52 posts
cfsolve claims to solve Cloudflare Turnstile with no headless browser. Reading its source offline shows it replays recordings of a real browser, and what that means for a site owner.
GitHub, npm and Open VSX all skip 2.1.313, yet Cursor says a malicious build with that number was served for 11 hours. What the reports agree on, where they conflict, and a read-only check of one PC.
A Hacker News post said Claude Code burns 15 to 20 percent of tokens on a status line. The status line is innocent. I reproduced the real culprit, an agent summary, and measured what it costs.
I modeled the KV-cache page reuse behind Cloudflare's GLM-5.3 leak. A fresh pool leaks nothing, a warm one hands over other customers' pages, and every reply still returns HTTP 200.
Node, .NET, curl and others each carry their own Brotli decoder. I counted the places on one Windows PC and tested which clients decode Brotli they never asked for.
I unpacked Python images layer by layer and rebuilt them with common Dockerfile habits. A later rm saved nothing on the wire, and half of pandas in the image was its own test suite.
I ran wall-clock, Lamport, hybrid and vector clocks through one simulated workload. Logical clocks removed the skew mistakes, yet 0.5% of writes still vanished. Only vector clocks kept them all.
I sent 200 requests across eleven kinds of hand-off with the OpenTelemetry Python SDK. Five lost every trace, batching lost 90%, and the failures you most want to trace were the ones left behind.
I simulated Google SRE Workbook burn-rate paging rules on a 99.9% SLO. They bound the budget lost before a page, but below about 900 requests a day they page for nearly every error.
On 1,536 LoCoMo questions, tagging each fact with its subject lifted search from 54.4% to 59.5%. Filtering the raw turns by speaker did the same, for free.
On 1,536 LoCoMo questions, Hermes-style FTS5 session search found the answer turn in its top 10 for 54%, ahead of a small embedding model. Nearly every hit came from an OR retry.
I sent three AI user-agent strings to 66 sites and compared robots.txt with what the server did. Sephora allows them on paper and returns 403. BBC denies them and serves the page.
Pele claims to be 500x faster and 6x smaller than Mermaid.js. I ran both on 39 diagrams from this blog: the speed holds at about 100x, the size only with an opt-in import.
Apple says decoding is memory-bandwidth bound. I checked that on my own machine, then fit Apple's tables: a fixed 9 to 16 ms per token explains why more Macs help less each time.
The BIPM made UTC continuous from 20 May 2027, citing the risk of a negative leap second. I fed 23:59:60 to six runtimes, and none turned it into a moment of its own.
Go 1.27 now drains an unread HTTP/1 response body so the connection can be reused. I counted dials on Go 1.26 and 1.27, and found where the fix stops working.
Anthropic replaced git for its monorepo. I measured what shallow, partial and sparse checkouts do to a 343,502-commit repository, and where stock git still runs out.
I graded the REA reverse-engineering tool against four tiny Electron apps with known answers. It got the main process right, missed script written inside HTML, and returned 125k tokens.
claude plugin validate rejects every way I tried to hide a mods API call, and the runtime blocks raw JavaScript escapes. But its output is a capability list, not a verdict.
A popular note says tiny Git commits cost kilobytes and big ones about 10% of the file. I repeated the tests: the kilobytes are filesystem blocks, and the 10% is only the first copy.
A blog post says WSL 3 beats WSL 2 by 5 to 60%. I upgraded my own machine from WSL 2.4.12 and repeated its tests: syscalls and first-touch memory got slower, not faster.
Why does Postgres skip indexes? On a 5M-row table the default planner picked a bitmap scan up to 1.85x slower than a sequential scan, and physical column order mattered more than selectivity.
I measured the performance claims behind 'push ifs up and fors down' in C. Hoisting an if is free at -O3, batching a lock saves 44x, and collecting into a batch made one loop slower.
I re-ran Igor Ostrovsky's 2010 processor cache experiments in C on a 2026 laptop CPU. The cache cliffs survived. The famous 'skipping is free' result did not.
DEDA decodes the hidden tracking dots laser printers stamp on pages. I generated my own pattern, decoded it, then corrupted one dot and watched the checksum fail while the decoder answered anyway.
