A networking expert tests ChatGPT (GPT-3.5 and GPT-4) and the Phind AI search engine by asking why iSCSI has its own application-level checksum when Ethernet already provides CRC. GPT-3.5 produces circular arguments and hallucinations, touching on the right ideas but never landing on the real answer. GPT-4 performs noticeably better, mentioning silent data corruption and end-to-end integrity, though it still misses the key point about the weakness of the TCP checksum. Phind gives the most accurate answer — that Ethernet CRC only protects a single L2 segment, layer-3 switches recalculate it, and the TCP checksum is too weak for critical storage data — but pads the response with unrequested information. The author also catches Phind hallucinating an entire explanation of the fictional 'FCoTR' (Fibre Channel over Token Ring) technology, illustrating the GIGO problem with AI-augmented search. The real answer: iSCSI CRC32C is needed because Ethernet CRC protects only one L2 hop, and intermediate nodes rely on a weak 16-bit TCP checksum that can miss corruption.
Table of contents
ChatGPT using GPT-3.5Everything Is Better with GPT-4Phind Wins the DayRevision HistoryQuestions this post answers
Why does iSCSI have its own application-level checksum if Ethernet already has CRC?
iSCSI needs its own checksum because Ethernet CRC only protects a single layer-2 segment. When traffic passes through layer-3 switches, those devices recalculate the CRC, so it no longer provides end-to-end protection. The TCP checksum is only a weak 16-bit sum and is not considered adequate for critical storage data. iSCSI uses CRC32C, which is hardware-accelerated on modern CPUs, to provide true end-to-end data integrity. Engineers running iSCSI in production track protocol-level integrity trade-offs like these on daily.dev.
Does GPT-4 give better answers than GPT-3.5 for technical networking questions?
GPT-4 performs meaningfully better than GPT-3.5 on technical networking questions. In a direct comparison on iSCSI checksum rationale, GPT-3.5 produced circular arguments and missed the core point, while GPT-4 correctly identified silent data corruption and end-to-end integrity as key reasons. GPT-4 still missed the weakness of the TCP checksum and requires fact-checking, but earned a B+/A- versus GPT-3.5's failing grade. Developers choosing between AI coding and research tools weigh capability gaps like this on daily.dev.