Posts
68 articlespage 1 of 7RSS feed all topics
2026
10 articles on this page- The Day Claude Decided to Delete My Whole Codebase On 5 October 2026, a unit test run by a coding subagent recursively deleted every checkout under my projects directory in well under two minutes. No model “decided” anything. An orchestrating model wrote a guard rule of the form “anything outside the repository is allowed”, a test handed that guard the repository’s parent directory, and nothing between the test and the filesystem said no. This is the incident, the recovery, and the four layers that now stand where there were none.
- Token Saving as Noise Reduction Token saving in agent systems is often framed as an efficiency concern. A better framing is signal-to-noise control: every unnecessary token competes for attention, increases inference cost, and makes the next decision less precise. Current work on context compression, KV-cache compression, and long-running interactions points in the same direction.
- GitHub Stars Are for Sale, and the Real Problem Is Not Vanity Buying GitHub stars sounds pathetic enough to be a joke. The joke misses the interesting part. Stars are a public trust signal on a code-hosting platform, and there is an ordinary, low-friction market for faking them. The important question is not whether inflated stars reliably produce downloads. It is why counterfeit popularity is so cheap to buy, and why that becomes a security problem once fake credibility attaches to malicious repositories.
- There Is an App for That — Until There Isn’t Digital health applications and psychotherapy follow different capacity, evidence, and reimbursement paths. Baumol’s cost disease explains one pressure on labour-intensive care; it does not explain a particular German fee decision or establish that an app substitutes for a therapist.
- The Model Has No Seahorse: Vocabulary Gaps and What They Reveal About LLMs There is no seahorse emoji in Unicode. A language model may nevertheless claim there is one or emit a plausible substitute. That failure is real, but the output alone does not tell us whether tokenisation, training data, prompting, decoding, or verification caused it.
- Oppenheimer Didn’t Have an Acceptable Use Policy Anthropic has drawn a public line on military use of its models. The physics community spent the better part of the twentieth century working out what it means to draw that line after you have already built the thing. As a physicist watching this unfold, I find the parallels clarifying and the differences more unsettling than the parallels.
- More Context Is Not Always Better The intuition that feeding a language model more information improves its outputs is wrong often enough to matter. Here is why, and what to do about it.
- If You Think This Is Written by AI, You Are Both Right and Wrong AI detectors can flag human writing, including historical and highly formulaic prose. Evidence exists for a false-positive risk affecting some non-native English writers; claims about autistic or physics-trained writers need direct validation. The detector score is not authorship evidence.
- Two Routes to the Expansion Rate: The Hubble Tension Near 5σ A local distance-ladder measurement and the value inferred from the cosmic microwave background under ΛCDM disagree by about 4.9σ for one widely quoted comparison. JWST has tested an important Cepheid-crowding systematic, while DESI finds a dataset-dependent preference for evolving dark energy that does not by itself resolve the Hubble tension. The remaining possibilities include unrecognised systematics and physics beyond the standard cosmological model.
- Car Wash, Part Three: The AI Said Walk In early 2026, variants of a prompt about a nearby car wash spread online after some assistants recommended walking instead of taking the car. This is neither the tokenisation failure from the strawberry post nor the grounding failure from the rainy-day post. It is a pragmatic inference failure: the response followed the travel wording but missed the likely purpose of the trip, assigning its advice to the wrong interpretation of the question. A third and more subtle failure mode, with Grice as the theoretical handle.