Hey Everyone 👋,
John Lindquist here with the 28th issue of AI Dev Essentials!
Claude Haiku 4.5 dropped with impressive cost savings and speed improvements, OpenAI continues their infrastructure expansion with a major chip partnership, and we're seeing remarkable progress in AI's ability to tackle complex scientific problems.
Personally, I've been improving MCP/Hooks workflows in Cursor (and Claude Code) by authoring custom libraries for mcp servers and advanced hooks. These enable a lot of creativity and power around cost-savings and speed by offloading to different models, improving prompts by fetching context in the background, and much more. I'm very excited to share more about this in my Cursor workshop tomorrow.
🚨 THE CURSOR WORKSHOP IS TOMORROW! TICKET SALES END TONIGHT! 🚨 https://egghead.io/workshop/cursor
🎓 New egghead.io Lessons This Week?
I recently released mcpez, a new library that simplifies MCP (Model Context Protocol) local development and could reshape how you approach agentic development. The library makes it dramatically easier to build custom tools and integrations for AI agents.
🔗 Check it out: https://www.npmjs.com/package/mcpez
I've also been talking to the Cursor team about they're hooks setup so that I can push them a little further. I've got a few more lessons lined up that should be released early next week.
🚀 Major Announcements
Claude Haiku 4.5: Speed and Cost Efficiency Revolution
Anthropic released Claude Haiku 4.5 on October 15, 2025, delivering near-frontier model performance at dramatically reduced costs and significantly faster speeds.
Performance highlights:
- Cost Efficiency: Up to 90% cost savings with prompt caching, 50% with Message Batches API
- Speed Improvements: 4-5x faster than Sonnet 4.5 for real-time applications
- Benchmark Performance: 73.3% on SWE-bench Verified, achieving 90% of Sonnet 4.5's capabilities
- Pricing: $1 per million input tokens, $5 per million output tokens
- Availability: Claude Code, API, Amazon Bedrock, and Google Cloud Vertex AI
The model excels at computer use tasks, making Claude for Chrome significantly more responsive while maintaining high coding performance standards.
(Anthropic Official, TechCrunch, AWS)
Definitely awesome at taking a plan and executing it. There's a certain feeling with a faster model and quicker feedback loops, where as a human in the loop, you can stay more engaged and more active in the project. So even if it does something wrong, you can give it feedback and quickly turn it around. It's definitely something you miss when waiting for larger, smarter models to work through problems. So I'm a huge fan of faster, smaller models.
OpenAI and Broadcom Announce 10 Gigawatt Chip Partnership
OpenAI and Broadcom announced a strategic collaboration on October 13, 2025, to deploy 10 gigawatts of custom AI chips designed specifically for OpenAI's infrastructure needs.
Partnership details:
- Massive Scale: 10 gigawatt deployment planned by 2029
- Custom Design: OpenAI-designed chips optimized for their specific workloads
- Phased Rollout: Manufacturing begins in 2026 with gradual deployment
- Strategic Collaboration: Combines OpenAI's AI expertise with Broadcom's chip manufacturing
- Infrastructure Independence: Reduces reliance on third-party chip providers
This represents OpenAI's most significant move toward vertical integration and infrastructure control.
(OpenAI Press Release, Broadcom Investor Relations)
10 gigawatts is like 5 Hoover Dams. Or 10x the power Doc needed to send the Delorean to the future. Or 2x the power of Los Angeles at its peak. All of that power going into AI. It's so easy to forget the sheer scale of the AI era when I'm sitting behind my computer complaining about AI-generated em dashes.
🛠️ Developer Tooling Updates
Claude Code Introduces Plugin System in Public Beta
Anthropic launched Claude Code Plugins in public beta on October 9, 2025, enabling developers to install and share curated collections of tools, agents, and automations.
