Jonathan Blow on ChatGPT and what real programming means

Technological concepts & product/tool analysis discussed

Rendering / engine optimization (beam wrapping on large levels)

• The speaker discusses a bug where beam segments fail to render when wrapping on very large levels.
• They describe an “AI-proposed” fix: force all eight offsets whenever a beam segment exceeds half of a level dimension.
• The speaker criticizes this as overkill (a blunt fix), noting potential problems such as:
  • Overdraw and extra batches
  • Potential “popping” right at the wrap threshold
• They contrast it with a more deterministic approach:
  • Compute per-axis wrap crossings
  • Draw only the necessary wrapped copies
  • Aim for at most 4 copies (though the speaker argues the specific AI claims are incorrect)
• Overall evaluation: the proposed hysteresis/threshold-avoidance fixes are described as “trash” / “shitty.” The speaker argues correct rendering should avoid pops rather than paper over the issue with hysteresis.

Debugging via AI suggestions for game/editor runtime errors (fog volumes)

• The speaker describes an engine/editor error: more than one fog volume is active on the same floor after sky level fast travel.
• They list example causes AI suggested, including:
  • A global fog AABB overlapping the sky zone
  • A floor resolver that doesn’t clear the previous floor’s fog
  • The sky zone inheriting the same floor ID
  • A spawner failing to unregister fog properly
  • The error message itself might be incorrect
• The speaker states the AI’s guesses were largely not true and recommends verification via instrumentation:
  • Run a “dump fog volumes” command to log active volumes and find duplicates
• AI proposes “safe fixes,” such as:
  • Making the fog system exclusive per floor
  • A “self-healing” function idea like fog system dot enforce exclusive for floor
  • Variants like disabling other volumes
• The speaker criticizes the AI’s reliance on hysteresis in floor detection as conceptually nonsensical and harmful—asking: “How does anyone ever get anything done with this?”

How ChatGPT-like systems work (and why the output is unreliable)

• The speaker argues the system’s core limitation is that it does not truly understand the codebase/game logic:
  • Even with more context, it may better mimic code patterns, but it still “bullshits” and generates nonsensical suggestions.
  • More context may improve formatting and coherence, but does not create true understanding
• They emphasize a view of programming correctness:
  • The solution space is such that most programs are “complete trash”
  • You can’t reliably “tune” a near-miss into correctness with small parameter changes (unlike an analog dial)
• Conclusion:
  • AI code suggestions can be useful for brainstorming/learning, especially for beginners
  • But they are not reliable for producing robust, correct production software
• They liken AI output to search:
  • It can provide ideas similarly to what you’d find by searching, with some convenience
• They argue people get misled because formatted, compiled-looking code can appear to be evidence of correctness.

Developer workflow / version control discussion (GitHub bans & liability)

• A side discussion references a company banning someone, with context implied to possibly involve sexual/political content or profile-related issues.
• The debate includes whether hosted platforms avoid giving reasons to reduce:
  • liability
  • legal/PR risk
• The speaker advises:
  • Don’t trust data fully on hosted services
  • Maintain backups

General workflow references

• Mentions the mindset: “ship the core early, aesthetics later.”
• References rapid prototyping concepts, such as:
  • Prove the idea
  • Use a limited time window
  • Avoid polish
• This frames a development approach that contrasts with AI’s unreliable “fixes.”

Key evaluations / takeaways (as stated)

• AI-suggested rendering fixes: described as incorrect, overcomplicated, or unsafe; hysteresis-based threshold behavior is rejected.
• AI-suggested debugging fixes: AI guesses are often wrong; verification requires engine instrumentation (e.g., dumping active fog volumes).
• Programming correctness: framed as needing precision/exactness; AI output doesn’t reliably produce it.
• Best use of AI: learning/brainstorming multiple approaches, not generating robust production code.

Main speakers / sources

• Jonathan Blow (primary speaker)
• References ChatGPT / GPT-like AI (as the tool being critiqued)
• Mentions version control systems: GitHub, plus Perforce, Mercurial, SVN