Overview
The talk investigates a practical limit that governs how complex AI skills and system prompts can become: how many simultaneous instructions can a model reliably follow? Reproducing the IFScale benchmark confirmed that frontier models from roughly a year earlier began degrading around 200–300 simple constraints. When the same test was applied to newer models, the original 500-instruction benchmark proved too easy, forcing the experiment to expand toward 10,000 required words. Depending on the model, degradation now begins around 750 to 5,000 or more constraints, representing an approximately tenfold improvement at the frontier. However, greater capacity does not imply dependable behavior. DeepSeek quietly forgets constraints, Claude may trigger safety refusals, Gemini can exhaust its reasoning budget, and GPT may produce a polished partial answer before abandoning the task. Additional research suggests that long-context reasoning can deteriorate before the context-window limit and that simple changes in instruction wording or order can radically alter compliance. The practical conclusion is that developers no longer need to aggressively compress or fragment every large skill file. Instead, they must balance completeness against latency and cost, choose models according to their failure patterns, and continuously evaluate complete outputs rather than assuming that confident-looking responses followed every rule.
Sections
Strategic Implications
Higher instruction capacity changes prompt architecture, but reliability remains probabilistic.
- Agent decomposition should now be justified by genuine specialization, tool boundaries, or workflow needs—not automatically by fear that one prompt cannot hold several hundred rules.
- As visible forgetting becomes less common, superficially convincing partial compliance may become the more consequential production risk.
- Model selection should consider failure observability as well as benchmark accuracy: an explicit refusal can be operationally safer than an undetected partial answer.
- Prompt-size guidance should be periodically retested because capability assumptions can become obsolete within roughly six months.
Benchmark and Experimental Details
The concrete methodology, measurements, and reported model behaviors.
- IFScale measures instruction density, n, by requiring a generated business report to contain a specified number of exact words; accuracy is the percentage of required words present.
- The replication tested GPT 4.1, Claude Sonnet 4, and Gemini 2.5 Pro because they were the surviving API-accessible models from the original benchmark set.
- The newer comparison tested GPT 5.5, Claude Opus 4.7, Gemini 3.1 Pro, and DeepSeek V4 Pro using the same prompt and keyword task.
- Because all newer models scored 100% on the original test, the benchmark was expanded from 500 required words to 1,000, 2,000, and ultimately a 10,000-word vocabulary.
- DeepSeek reportedly began forgetting around 750 rules and omitted nearly half by 2,000, while GPT 5.5 reportedly maintained 99% accuracy through 5,000 rules.
- The complete experiment reportedly involved approximately 2,300 calls across seven models and cost $29.
Risks and Limitations
Factors that prevent benchmark capacity from translating directly into production reliability.
- Keyword inclusion is only a proxy for realistic instructions and does not test complex reasoning, conditional logic, or conflict resolution.
- A model may return a polished but incomplete response that appears successful unless the entire output is inspected.
- Long inputs may suffer context rot before reaching the advertised context-window limit.
- Small changes in wording or instruction order may cause large changes in compliance.
- Large prompts increase inference cost and latency even when they remain within the model's instruction capacity.
- Safety classifiers may reject benign prompts when large vocabularies contain suspicious combinations of words.
Mentioned Research and Tools
Benchmarks, research projects, and evaluation systems referenced in the talk.
- IFScale — the original benchmark used to measure compliance with increasing numbers of discrete instructions.
- Chroma context-rot research — a reported study across 18 models examining accuracy degradation on long inputs.
- Revisiting the Reliability of Language Models in Instruction Following — a reported paper testing instruction-following stability across 46 models.
- FireBench, CCR-Bench, and GuideBench — newer benchmarks intended to measure compliance with numerous realistic and messy constraints.
- Arize AI — the speaker's company, described as providing LLM-based output monitoring and evaluations.
- OpenAI safety filter — used during the experiment to remove vocabulary likely to trigger Claude's safety classifier.