Shikhar Singh

Founder, KedByte Technologies. I write about software correctness.

[Shikhar Singh @_singhshikhar]

Free, no signup: https://t.co/VjNg9ZAiMG

[Shikhar Singh @_singhshikhar]

Chapter 3 gave Cat 6 twist rates as one to three twists per centimetre.

The agent checking it pulled US patent 7,015,397 B2 — Belden, Table 2, lay lengths of 0.394, 0.809, 0.551, 0.898 inches.

Three twists/cm implies a 3mm lay. Off by an order of magnitude.

[Shikhar Singh @_singhshikhar]

"An AI wrote it" isn't the interesting question.

The interesting question is whether anything checked it, and what happened when it did.

[Shikhar Singh @_singhshikhar]

Every marker had to resolve to a primary source, or the claim was deleted. Final count: zero.

The pass found the book wrong about my own subject.

[Shikhar Singh @_singhshikhar]

Book three ran that version.

Some of its corrections may have vanished without anyone noticing.

[Shikhar Singh @_singhshikhar]

One agent per chapter, not one per claim — I got that wrong the first time.

The obvious design gives every claim its own agent. That meant six agents editing ch55.md at once. Concurrent in-place edits silently lose each other.

[Shikhar Singh @_singhshikhar]

Then a second pass: 43 agents, one per chapter.

Each was told it was the sole writer of its file, and required to run grep -c UNVERIFIED on that file and report the count before it could finish.

[Shikhar Singh @_singhshikhar]

Here's what makes that defensible, or not.

Writers were forbidden from guessing. Anything a writer couldn't confirm had to be marked [UNVERIFIED] in place rather than smoothed over.

Book four came out of drafting with 84 markers across 43 chapters.

[Shikhar Singh @_singhshikhar]

No human wrote these chapters. No human read them all.

That's the uncomfortable part, so I'll start there.

The KedByte Library is 4,597 pages of free technical books. Agents wrote them. I designed the pipeline and the constraints.

[Marc Andreessen 🇺🇸 @pmarca]

The new American credentialing system: Degree from MIT < Drop out of MIT before graduating < Drop out of MIT on the first day < Accepted to MIT but never go for even a day < Drop out of MIT because AI is going to kill us all.

[Martti Malmi @marttimalmi]

My email correspondence with Satoshi in 2009-2011: https://t.co/jyoX8gXckp