Boyd's Blog

I use the various language models a lot for my research, and this means I end up seeing quite a lot of their failure modes.

Poisoned Here’s a simple idea: what if we trained in a string of characters that caused an LLM to emit the end of sequence token <|eos|>, regardless of where that string was in the LLM’s context window?

| Read on LessWrong | Comment on HN | Large language models often take actions running on one computer (via an agentic harness such as Claude Code or Codex), however the LLMs’ responses to prompts are computed on a different computer with GPU access.

About Me I like writing about expertise, using computers to solve problems, and bringing novel ideas into existence.

| Read on LessWrong | The problem This tweet recently highlighted two MATS mentors talking about the absurdly high qualifications of incoming applications to the AI Safety Fellowship: Another was from a recently-announced Anthropic fellow, one of 32 fello…

| Read on LessWrong | Claude & I vibecoded an interface for the extropians mailing list.

| Read on LessWrong | American bald eagle caws angrily in the distance Whoa whoa whoa, just hear me out.

| Read on LessWrong | I’m a MATS 9 extension fellow, and usually my week is spent trying to find better ways of evaluating Large Language Models.

| Read on LessWrong | Basically every meeting is scheduled with an explicit duration.

| Read on LessWrong | Around July last year I decided I was going to go all in on technical AI safety research.

Hey!

Often, I’ll find myself wanting to reference AI safety papers or findings, knowing vaguely that something’s been shown to be true, but not being able to find the source paper.

| LessWrong post | I couldn’t find a recent write-up from a MATS alum about what attending MATS was like, so this is the thing that I wish I had.

I’m a MATS fellow, having completed cohort 9 from Jan-Mar 2026 in Berkeley, California.

This is a public version of my comments sent to the South African Department of Communications and Digital Technologies on the Draft South African National AI Policy.

Many card games are similar to Cards Against Humanity in that they fit the template of: a large (50+) deck of cards each card has some creative prompt or phrase generally fun at parties (until the novelty wears off) some easy-to-learn rules about how to p…

In an effort to write more and to think less, I’ll be publishing something every day during November.

Right now I’m interested in and working on: MATS 9 and MATS 9 extension with Alex Turner (GDM) and Alex Cloud (Anthropic), working on methods to detect deceptively misaligned AI.

You can download my CV here.

| Read on LessWrong | I’m not quite sure how unequal the world used to be, but I’m fairly certain the world is more equal (in terms of financial means) than the world was, say, in the 1600s.

| Also on LessWrong | I’m generally very bad with names, and especially struggle when I’m quickly introduced to large numbers of people.

| Read on LessWrong | We don’t have AI that’s smarter than you or I, but I believe we do have something that’s somewhat similar, and analysing this thing is useful as an argument in favour of ASI not being aligned to humanity’s interests by default.

The Provenance Protocol allows anyone to verify the source of an image, PDF, or piece of text, thereby making AI-generated images nearly impossible to pass off as real.

This page is mostly used as a reference, so that I can create links in different posts and have consistent definitions.

Can't you just turn it off?

This mini-project was inspired by this tweet: After which I spent about two hours making a small script that grabs data from the rust package repository crates.io, and analyses the data to see which crates are downloaded mostly on the weekends (indicating…

Around August 2022 for my 4th year Honours project at Stellenbosch University, I scraped over 26k TikTok videos taken over a two-week period with the goal of analysing what makes a video go viral, and what that virality actually looks like.

Eskom-calendar was an open source project I created and maintained for ~2 years until (thankfully) its raison d’être ceased to be.

So I agree with the concept of an /ideas page as a means to finding people who you can collaborate with (read the about-ideas-now manifesto), but it also seems like the perfect place for people to poach good ideas from.

YAML is just JSON.

Elevation maps are a project I’ve been working on-and-off for a while.

I recently made git-activity, a small shell script/git alias that lets you view a github-style activity graph in your terminal.

ListenToAnything.com lets you convert any website, PDF, blog post, or article into a podcast episode that gets uploaded to your own custom podcast feed.

Something that’s painfully understudied is how experts are more efficient than novices while achieving better results.

I think I know vim pretty well.