IR for JavaScript - MLIR

 [RFC] JSIR: A High-Level IR for JavaScript - MLIR - LLVM Discussion Forums

This RFC introduces JSIR, a high-level IR for JavaScript:

• JSIR preserves all information from the AST and supports high-fidelity round-trip between source ↔ AST ↔ JSIR;
• JSIR uses MLIR regions to represent control flow structures;
• JSIR supports dataflow analysis.

JSIR is developed and deployed in production at Google for code analysis and transform use cases.

JSIR is open source here: GitHub - google/jsir: Next-generation JavaScript analysis tooling · GitHub

Industry trend of building high-level language-specific IRs

The compiler industry is moving towards building high-level language-specific IRs. For example, the Rust and Swift compilers perform certain analyses on their high-level IRs before lowering down to LLVM. There are also a number of ongoing projects in this direction, such as Clang IR, Mojo, and Carbon.

The need for a high-level JavaScript IR

Why do we need a high-level IR for JavaScript specifically? While much of JavaScript tooling relies on ASTs (like ESTree), complex analyses require a control flow graph (CFG) and dataflow analysis capabilities, which JSIR provides by using the MLIR framework.

[2024 LLVM DevMtg] JSIR - Adversarial JavaScript Analysis with MLIR PDF

This RFC (Request for Comments) introduces JSIR, a high-level Intermediate Representation (IR) for JavaScript developed by Google and built on the MLIR framework.

What is JSIR?

JSIR is designed to bridge the gap between abstract syntax trees (ASTs) and low-level IRs. Unlike typical IRs that lose source-level information during lowering, JSIR is "reversible," meaning it supports a lossless round-trip between:

Source $\leftrightarrow$ AST $\leftrightarrow$ JSIR

Key Features

• High-Fidelity Round-tripping: It preserves enough information to lift the IR back into valid JavaScript source code, achieving a 99.9%+ success rate in internal Google evaluations.

• MLIR-Powered Analysis: It uses MLIR regions to represent JavaScript control flow structures (like if, while, and logical expressions) as nested blocks rather than flat graphs.

• Enhanced Dataflow API: It provides a simplified wrapper over MLIR’s dataflow analysis, making it easier for developers to define lattices and transfer functions without manually managing worklists.

Primary Use Cases

Google currently uses JSIR in production for:

• Decompilation: Lifting Hermes bytecode back into readable JavaScript.

• Deobfuscation: Transforming obfuscated code into a clearer format, sometimes in combination with LLMs like Gemini.

• Code Transformation: Performing complex refactoring or optimizations that require both dataflow insights and the ability to output source code.

Current Status & Future

• Open Source: The project is hosted on GitHub.

• Community Integration: The authors are exploring upstreaming JSIR to the LLVM/MLIR project, though they note practical challenges regarding dependencies like QuickJS (for constant folding) and Babel/SWC (for parsing).

Tesla Model 3 Premium: $29K, in Canada, from China?

Tesla SLASHES Price of Model 3 Premium By OVER 50% in Canada - YouTube

Tesla has drastically reduced the price of the Model 3 Premium in Canada, lowering the entry-level price to approximately $29,000 USD (under $40,000 CAD). 

This move represents a reduction of over 50% from previous pricing. The Model 3 Performance also saw a 17% discount. 

The host suggests this strategy is likely intended to stimulate demand, potentially by leveraging vehicles sourced from China, and expects it will force competitors to lower their prices to remain relevant in the Canadian market.

Blender: 3D editor tool

Download — Blender

Blender (software) - Wikipedia

Blender is a free and open-source 3D computer graphics software suite. It runs on Windows, macOS, Linux, BSD, Haiku, and IRIX and is used for general 3D modeling, animation, digital sculpting, visual effects, rendering, compositing, motion tracking, drawing, 2D animation and video editing. It is widely used to create animated films, visual effects, art, 3D-printed models, motion graphics, interactive 3D applications, virtual reality, and 3D models for video games.

Blender was used to produce the Academy Award-winning film Flow (2024).[13]

Thinking as as scientist: book Third Millennium Thinking

excellent podcast interview!

