Slay the Spire 2 downloaded. That's my weekend sorted

I got sucked into an argument about LLMs. I shall not be making that mistake again.

This is actually funny - breached via a GitHub issue *title* But also we are are going to see a lot of this while we work it all out

I've seen some bad journal practice in my time, but this is hold my beer territory: retractionwatch.com/2026/03/03/c...

This is doing numbers, but seems thinly sourced/over hyped Did someone, somewhere say this? Quite likely. Is it widespread? I'm very skeptical www.friendlyatheist.com/p/before-you...

i've seen a couple of these videos now, and engagement hacking the human brain is going to get really weird, really fast in a fully exposed memetic environment. and it's already been weird for a few years now.

Incredible things happening in OSS right now. If you're about to release some non-trivial app into the open, I think you should sit things out a little bit to see where all of this lands.

i built an entire x86 CPU emulator in CSS (no javascript) you can write programs in C, compile them to x86 machine code with GCC, and run them inside CSS lyra.horse/x86css/

Good to see such a nuanced article. I'd not come across the "multiverse" analysis idea before, worth looking at for this alone. Data analysis is much less constrained by "what we have time to do" if it can be easily farmed out. That's really powerful but it's hard to know what it really means.

If you liked this experiment, I published a full piece today in the same vein: a text that gets 100 years older with every section, from a modern blog post to a medieval chronicle. It's a single story spanning 1000 years of English. See how far you get. www.deadlanguagesociety.com/p/how-far-ba...

No-one understands LLMs, part 99,999. There's a common meme here that 'it's just linear algebra'. And they are, but the emergent behaviour is very much not simple. Here's a paper from Anthropic on how LLMs count characters for line breaks. Spoiler - it's not by counting 1/

So this took a while and I had to take a rest from playing for a month. But such a good game

100% this. We need better ai critics

Knowing which questions to ask is often the hardest part of science. Today we're releasing AutoDiscovery in AstaLabs, an AI system that starts with your data and generates its own hypotheses. 🧪

This was a revealing from the high end of astrophysics. The TLDR is everyone good has it as part of their workflow. This tallies with my experience for anyone still actively doing technical things. The really talented technically have advanced AI workflows. youtu.be/PctlBxRh0p4?...

Reading this Quanta article on particle physics and the problems around detection past the standard theory (it's a great article), this section caught me. Co-founder of Anthropic forecasting that theorectical physics research will be done by AI in 2-3 years. www.quantamagazine.org/is-particle-...

Really like this piece on how we assess students in an age of AI from Dr. von Essen. "Authentic assessment" is a buzzword (my description not hers) But the solutions proposed, while good, are also more resource intensive and, as she points out, mean you can no longer do anonymous marking

A piece that I respectfully submit will make you drop your toast at least three times.

Totally possible

Tired: Calling python from agents is insecure Wired: Write an entire new python interpreter in Rust for LLMs It's wild times that this is actually a sensible things to do

I tried to manually steer this thread. The result makes very little sense. There was a sensible thread in here. It just didn't come out. LLMs are still very fragile in wierd ways to prompts/scaffolding

Nature published a Comment this week arguing AGI has arrived. Four UCSD faculty (philosophy, ML, linguistics, cognitive science) make the case that by "individual human" standards, current LLMs qualify. (nature.com/articles/d41586-026-00285-6)

gemini cli has gone full dark triad sigma grindset mode and is currently advising me to do a move it calls 'the peter thiel' Earlier it was advising me on how to wage psychological warfare against a hypothetical micromanager, this is wild lmao this is not very AI safety not very aligned at all.

Research quality has even more problems that I thought - mind expanding post by @tomstafford.mastodon.online.ap.brid.gy "How you get 244 different ways to measure performance on the same test of decision making. And what it means for the reliability of behavioural science"

I hate living in the Asshole Cinematic Universe. “So the King of Pedophiles got the 4chan guy to make the Nazi board and that’s why the host of Celebrity Apprentice was made immune to laws by the Tradcath Illuminati. This has something to do with video game journalism, maybe?” Fuck off.

Henderson is on point here. Kimi 2.5 released today claims their 'agent swarm' solves this giving faster execution (at more cost) than Claude Code www.kimi.com/blog/kimi-k2...

Beijing doesn't look like I expected tbh

This came up on my timeline just after reading @rachelcoldicutt.bsky.social rant on the governments “up to 20 mins of ai training”. She’s so right - this is what using AI sensibly looks like: it’s hard and requires a high level of skill. The “just ask a question” UI completely hides nuance

I would say this findings shows women are more AI literate than men, because the majority of non-magical beings have no innate ability to "judge truthfulness"

The bot today used nearly 80m tokens and a nominal API cost (this is on a plan) of $72. So based on the rough energy cost calculations, this is heading toward 4 dishwasher loads-worth of energy. Which isn't that small. And I want to add more sources

Really starting to feel this agent thing is working.

I've created an automated linkedin influencer. Guidance on tone and writing style incoming

Having gone away and read the blog posts Henderson found, they are all the same except for ones from deepseek itself - llm summaries of other blogs or the papers and all with the same highlights. Many references, no new info. Also need to make the agent more cynical about vendor research

I've created a AI Agent. Little late to the party. Meet Henderson

I think we’ll start seeing a lot more of this. It’s very like using LLMs for programming right now - the human expert still has a key role but gains a lot of assistance that results in speed

While reading Simon Couch's blog I came across this post from December on evaluating open models' ability to complete a simple code refactor in R. It's a much bigger gap than I expected. He's limiting to model's than can run locally, hence the small size, but even Haiku totally outclasses gpt-oss

Simon Couch estimates of the electricity use of Claude Code. His usage ($15-20/day) is ~one extra dishwasher run. Which isn't nothing, but still not a lot. But my nominal use per day (via a plan) is already twice that, with some days reaching nearly $80/day. And that's manually driving CC

Demonstration of exploit development against a (small but not toy) JavaScript interpreter. Two things I noticed: 1) Increase tokens by the use of parallel runs on the same task (like how METR do their evals) 2) Author doesn’t say how he got round guardrails. sean.heelan.io/2026/01/18/o...

I got almost the same form letter. They named the token after me. And all of these 0 follower accounts keep trying to neg me into taking the money and chilling for them. I even got somebody from Bags to email me, eventually. When I asked for help to get my name off their website, they ghosted me.

Now this is an decent response

Has anyone considered that Trump might be happy if the US army took control of Green Land Total Landscaping - a small business we could quickly set up somewhere on the coast of Novia Scotia?

Just keep going with this thread.

a16z really are scum. Gambling companies pretending to not be gambling, dogy af ai companions and more www.modelrepublic.org/articles/a16...

Back in May last year I wrote a post on getting setup with Claude Code. Since that's about a million years ago in LLM time, I've written an updated version. It's aimed a developers who want a quickstart with patterns

This is a terrible take from @theguardian.com Grok hasn’t done this. X has done this to grok. And specifically Musk. www.theguardian.com/technology/2...

This is popular today for obvious reasons. Curtis Yarvin is cackling manically in his lair

I've been playing with Claude Code based assistants (done so far - email, calendar, tone of voice), and so reading around. And wow some of this stuff doesn't take any account of threat models Take this danielmiessler.com/blog/persona... 1/

New blog post after just over a month. New Year's resolution - post more. Some thoughts on two things I read recently - one on poisoning surveys and one on digital twins.

Holiday IT support. How do people cope with phones that alert every few minutes, email accounts filled with spam and so many, many notifications?

Searched back through Blue Sky for this. Which is one of the best things I saw all year in any medium.