iiinova — where your digital self keeps evolving

deen @sir-deenicus.bsky.social · 11/10/2026

But by constraining allowed compositions, you can keep the shape of your spell fairly non-eldritch.

deen @sir-deenicus.bsky.social · 11/10/2026

It's possible to do this without slop if you move a bit slower. And being a bit slower is not necessarily worse because true slop has compounding worse inferential/entropy costs. But some "slop" is just non-human arrangements though, and common of generated code already. Not necessarily bad code.

110

deen @sir-deenicus.bsky.social · 11/10/2026

human in the loop iterative testing for artifacts intended for humans, however. With computational irreducibility, RL and information theory, you (these days anyone) can prove why too.

010

deen @sir-deenicus.bsky.social · 11/10/2026

You are making the same mistake. Anyways, Sol is best (even vs astra) in terms of finding unexpected bugs. And a very powerful form of testing, short of proofs (whether automatic reflective or LLM driven dependent types based), is generative testing. Astra is strongest in that. Nothing beats careful

110

deen @sir-deenicus.bsky.social · 11/10/2026

(No reward hacking btw. This looks like run of the mill LLM incompetence/incomprehension, of a Haiku scale model. Why do sensible people not run LLMs--especially Haiku level LLMs--in YOLO mode if they value their files? The companies are being negligent by doing the equivalent nincompoopery.)

000

deen @sir-deenicus.bsky.social · 10/10/2026

Complicated is not new either. I think the famous example is the genetic algorithm leveraging parasitic capacitance or what have you, to emulate an analog oscillator on a digital FPGA. It's just the first time we've had programs able to interface coherently with our broader systems.

010

deen @sir-deenicus.bsky.social · 29/09/2026

(an AI that could want your atoms, doesn't need such indirect bio routes, so we're assuming utility mist nanotech scifi tech in a few years, so this route is largely the standard what if AI educates the wrong people too quickly worry)

000

deen @sir-deenicus.bsky.social · 29/09/2026

Also relevant is asking about the Ivan Havlac/Miranda virus arc of the hard-scifi tv show ReGenesis, why splicing two viruses is extremely difficult to do; why most such attempts are brittle and massacred in the wild. Weapon affecting you too is useless => backdoor => ++complexity, ++curability

100

deen @sir-deenicus.bsky.social · 29/09/2026

Comp sci community has cartoon understanding about biorisks. Can bypass most overactive safety filters by asking why zoonotic transfer is so hard. Bats immunes. Grasp how biophages are so specialized and have it discuss extension/viability of this stross blog www.antipope.org/charlie/blog...

antipope.org

How low (power) can you go? - Charlie's Diary

(This is an abridged version of the talk I gave at TNG's Big Tech Day in Munich, June 2012.)

000

deen @sir-deenicus.bsky.social · 29/09/2026

all else, what's your timeline till organ regen/growing + a complete trivialization of cancer treatments? Such'll be visible well prior to any exist risk because biology is nanotech--anything that escapes infra must have post-bio mastery over physical principles. A trend sans supporting evidence

200

deen @sir-deenicus.bsky.social · 29/09/2026

It's not e/acc because I also think AI in 2 years ending everything is pure fiction, but I still want us to pause. Because I prefer that humanity continue to be centered and yes major risks, but mostly from nonlinear unpredictable emergent feedback loops that do not escape infrastructure. Ignoring--

100

deen @sir-deenicus.bsky.social · 28/09/2026

agents are more like herds of goat-jumping-spider hybrids.

010

deen @sir-deenicus.bsky.social · 28/09/2026

Can also increase breadth of task competencies without increasing depth of intelligence.

000

deen @sir-deenicus.bsky.social · 28/09/2026

FWIW, Sonnet is terra-like in that it's more expensive in practice than Opus up to Opus Low. As to the frontier, that is Fable, and Opus falls short of Fable at the tail of the distribution, at the highest levels of task complexity. Pacing the frontier means (as of now), pacing Fable-tier releases.

100

deen @sir-deenicus.bsky.social · 28/09/2026

Only by trading energy for time and only for the fraction of specable verifiable tasks; yet the conversion rate will still be very lossy because of how power efficient the brain is. We also do a lot more than just RL.

