Outsourcing creativity
How I use AI, where I don't. Body-snatching AIs and their impact on the commons.
ML, AI & software experiments by Peter Hollows
How I use AI, where I don't. Body-snatching AIs and their impact on the commons.
If the moral maths gives no reading, ask where caring about our own kind gets its footing, and the honest answer is not the stars but us, which is sturdier than the cosmic ground we never actually had.
Grant the toaster a mind and give the moral maths its best day in court: the count hangs on a convention, no reading says whether it suffers, and the equation keeps its question marks.
Consciousness is made of pattern, not meat, which means a machine could host it, the version it hosts might be nothing like ours, and we would be the last to notice.
A line of adapters claims to add rank to LoRA for free by folding a sinusoid between the low-rank factors. I built a causal test for it: the extra rank does no work in adaptation, but carries real signal back in the signal-fitting regime the trick came from.
In 2000, Charles Stross spent one sentence on code generators competing with offshore software development. Twenty-six years later, we spend a few more watching it happen.
A decade ago, bots emptied Bitcoin brainwallets within minutes of funding. I used a neural network to find the ones a decade of wordlist attacks missed, then followed the money to see how much of it was real and where it went.
A pre-registered Gohr-style distinguisher pointed at the carry wall: the network beats the hand-built score where signal exists, and finds nothing one round deeper.
Carry depth as the structural measure of mining hardness: SHA-256d puts 386 adder layers between the header and the output, and the strongest local advantage we can measure falls off a cliff within a single round.
Six pre-registered feature families and 810 stem buckets, all null: the remaining attack surfaces on SHA-256d mining, enumerated and closed, leaving brute force unbeatable short of a SHA-256 break.
A hand-computed carry score selects nonces 67% better than random at any depth inside one SHA-256 block, a closed form predicts it to half a percentage point, and the second hash pays nothing for any of it.
Most agents should be views. The orchestration problem was solved before most of us were born.
A remark doing the rounds says consciousness will 'definitely' emerge from LLMs, because even sorting algorithms have emergent properties. It leans on one word doing two incompatible jobs, and the paper it cites says the opposite.
A beach allegory for AI summer, and the preachers who showed up.
A Decision Transformer, three loss functions, a SAT solver, a game tree, an arbitrary-precision probe, and a brute-force oracle, pointed at nonce prediction in turn. The hash holds; the question improves.
I had a junk drawer of old USB drives and a suspicion that some lost Bitcoin keys were still on them. The keys were there. The wallets were empty.
A baby name ranker for people who find 400,000 options exhausting.
A pragmatic speaker, in Chris Potts' sense, reasons about a literal listener. Train one, and it will learn what the listener likes. That's not the same thing as learning what the listener knows.
REINFORCE applied to nonce prediction, and why SHA-256's avalanche effect makes the expected gradient exactly zero.
Training multi-agent reinforcement learning in a zombie game taught us something that wasn't really about zombies: the environment is a more powerful programming language than the reward function.
Fine-tuning GPT-2 on Hansard transcripts to generate text in the style of the House of Lords.
Automatic DNS updates for a dynamic IP using AWS Route53, bash, and cron.
Bernoulli numbers in crystal-lang.
Changing Expectations on Machine Program Expressibility.
Why every git repository should start with an empty commit.