AI Models Show Anchoring Bias; Fast Responders Most Affected, Thinkers Less So

·7 views

A Kaggle benchmark tested 13 AI models for anchoring bias — a cognitive tendency where irrelevant numbers skew estimates — inspired by Kahneman and Tversky's 1974 wheel-of-fortune experiment. Each model answered 40 estimation questions paired with artificially low or high anchor numbers across five prompt formats. Fast-responding models showed the strongest anchoring, with GPT-4.5 nano scoring 0.52, close to the human benchmark of 0.49. Models that reasoned before answering were largely resistant, scoring between 0.02 and 0.07, though Qwen3 Thinking was a notable exception, anchoring as much as its instant counterpart despite using roughly 2,000 tokens of reasoning. The study highlights a practical concern: AI models embedded in real applications may silently skew outputs like time or cost estimates when unrelated numbers appear nearby in prompts.

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)

PostgreSQL Advisory Locks Offer a Built-In Fix for Distributed Race Conditions

Race conditions in clustered applications occur when multiple pods process the same request simultaneously, leading to errors like duplicate payouts or double-credited rewards. Common fixes such as Java's synchronized keyword fail in multi-pod deployments, while row-level database locks require a pre-existing row and can cause storage overhead. Introducing a separate Redis cluster adds operational complexity and potential split-brain failure risks. PostgreSQL Advisory Locks provide an alternative by letting applications define integer-keyed locks held directly in the database server's memory, with no row or table attachment. Since most teams already rely on PostgreSQL as their primary datastore, this approach eliminates the need for additional infrastructure while effectively serializing critical operations across pods.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CrawlForge MCP fixes residential proxy routing across six releases in eleven days

Developer tool CrawlForge MCP released six updates between September 16 and 27, 2026, after discovering that residential proxy routing was misconfigured and failing to bypass Cloudflare bot detection. An internal benchmark against version 6.6.2 exposed the flaw, triggering a rapid series of fixes. The updates introduced a multi-stage escalation ladder that progresses from a plain fetch to a Chrome TLS handshake to a stealth browser, stopping at whichever step successfully retrieves the page. A clearance jar was added to cache solved Cloudflare challenges, preventing repeated solving on subsequent requests. No tools were renamed, no prices changed for most tiers, and the team stated it would not implement CAPTCHA solving or token forgery, with robots.txt checked at every escalation stage.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How developers can convert e-commerce reviews into actionable defect tables for free

A software developer has shared a technical walkthrough for transforming unstructured product reviews from platforms like Amazon and TikTok Shop into structured defect tables without paid APIs or scraping tools. The method addresses common pitfalls such as lazy-loading review widgets, pagination loops that return duplicate results, and CSV encoding issues that corrupt non-Latin text. Instead of relying on star ratings, the approach clusters one-to-three star reviews into six complaint categories — including packaging damage, wrong sizing, and counterfeit concerns — using simple keyword matching rather than machine learning models. The resulting defect table allows merchants to benchmark their products against competitors across specific complaint types. The developer argues this turns vague quality feedback into concrete, prioritised product improvements.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Dev builds bilingual browser cooking game using Astro and Cloudflare edge caching

A developer has launched Cá Viên Chiên, a mobile-friendly browser game about Vietnamese fried fish ball street carts, available in both English and Vietnamese. Built with Astro, the project uses a single vanilla JavaScript island of around 26 KB with no framework runtime, keeping the site lightweight and SEO-friendly via static HTML pages. Game progress is saved to localStorage and transferred between devices using a copy-paste save code, requiring no backend or user accounts. The main technical challenge was caching HTML at Cloudflare's edge, as pages initially showed dynamic responses with high origin latency despite cache rules being in place. The fix involved using the CDN-Cache-Control header to set a 24-hour edge TTL independently of browser cache headers, with cache purges triggered on each deployment to ensure fresh content.

0 comments Read more at DEV Community