subseciva · 2026-10-21

"In the imperial period, there was great state control over trade in order to guarantee supply (the annona system) and even a state merchant fleet, replacing the system during the Republic of paying subsidies (vecturae) to encourage private shipowners."

https://www.worldhistory.org/article/638/trade-in-the-roman-world/

Trade in the Roman World — Regional, inter-regional and international trade was a common feature of the Roman world. A mix of state control and a free market approach ensured goods produced in one location could be exported far...

"The state taxed the movement of goods between provinces and also controlled many local markets (nundinae) - often held once a week - as the establishment of a market by a large land-owner had to be approved by the Senate or emperor."

"what have the Romans done for us"

At one point the British East India Company accounted for over half of all trade, not in the British Empire, but in *the entire world*.

actually the british east india company is fascinating cause it starts out as private trade and slowly metastasizes into empire; but was actually, in terms of the empire operating as declared, held separate. it almost sort of became a vassal state

the boundary between "trade" and "protection racket" is blurry

we should do the best possible. if the best possible happens to be an aligned asi disassembling the universe to make a giant dense computronium bubble, we should do that

But I'd rather go transhuman -> posthuman of equivalent power to an ASI than hand over control to a separate ASI.

Even if the end point was almost the same.

Unfortunately I suspect human alignment at that scale is a lot harder than ai alignment.

it's a lot easier for a universe like that to end up broadly bad.

I think I would take an ASI over a human of the same intelligence

the ASI has a chance of acting like how humans talk about themselves

whereas I'm confident the human will just be terrible if given absolute power

Really, we shouldn't give anyone absolute power

Multipolarity forever

i.e., age of em/accelerando/quantum thief.

permanent underclass sort of universe.

Which I also don't want lol

how do we protect the weak from the strong?

If there's only one strong thing in the universe, we have to get the one strong thing to want to do it

and that's alignment :P

if we're weak, and life's not shit, we've succeeded

But I don't have any faith in alignment actually doing what people want it to do.

And even less faith in all the proposed alternatives.

Alignment is very very hard.

it's not even actionable. I just think everything else is either obviously terrible or just genocidal

and "make the strongest thing, make it want to protect weak things" is at least an outline of a plan. this is all straight out of scott moloch.

It's a bad plan

But I know no better.

"Something needs to be done, this is something, therefore this needs to be done"

"Something needs to be done, everything else is worse, therefore this needs to be done"

But I do think we should investigate the "everything else". See if we've missed anything.

"Disagree? If an AI aligned to your values that was smarter than you were, and had devoted much more time to thinking about the question, told you that your disagreement was wrong, and actually your values would be better served by adopting this system, would you believe it?" no, because it's disagreement demonstrates that it's not aligned to my values

"aligned to your values" seems to have become a thought-terminating cliche

"i want you to join me in the torment nexus" "no" "but an AI aligned to your values said you should"

somehow an AI aligned to my values always tells me to do whatever scott thinks is best, what a weird coincidence

https://hachyderm.io/@madengineering/117475782040902491 ... he jokes but literally https://theapplewiki.com/wiki/Dont_Steal_Mac_OS.kext

Mad Engineering (@madengineering@hachyderm.io) — @Viss@mastodon.social It's a constant surprise how much of our IT infrastructure is the equivalent of leaving our front door wide open with a cardboard sign that says "These are my things, please don't steal." And then …

I just hope they patch all PHP versions and not just the latest one

I think local models are only worth it for privacy, learning, low latency, and messing with the model (fine-tunning, steering, etc.)

Or if you have unreliable Internet

They don't pay for the hardware in inference cost

It is a fun hobby running them

And I believe we are getting close to where you'll be able to run a pretty capable model on the CPU only with 48G RAM because the model will have only a few billion active parameters

Qwen3.6-35B-A3B is still slow on the CPU. https://www.liquid.ai/blog/lfm2-5-24b-a2b is closer.

LFM2.5-24B-A2B: Built for Fast On-Device Agents | Blog — LFM2.5-24B-A2B: an open-weight 24B reasoning model with 2B active parameters, built for fast, private on-device agents. Beats Qwen3.6-35B-A3B on LiveClaw.

You can try the free models on OpenRouter. Some don't train on your data.

Not paying for any subscriptions ever is a big handicap to take on, in terms money and productivity

Thinking of paying $10/mo for https://opencode.ai/go

OpenCode Go | Low cost coding models for everyone — Go costs $10/month, with generous usage limits and reliable access to leading coding models. Go Plus costs $40/month with higher limits.

It's on my to-do to try https://commandcode.ai/

Command Code - The best coding agent for open models — The coding agent that made open models work. Tool call repairs on every model, best in class token efficiency on read and shell tools, 99%+ cache hit rates, and taste-1, which learns your coding taste. 100K+ developers,…

I have defected so far from optimum on the not giving money to AI companies front ;_;

Carbon offseting, but for AI risk

https://www.greaterwrong.com/posts/4uBEnYsmwxDw48fb6/donation-offsets-for-chatgpt-plus-subscriptions The only substantial discussion I can find on LW is for 2023

Donation offsets for ChatGPT Plus subscriptions — I've decided to donate $240 to both GovAI and MIRI to offset the $480 I plan to spend on ChatGPT Plus over the next two years ($20/month).I don't have a super strong view on ethical offsets, like donating to anti-factor…

I guess it's easy if what you're worried about is power concentration: for every American AI subscription, buy one European and one Chinese

It is not really RSI unless you work on autonomous AI research

There is a cap on the difference a harness can make

It isn't close to training new models

That said, my harness work was using AI to run AI experiments to use more AI, and the result is that I am much more interested in running AI experiments

RSI is bad because of the exponential.

While orchestrator work is recursive, its benefits probably follow something like a sigmoid curve, so the bad thing about RSI is removed

They are also much smaller in absolute terms

"Calculating... Difficulty 20" sourcehut is retarded

https://stencil.so/blog/the-harness-problem possibly the best introduction to hashline, even if LLM-written

We improved 15 LLMs at coding in one afternoon. Only the harness changed. · Stencil — Across 16 coding models, hashline’s content-hash line anchors beat patch for 14 models; the weakest models gained the most.

https://en.wikipedia.org/wiki/Tactile_paving

Tactile paving - Wikipedia

"An (𝑚, 𝑛)-SuperHyperGraph is a mathematical structure in which each vertex corresponds to an (𝑚, 𝑛)-superhyperfunction defined on a base set, while the hyperedges group such functions together to represent higher-order relationships and contextual connections." https://z-lib.sk/book/kzPZXLa45G/hypergraph-and-superhypergraph-theory-with-applications-i-foundations-definitions-and-theoretic.html

https://z-lib.sk/book/6qBy3obMzJ/the-infinity-machine-demis-hassabis-deepmind-and-the-quest-for-superintelligence.html

https://z-lib.sk/book/Gz3XwNJMqE/revolutionary-mathematics-artificial-intelligence-statistics-and-the-logic-of-capitalism.html

https://x.com/repligate/status/2110275217348321687