Large language models, prompting, agents and evaluation.
35 items
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.
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.
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.
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.
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.
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.
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.
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.
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.
Prefill and decode have opposite performance characteristics, and almost everything confusing about LLM latency, cost and context limits falls out of that one split.
ego lite and Aside are both browsers built for AI agents, but one has no agent of its own and the other ships with one. A guide to what each is, with the same page-reading test run on both.
Index and reading guide for an eight-part series that goes from a single neuron to the transformer, with one NumPy measurement per idea, the surprises along the way, and how to rerun everything.
Turn "how wrong is the network" into one number, then lower it by small steps downhill. Measured on a real digit network: five starts, a learning-rate sweep, and what a gradient entry means.
How a network gets all 13,002 gradients from one backward pass instead of 13,002 experiments, checked against brute force, plus why training takes small noisy steps instead of exact ones.
Run the learning loop for real and test the usual stories: sigmoid versus ReLU (the shallow result surprised me), deep networks that cannot learn, noise read as digits, and memorized random labels.
How text becomes vectors and vectors become next-word probabilities, with directions measured in a real embedding space, softmax and temperature on real counts, and GPT-3's first parameter lines.
How attention lets each token look at the others, built in NumPy, with a leak test for the mask, a tiny model trained to 100 percent by hand-written backprop, and what the n by n pattern costs.
A feed-forward block as a table of if-then rules, built by hand and trained, with the ReLU AND gate measured, how many neurons it takes to memorize 900 facts, and the finished GPT-3 parameter ledger.
Why a 12,288-dimensional space can hold far more than 12,288 features: the geometry of almost-perpendicular vectors, a read-back test, and a tiny network that packs 20 features into 5 dimensions.
Start from the idea that a language model is a function from text to next-word probabilities, then build the simplest neural network by hand and count its 13,002 parameters.
Augmented engineering has no standard definition, while agentic engineering does. This guide sorts out the usages, compares the two working styles side by side, and says when each one fits.
Claude Duet is Anthropic's proactive personal agent. This guide explains what it does, how it asks for approval, and how it differs from Meta Muse and OpenAI dots, using the vendors' own pages.
A beginner-friendly walk through what happens between pressing send and seeing the first word, checked with a 135M-parameter NumPy model and real tokenizer counts.