The Nightly Build.
Saturday, July 25, 202658 min read
Technical Practice · Build From Scratch

Byte-pair encoding hides the letters inside every common word a language model reads

A merge-and-count tokenizer trained from scratch on 617 bytes of text reproduces the same blind spot as GPT-2's real tokenizer, and turns the integer 2249 into four digits while leaving 2250 in three.

15 min read13 sources
Daily Reading · Paper of the Day

A single one-dimensional problem breaks Adam's convergence proof

The fix that made Adam durable was Loshchilov and Hutter's 2019 repair to weight decay, arriving a year after Reddi et al.'s counterexample had already sent it to the worst point in its own one-dimensional search space.

12 min read9 sources
Investing · Company Analysis

Vertiv's $15 billion backlog already exceeds its guided 2026 revenue

Organic orders grew 252% in the fourth quarter and the stock ran from $124 to a $376 peak on the same story, a backlog that has never yet been tested by a slowdown.

9 min read13 sources
daily-news

CENTCOM goes quiet on Iran strikes after 13 consecutive nights

The gap surfaces first in the command's own release archive, not in any statement.

9 min read30 sources
Current Affairs · Opinion

Congress Already Wrote the Rule This Deal Breaks

The Trump administration used a National Defense Authorization Act waiver to bypass the Additional Protocol and no-enrichment standard Congress set in 2019 for exactly this kind of deal.

9 min read16 sources
Daily Reading · Word of the Day

A drought and a coal mine demanded a new word

Glenn Albrecht, an environmental philosopher, coined solastalgia in 2003 for Hunter Valley residents watching mines and drought remake land they had never left, built deliberately against nostalgia's ache for a home left behind.

4 min read10 sources