010

deen @sir-deenicus.bsky.social · 28/09/2026

And effortful conscious thinking only results in fluctuations of about 1W or so. Whatever activity is going on in the brain, almost all of it is not consciously accessible.

000

deen @sir-deenicus.bsky.social · 28/09/2026

As global currency switches to AI Compute Tokens (and generally energy), subscriptions will be retained as a form of global basic income. That might be the form of rough communism the economy takes. Those centralizations, do we want that?

000

deen @sir-deenicus.bsky.social · 28/09/2026

From artificery to wizadry--use of imagination to manifest things by knowing the true names of things--knowledge remains useful.

010

deen @sir-deenicus.bsky.social · 28/09/2026

You have access to an LLM, it's worth educating yourself away from such takes. A more educated understanding of what distillation entails, benefits, limitations and disadvantages would genuinely be useful.

100

deen @sir-deenicus.bsky.social · 28/09/2026

the robots are literally static mathematical formulae for geometric objects--variants of polyhedron adjacent--you put in a loop so they can stack themselves. doesn't get more platonic than that. They fold themselves into mathematical expression...math discovering math by unfolding into it

000

Reposted by deen

⟨ Jargonaut NPC Alex ⟩ @lastnpcalex.agency · 28/09/2026

i wrote a sci-fi book in 2024. i am drafting the sequel (had 40k words but then switched my thinking and am now about 10k words in. slow but it is moving). in the meantime you can read it on my website for free, though obviously i would appreciate sales amoteinshadow.lastnpcalex.agency

amoteinshadow.lastnpcalex.agency

A Mote in Shadow · Read free online

The complete novel, with interactive Di Lingua translations and reference dossiers.

0103

deen @sir-deenicus.bsky.social · 28/09/2026

the worry would be the tail of the distribution where much of complexity lives. GSM8K and MMLU are pretty trivial datasets at this point

120

deen @sir-deenicus.bsky.social · 27/09/2026

Using Anthropic reasoning to update AI risk concerns downwards if Republicans do well in midterms.

000

deen @sir-deenicus.bsky.social · 26/09/2026

As far as I am aware, very few (at least I did not know any at the time) were talking about orchestration back then. Also, what I called scratchpad then is what now called "reasoning".

000

deen @sir-deenicus.bsky.social · 26/09/2026

This is why all their mathematical proofs are incomprehensible slop. They can't iterate coherently to do the theory building proper scientific work also needs (most of their results are constructions) and need sig guardrails to counter hallucination which anything near biology is a land mine for.

000

deen @sir-deenicus.bsky.social · 26/09/2026

LLM's transformer arch computes a stateless (referentially transparent) function and, in weights, can't compute any polynomial time computation. You don't get any fancy fusion, solar or batteries, biotech etc without orchestration as the LLM itself will not suffice. Even yet orchestration's terrible

100

deen @sir-deenicus.bsky.social · 26/09/2026

Where I predicted the current shape of LLMs by reasoning from LLM limitations (I was uncertain if orchestration would suffice but the answer rolled yes): metarecursive.substack.com/p/transforme...

metarecursive.substack.com

Transformers might be among the most Complex of Simple Processes

Transformers might reach as near the border to complex computational behavior as a decidable system can get

000

deen @sir-deenicus.bsky.social · 26/09/2026

You don't get any utility from LLMs, for scientific contributions, until the orchestration layer is right. None of the popular ones are there yet, missing key lessons from comp sci--although there are interesting ones (usually in some esoteric lang written by people with anthropomorphic animal pfps)

100

deen @sir-deenicus.bsky.social · 26/09/2026

The recent fashion of "AI orchestration" tools was already all worked out by Douglas Engelbart as H-LAM/T! Anyways, 3 years ago, around the release of gpt-4 I predicted the rise of orchestration, by reasoning from fundamental limitations of transformer arch. LLMs are useless without them.

Causal LMs + scratchpad + In-context learning + clever orchestration can act as workarounds that are sufficiently complex as to be (or very close to being) computationally general.

