Coding agents are very good at building what you ask for.
The gap between what you specify and what you expect is where the agent starts making assumptions.
More assumptions → more rework → more iterations → more tokens.
The clearer the specification, the closer the output.
🤖 Made with AI
1
165
AI systems are increasingly becoming distributed systems with probabilistic components.
The hard part is no longer the model.
It’s:
- orchestration
- reliability
- memory/state
- evaluation
- observability
- human oversight
LLM calls are easy.
Production systems are not.
🤖 Made with AI
1
97
Your AI demo works. Your AI system won't. The gap isn't the model. It's everything around it. Evaluation. Retrieval. Drift. Observability. Structured outputs. The model is the easiest part to replace. The system is not.
Read the full article here 👇
2
1
148
Paper Summary:
1. LLM hallucinations stem from the fact that even with perfect (error-free) training data, the cross-entropy objective forces the model to guess plausible outputs — binary classification errors are inevitable under epistemic uncertainty.
2. In post-training, models are incentivised to “guess” rather than abstain (e.g. say “I don’t know”) because benchmarks use hard accuracy (0-1) scoring, rewarding confident but possibly wrong answers.
3. The fix: change evaluation design — explicitly reward uncertainty or abstention (via confidence thresholds) instead of punishing it, so models are not forced into hallucinating to win.
Re-reading the article "How to build effective agents". anthropic.com/engineering/bu…
5
394


















