Overview of 1021: We got addicted to an AI model we can't talk about
In this episode of Syntax, Wes Bos and Scott Tolinski talk with Dax Raad, creator of OpenCode, about the rapidly changing world of AI-assisted software development. The conversation covers remote dev environments, always-on machines, OpenCode’s 2.0 roadmap, model routing, token usage, open-source vs frontier models, Claude Code policies, MCP/tools, and how voice prompting is changing the way the team works. A major theme is that the newest generation of models feels unusually good to use — to the point that Dax says his team got “addicted” to one preview model they were temporarily given access to.
Remote Development Setups and Always-On Machines
Dax explains his long-running setup of using a beefy bare-metal server in the cloud as his primary dev machine instead of a local laptop.
Why he prefers it
- Better performance and easier upgrades
- No need to constantly buy/sell hardware
- Seamless switching between devices with SSH + Tmux
- Useful for coding agents, which benefit from always-on compute
Team rollout
- His team is now experimenting with the same pattern
- They’re using bare-metal servers sliced into VMs for each developer
- They’re aiming for servers with:
- Fast NVMe storage
- Modern CPUs
- Enough performance to compete with a MacBook
Providers mentioned
- Personal setup: a lower-cost random local provider, chosen mainly by CPU and price
- Team setup: Latitude.sh
- Related turnkey service: exe.dev
OpenCode 2.0: What’s Changing
Dax says OpenCode is heading into a major 2.0-style rewrite after learning what really matters in the space.
Core upgrades
- A redesigned API
- Service-first behavior: OpenCode runs as a persistent local service by default
- Better syncing across desktop/web clients
- Scriptability: users can write their own apps and automation against it
- New plugin API
- Major performance and UI improvements
Timeline
- Beta expected by the end of the week of recording
- Official release planned about a month later
Desktop app
- The desktop version has been in beta/“beta-beta” territory for a long time
- It is moving fully to Electron
- New 2.0 APIs, performance fixes, and a brand-new UI are in progress
How OpenCode Is Used Day to Day
Dax describes OpenCode as an always-on assistant that can operate across multiple machines.
Practical workflows
- One Tmux session per project
- Consistent window/pane layout for strong muscle memory
- OpenCode sessions for:
- Work projects
- Workout/training logging
- iMessage integration
- Cross-machine tasks via Tailcale-connected hosts
A memorable example
He tells a story about trying to use OpenCode to help buy a gift for his wife. The assistant kept escalating the interaction, and his wife reacted so negatively that the Cloud model essentially shut itself down after detecting the conflict.
Model Quality, Model Routing, and “Addictive” New Models
A big chunk of the conversation is about model behavior, usability, and where the best gains are actually coming from.
Key takeaways on models
- Dax and his team are not “AI hype” people; they’ve been cautious for years
- The newest preview models felt dramatically better in real workflows
- The improvement wasn’t just intelligence — it was usability, trust, and responsiveness
- The team lost access to one preview model and felt like they were mourning it
Why the newer models stand out
- They follow instructions better
- They pick up on context more reliably
- They feel like a better partner in coding, not just a smarter autocomplete
Model routing thoughts
- “Routing” at the API-middleman layer is somewhat limited
- More promising is the orchestrator pattern:
- One expensive “primary” model
- It spawns cheaper sub-agents for exploration and code changes
- Background work can happen in parallel
- Dax thinks this is where routing actually makes sense in practice
Cost, Inference, and the Economics of AI Tools
Dax gives a pretty candid look at model economics from an inference-provider perspective.
What their token usage says
- The team’s monthly token usage recently increased by about 5x
- That growth reflects how much more effective and useful the models have become in real work
Cost breakdown
- For their team, AI usage is about 15% of payroll
- He argues that’s manageable for a tech company
- He also believes frontier model inference margins may be extremely high
Open-source models
- He says open-source models are getting better fast
- In some cases, cheaper models are now “good enough” to replace older frontier models
- The gap between frontier and open-source is shrinking, though frontier labs still have an edge
Claude Code, OpenCode, and Platform Policy
The discussion also covers the messy relationship between model providers and third-party tools.
Current status
- OpenCode’s old integration that let users use their Claude Max plan is not allowed
- Using Claude’s headless SDK in third-party wrappers is currently in a gray area
- Dax says OpenCode won’t wrap Claude in a way that defeats the point of the product
Why providers may resist
- Consumer products can pull users into a company’s funnel
- Enterprise/API usage may be more profitable or strategically important
- Compute allocation becomes a business and sales decision, not just a technical one
Safety, Regulation, and Public Perception
The conversation briefly turns to AI safety and how governments may react to frontier models.
Dax’s view
- There are legitimate safety concerns
- Some kind of review process before major model releases makes sense
- The dangerous part is if access becomes unequal, corrupt, or overly restrictive
- He worries that hype and “nuclear weapon” style rhetoric could trigger irrational political responses
Tools, MCP, Artifacts, and UI for AI Agents
The team’s internal workflows use a lot of tools, connectors, and agent infrastructure.
What they rely on
- MCP servers/tools connected to company systems
- A Discord bot that acts like a supercharged internal assistant
- Access to their full data lake
- “Gang grills” as a collaborative prompt/decision-making pattern
Future UI direction
- OpenCode will likely add artifact support so agents can generate docs, HTML, SVG, and visual outputs
- Dax is bullish on richer interfaces for non-technical users
- He’s less convinced that MCP UI specs are fully ready yet, but expects support in the desktop app
Voice Prompting Is Taking Over
One of the more surprising themes is how much the team has moved toward voice.
Why they like it
- Faster than typing
- More natural for rough, messy thought
- The model is good at interpreting imperfect speech
- It reduces friction for prompting and internal communication
Hardware and apps
- Dax uses a foot pedal and mouse shortcuts to trigger dictation
- He uses:
- Hex on Mac
- Handy on Linux
- The team is increasingly using voice even in Discord messages
Notable Observations
On benchmarks
- Dax says benchmarks matter less than real-world feeling
- He trusts:
- Qualitative user feedback
- Team productivity
- Token usage trends
- The main question is whether people are happier and getting more done
On local models
- He supports running things locally for privacy
- But he argues local inference usually isn’t a real cost advantage
- If a model gets cheaper locally, it typically also gets cheaper in the cloud
Sick Pick and Plug
Sick pick
- exe.dev — Dax highlights it as a smart product for renting performant cloud dev machines with persistent disks
Shameless plug
- OpenTUI — the TUI framework powering OpenCode, written in Zig
- He says it’s a strong base for building polished terminal apps, and it’s becoming widely used
Main Takeaways
- Remote dev environments are becoming more practical, especially with coding agents.
- OpenCode 2.0 is a major redesign focused on persistence, extensibility, and polish.
- The latest AI models feel materially better to use, not just marginally smarter.
- Orchestrator-style workflows with sub-agents are likely the most useful form of “model routing.”
- Voice prompting is quickly becoming a real productivity multiplier.
- The biggest wins in AI tooling come from good primitives, good UX, and careful workflow design — not just bigger models.