting Humans Against AI on Expert-Level Academic Questions

·1 views

A developer has created 'Humans vs. HLE', a web-based quiz that challenges users to compete against frontier AI models on Humanity's Last Exam, a benchmark of expert-level academic questions designed to stump even advanced AI systems. Players answer multiple-choice questions drawn from the HLE dataset, and are eliminated after four wrong answers. Scores are displayed alongside published AI model accuracy rates, and a persistent leaderboard tracks how human participants perform over time. The app runs on Cloudflare Workers and uses AES-GCM encryption and HMAC-signed result tokens to keep answer keys server-side, preventing users from cheating via browser tools. The project was built using Claude and tracked with a tool called Entire, with 140 curated questions bundled directly into the application.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

A Three-Question Decision Tree to Pick the Only 2 Refactoring Tools You Need

A practical framework published on DEV Community argues that developers can cut through refactoring tool overload by answering just three questions: the scope of change, how type-heavy the codebase is, and who will run the tool. Based on those answers, the guide narrows the choice to two tools across three common scenarios — solo TypeScript developers are pointed to VS Code and ts-morph, small teams to WebStorm and jscodeshift, and multi-repo organisations to ast-grep and Sourcegraph Batch Changes. The article advises skipping tools like Babel, Comby, OpenRewrite, and ReSharper unless a developer is already working within their specific ecosystems. The author frames tool sprawl as a symptom of a deeper organisational problem: teams debating ten options have usually not defined what kinds of changes are worth automating. This guidance comes as the refactoring-tools market is projected to grow from $1.74 billion to $5.82 billion by 2033, a trend the article warns will produce more tools rather than clearer choices.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Google's Gemma 4-26B Lets Indian Firms Run Powerful AI Locally on Budget Laptops

Google's Gemma 4-26B, a fourth-generation open-weight AI model with 26 billion parameters, can run on a standard consumer laptop with 16GB RAM and around 14GB of disk space, requiring no cloud subscription or GPU. Indian startups and SMEs can deploy it on laptops costing between ₹35,000 and ₹50,000, avoiding recurring API fees that can reach ₹40,000 per month for mid-sized firms. Running the model locally addresses key concerns around data privacy, regulatory compliance under frameworks like SEBI and RBI guidelines, and operational reliability during internet outages. Practical use cases span financial analysis, options trading research, contract review, customer support bots, and Python-based data pipeline development. The model runs at roughly 8–12 tokens per second on a modern CPU using 4-bit quantization, making it accessible without specialised hardware.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How One Recruiter's Bias Quietly Made an AI Hiring System Less Effective

A company's AI interview platform showed consistently improving metrics over six months, with pass rates, completion rates, and candidate satisfaction all trending upward. However, a senior engineering manager noticed that despite more candidates clearing interviews, the actual on-the-job performance of new hires showed no improvement over previous cohorts. Investigations ruled out AI tool usage by candidates, question leakage, and model drift as causes. The team eventually traced the issue to a feedback loop in the system's design, where human reviewer approvals were used monthly to recalibrate AI scoring thresholds. One recruiter who consistently approved nearly 90% of AI recommendations — far above the team average of 55–65% — had inadvertently skewed the model toward her preferences at scale, gradually lowering the effective bar for all candidates.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Google Translate Silently Breaks React Apps and What Developers Can Do

A subtle but serious bug can cause React applications to crash with a blank screen for users whose browsers auto-translate web pages. The issue arises because Google Translate rewrites the DOM by splitting and wrapping text nodes, while React simultaneously relies on its own internal model of the DOM structure without re-checking it. When React attempts its next update, it finds the DOM no longer matches its expectations and throws an unrecoverable error, often leaving users with a white screen. The conflict is not a flaw in either React or Google Translate individually, but stems from both systems trying to control the same DOM elements at the same time. International users and anyone with browser auto-translate enabled are most likely to encounter this issue, making it difficult to reproduce in standard development environments.

0 comments Read more at DEV Community