Many companies are still paying an AI “learning tax.”
Teams are sending duplicate requests, carrying enormous context windows between interactions, and using expensive agentic workflows for tasks that deterministic software could handle better.
The answer is not simply tighter
New podcast drop!
@BrandonMathis talks with @cmgriffing from @GitKraken about #GitBench, a benchmark testing how well LLMs handle real Git tasks such as merge conflicts, worktrees, and commands where a bad decision can be costly.
The key takeaway: match the model to the job,
#AI may write more implementation code, but engineering still owns the system around it.
Next Friday at @zurichjs, I’ll explore specifications, orchestration, validation, and frameworks for autonomous engineering.
conf.zurichjs.com