How can we improve political dialog? with Saul Perlmutter – David Eagleman

The One Skill to Fix Politics - YouTube

This conversation between David Eagleman and Saul Perlmutter explores how the scientific mindset can be applied to solve societal polarization and improve public discourse. Here are the key takeaways:

• The Scientific Approach to Uncertainty: Science is not just a collection of facts; it is a rigorous process of navigating uncertainty. Scientists assume their current models are incomplete or potentially wrong and actively seek out evidence that challenges their theories.
• The Power of Disagreement: In science, disagreement is a valuable tool rather than an obstacle. By inviting criticism and working with those who hold opposing views, scientists can identify weaknesses in their own thinking and refine their understanding of the world.
• Moving Beyond Binaries: Human beings naturally tend to simplify complex issues into binary "us vs. them" narratives. A scientific approach encourages moving away from this tribal instinct toward probabilistic thinking, where we acknowledge that many issues are complex balancing acts rather than simple "right or wrong" scenarios.
• The Role of Intellectual Humility: A core principle of scientific thinking is intellectual humility—the willingness to admit that one might be wrong. This is presented as an essential skill for modern citizenship, allowing individuals to update their beliefs as new information becomes available.
• Structured Deliberation: Instead of traditional, debate-style interactions that prioritize "winning," society could benefit from structured deliberation. Techniques like citizen assemblies or facilitated, calm conversations (including potential assistance from AI) can help people focus on finding common solutions rather than protecting their personal identities.
• The "Meta-Problem": The speakers suggest that humanity's greatest challenge is not a specific technical problem, but a "meta-problem": how we communicate and navigate disagreement. If we can improve our collective ability to think through problems together, the evidence suggests that many other seemingly insurmountable challenges can be solved.

Saul Perlmutter at NobelPrize.org

Saul Perlmutter at Wikipedia

Third Millennium Thinking: Creating Sense in a World of Nonsense: Perlmutter PhD, Saul, Campbell PhD, John, MacCoun PhD, Robert: 9780316438100: Amazon.com: Books

Based on a wildly popular UC Berkeley course, how to use scientists’ tricks of the trade to make the best decisions and solve the hardest problems in age of uncertainty and overwhelming information.

In our deluge of information, it’s getting harder and harder to distinguish the revelatory from the contradictory. How do we make health decisions in the face of conflicting medical advice? Does the research cited in that article even show what the authors claim? How can we navigate the next Thanksgiving discussion with our in-laws, who follow completely different experts on the topic of climate change?

In Third Millennium Thinking, a physicist, a psychologist, and a philosopher introduce readers to the tools and frameworks that scientists have developed to keep from fooling themselves, to understand the world, and to make decisions. We can all borrow these trust-building techniques to tackle problems both big and small.

3D design: USD: Universal Scene Description

Universal Scene Description - Wikipedia

Universal Scene Description (USD) is a framework for interchange of 3D computer graphics data. The framework focuses on collaboration, non-destructive editing, and enabling multiple views and opinions about graphics data.[1] USD is used in many industries including visual effects, architecture, design, robotics, CAD, and rendering.[2][3]

It is developed by Pixar and was first published as open source software in 2016, under a modified Apache license.[4] Pixar, Adobe, Apple, Autodesk, and NVIDIA, together with the Joint Development Foundation (JDF) of the Linux Foundation, announced the Alliance for OpenUSD (AOUSD) on August 1, 2023 to "promote the standardization, development, evolution, and growth of Pixar's Universal Scene Description technology."

OpenUSD Home — Universal Scene Description 26.03 documentation

USD is a high-performance extensible software platform for collaboratively constructing animated 3D scenes, designed to meet the needs of large-scale film and visual effects production.

Running AI locally: Ollama in Docker container

Ollama is not "free only" anymore, has paid versions.
But the installer is not digitally signed, and is HUGE, 2 GB.
So Windows 11 rejects it for security reasons.

Download Ollama on Windows

Pricing · Ollama

Ollama can be run in Docker using its official Docker image. 

