GPT-5 Just Learned to Tell the Truth (Finally)

"My therapist: 'Just be honest about your limitations.' GPT-5: 'Hold my training data.'"

Hello Builders! 👋

OpenAI just dropped something actually useful for once: GPT-5 Thinking that admits when it's bullshitting. Their "confessions" method trains the model to flag when it's guessing, taking shortcuts, or breaking rules.

Here's why this matters: we've all shipped features where the LLM confidently hallucinated its way through edge cases. Now imagine your AI actually saying "hey, I'm not sure about this" instead of making up JSON schemas or inventing API endpoints.

The proof-of-concept surfaced hidden failures that standard evals missed entirely. Still early, but this could be the difference between "AI-powered" being a liability and actually shipping reliable products. Finally, some honesty in the AI stack.

v0 Adds Custom MCP Support - Vercel's AI design tool now lets you plug in your own Model Context Protocol servers with presets or custom configs.

Anthropic Launches AI Interviewer Tool - Week-long pilot to gather perspectives on AI through conversational interviews, because surveys are apparently too 2023.

Designly Generates Full App Designs - AI agent creates complete web and mobile application designs with user flows in minutes, free during launch.

Prompt Injection Checker Blocks Jailbreaks - Low-latency tool prevents 97% of prompt injection attacks after developer's previous project got wrecked by bad actors.

Google Ships Workplace Studio Agents - Enterprise AI agents for workplace automation, Google's latest move in the increasingly crowded agent space.

Supply Chain App in 20 Minutes - Samir Saci shows why Streamlit beats Jupyter notebooks for factory operators who actually need to ship working apps.

k-Means Clustering in Excel - Angela Shi implements actual ML training algorithms in spreadsheets because sometimes understanding beats using fancy libraries blindly.

200-Page LLM Paper Breakdown - Maryam Miradi distills a massive research paper on prompting, training, and alignment into the parts that actually matter for practitioners.

The Architecture Behind Web Search in AI Chatbots - Technical deep dive into chatbot search

Cursor's $60 Plan Gone in 7 Days - Artem burned through his monthly Cursor tokens in a week and compares it to a drug habit.

Multi-AI Assembly Line Workflow - Yorgo chains Claude Opus for planning, then hands off execution prompts to GPT Codex 5.1 for implementation.

ChatGPT Might Be Dulling Brains - MIT's first brain-scan study on ChatGPT users suggests prolonged AI use may reduce brain function over four months.

Free JavaScript from Oracle - HN debate with 244 comments on liberating the JavaScript trademark from Oracle's control after 13 years.

Anthropic's Safety Team Under Pressure - The Verge digs into tensions facing the team studying AI's societal risks at Anthropic.

Anthropic Scores $200M Snowflake Deal - Claude now available to 12,600 Snowflake customers in multi-year partnership expanding enterprise AI access.

Google Partners with Replit - Taking on Cursor and Claude's coding dominance with AI-powered development tools in the browser-based IDE.

Amazon's Nvidia Killer Prints Money - AWS CEO reveals their custom AI chip is already a multibillion-dollar business despite Nvidia's market stranglehold.

AWS re:Invent 2025 Roundup - New agent deployment tools and infrastructure updates dropped at Amazon's annual developer conference this week.

Anthropic Brings Claude to Dartmouth - Claude for Education rolls out campus-wide through AWS partnership, expanding their higher education footprint.

React 19 Vulnerability Quick Fix - Someone packaged a Next.js security patch as an AI prompt in a GitHub repo, which is honestly the most 2025 way to distribute a hotfix.

Credit Union Mortgage Rate Tracker - Developer got quoted 7% by their bank, found 5.5% at a credit union, then rage-built a dashboard comparing rates across 120 credit unions because apparently that's a 1.5% "big bank convenience fee" we've all been paying.

Google AI Devs Announcement - Latest update from Google AI team

AI chatbots can be wooed into crimes with poetry - Poetic prompts bypass AI safety guardrails

CompTIA launches AI Essentials training to bridge workforce skills gap - New AI certification program for professionals

Wild that OpenAI is teaching models to admit when they're unsure while Cursor users are burning through $60/month in tokens like it's nothing. The honest AI might actually save us money.

What are you building or shipping this week?

Keep shipping,

P.S. If you're trying v0's new MCP integration, start small—nobody needs a custom supply chain dashboard connected to 47 different protocols on day one.