When Blog Corrections Are Worth Recording

Every long-running blog accumulates errata, but most of it is noise: typos, off-by-one dates, and small slips that don't change the argument. This post is not about those. It is about the subset of errors that are genuinely instructive—mistakes that reveal something about how technical arguments go wrong, and why even careful writers can publish them.

The list below is not exhaustive, nor are the categories clean. The errors are grouped by the kind of failure they represent, from insufficient verification to outright blunders. Several of these are from the first few months of blogging in 2013, which is arguably a good sign, but the more recent entries show that the failure modes do not disappear with experience.

Failures of Verification

Two entries stand out as cases where the author simply did not check the numbers closely enough.

  • 2016, “Building a search engine isn't trivial”: The overall thesis holds, but one piece of supporting evidence came from figures given verbally by someone who had worked on a search engine. When the author later measured real numbers, the figures turned out to be off by multiple orders of magnitude.
  • 2022, “Futurist predictions”: This is a more subtle case. The author misread a summary of a sensitivity analysis, even though he initially agreed with the correct interpretation on first reading. A second reading produced the wrong conclusion, which then made it into the post. The errors were introduced in a last-minute edit made under time pressure—a decision that should have instead led to a delayed publication. The broader point of the post did not hinge on the exact numbers, but the process was bad: publishing with known uncertainty, without flagging it, in a post that was itself about the dangers of misreading evidence.

Simple Arithmetic Blunders

Two errors fall into this camp, though they differ in severity.

  • 2015, “Checking out Butler Lampson's review of what worked in CS”: The claim that capabilities were a “no” in 2015 was wrong given their demonstrated effectiveness on mobile platforms. The author calls this a blunder—something that was obviously incorrect even at the time, not a case of insufficient depth of thought.
  • 2024, “Diseconomies of scale”: A division mistake in the arithmetic produced a percentage error of multiple orders of magnitude. It was caught by a reader within hours and fixed immediately, but it remains a silly error.

Naive Assumptions

Some errors are less about facts and more about worldview. Two posts suffer from this.

  • 2013, “The institutional knowledge and culture that create excellence”: Written before the author had worked in software, this post assumed that software companies succeed with young teams because the work is easier. In hindsight, the difference is not that successful software companies produce reliably good software—they frequently do not. It is that product-market fit and network effects matter so much that poor quality often does not sink a product.
  • 2015, “Why people don't read citations”: The author found it mysterious that type system advocates would cite non-existent evidence. Unlike typical citation errors, where an incorrect result gets passed along, this seemed to be a case of inventing the existence of evidence entirely. In retrospect, the mystery was a sign of naivete: people can simply fabricate the notion that evidence supports their belief, and that fabrication can propagate without any actual evidence to anchor it.

Ideas That Needed More Thought

  • 2013, “Using random algorithms for HN story selection”: This was a proposal, not a claim, so it was not technically wrong. But the author feels he should have considered human-in-the-loop solutions. He avoided them because of Digg’s well-known backlash over human editorial influence. In hindsight, that outcome was not inevitable, and the possibility of a backlash-free implementation was predictable with more thought.

Explanations That Were Clearer to the Author

Two posts suffered not from wrong facts but from poor presentation.

  • 2013, “How data alignment impacts memory latency”: The main plots used a ratio of latencies, an indirection that confused many readers.
  • 2017, “It is easy to achieve 95%-ile performance”: The most common objection was not a technical disagreement but a failure of calibration. Readers argued that reaching a 2200 chess rating or playing in the NBA is hard, missing the point about percentiles. The author should have provided concrete examples of what various performance levels actually look like. One critic claimed a 2200 rating is the 95th percentile for anyone who has ever played chess online—a figure off by perhaps four orders of magnitude.

Erroneous Predictions, Corrected by Reality

  • 2015, “Blog monetization”: The author grossly underestimated his potential Patreon earnings by benchmarking against a few other writers. He expected to earn less than each of them, which turned out to be wrong. The key lesson was that long-form, technical writing monetizes better than expected. That realization is now so widely shared that it is hard to recall it was once non-obvious—but the fact that no one built a tool like Substack until founders did suggests the error was not unique to this author.

Thanks to Anja Boskovic and Ville Sundberg for comments, corrections, and discussion.