Running Ollama in a container provides a clean, isolated environment and simplifies deployment across different systems. [1, 2, 3, 4]

CPU-only:

docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama


Nvidia GPU support (Requires NVIDIA Container Toolkit):

docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama

AMD GPU support:

docker run -d --device /dev/kfd --device /dev/dri -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama:rocm

ollama/ollama - Docker Image

1.4 GB

ollama/ollama: Get up and running with Kimi-K2.5, GLM-5, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. @GitHub

Human & AI thinking: science & engineering

Can we engineer human thought? with Tom Griffiths – David Eagleman

"Can we engineer human thought?" with Tom Griffiths | Inner Cosmos with David Eagleman - YouTube

cognitive scientist Tom Griffiths... whether we're moving towards laws of thought.

Three Mathematical Lenses for Cognition

• Rules and Symbols: A foundational approach stemming from logic that aims to define the structure of thinking through formal rules, similar to computer programming or arithmetic.
• Artificial Neural Networks: These represent concepts as points in a multidimensional space, capturing the "fuzziness" and graded nature of human concepts that rigid logic often misses.
• Probability and Statistics: A framework for reasoning under uncertainty, allowing for the revision of beliefs as new evidence is acquired (Bayesian inference).

The Intersection of Humans and AI

• Data Efficiency: A primary distinction between humans and machines is the amount of data required to learn. Humans possess strong "inductive biases" that allow for rapid learning from few examples, whereas current AI models typically require massive datasets.
• Jagged Intelligence: Modern AI often exhibits inconsistent performance—brilliant in some areas and nonsensical in others—which is largely a result of being trained on different data and lacking the same inductive biases shaped by human biology and evolutionary constraints.

The Future of Cognitive Science

• Hybrid Modeling: The most promising path for understanding the mind lies in integrating these three lenses. Logic and probability help characterize the problems the mind solves, while neural networks provide a system for learning how to approximate those solutions.
• Engineering Thought: By moving toward a mature "physics of thought," researchers hope to eventually support human decision-making, optimize learning environments, and build more robust, generalizable AI systems that complement rather than just mimic human intelligence.

Key Takeaway Understanding human thought is not about finding a single "silver bullet" theory, but recognizing that different mathematical formalisms illuminate different facets of the mind. Viewing intelligence as an adaptation to computational constraints—rather than a one-dimensional measure—provides a more nuanced perspective on the future of both human and artificial cognition.

Amazon.com: The Laws of Thought: The Quest for a Mathematical Theory of the Mind eBook : Griffiths, Tom: Kindle Store

Everyone has a basic understanding of how the physical world works. We learn about physics and chemistry in school, letting us explain the world around us in terms of concepts like force, acceleration, and gravity—the Laws of Nature. But we don’t have the same fluency with concepts needed to understand the world inside us—the Laws of Thought. While the story of how mathematics has been used to reveal the mysteries of the universe is familiar, the story of how it has been used to study the mind is not.

There is no one better to tell that story than Tom Griffiths, the head of Princeton’s AI Lab and a renowned expert in the field of cognitive science. In this groundbreaking book, he explains the three major approaches to formalizing thought—rules and symbols, neural networks, and probability and statistics—introducing each idea through the stories of the people behind it. As informed conversations about thought, language, and learning become ever more pressing in the age of AI, The Laws of Thought is an essential read for anyone interested in the future of technology.

AI columnar db: ClickHouse

 Fast Open-Source OLAP DBMS - ClickHouse

ClickHouse is an open-source, column-oriented database management system (DBMS) designed for high-performance, real-time online analytical processing (OLAP). Written in C++, it enables fast SQL queries on massive datasets, making it ideal for analytics, business intelligence, and log management. 

Key Benefits & Features:

• Blazing Speed: Often 100x faster than traditional row-oriented databases for analytics, capable of processing billions of rows per second.
• Columnar Storage: Efficient compression and disk I/O, optimized for queries that analyze specific columns rather than whole rows.
• Real-time Ingestion: Supports high-speed data insertion while maintaining immediate queryability
  .
Scalability: Supports distributed processing, sharding, and replication.
Standard SQL: Uses SQL dialect with support for approximate calculations (e.g., uniqCombined, median) to increase speed further. 

