Pac-Bench tests AI models ability to build Pac-Man in one HTML prompt

·1 views

A new benchmark called Pac-Bench evaluates how well AI language models can generate a functional Pac-Man game from a single prompt. Each model receives exactly one instruction — 'Create a Pac-Man game in a single HTML page' — with no follow-up prompts or corrections allowed. The project uses the Harness framework to run and compare model outputs consistently. It was shared on Hacker News, where it drew early community attention. The benchmark aims to measure one-shot coding capability across different AI models in a fun, visual way.

Read the full story at Hacker News

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

Discussion (0)

ClientN Shows How to Add Passkey Login Without Collecting User Emails

A developer building ClientN has published an open-source reference integration demonstrating how small websites can implement passkey-based sign-in without storing user email addresses. The system assigns each user a site-specific clientn_id after passkey confirmation, meaning no password or email is ever transmitted to the site's server. The five-step server flow covers session creation, browser binding via HttpOnly cookies, callback signature verification, and single-use session consumption to prevent race conditions. The reference code is available in Node.js and PHP under an MIT licence, with a test suite covering attacks such as cross-site requests, forged callbacks, and expired logins. Sites must register a domain at clientn.com and receive signed callbacks over HTTPS, making localhost testing unsuitable.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Five Common Reasons Telegram Store Bots Fail Under High Traffic

Telegram bots have become a popular low-cost storefront for sellers in MENA, LATAM, and crypto markets, but most struggle when order volumes increase. Key failure points include storing cart data in memory rather than a database, causing data loss on redeployment, and lacking idempotency keys for payment webhooks, which can result in duplicate deliveries or missed payments. Tying product fulfilment directly to the bot process also causes slowdowns, as a single slow upload can block all other users. Additional risks include missing reconciliation jobs that silently drop undelivered orders and ignoring Telegram's rate limits, which can get a bot muted. Developers are advised to treat the bot as a thin interface and offload orders, payments, and fulfilment to resilient, restart-safe systems.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Custom Build or Template? A Scoring Framework to Guide Web Project Decisions

Choosing between custom web development and templates is often decided by opinion rather than structured criteria, but a weighted scoring approach can make the decision more objective. Templates suit content-focused sites with tight budgets and non-technical teams, while custom builds are justified when projects require complex business logic, third-party integrations, or strict performance control. A numerical scoring tool assigns weights to key project signals, with scores below 40 favouring templates, 40–70 suggesting a hybrid CMS-plus-custom-components setup, and above 70 pointing toward a full custom build. A hybrid approach — using a headless CMS for content alongside a custom front end — is increasingly common, letting editors and developers each work in their preferred environment. Regardless of the chosen route, enforcing a Lighthouse CI performance budget ensures neither templates nor custom code silently degrade site speed.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How AI Agents Can Retain, Recall, and Cite Evidence Without Hallucinating

A developer walkthrough published on DEV Community demonstrates how a working AI agent can store and retrieve community messages and product updates as timestamped memories using a unified memory layer. The system is designed to return an INSUFFICIENT_EVIDENCE status when relevant data is lacking, treating uncertainty as a formal outcome rather than masking it with model confidence. All generated FAQ answers are assigned a DRAFT status and require human moderator approval before publication. The agent is also required to distinguish observations from interpretations and explicitly flag weak or conflicting evidence in its output. A key design principle highlighted is that every published answer links back to specific memory IDs, allowing the system to show its reasoning rather than simply assert conclusions.

0 comments Read more at DEV Community