Wes Bos: How Developers Stand Out When AI Writes the Code

AI is changing how developers prove skill

• With AI agents writing code, “the code doesn’t speak for your ability anymore.”
• Developers need other signals to stand out, such as:
  • Personal branding
  • Demonstrated problem-solving
  • Community involvement

Generative UI: promising but not “cracked”

• The discussion centers on LLMs generating interfaces using pre-configured component sets filled with context.
• The group argues that “generated design from an LLM” hasn’t yet reached the quality bar set by strong human UI/UX designers.
• Example limitation: LLM-generated overlays (e.g., podcast lower-thirds) can still feel cheesy or insufficiently polished.

Context-aware UI and a “middle ground” approach

• Future UIs could be generated/updated using rich personal context, such as:
  • Smart home state (cameras, lights, interactions)
• Instead of manually designing every screen, the UI could be:
  • Auto-generated, or
  • Adaptively populated based on context
• A key approach is generative UI with tweakability:
  • Use LLMs to scaffold from components
  • Allow humans to adjust to taste
• The group criticizes “pure” user-driven generation when it produces low-quality experiences—while acknowledging users can often know what they want.

Agent workflows and “loop engineering” (agentic iteration)

• People are moving from simple prompting to engineering the workflow/loop around the model—often described as loop engineering.
• Reason: websites/apps are dynamic:
  • errors
  • user complaints
  • logs
  • updates
• Effective agents may need self-healing behavior, but fully autonomous continuous improvement is currently too expensive (token costs / compute).

Practical looping pattern

• Use LLMs iteratively:
  • collect context
  • run a bounded prompt
  • re-check against sources (API docs/examples)
  • repeat in loops
• Example described:
  • Building a UI while repeatedly validating coverage against an API + examples worked better than trying to do everything in one pass.

Cost pressure and choosing “harnesses” / tools

• Speakers compare using different coding “harnesses” (AI IDE integrations).
• Key points:
  • Some teams/models are expensive, and token limits are hit quickly.
  • Wider adoption may come from cheaper models that are “good enough,” not only top-tier models.
• Tooling behaviors mentioned:
  • Letting the model handle more of the workflow end-to-end, including Git interactions (commit/push/rebase).
  • Keeping the developer’s workflow unified in a single interface to reduce context switching.

Local vs hosted models (hardware optimism, timeline uncertainty)

• Open-source/local models are desired, but the speaker doesn’t expect truly common local deployment soon.
• Hope (speculative timeline):
  • within a few years (or possibly never), people can buy a ~$15k–$30k machine to run local models faster
  • analogy: like video editors/photographers investing in equipment

Education changes: from watching tutorials to using agents

• Training is shifting:
  • Instead of long courses, learners can ask an agent to explain and guide quickly.
• Skill emphasis is changing:
  • less syntax memorization
  • more high-level problem solving and system design thinking
• Concern:
  • agents can help, but may not fully replace the ability to solve entirely new problems independently.

When AI isn’t enough, deep technical knowledge still matters

Examples show developers still succeed by applying real engineering skills:

• Reverse engineering Bluetooth mesh protocols (using an ESP32, writing C firmware) to control smart lights via home assistant.
• Building automation/plugins where the model can’t handle unknowns, then coding the integration manually.
• Improving video-editing workflows (e.g., plugins that remove silences), even if the speaker doesn’t understand every underlying tool orchestration detail.

Conferences as a differentiator

In-person conferences are valued mainly for:

• Networking and hallway conversations
• Meeting diverse practitioners (including surprising domains like elevator software)

They contrast this with online conferences, which lacked feedback and relationship-building.

Personal branding becomes more important

• Since code output is easier to generate, standing out requires communicating expertise and impact.
• Suggested channels/formats:
  • social media presence
  • blog posts/videos/podcasts
  • community engagement
• Recommendation:
  • choose a medium and avoid spreading thin across everything
• Content discovery note:
  • algorithms control reach on many platforms, so newsletters can be a more reliable direct channel.

Caution about AI-generated content

• Automatically writing emails with AI to get attention is viewed negatively (insulting/inauthentic).
• AI can help with editing or automating administrative tasks, such as:
  • posting conference pages
  • aggregating links
  • trimming videos
• But not for the speaker’s core original thinking or slide narrative.

Main speakers/sources

• Wes Bos (primary speaker; referenced throughout as “West Bos”)
• Other participants (unspecified; discussed as a group / interviewer-style prompts, but not individually named in the subtitles)