In contrast to intensional type theory, which is a type theory containing an identity type x=yx=y whose elements are identifications of xx and yy, unintentional type theory has a mistaken identity type x≈?yx \overset{?}{\approx} y whose elements are inadverten…
I've made a new video. It is just one minute. You will like it. It's a good video. You can also share it because it is a good video. Part two is 2 minutes long.
When you disagree with someone, there are two paths forward. The first is to try to engage with what they’re saying: what their reasons are, whether their…
"System one" decision models are models that infer and respond with calibrated probabilities or every allowed answer.
Consider your everyday language model, to get typed output from it (JSON), you may use Structured Output to constrain the output to guarantee…
I remember hearing a while ago that Dwarf Fortress didn't use version control. That fact has been lodged in my head, because given its inherent complexity I have trouble imagining …
Bit of a niche link, but I've been waiting for this for a couple of days - I even told ChatGPT to check for it: Clone https://github.com/actions/python-versions and git pull …
this is kinda a sequel-ish to my previous post, in which i go over the math_errhandling macro, and how math error handling is done in glibc and musl (as well as how it's specified in the standard).
Introducing BreadBowl-Embed, a new late-interaction architecture for embeddings: 16 routing–value slots per passage, so the same index can retrieve and rerank, with no cross-encoder and no second pass over the text.
Why a load exclusive (ldaxr) raised a Data Abort on real Raspberry Pi 5 hardware but not in QEMU, and what cacheability has to do with AArch64 exclusive accesses.
In our previous article, we trained a small selector for a single decision inside dksplit, our word splitter for domain names, and tested Jev on the same task: it does...
In contrast to intensional type theory, which is a type theory containing an identity type x=yx=y whose elements are identifications of xx and yy, unintentional type theory has a mistaken identity type x≈?yx \overset{?}{\approx} y whose elements are inadverten…
I’ve been hiring software engineers for the last 8 years. I’ve reviewed hundreds of home assignments as part of hiring process. In the recently introduced AI era, it is more and more believed that hiring assignments are useless because candidate can just feed …
I cancelled my Claude subscription today because I can't use my purchased tokens in my harness of choice. I also couldn't use them with an LLM proxy server, either. Sure, I can do some hacky things to make it work, but it's not supported and I risk
In the last few years I've seen people struggling to offer some value through the use of AI Agents. Across examples I noticed a common shape in them, a...
How every 3B workflow step gets a durable, versioned filesystem, from a gVisor gofer that hands the sandbox a local file to a storage engine with snapshots, forks, an LSM index, and S3 as a cold tier.
Two ways to put Jev inside an agent harness: checking step completion and making bounded business decisions. Experiments with Proceda cut main-LLM calls by almost a third.
A load balancer forgets an idle connection, a connection pool remembers it, and the first request after a quiet stretch gets a TCP reset the client logs as a 502 from a service that never saw it. AWS documents the behavior at 350 seconds, without a FIN. We rep…
Can LLMs make decisions without extra fine-tuning? Local tests of Qwen, Gemma, Jeff, and Laya explore whether a workflow needs a dedicated decision model.
I got curious about how Octopath Traveler calculates Therion's steal rates, so I dug into the game's code to find out exactly what's going on behind the scenes.
The checkable rules each AI host applies to an MCP server, from tool-name length to mandatory annotations and timeouts, where they conflict, and which tools validate against them. As of 1 October 2026.
[6 min read] A couple of weeks ago, I gave my keynote, “The Most Valuable Engineer Isn’t Shipping Features,” at Rails World 2026 in Austin, Texas.
If you ...
Changelog 32 - Slick Seventeen, Awesome Applications, Universal USB Tools
Service Health & Status Dashboard
Don't diagnose your opponents
Build your own decision model
One vector is too few. One per token is too many.
Only mock your own interfaces 🦜
Jev and the Return of Fuzziness
Bag of Decisions Reranker
Using Jev in Domain Abuse Review: A Workflow Design
We Should Still Send Home Assignments in Engineering Interviews
Put a price on breakthroughs
Llms Arent Inevitable | deadSimpleTech
If LLMs Can Decide Without Fine-Tuning, Do We Still Need Models Like Jev? · 欧的Leo
Reverse Engineering Therion's Steal Success Rate in Octopath Traveler · Alessandro Cuzzocrea
On the Nature of the Swarm
Claude Code vs Codex: 10 MCP features tested | SineFrame M3
MCP server rules: Claude, ChatGPT, Codex, VS Code and more
The Keynote I Thought I Was Going to Give