Some smoooooooth template literal types spotted in Vercel's Chat SDK Gives you automatically typed error codes for multiple features at once.

Damn, I really want to make a "UX Patterns For AI Applications" course All the hard stuff like stream resuming, generative UI, properly streaming markdown, editable plans, showing LLM progress

Hilarious use case for a mapped type: Use it to rename an object key your backend developers consistently misspell parms -> params

Some days my brain doesn't work. Today is one of those days

Spotted in the wild today: a Mutable<T> type. Does the opposite of Readonly<T>. Weird that TS doesn't have this built-in.

I've made a breakthrough today - come up with a new style of course that's ready for the AI-assisted developer era. It's just unbelievably cool, can't wait to show you.

Two cool things in this snippet: 1. You can use 'glob' from Node instead of a library! (still experimental, available in 22+) But glob returns an async iterator, which can be annoying to work with. So... 2. Array.fromAsync lets you turn an async iterator into an array. Beautiful.

Want to enter the world of AI, but don't know where to start? I've built a roadmap on how to become an AI Engineer. We cover: - What AI engineers are - What LLM's are used for - Evals - A directory of 17 techniques It's the perfect way to get ahead of the curve. www.aihero.dev/ai-engineer-...

"The AI Engineer Roadmap" dropping soon. Covering everything you need to enter the AI world: - What LLM's are used for - The AI engineer mindset - Evals - 17 techniques for improving your LLM-powered apps

This is unbelievably cool Just keep pressing 'space' to learn how an LLM works

I have worked with TypeScript for 7 years. I have written a book, produced 5 courses, made countless videos and articles, and delivered workshops about TypeScript. And today I learned that TS can do this:

It's the evals, stupid

Trying to define an agentic loop in the fewest amount of words.

"In a world where folks are throwing LLMs at every problem, being able to sniff out when not to use them is a valuable skill."

LIVE NOW with Anders Hejlsberg: www.youtube.com/watch?v=NrEW...

OK, Bluesky. Time for an early preview. (X hasn't seen this yet) I've shipped a 9-lesson tutorial on the Model Context Protocol - the USB-C of AI. It'll take you from "WTF is MCP" to shipping your own servers to the world. And, it's free. Check it out: www.aihero.dev/model-contex...

Nearly ready...

Give me your favourite MCP client. Claude Code is fine. Cursor is OK. I want to branch out.

TypeScript announces rewrite to Go, with a 10x editor and CLI speedup. It's the biggest TS announcement I can remember. www.totaltypescript.com/typescript-a...

Here's how to build an MCP server with a single TypeScript file, and connect it to Claude Code. No build step. No bullshit. Totally transformative.

Next.js, do you even know who I am?

Shower thought: The system prompt is the LLM's UI. In an agentic loop, elements of the system prompt have state, as well as props - lists of failed attempts, external data, exemplars. So, a React for system prompts makes sense.

Jina's open source DeepSearch implementation is fascinating, but SUPER hard to read. So, as a research project, I refactored the code to make the underlying logic clearer.

Has anyone written a JSX -> system prompt library? node-deepresearch would genuinely benefit so much from it. Endless string interpolation, impossible to grok what's going on. A proper component model for system prompts would be pretty convenient.

Folks who are building Deep Research, or Deep Research-type products... What do your evals look like? Do you e2e the entire system? Do you do any mocking? Do you eval individual parts of the system? (query rewriter, etc)

And @nor.dev is go. Kicking off with The Future of #TypeScript by @mattpocock.com #norDevCon25

VSCode catching up with Cursor. Add this to get Cursor-like tab completions for suggesting the next edit.

One fascinating decision with building LLM integrations is Agent vs Workflow. Agent: give the LLM control of execution. Slower, but better at a wider range of things. Workflow: use code to control execution. Faster, but more specialised. Ultimately, it comes down to the product you're building.

