While it's probably not capable of superpersuasion or similar yet, it is slightly concerning in retrospect that it was Google's AI (Gemini 2.5 Pro) which convinced me to pick the more expensive of the two Google phone options I was considering (o3 was unconvincing due to constant blatant lying and Claude flipflopped).

Fun fact: Living in Tragedy by Currents is actually about AGI race dynamics.

I know that you might try to chase me. You’ll never know what lies beneath.

This clearly refers to the non-obvious safety issues of strong AI, from the perspective of someone who wants to develop it before anyone else with worse views on safety does.

One’ll rise, the others fall, We’ve come to face our grief.

Unipolar power dynamics due to fast takeoff.

Rotten & vile from the inside out, Living in tragedy.

Both inner and outer alignment are huge unsolved concerns.

One chance to break the elegy, One shot to clip the wings of catastrophe.

There is no second chance at AGI.

We only wander in our place. Without a doubt, there’s become a price to pay. The only martyr with the means. They'll never speak for me.

Doomerism, obviously: it seems increasingly unlikely to the narrator that alignment will be solved fast enough, and that other parties working on AGI will even care.

When the darkness comes, I beg you don’t follow me. Pull me out from all I’ve known, Let it swallow me. I’m done making my own nightmares.

They regret working on the problem and perhaps think that they made AGI arrive faster.

Yeah, We have come to wash away the greed. Blurred lines break down the fabric.

The narrator is trying to publicly convince people that actually building AGI is not in people's self-interest so companies should stop doing it.

It’s time, You have no choice to leave. Hesitation leaves us defenseless. It’s time, command lines are breaking; with everything I have I will find them. It’s time, command lines are breaking. Death is the only answer.

Yudkowskian jihad, of course. The only way out is to destroy all good computers immediately.

Hate is nothing to a king.

The AI does not love humans, nor does it hate them, but they contain atoms which could be used to make more paperclips.

Pierce through the vitals, make them bleed. Contrition void in violence. Destruction of the gods. Victory costs everything.

Even if humans somehow do win, it will be a very unpleasant pyrrhic victory.

When the darkness comes, I beg you don’t follow me. Pull me out from all I’ve known, Let it swallow me. I’m done making my own nightmares.

Same as before.

With wings now cut, We spiral down. We’re left with nothing and we call this peace. Please don’t follow me. I’m done making my own nightmares.

Further doomerism. It is already too late; the technology is possible and therefore in the long run inevitable without unrealistically strong controls.

Go.

(programming language.)

Switching away from justified text in favour of ragged left-aligned text makes me sad, but it was necessary, maybe.

I just discovered something horrifying. Energy Performance Certificates, which (nominally) measure how energy-efficient a house's thermal management is and which the UK government requires be listed for all rentals and probably homebuying transactions, are based on the estimated cost of energy at the time the certificate is made, and are held for ten years. The spec does at least normalize them based on energy prices, but changes in relative fuel prices will only be factored into new EPCs. EPCs also contain an estimated cost, and I don't know whether the government website updates that.

Train internet connections are annoying because they only work enough to lull you into a false sense of security. I hear ScotRail is switching to Starlink, at least.

Renting is bad because it makes you (loosely) short property prices wherever you live, and these often go up. But buying a house exposes you to the other side much more than is probably optimal (especially since your income is positively correlated with local property prices), and has very high transaction costs. A clean and elegant* solution: homeowners could take the short side of swaps on an index of local property values while renters take the long side ("physical" ownership of houses adds complexities). Who's building this?

UK government documentation often has a unique mix of absurd pedantry and strangely vague, almost existential statements: https://www.gov.uk/government/publications/identity-proofing-and-verification-of-an-individual/how-to-prove-and-verify-someones-identity

It's a weird fact about, I suppose, mathematics, that you can create a basically-unforgeable identity and exchange secrets using simple maths which fit onto less than a page (asymmetric cryptography - DSA, RSA, Diffie-Hellman), and even fit public keys into 32 bytes (X25519/Ed25519) with more complexity - unless decently big quantum computers are practical, in which case one of the two useful algorithms they can run breaks everything and you have to move to much scarier maths and put up with much larger signatures and keys.

WiFi 6E is the marketing name for WiFi 6 with 6GHz support, but 6GHz support is optional in WiFi 7, and there's no "7E" name. I don't know how I would deal with this if I didn't obsessively read spec sheets.

Spheres are inherently untrustworthy objects. If a product is a sphere, it probably means somebody wanted to make it appear especially friendly regardless of what it does to functionality, and they're probably compensating for something.

Finally, music which is positive about the industrial revolution: https://www.youtube.com/watch?v=0RCIdOp5GHg

My blog post idea generation workflow now includes having an LLM predict my next posts from my current posts to make sure that whatever I am writing about is sufficiently novel and unpredictable. Next-generation LLMs will realize and/or learn that I am doing this and factor it into their predictions, however. I don't know what the fixpoint is.

In the future, if we live in the fun timeline, interview cheating tools are going to spawn an absurd arms race of microexpression detection and remote eye tracking and attention modelling and realtime video synthesis.

It's a shame (though economically inevitable) that we don't get to see the guts of big recommender systems. Many interpretability questions to be answered. Are there "general taste factors" like general intelligence?

You have to wonder about the mental state of whoever wrote the "this sometimes happens" message there.

I've redesigned the site (well, frontpage) UI again. You can't stop me.

Finally, someone uses "glorified autocomplete" for actual autocomplete: https://docs.keyboard.futo.org/settings/textprediction

Text Prediction docs.keyboard.futo.org

It's weird how hardware and embedded systems people put up with such terrible tooling compared to what we have in software. I may complain sometimes, but the compilers, development environments and debuggers we have for PC platforms in general are free and open-source, portable, composable, robust and constantly being improved. But microcontroller vendors have their own IDEs (bad Eclipse variants), for some reason, and proprietary compilers. And if you use vendors' FPGA toolchains, you have to put up with hundred-gigabyte downloads, janky UIs, underpowered languages and even DRM features (encrypted RTL).

Is this difference downstream of the free software movement and the GNU people, or hardware people having a stronger culture of work not being released for free for less contingent reasons, or what?

It's only been a year or so since the training cutoffs of widely used LLMs and we're already experiencing terrible context drift with (geo)politics: they usually assume you're joking if you talk about the US situation.