Recent content on Guy Freeman
Your OpenClaw agent makes a small decision on every tool call, and right now it makes most of them badly. It sends each call to whichever model you hard-coded, whether the task needed the dear one or the cheap one. It re-runs calls it already ran a turn ago, and pays for them again. And when a prompt injection rides in on a document it was only meant to read, nothing stands between that and a real…
I built a machine that decides whether to answer. Most of the time, it decides not to. Then I built a second machine to decide whether the first one was worth keeping, and it told me no. Then I did the one thing that could test that no: I built the first machine properly and ran it over my own files. It answered almost nothing. This post is about why all three of those are the same result, and why…
Update (June 2026): a lot has shipped since this post. Routing each call to the cheapest capable model is now on by default, and shadow mode will audit the savings on your own traffic before it changes anything. The current announcement is Make Your OpenClaw Agent Cheaper, and Measure It Yourself ; this one stays as a point-in-time record. OpenClaw makes tool calls all day, and two kinds of them…
In the last post I built a governance layer for a coding agent’s tool calls: a body that hooks the agent’s tool_call event, extracts a few features, and dispatches ask, proceed, or block; and a brain , a Julia daemon that holds a belief and maximises expected utility. The commitment that held it together was that the brain is opaque to the body. The wire carries observations and named…
This is Part 5 of the Velotix Files. Part 1 covers the full investigation. Part 2 profiles the CEO. Part 3 examines the marketing machine. Part 4 documents the DMCA takedown notice. The Comeback On 3 June 2026, the financial section of Maariv carried the good news under the byline of Ariel Feiglin: The founder injected 11 million shekels — the cyber startup avoided bankruptcy Velotix, the story…
This is Part 4 of the Velotix Files. Part 1 covers the full investigation. Part 2 profiles the CEO. Part 3 examines the marketing machine. Part 5 covers the comeback. The Notice On 18 May 2026 at 19:43 UTC — fifty-nine days after Part 1 was published — Cloudflare’s Trust & Safety team forwarded a DMCA copyright infringement complaint to the abuse address registered on the account that fronts…
There is a coding agent — pi , the AI tool I use to write code — and it makes tool calls all day. It runs bash , edits files, opens HTTP requests, queries databases. Most of what it wants to do is fine. Some of it isn’t. The question I have been circling for months is: who decides which is which, and how, and on what basis? The current answers are unsatisfying. The agent’s own RLHF…
The production question about LLM agents, once you have gotten past whether they work at all, is how much they cost. A customer-service agent that answers well but costs eight cents per turn is not a customer-service agent; it is a charity. The conventional answer is to pick a cheaper model and hope it is good enough. The less conventional answer is to treat model selection as a decision problem.
The standard way to evaluate a question-answering system is to measure how often it gets the right answer. This seems reasonable. It is, in practice, a trap. I ran an experiment to demonstrate why. A Bayesian decision-theoretic agent — built on the Credence DSL, using Beta-Bernoulli reliability tracking and value-of-information calculations — competed against several LLM agents on a…
The accuracy paradox post reported the headline: a Bayesian agent scoring +129.5 against an LLM agent’s +10.8, despite lower accuracy. This post is about the LLM side of that experiment — what was tried, what helped, and where the ceiling is. The Three Variants Three LLM agents were tested on the same 50-question benchmark. They differed only in prompting: LLM Bare. The model receives…
The previous post in this series described what I called Tier 1 of the Credence architecture: a DSL for Bayesian decision agents with three types, four axioms, and a constitution forbidding everything else. That post ended with a program the user had to write by hand — a short S-expression encoding a hypothesis about what the environment was like and how to act in it. Hand-written programs…
This is Part 3 of a series. For the axiomatic foundation, see Part 1: Three Types and a Funeral . For the VOI-gated text adventure agent, see Part 2: Teaching Zork to a Bayesian . Every reinforcement learning agent that has ever played a text adventure has, at some point, tried to take the lantern fifty times in a row. Not because it’s stupid. Because its state representation makes…
This is Part 2 of a series. For the axioms and types underneath, see Part 1: Three Types and a Funeral . For the state-representation consequences, see Part 3: The Loop Problem . Every AI agent demo involves web search, retrieval, or API calls — tasks where querying everything is merely expensive. A LangChain ReAct agent that hammers all four tools on every question wastes money but still…
This is Part 1 of a series on Bayesian decision-theoretic agents. This post describes the Credence architecture as it stood in March 2026, when the system used the standard Kolmogorov definition of probability — measures over sample spaces. Since then, the foundation has been reconstructed around de Finetti’s definition, where expectation (the prevision ) is the primitive and…
