Overview of Syntax’s State of AI discussion
Wes Bos and Scott Tolinski break down the latest State of AI survey from Sacha Greif (the creator of the State of JS/CSS surveys) to examine how developers are actually using AI, which tools they prefer, what they pay for, and what concerns are growing. The big takeaway: AI is now a normal part of many developers’ workflows, usage is climbing fast, and the industry is still figuring out how to measure the real impact beyond “which chat app are you using?”
Biggest survey takeaways
AI usage is rising sharply
- A large share of respondents now use AI heavily in day-to-day coding.
- Notable self-reported usage levels:
- 18% said around 75% of their code is AI-assisted
- 19% said 85–90% of their code is AI-assisted
- 21% said they constantly refactor with AI, up about 10 points year over year
AI is used for much more than code generation
Respondents said they use AI for:
- Code generation — 90%
- Code review / assistance — 68%
- Learning / research — 68%
- Debugging — 67%
- Image generation — 37%
AI is becoming a workflow default
- 72% now consider AI an integral part of their workflow
- Overall happiness with AI ticked up slightly from 3.3 to 3.4 out of 5
Models, agents, and tools people actually use
Most recognized models
- ChatGPT/OpenAI remains the most widely known and used
- Claude has noticeably better sentiment:
- 2% negative sentiment
- stronger positive sentiment than ChatGPT in the survey
- The hosts note that many people still don’t clearly understand the difference between:
- the app
- the model
- the provider
Most used paid agents / subscriptions
- Claude Code leads paid agent usage at 58%
- Copilot is next at 42%
- OpenAI Codex at 24%
- JetBrains at 5%
- Cursor appears as a write-in and is clearly undercounted in the survey
Important tools missing or underrepresented
The hosts argue the survey should explicitly include:
- Cursor
- OpenCode
- Pi
- Kiro
- other emerging coding agents and IDE-adjacent workflows
Their point: a lot of developer behavior is now about how the agent is orchestrated, not just which model is selected.
How Wes and Scott are using AI
Wes
- Uses a mix of:
- Claude Code
- Cursor
- Codex
- Copilot models in OpenCode and Pi
- Prefers:
- Claude Code for most projects
- Pi for system-level tasks like server/NAS maintenance
- Likes the terminal for quick operations and agentic tasks
Scott
- Has increasingly moved toward Claude Code
- Recently canceled his ChatGPT subscription
- Finds Codex too slow and too verbose for frontend work
- Still likes IDE-based visibility in tools like Cursor
- Uses AI for tasks like:
- Docker/image updates
- NAS maintenance
- file operations
- quick automation
Image, video, and app generation
Image generation
- Nano Banana appears to be the top image model in the survey
- ChatGPT image generation also sees usage
- The hosts prefer AI image tools for:
- assets
- thumbnails
- icons
- background images
- replacing bad stock-photo searches
They also mention:
- Midjourney for higher-end creative assets
- OpenRouter and Replicate for pay-as-you-go image/model usage
Video generation
- The hosts are skeptical of practical use cases
- They mostly see it as novelty or social-media bait
- Their view: video generation is not yet a must-have developer workflow tool
App generation
- v0 and Lovable show some usage, but the hosts think the initial hype has cooled
- They suggest much of the early buzz may have been paid promotion rather than durable adoption
Risks and pain points
Top concerns
Survey respondents most often flagged:
- Hallucinations / inaccuracies
- Code quality
- Job displacement
- Military use
- Environmental impact
- AI slop
- Negative cognitive effects
- Rising AI costs
What the hosts emphasize
- AI-generated code often needs strong review and refactoring
- Good AI use depends on:
- linters
- tests
- proper error handling
- security checks
- strong human oversight
- They feel AI often makes codebases more fragmented, not less, unless handled carefully
Local AI: interesting, but likely misunderstood
- The survey says a large chunk of respondents use local AI
- Wes and Scott think this number is probably inflated or misunderstood
- Their view:
- many people do not really know what “local AI” means
- truly useful local AI often requires expensive hardware
- most people who want cheaper or private workflows are moving toward smaller/cheaper cloud models instead
Survey critique: what’s missing
The hosts think the survey is valuable, but too narrow. They want future surveys to ask about:
Workflow and orchestration
- Are people using:
- MCP servers
- skills
- subagents
- swarms
- multi-agent systems?
- How many agents are running at once?
- How are people monitoring them?
Developer strategy
- Are people using:
- markdown docs
- ADRs
- task managers
- planning layers
- How much code is reviewed by hand?
- How much AI output is actually read carefully before shipping?
Psychological impact
They also want questions about:
- stress
- fragmented attention
- confidence
- mental fatigue
- whether AI makes developers feel more productive or more anxious
Their core critique: the survey measures tools and sentiment, but not enough about how AI is actually reshaping work.
Final takeaway
The episode’s main message is that AI coding is no longer niche. It’s becoming standard practice, especially for developers using tools like Claude Code, Cursor, Copilot, and OpenCode. But the industry is still in a messy transition:
- usage is rising fast
- costs are rising too
- the workflow is changing
- and the next big question is not “Do you use AI?”
It’s: How are you actually using it, and how much of your workflow is now built around it?