Plugin capabilities:
- Comprehensive Tooling: Slash commands, agents, MCP servers, and hooks in unified packages
- Community Marketplace: Share and discover plugins through centralized registry
- Easy Installation: One-command plugin installation via
/plugincommand - Developer Experience: Streamlined setup for complex automation workflows
- Team Collaboration: Share standardized toolsets across development teams
The plugin system addresses the complexity of managing multiple AI development tools and configurations.
Finally! Managing all the different MCPs, hooks, and custom commands was becoming a nightmare. Having a plugin system means I can package up my entire development workflow and share it with the team without everyone having to manually configure dozens of different tools.
Vercel v0 Integrates with Slack at Dreamforce 2025
Vercel demonstrated v0's native Slack integration at Dreamforce 2025 on October 15, 2025, enabling AI-powered application development directly within Slack channels.
Integration features:
- Native Slack App: Build and iterate on applications without leaving Slack
- Team Collaboration: Plaintext suggestions from team members directly actionable by v0
- Rapid Prototyping: Create landing pages and applications in conversational interface
- Agentic OS Vision: Part of Salesforce's broader AI platform strategy for workplace automation
- Workflow Integration: Seamless handoff between design, development, and deployment
The integration represents a significant shift toward conversational development workflows.
(Slack Blog, Salesforce Press Release)
I'm really curious what designs might look like when the barrier to feedback is lowered to posting a message to Slack on your phone. It's a little concerning that execs might start stepping on the designers toes, but the good execs should now be able to give much better actionable feedback. I love enabling people to share their ideas and anything we can do to move those ideas from text to visuals is a huge win IMO.
🤖 AI Ecosystem Updates
AI Models Achieve Gold Medal Performance in Astronomy Olympiad
GPT-5 and Gemini 2.5 Pro achieved gold medal performance in the International Olympiad on Astronomy and Astrophysics (IOAA), marking a significant milestone in AI's scientific reasoning capabilities.
Performance metrics:
- GPT-5 Score: 85.6% accuracy on IOAA competition problems
- Gemini 2.5 Pro Score: 84.2% accuracy, also achieving gold medal standard
- Human Comparison: Outperformed average human participants in complex astrophysics calculations
- Problem Types: Theoretical astrophysics, observational astronomy, and mathematical modeling
- Research Validation: Results published in peer-reviewed arXiv papers
This represents AI's growing capability in advanced scientific domains requiring complex mathematical reasoning.
(arXiv Research Paper, Google DeepMind Blog)
AI is smarter than me. I get it. But I think it's incredible that the leading labs struggle to find people smart enough to check on the work from the AI. I know they keep promising to "cure cancer" and I take those statements with a healthy dose of skepticism, but what if they do...
Google's C2S-Scale 27B Model Generates Novel Cancer Research Hypothesis
Google and Yale University's C2S-Scale 27B parameter model generated and experimentally validated a novel hypothesis about cancer cellular behavior, demonstrating AI's potential for scientific discovery.
Research breakthrough:
- Novel Hypothesis: Identified drug combination (silmitasertib) to enhance tumor visibility to immune system
- Experimental Validation: Successfully tested hypothesis in living cancer cells
- Model Architecture: 27 billion parameters trained on biomedical research data
- Collaborative Research: Joint effort between Google DeepMind and Yale medical researchers
- Open Science: Model and findings made available to broader research community
The work represents a significant step toward AI-driven scientific hypothesis generation and testing.
(Interesting Engineering, Google Blog)
This is exactly the kind of AI application that gets me excited. Rather than just optimizing existing processes, the AI actually generated new scientific insights that researchers could test and validate. It's a preview of how AI might accelerate scientific discovery by exploring hypothesis spaces that humans might not think to investigate. We don't know what we don't know!
NotebookLM Introduces Visual Styles for Video Overviews
Google launched six new visual styles for NotebookLM's Video Overviews on October 13, 2025, powered by the Nano Banana image generation model.