I have about eleven million words of personal documents. Contracts, invoices, court filings, medical notes, research papers, travel itineraries, conversation transcripts, CVs of various vintages, takeaway menus from restaurants that have since closed. A decade of Syncthing directories and Dropbox archives and email attachments saved twice because I wasn’t sure which copy was authoritative. I…
“Help me keep that coding agent pup on the straight and narrow, as it likes to be ‘pragmatic’.” That was the instruction I gave Claude.ai a week before any code was written for a personal-knowledge-management project . Its companion describes what that foundation is and why none of the existing tools fit. This one is about how the foundation got built: two AIs, one spec as…
This is Part 3 of the Velotix Files. Part 1 covers the full investigation. Part 2 profiles the CEO. Part 4 documents the DMCA takedown notice filed against Part 1. Part 5 covers the comeback. The Plumbing When Israeli retail investors encountered advertisements for the Velotix angel round in January 2026, they saw what appeared to be independent coverage across multiple platforms. In reality, the…
This is Part 2 of the Velotix Files. Part 1 covers the full investigation. Part 3 examines the marketing machine. Part 4 documents the DMCA takedown notice filed against Part 1. Part 5 covers the comeback. The CV Dr Adi Hod’s marketing biography, as it appeared across investor materials, conference profiles, and press coverage, describes a figure of considerable accomplishment. A researcher…
Disclosure I am a former Velotix employee (October 2024 – May 2025). I filed the complaint with the Ministry of Labour referenced in this article. This piece is based on official government documents, court filings, archived web pages, and publicly available information. Where I draw on my own experience, I say so. This is Part 1 of the Velotix Files. Part 2 profiles the CEO. Part 3 examines the…
I wrote an essay arguing that decision theory had been quietly abandoned by mainstream AI — not because it stopped working, but because deep learning absorbed all the oxygen. I posted it to Hacker News . A commenter informed me I was “annoyed at the Bitter Lesson.” I hadn’t read the Bitter Lesson. This proved awkward for approximately forty-five seconds, after which it…
Someone asked a good question. I’d written a post arguing that what the industry calls “AI agents” are flowcharts with good marketing, and that the mathematics to do better has existed since the 1960s. A commenter on LinkedIn replied: “So why did it stop being widely used?” I sat with this for a day. It deserved a proper answer, not least because I’d spent a…
I’ve spent the last few months building agents that maintain actual beliefs and update them from evidence — first a Bayesian learner that teaches itself which foods are safe, then an evolutionary system that discovers its own cognitive architecture. The experience has given me a certain clarity about the industry’s use of the word “agent,” in much the same way that learning…
The companion essay argued that LLM-based “agents” don’t earn the title. No beliefs, no uncertainty quantification, no principled mechanism for deciding whether a tool query justifies its cost. This post supplies the technical scaffolding for that claim — the mathematics and code behind Credence , the benchmark I built to test it. Think of it as the receipts. For the…
In Part 1 , I built an agent that learns which foods are safe through Bayesian inference. It starts ignorant, observes outcomes, updates its beliefs using exact conjugate mathematics, and eventually acts with something resembling competence. Clean code, sound theory, and those belief distributions converging in real-time remain genuinely satisfying to watch, in the way that all correctly…
I step on a bathroom scale and my body fat percentage gets beamed to a server in Shenzhen. This is the arrangement. In exchange for this intimacy, the app suggests I upgrade to premium, which I find touching in a way the developers probably didn’t intend. I don’t have a problem with companies making money — I have a problem with the default. The default is: your body’s…
You’re stranded somewhere unfamiliar with twelve types of food scattered around. Some provide energy. Others are toxic. You don’t know which is which, you’re losing energy with every step, and nobody left a manual. The question is whether you can learn fast enough to survive. This is the exploration-exploitation tradeoff, and it’s one of those problems that sounds like a…
Body composition scales are, on their own terms, genuinely useful devices. Step on, wait a few seconds, receive a small dossier on your own physical form: weight, body fat percentage, muscle mass, and several other numbers whose accuracy I’m diplomatically not questioning here. The problem is what happens next. Your data vanishes into whichever proprietary app the manufacturer saw fit to…
Every GTD app I’ve tried commits one of two sins. Either it’s a Byzantine fortress of features — Kanban boards, team collaboration, productivity gamification, the full catastrophe (looking at you, OmniFocus) — or it’s simple but demands you learn its particular syntax for categorising things, which rather defeats the purpose of a system meant to get tasks out of your…
The madpersons behind the retro-chic data analysis tool VisiData have created an alternative to the infamous “Advent of Code”. Instead of programming to solve yet more optimisation puzzles, and instead of the hackneyed Christmas theme, the Hanukkah of Data requires data analysis to solve a murder mystery rug-finding mission. I only just found out about it, so although one puzzle was…