Week of May 3, 2026

This Week's Top Videos

Bring your work into Codex in a few clicks

By OpenAI

OpenAI's Codex now integrates with your entire work stack—Gmail, Slack, Google Drive, calendar—to create personalized AI workflows in under 60 seconds. The seamless plugin architecture lets it search across all your tools and produce meeting briefs automatically. This is the first glimpse of truly connected AI assistants that builders can implement today.

• Codex offers one-click project migration from other AI tools like Claude, preserving your existing workflows while upgrading capabilities
• The plugin architecture connects to 7+ core business systems simultaneously—browser, computer, calendar, email, Slack, Google Drive, plus document creation
• Personalization starts with telling Codex your work type, enabling customized experiences rather than generic AI responses
• Cross-platform search functionality allows Codex to pull context from Gmail, Google Calendar, and Slack simultaneously for comprehensive briefs
• Complete setup and first useful workflow execution takes under 60 seconds, making adoption friction nearly zero

AI-Personalized Medicine

By Y Combinator

Genome sequencing costs are falling faster than Moore's Law while FDA opens doors to N-of-1 personalized medicines using mRNA delivery vectors. AI agents can now analyze EHR data, wearables, and genomics to deliver user-specific medical suggestions. This convergence of cheap diagnostics, personalized drug manufacturing, and intelligent data analysis creates massive opportunities for healthcare startups right now.

• Genome sequencing costs are dropping faster than Moore's Law, making personalized diagnostics economically viable for widespread adoption
• N-of-1 genetic medicines can now be manufactured at scale using mRNA delivery vectors, enabling truly personalized drug treatments
• FDA is showing increased openness to personalized medicine procedures, reducing regulatory barriers for innovative treatments
• AI agents like Cloud Code can analyze multiple data streams (EHR, genomics, wearables) simultaneously for highly accurate personalized health recommendations
• New diagnostic tools entering the market enable early detection of various health signals, expanding preventive care opportunities
• The convergence of abundant health data and AI intelligence will democratize access to treatments for serious illnesses

Andrej Karpathy: From Vibe Coding to Agentic Engineering

By Sequoia Capital

Andrej Karpathy felt 'more behind as a programmer' than ever after December 2024, when AI agents stopped needing corrections and 'vibe coding' became reality. Software 3.0 isn't faster programming—it's prompting LLMs as programmable computers, making traditional apps obsolete. Builders must shift from writing code to crafting context.

• December 2024 marked a stark transition where AI code chunks consistently came out fine without corrections, enabling true 'vibe coding' workflows
• Software 3.0 paradigm: Programming becomes prompting, with context windows as your control lever over LLMs acting as interpreters
• OpenClaw installation demonstrates the shift—instead of complex bash scripts, users copy-paste text instructions to agents who intelligently adapt to different environments
• Traditional apps become obsolete: Karpathy's menu photo app was replaced by simply asking Gemini to overlay images directly onto menu photos
• The shift extends beyond coding to general information processing automation, working with unstructured data that couldn't be programmed before
• Builders must reframe from 'speeding up existing things' to recognizing entirely new capabilities are now possible with LLMs

Insider Tips for Prompting with AI

By Google

AI Studio builds complete games from prompts, but most founders fail by being too generic—'make me a racing game' vs. bullet-pointed specs with specific libraries like three.js can mean the difference between garbage and production-ready code. The key insight: ask the AI what technologies to use, then reference them in your prompts for dramatically better output.

• Specific prompts with bullet-pointed requirements drastically outperform generic requests—'blue car, round track, boost pickup' generates usable games vs. vague 'racing game' prompts
• Mentioning specific libraries like three.js in prompts triggers better model performance because AI models are trained extensively on popular frameworks
• Ask the AI itself what technologies to use before prompting—'what are the best web technologies to make 3D games' reveals optimal keywords for better outputs
• Image references alongside text prompts provide visual guidance that dramatically improves AI output accuracy for complex visual requirements
• Iterative prompting after testing the initial output allows fine-tuning specific behaviors like car handling or boost duration for production quality
• AI Studio's app gallery provides templates and examples that can be remixed, offering proven prompt patterns for faster development

OpenAI Codex Masterclass — Vaibhav Srivastav & Katia Gil Guzman

By ai.engineer

OpenAI's Codex isn't just a coding assistant—it's a full software engineering agent that runs tests, explores codebases, and integrates with Slack/GitHub workflows. New websockets deliver 1.75x faster tokens, with fast mode adding another 2x speed boost. This matters now because Codex can replace entire junior dev workflows while senior engineers focus on architecture.

• Codex functions as a complete software engineering agent beyond just coding—it runs commands, tests, explores codebases, and handles everything a software engineer colleague would do
• New websockets connection provides 1.75x faster token delivery, with fast mode adding another 2x speed improvement on top for dramatically faster responses
• Native work trees support in Codex app enables working on multiple features simultaneously within the same project without context switching or task interference
• Codex integrates directly into existing workflows through Slack, GitHub, CLI, IDEs, plus tools like Figma, Linear, and Notion for seamless adoption
• Automations allow Codex to run scheduled processes like daily briefings, calendar reviews, or routine tasks without manual intervention
• GPT-4 Mini and Nano models specifically designed for short-running tasks and sub-agents provide faster, more efficient processing for smaller workflows