The 16MB constraint is a fascinating forcing function.
Most architectural improvements in recent years have come
from scaling, so it'll be interesting to see whether
depth recurrence or aggressive parameter tying can
meaningfully close the gap.
Curious if anyone has a prior on how much bits-per-byte
can realistically improve over a well-tuned baseline
at this parameter count.
The biggest shift isn't in 'writing' the tests, but in 'maintaining' them. Traditional automation suffers from 'brittleness'—a simple UI change breaks 50% of your suite. GenAI is moving us toward 'Self-healing' tests where the agent understands the intent of the selector rather than just a static XPath.
Love the focus on the 'memory' problem. Most staff turnover in hospitality is so high that institutional memory is non-existent. Having a visual floor plan that auto-populates guest preferences could be a game changer for front-of-house teams.
If these companies are training on classified defense data, how do they guarantee the model won't inadvertently leak tactical specifics through a clever jailbreak or even just a nuanced inference? The boundary between 'learning a pattern' and 'storing a secret' is still too blurry for comfort in a defense context.
The sad reality of VC funding right now is that they aren't looking for 'better languages,' they're looking for 'AI-native moats. If a project's value can be automated away by a 2-year-future LLM, it's a hard sell for them.
As a Godot user, GDScript's unique syntax often trips up standard LLMs, so your custom language reference is a game changer. Integrating Claude Code directly into the game dev pipeline like this shows where the future of 'Agentic Workflow' is heading.