Red Hat benchmarked Jev against 8 other guardrail models and found a local classifier beats it on speed by 6x. I tried to reproduce that result myself, on my own task.
liao.gg argued vector-based agent memory cannot tell correct from stale, without a benchmark. I tested the claim on this blog's own corrected numbers, with real embeddings.
Cloudflare launched an OHTTP gateway that hides IPs from app servers. The RFC and Cloudflare both say privacy depends on relay and gateway never sharing an operator, a rule no cryptography enforces.
Git is replacing SHA-1 with SHA-256 for every commit and file it tracks. A GitHub co-founder argues the attacks motivating the switch do not threaten how Git actually works.
Netlify replaced the lightweight sandbox running its edge functions with full virtual machines, and median latency dropped from about 30ms to 5 to 6ms.
An open-source benchmark set out to catch Claude quietly getting worse after launch. Its own validation test, swapping in a different model, came back statistically indistinguishable from noise.
The same classification task, run through Jev and Gemini 3.8 Flash with real API calls: identical accuracy, four to five times the latency, eight to thirty-four times the cost.
Ten processes update the same git ref at once. Git returns success for all ten, and by default it hides which nine of them never actually stuck.
DeepSeek's DSec runs 380,000 concurrent AI-agent sandboxes a day. Not by caching one big image, but by admitting most of a sandbox can never be reused, and building around that instead.
Go 1.27 ships a portable SIMD API with zero published benchmarks. Measured on real hardware: a plain loop wins below 256 elements, then SIMD holds a steady ~2x, not the 8x its lane count implies.
Gambit Security recovered the staging server behind a three-agent breach of 27 companies. The same ungated pipeline destroyed its own stolen data — the exact failure mode this blog already mapped.
Docker Sandboxes isolates AI agents in microVMs, not containers. Two bugs patched in the same release both broke it at the one place the guest is allowed to talk back to the host.
Three companies shipped a standalone agent browser. One matched Windows to macOS day one. One took 83 days. One never got there, shut down, and retreated into an extension.
I gave one reviewer agent nothing and another a 210-word handoff note. Cognition says blind works better. The one with the note caught the bug that actually mattered.
Two subagents build the two halves of a receipt module in total isolation. One half won't compile. The fixed version compiles, runs, looks correct — and is silently wrong by a cent.
Two subagents with identical toolboxes and different system prompts boot 104 tokens apart. Add four tools and the gap is 17,252. The boot bill is the toolbox, and two tools are most of it.
A subagent that runs one wc -w burns 34,172 tokens. Split the work across five and the bill is 4.85x for a 1.79x speedup - the multiplier is a boot sequence, paid again every time you fan out.
Container isolation for a fleet of agents costs 48 seconds once and 4.1 kB per agent after — until the bind mount that makes it convenient costs 465 MB each and hands the sharing back.
An ADE gives every agent its own git worktree. Here that isolates 44 files and leaves 23,199 shared — and the usual fix is a symlink, which is the opposite of isolation.
Umami is MIT-licensed, so the identification path is readable. Where, who and when turn out to be a CDN header, a hash that expires on the 1st, and a bucket quantized to the clock hour.
The web's oldest isolation guarantee is a string comparison. Put an agent inside the browser and it becomes a model's ability to resist a prompt injection.
A multi-turn agent replays its entire history on every request. Caching, context editing and compaction each attack a different part of that bill — and one of them, at its defaults, made it worse.
Prometheus reports fractional counts for counters that only ever increment by one. The reason is a specific extrapolation rule, and knowing it tells you exactly when the number can be trusted.
Hermes Agent says it improves its own skills. In the source, that loop is a tool-call counter, a forked LLM review told to be active, and a curator that archives what nobody uses.
Rendering is a separate, queued, stateless stage between crawling and indexing. Almost everything confusing about JavaScript and search comes from ignoring that.
Prefill and decode have opposite performance characteristics, and almost everything confusing about LLM latency, cost and context limits falls out of that one split.
Cardinality does not creep up on you — it multiplies. Here is the arithmetic behind an exploding metrics bill, the four places the multiplication hides, and what actually shrinks it.