1024: Open Models Replace Big AI

Summary of 1024: Open Models Replace Big AI

by Wes Bos & Scott Tolinski - Full Stack JavaScript Web Developers

1h 7m•July 27, 2026
Listen to Episode

Overview of 1024: Open Models Replace Big AI

This episode of Syntax covers a wide mix of web-dev news and tools, with the biggest themes being Vue 3.6’s performance upgrades, new ways to manage Git worktrees, browser platform tracking, and a big discussion around open-weight AI models that are quickly catching up to frontier models. The hosts also dig into practical developer apps, a nostalgic look at Jurassic Park’s real computers, a live Tokyo train visualization, and a “funeral” for Stack Overflow as AI changes how developers search for answers.

Vue 3.6 RC: Faster Reactivity and Vapor Mode

Alien Signals under the hood

  • Vue 3.6 has entered release candidate status.
  • A major internal refactor uses Alien Signals, a very fast signals library.
  • The result: Vue apps should get a performance boost without a major version bump or breaking changes.

Vapor mode

  • Vapor mode removes the virtual DOM from the runtime for opt-in components or full apps.
  • Similar in spirit to Svelte and SolidJS:
    • compiles down to direct DOM update calls
    • ships less runtime JavaScript
    • better for highly dynamic UI like tables or dashboards
  • It can be enabled:
    • per component
    • per app
  • The hosts are excited because it gives Vue a path to much better performance for specific use cases.

Rift: An Experimental Alternative to Git Worktrees

Why worktrees are awkward

  • Git worktrees are useful for parallel branches and agentic coding.
  • But they can be slow and confusing:
    • lots of folder setup
    • dependency installs in each worktree
    • planning overhead for multiple agents

What Rift does

  • Rift is an experimental tool from the OpenCode team.
  • It uses copy-on-write filesystem primitives instead of copying every file.
  • Benefits:
    • much faster workspace creation
    • less duplication overhead
    • still supports isolated writable workspaces

Key limitation

  • It’s still experimental
  • You still need separate installs in workspaces, though cached dependencies help
  • The hosts’ takeaway: it works well enough to be interesting and may be a better default than traditional worktrees for some workflows

Staying Current with Browser APIs: MDN Updates Page

A hidden but valuable resource

  • Scott highlighted MDN’s updates page:
    • developer.mozilla.org/en-US/updates
  • It tracks what lands in:
    • JavaScript
    • HTML
    • CSS
    • browser APIs
    • Node-related web platform changes

Notable APIs and features mentioned

  • text-fit: automatically scales font size to fit a container
  • FlexWrap Balance / FlexFlow Balance
  • popover hint
  • rule row
  • sanitizer API entries

Why this matters

  • It’s a strong source for:
    • keeping up with browser support
    • feeding data into tooling or AI agents
    • discovering platform features before they become widely discussed
  • The hosts note the page is dense, but extremely useful once you know it exists.

Open-Weight AI Models Are Catching Up

The headline models

The episode spends a lot of time on a major wave of new open-weight model releases:

  • Kimi K3
  • Qwen 3.8
  • Inkling from Thinking Machines

Kimi K3

  • A massive 2.8 trillion parameter model
  • 1 million token context window
  • multimodal: can handle images, PDFs, and documents
  • Strong coding benchmark performance, reportedly competitive with top frontier models
  • Open weights are expected to be released soon, but the initial API use is hosted in China

Qwen 3.8

  • Another huge model from Alibaba
  • Similar scale and benchmark strength
  • Also not yet broadly available for local deployment at the time discussed

Inkling

  • A U.S.-based model from Thinking Machines
  • Trained from scratch
  • Released as open weight
  • Strong performance in design-focused benchmarks
  • Weaker than the top models on some software benchmarks, but still notable because it’s a serious U.S. open model

Why this is a big deal

  • Open-weight models are now getting close to frontier quality
  • They’re often much cheaper than proprietary models
  • Enterprises may prefer them for:
    • privacy
    • compliance
    • cost control
    • self-hosting
    • fine-tuning on internal data

Open weight vs. open source

  • The hosts clarify an important distinction:
    • Open source refers to software/source licensing
    • Open weights means you can access the trained parameter values
  • You can inspect and run the weights, but not necessarily see:
    • the training process
    • the full dataset
    • the architecture choices behind the training
  • Licensing still varies and may restrict commercial use or redistribution.

