Andrew Ng's claim that 90% of agent deployments fail is still being cited everywhere, but what our editorial team thinks is more worth noting this week is Simon Willison using GPT-5.6's voice mode to "talk to himself" through a complete prototype design: 1,000 document revisions originally totalling 20.4MB compressed down to just 80.3KB — roughly a 254x compression ratio.
What This Is
Willison is a well-known independent UK tech blogger with a long focus on data tools. While on a walk, he revisited an old problem: how to efficiently store "every historical version" of a text that keeps getting edited. The conventional approach is one row per version in a database — for any reasonably long document, each edit costs tens of KB. His proposal is straightforward: stuff all historical versions into a JSON string array, compress the whole thing with zstd (an algorithm with better compression ratios than traditional zip), and pair it with a timestamp array column. GPT-5.6 helped him run 38 minutes of experimental code to validate the approach.
Technically, this isn't complex — zstd's compression of repetitive text is well-established. Two things are worth noting: first, the segmentation strategy (split every 128 revisions or every 3MB) avoids having to decompress the entire array on every edit; second, he used ChatGPT's new voice mode to talk through the design while walking, then switched to text mode to have the model write the code.
Industry View
Supporters see this as an early template for "natural language programming" — humans handle divergent thinking and intent articulation, models handle implementation details. Willison himself describes the whole process as "chatting with a pair programmer who never gets tired of you," a use case that wasn't viable last year (voice-to-text latency and accuracy weren't good enough).
Skepticism is equally present. Compression isn't new, and zstd isn't new either — what's genuinely scarce in this prototype is the engineering closed loop of "one person intuitively conceiving it and validating it end-to-end within two hours." This means lower barriers for independent developers, but the opposite for enterprise IT — the gap between "any prototype anyone can build" and "a system that holds up in production" is bridged by concurrency, disaster recovery, permissions, and auditing — the messy work that GPT can't write for you. Another risk: Willison ran simulated data; in real scenarios, text edit locality may be much worse, and the compression ratio may not reproduce.
Impact on Regular People
For enterprise IT: Business units will soon come with prototypes built in two days à la Willison's, asking for budget to put them into production — IT departments need to think ahead about how to bridge the gap from prototype to production, rather than simply saying "no."
For individual careers: The ability to "articulate vague ideas clearly via voice" will shift from a nice-to-have to a baseline skill — especially for product managers, consultants, and others whose work has higher thinking density than typing density.
For the consumer market: ChatGPT's voice mode transcription quality has crossed the threshold to "usable as a work tool." Overseas, podcasters have already started using similar workflows for topic selection; the Chinese-language equivalent hasn't appeared yet — a product gap domestic large-model companies haven't filled.