New features:
- Visual Styles: Watercolor, Papercraft, Anime, Whiteboard, Retro Print, and Heritage aesthetics
- Video Formats: New "Explainer" and "Brief" options for different content needs
- Nano Banana Integration: Leverages Gemini's image generation for consistent visual themes
- Rollout Timeline: Available to Pro users first, expanding to all users in coming weeks
- Content Personalization: Customizable presentation styles for different audiences
The update significantly enhances NotebookLM's content creation capabilities for educational and professional use.
(9to5Google, Google Labs Blog)
My main struggle with NotebookLM (which I love dearly) is the fact that I'm learning so many new things all at the same time. I never make time to follow-up or study to help solidify the knowledge. I just expect the AI to be able to recall it for me. I'm loving the video overviews to help the knowledge stick more during the first pass.
⚡ Quick Updates
OpenAI Announces Relaxed ChatGPT Content Restrictions
- Policy Changes: Reduced mental health restrictions for general users
- Personality Options: More human-like interactions and emoji usage coming soon
- Adult Content: Erotica allowed for verified adults starting December 2025
- User Choice: Opt-in system for personality changes, not default behavior
Amp Code Becomes Free with Ad-Supported Model
- Free Tier: All core coding features available without subscription
- Ad Support: Revenue through developer tool partner advertisements
- CLI Access: Available via
/mode freecommand or editor extension - Community Focus: Expanding access to agentic coding tools
Andrej Karpathy Releases NanoChat Training Pipeline
- Open Source: Complete LLM training pipeline in 8,000 lines of code
- Cost Effective: Train a ChatGPT clone for as little as $100 in 4 hours
- Educational Focus: Designed for LLM101n course and research harness
- Full Stack: Covers tokenization, pretraining, fine-tuning, and inference
Microsoft Hires Jared Palmer as VP of CoreAI
- New Role: VP of CoreAI and SVP of GitHub focusing on developer experience
- Strategic Focus: Building integrated AI tools across GitHub, VS Code, and Azure
- Industry Background: Former creator of popular React development tools
- Platform Integration: Emphasis on unified developer experience across Microsoft AI platforms
n8n Raises $180M Series C at $2.5B Valuation
- Funding Round: Led by Accel with multiple strategic investors
- Total Funding: $240 million raised to date
- Market Position: Leading workflow automation platform for AI integration
- Growth Focus: Expanding AI-native automation capabilities
Google Launches Gemini Enterprise for Workplace AI
- Enterprise Platform: AI agents for workplace automation starting at $30/user/month
- Integration Support: Works with Google Workspace, Microsoft 365, Salesforce, and SAP
- Agent Capabilities: Custom AI agents for department-specific workflows
- October Launch: Generally available as of October 9, 2025
✨ Workshop Spotlight
🚨 TODAY IS THE LAST DAY TO BUY TICKETS! 🚨 Cursor Workshop - Friday, October 17th, 2025
Become a Cursor Power User
In just a single day you'll set up shared rules, optimized workflows, and custom agents to help you ship AI-generated code with confidence.
Date: October 17, 2025 Time: 9:00 AM - 2:00 PM (PDT) Platform: Zoom
What You'll Learn:
The AI development landscape continues evolving rapidly with new models, tools, and best practices. This workshop teaches fundamental principles and repeatable workflows that work regardless of which tool you use—Cursor, Claude Code, WindSurf, or the next breakthrough.
Essential Skills:
- Planning & Cursor Rules: Refine workflows with custom rules and systematic planning
- Context Engineering: Learn the best practices around gathering context to solve the toughest problems
- Logging & Feedback Loops: Capture the output of your apps to create feedback loops
- UI Design: Using the Chrome Dev Tools MCP to iterate on design and debug the browser
- Agent Pipelines: Create specialized workflows with hooks and mcps to automate difficult tasks
Register: https://egghead.io/workshop/cursor
Read this far? Share "AI Dev Essentials" with a friend! - https://egghead.io/newsletters/ai-dev-essentials
- John Lindquist