Common Use Cases:

• Real-time analytics dashboards and user behavior analysis.
• Time-series data, such as observability, monitoring, and log analysis.
• Financial analytics and fraud detection.
• AdTech and marketing technology (MarTech). 

Protocol & Access:
ClickHouse supports native, HTTP REST API, MySQL, and PostgreSQL wire protocols, allowing easy integration with existing tools. 

Deployment Options:

• Self-Managed: Deployable on-premise, in cloud VMs, or via Kubernetes.
• ClickHouse Cloud: A fully managed database-as-a-service (DBaaS). 

Originally developed by Yandex in 2009 and open-sourced in 2016, ClickHouse is now developed under the Apache 2.0 license. 

 +1The leading
database for AI

Powering agentic systems with millisecond queries at petabyte scale.

portable GUI: Avalonia, Impeller, Flutter

 Impeller rendering engine

Impeller provides a new rendering runtime for Flutter. Impeller precompiles a smaller, simpler set of shaders at engine-build time so they don't compile at runtime.

Understanding Impeller — Flutter’s Modern Rendering Engine | by Ammar Yasser | Medium

Avalonia UI – Open-Source .NET XAML Framework | WPF & MAUI Alternative

.NET Rocks! Avalonia 12 with Mike James & Matt Lacey

Avalonia, the open source UI framework for building cross-platform applications with .NET. Mike's conversation with the Google Flutter team has led to replacing the Skia rendering engine in Avalonia with the newer Impeller Rendering Engine that Flutter itself depends on. This opens the door to excellent smartphone implementations with Avalonia, alongside its usual desktop and embedded roles.

Stitch - Design with AI

Local AI PCs

 The Cheapest 4TB DGX Spark Alternative… ASUS GX10 - YouTube

The Most Powerful APU on Earth - YouTube

$2,000 cheaper Beelink GTR9 Pro through LLM benchmarks against the Mac Studio M3 Ultra and other AMD Ryzen AI Max+ 395 (Strix Halo) boxes

Amazon.com: Beelink Mini PC, GTR9 Pro AMD Ryzen AI Max+ 395 CPU (126 Tops), 128GB RAM 2TB Crucial SSD, Mini Computer 10GbE Dual LAN/WiFi 7+BT5.4/8K Quad Display/USB4.0 * 2/SD Card Slot/DeepSeek 70B : Electronics
$2999

Cute, but powerful: meet NanoCluster, a tiny supercomputer - YouTube

AI Agent Skills

An AI agent skill is a portable, reusable bundle of instructions, scripts, and context (a skill.md file) that extends an agent's capabilities to perform specific tasks.

These skills act as "modular tools" that provide domain expertise, ensure consistent workflows, and reduce token usage by being loaded only when needed

 Agent Skills Overview - Agent Skills.io

A standardized way to give AI agents new capabilities and expertise.

Agent Skills are a lightweight, open format for extending AI agent capabilities with specialized knowledge and workflows.At its core, a skill is a folder containing a SKILL.md file. This file includes metadata (name and description, at minimum) and instructions that tell an agent how to perform a specific task. Skills can also bundle scripts, reference materials, templates, and other resources.

my-skill/
├── SKILL.md          # Required: metadata + instructions
├── scripts/          # Optional: executable code
├── references/       # Optional: documentation
├── assets/           # Optional: templates, resources
└── ...               # Any additional files or directories

Agent Skills – Codex | OpenAI Developers

Agent Skills | Microsoft Learn

code.claude.com/docs/en/skills.md

You're likely missing out on agent skills true potential! - YouTube
by Maximilian Schwarzmüller - YouTube

Agent Skills 101: Never Explain Twice Again... - YouTube
by NeuralNine - YouTube

Agent Skills :Standard for Smarter AI | by Plaban Nayak | Medium