Distillation and model training

  • They explain how many models are likely being built through distillation:
    • prompt a stronger model many times
    • collect outputs
    • train a new model on that generated data
  • The hosts note this is ethically murky, especially when it piggybacks on work from companies like Anthropic or OpenAI.

Mixture of experts

  • Some of these models use mixture-of-experts architecture.
  • Not all parameters are active at once.
  • That makes very large models more practical than their raw parameter counts suggest.

Stack Overflow’s Decline

The data trend

  • A graph from Stack Overflow’s public data shows a huge drop in question volume:
    • strong usage in the pre-AI era
    • steep decline after ChatGPT-era adoption
    • now under 1,000 questions per day in the example discussed

Why the hosts care

  • They see this as a loss for the web because:
    • questions are no longer indexed and reusable
    • people solve problems in private chats instead of public forums
    • the same questions get re-answered repeatedly by AI
  • Wes notes that Stack Overflow often forced people to ask better questions and think more deeply about the correct approach, not just the quick fix.

The bigger concern

  • Less public knowledge sharing means:
    • less searchable learning for future developers
    • more duplicated compute and effort
    • more AI-generated “slop” if these systems aren’t curated well

Tools and Apps Worth Checking Out

Mole

  • A Mac cleanup app/CLI for:
    • finding large files
    • removing caches
    • uninstalling stuff cleanly
  • The hosts like that it’s:
    • polished
    • available as both CLI and GUI
    • a one-time purchase instead of a subscription

Moshi

  • A mobile terminal app for iOS and Android
  • Supports:
    • touch interaction
    • Mosh-based connectivity
    • Herder integration
  • Useful for managing terminal workflows from a phone without fighting tiny keyboards and awkward shortcuts

Foot pedal workflow

  • Scott describes using an Elgato foot pedal for:
    • dictation
    • Enter key input
    • switching terminal tabs / Herder workflows
  • It’s part of his broader productivity setup with terminal apps and session managers.

Other Fun/Notable Links

Decoy Font

  • A font designed to make text harder for AI systems to read from images
  • The idea:
    • humans read one message
    • AI sees a different or garbled one
  • Intended for:
    • PDFs
    • images
    • anti-training / anti-scraping use cases
  • The hosts joke that it sacrifices accessibility, since it’s image-based rather than real text.

Jurassic Park deep dive

  • A retro-tech article about the computers shown in Jurassic Park
  • The movie used real computers and tried to be accurate for nerd viewers
  • The hosts discuss film camera shutter rates and the technical detail involved in filming screens.

Tokyo train voxel viewer

  • A real-time 3D voxel visualization of Tokyo transit
  • Lets users ride and follow trains using live transit data
  • Wes and CJ are excited because it’s both fun and useful for travel planning.

Main Takeaways

  • Vue is getting meaningfully faster with Alien Signals and Vapor mode.
  • Git worktrees may have a more ergonomic future with tools like Rift.
  • MDN’s updates page is a great way to track browser platform changes.
  • Open-weight AI models are becoming competitive with frontier closed models, especially on cost and customization.
  • Stack Overflow is in steep decline, and that may be bad for the long-term quality of public developer knowledge.
  • A bunch of smaller tools and apps—Mole, Moshi, Decoy Font—show how much developer tooling is still evolving quickly.

Resources Mentioned

  • Vue.js 3.6 RC
  • Alien Signals
  • Rift
  • MDN browser updates page
  • Kimi K3
  • Qwen 3.8
  • Inkling
  • Stack Overflow data explorer
  • Mole
  • Moshi
  • Decoy Font
  • Tokyo transit voxel viewer