Did some refactoring on Evalite (my evals runner for AI-powered apps) and made two big calls: - Commander -> stricli - Remix -> TanStack stricli is a gorgeous, TypeScript-first CLI library that just works. Easy to test, easy to run, and full type-safety for parameters and named flags

Looking to up my Cursor game. The feedback loop is extraordinary but the UI is strange. Drop me your best tips.

STORM is an amazing paper for anyone interested in AI workflows. It's a Rube Goldberg machine for generating long-form articles. And it's really freaking cool. Here's my breakdown:

Is there a decent chunking algorithm library on NPM? I know Langchain and LlamaIndex have some, but figured there were probably some unbundled from frameworks. Chunking: chunking text documents to be fed into a RAG system.

TIL about process.exitCode = 1; Useful if you want to mark a process as failed without immediately exiting it

Deepseek R1 is pretty good at TypeScript, y'all. These are its reasoning tokens, not the text emitted to the user.

I'm reviewing classic papers from the field of AI agents. And I'm starting with a nice intro to the field - Google's whitepaper on agents. It's particularly interesting because they define two different types of tools: extension and functions. Enjoy! https://www.aihero.dev/google-agents-white…

Give me the best API's to hook up to AI agents. I want stuff that's frictionless to get an API key for, like Serper for Google Search.

Planning to look at these papers for my AI Agent papers (free!) mini-course: ReAct Tree Of Thoughts Reflexion Google's Agent Whitepaper Anthropic's Building Effective Agents Agents That Matter Which ones am I missing?

Thinking about making a course covering the most important papers for AI engineers, in language you don't need a PhD to understand. And not the most important theoretical ones, but the most useful practical stuff to build great apps. ReAct, Reflexion, Prompt Report etc. What do you think?

I've been messing about with a pattern where the agent persists small snippets in its working memory. This small change has a pretty huge impact on the user experience. Feels like your agent is really getting to know you.

I asked a few thousand folks (my AI Hero mailing list) what they want to learn about AI. I saw this sentiment a LOT: "Most material out there goes deep on how these models work. I see that as a long-term goal. Right now I want to learn to build stuff with them." Do you agree? I do.

Give me the best resources you've ever read on AI agents. Go.

just wrapped up a vector search-driven recommendation system for @mattpocock.com's aihero.dev. 🎯 here’s a sneak peek at the ui: a “RECOMMENDED NEXT” overlay dynamically suggesting your next resource.

My first AI course is complete... And it's free! It's a 16-lesson tutorial on Vercel's AI SDK. It's the missing piece for working with AI in TypeScript. No bullshit, no hype. Just solid, practical advice. https://www.aihero.dev/vercel-ai-sdk-tutorial

Right, I'm building an AI agents course. Tried scripting the first video today and realised I can't talk about agents without: - Client/server boundaries - Human in the loop - Routing & classification I've got my work cut out

Folks who've been enjoying my AI SDK tutorial... What should I cover next? Agents? RAG? Evals?

What are you using for observability? If your answer is boring, write it anyway. I want to know the dull, everyday tools that folks actually use in prod.

My first AI course is complete... And it's free! It's a 16-lesson tutorial on Vercel's AI SDK. It's the missing piece for working with AI in TypeScript. No bullshit, no hype. Just solid, practical advice. https://www.aihero.dev/vercel-ai-sdk-tutorial

What tricks have you used for getting round the soft limit on the number of tools that a LLM can use? - Making tools more polymorphic - Orchestrator LLM's What else?

However much I focus on AI, the TS stuff keeps being relevant. This time, it's the AI SDK not providing a built-in type for tool calls - so I've got to extract it using: - Awaited<T>, to unwrap the promise - [number], to turn the array of tool calls to a single member

🔥 Massive Content Drop 🔥 I'm open-sourcing AI Hero, the repo of material I'm putting together for my course. So far, there are 16 runnable examples on Vercel's AI SDK. Full tutorial dropping any day now. https://github.com/ai-hero-dev/ai-hero/