Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
My AI Died 10,000 Times and All It Learned Was Physics
An AI learns a game without rewards, discovering controls and physics from its own actions. It reasons by predicting outcomes, updating its understanding when predictions fail.
An agent that learns to play a side-scrolling game it has never seen — with no gradient updates and no reward function — by discovering the controls, recovering the game’s physics from watching itself, and then reasoning about each move by imagining the outcome before it commits. Its entire policy is plain-English notes you can read out loud and argue with, not a set of weights.
Live, I’ll run a working browser app and walk through it thinking: it presses the buttons to work out which does what, induces the physics from its own jumps (and shows those recovered beliefs in readable language), then imagines a jump’s arc — drawn on screen as a ghost trajectory the instant before it leaps — to derive the move that clears each obstacle. You’ll see it fail, update its model, and write a new note in its own words, and you’ll be able to read its whole “notebook” and even race it yourself.
- TypeScriptTypeScript is an open-source superset of JavaScript: it adds static typing and compiles to clean, standards-based JavaScript.TypeScript is a high-level, open-source language developed by Microsoft: it acts as a superset of JavaScript, adding a powerful static type system. This system enables compile-time type checking, catching errors before runtime (a critical benefit for large-scale applications). The TypeScript Compiler (TSC) reliably transpiles all code into clean, standards-based JavaScript (ES3 or newer), ensuring compatibility across any browser or host environment (Node.js, React.js, etc.).
- OllamaDeploy and run open-source Large Language Models (LLMs) like Llama 3 and Mistral locally on your machine: achieve private, cost-effective AI via a simple command-line interface.Ollama is the essential tool for running LLMs locally: consider it the Docker for AI models. It packages complex models and dependencies into a single, easy-to-use application for macOS, Linux, and Windows systems. You get immediate access to models like Gemma 2 and DeepSeek-R1 via a straightforward CLI or REST API. This local-first approach guarantees data privacy and security, eliminating cloud dependency and high API costs. Ollama also optimizes performance on consumer hardware using techniques like quantization, ensuring efficient execution even on standard desktops.
- ViteVite is the next-generation frontend build tool: instant server start with native ES modules and lightning-fast Hot Module Replacement (HMR).Vite (French for 'quick') is a modern build tool created by Evan You (Vue.js's creator) that delivers a dramatically faster development experience. It operates in two major parts: a dev server that serves source code over native ES modules, eliminating the initial bundling step for instant startup; and a production build command that leverages Rollup and esbuild for highly optimized static assets. This architecture provides consistently fast HMR, even for large applications, and offers out-of-the-box support for TypeScript, JSX, and CSS via a universal plugin API.
- NodeNode.js is a high-performance JavaScript runtime built on the V8 engine for executing scalable network applications.Ryan Dahl launched Node.js in 2009 to rethink server-side concurrency. It utilizes an event-driven, non-blocking I/O model to manage thousands of concurrent connections on a single thread. The system runs on Google's V8 engine (C++) and provides access to npm (a registry with over 2 million packages). Companies like Netflix and LinkedIn use it for its speed and scalability: it remains the top choice for real-time data streaming and microservices.
- HTML CanvasHTML Canvas is a powerful HTML5 element providing a bitmap-based drawing surface for rendering 2D and 3D graphics, animations, and real-time data visualizations via JavaScript.The HTML `<canvas>` element is your dynamic, pixel-level drawing board, introduced with HTML5. It functions strictly as a container, requiring JavaScript and the `getContext('2d')` method to access its powerful 2D rendering context (or WebGL for 3D). Unlike SVG, Canvas is bitmap-based, making it the go-to solution for high-performance tasks: complex game graphics, real-time video processing, and interactive data visualization (e.g., charts). Developers utilize methods like `fillRect()`, `arc()`, and `drawImage()` to programmatically manipulate every pixel, enabling creation of rich, custom web experiences.
Related talks
More from the community
Building a System That Knows Why It Changed
Nashville
This talk shows a system that learns from failures by committing to expected outcomes and using structured debug…
Watch a Governed Multi-Agent System Block, Verify, and Replay AI Code — Live
Montreal
See a live demo of aming-claw, a system that lets AI agents write code under enforced contracts, including…
Ask the Oracle Before You Decide : Simulating the Future with LLM Agents
Toronto
Simulate future decisions with Pythia, an LLM engine modeling cognitive biases and crowd psychology. Watch AI agents interact,…
Stop Re-Reading Your Conference Notes: Build a Spaced-Repetition Course Instead.
Columbus
Learn how to build a spaced-repetition course from conference notes using Claude agents, turning transcripts, audio, and photos…
Words to World, AI learning strategies for building world models.
San Diego
Explore AI that builds physically real worlds from text, navigates them with robots, and learns through self-modifying code…
Your AI is Two Versions Behind
Zürich
Learn how to fill AI model knowledge gaps by structuring recent changes into markdown folders. See a Go…
Compose Email
Loading recent emails...