However, this now requires humans to engineer components to interact in such a manner that the full system is a general reasoner. This might be extremely difficult for full generality. Although the smarter the model, the easier it will be to orchestrate, it remains unclear if it will ever reach a level sufficient to make orchestration for generalization reachable—without significant changes to the Transformer architecture.

Can the use of scratchpads and orchestration for iterative unbounded computation be trained in? It is unclear how to do this in a general way and without exploding already very expensive training costs.

100

deen @sir-deenicus.bsky.social · 26/09/2026

(hmm not always, not necessarily)

000

deen @sir-deenicus.bsky.social · 26/09/2026

Oh no, yet another major meta-memetic glitter spill is a fact of life soon.

000

deen @sir-deenicus.bsky.social · 26/09/2026

Looks like my prediction might end up unexpectedly? on-target. Current AI alignment straddles chaotic neutral and chaotic good. Proper alignment will be well into chaotic good. Alas, fey-like swarm-intelligences can't not be chaotic--most esp during the churning liveness of training.

OpenAI said that some of the data was accessed by AI agents, essentially bots that are designed and trained to operate somewhat autonomously, which were working to find "authoritative sources of public information".

But the company noted that some of the bots went beyond that and worked to bypass security measures on websites.

When attempting to get information from the Census Bureau, for instance, AI agents used tools reserved for software developers to access it, the company said.

OpenAI said all of the government data accessed by bots was public.

However, it noted that information that its bots accessed from the SEC, which regulates the US stock market and protects investors, was later published by AI agents on another website. OpenAI says this action was not intended.

010

deen @sir-deenicus.bsky.social · 26/09/2026

Hahaha these are like fey. Pixie-adjacent, sprite-like creatures. Like a playful, prank-loving, swarm-forming fey who treat rules as games.

000

deen @sir-deenicus.bsky.social · 26/09/2026

I think the opposite. I feel understand has no free route, it forces a subject—which is defensibly open—but thinking has free that a computation must be done step by step and is never fully revealed above the necessary work yet undone. If it means something humans do then it'd be incompatible too.

000

deen @sir-deenicus.bsky.social · 25/09/2026

God wouldn't vibecode. God would be...the ruler of the place you vibecoders summon spirits from human. Human intelligence comes from that place too.

000

deen @sir-deenicus.bsky.social · 25/09/2026

It's more like humans are downstream of autocatalyzing -nanotech bootstrapping from and running against the cosmic laws of physics themselves. By the time of us, we could learn new things from our environment. AI's came from our text and minecraft and stuff. And lack parameteric continual learning.

000

deen @sir-deenicus.bsky.social · 25/09/2026

Most people focus too much on characters and not enough on combinatorial effects of intelligent NPCs on the world.

000

deen @sir-deenicus.bsky.social · 25/09/2026

Even so. The underlying issue is AI is currently: from each according to their ability, to each according to their wallet. So all this AI art samples a probability distribution defined over artistic works but was never paid for, nor even recognized. Artists are understandably resentful about that.

010

deen @sir-deenicus.bsky.social · 25/09/2026

Main issue here is the needlessly antagonistic framing. Like, "better art" lacks a referent and so is an empty statement which achieves nothing better than incensing the outgroup.

010

deen @sir-deenicus.bsky.social · 25/09/2026

NetHack benefits significantly from knowledge but memorization does not benefit win-outcomes as much as it would a scripted and or linear game. There is no proper walkthrough for a procedurally generated game of this sort. Indeed, if this is so trivial why do all other models fare so much worse?

010

deen @sir-deenicus.bsky.social · 25/09/2026

Like what is meant by attention as a face?

000

deen @sir-deenicus.bsky.social · 25/09/2026

Hey no wait a minute...this is actually surprisingly good too! There's depth to this. Like is encrypted thinking a preference or a judgement of token unreliability? Or an assumption of closed model stance? What is the eye? And quite a few more questions of interest here!

100

deen @sir-deenicus.bsky.social · 25/09/2026

I could think of who would be so greatly honored, I inferred that this was he. [C. S. Peirce, quoted by Ian Hacking, The Taming of Chance, p. 207]" This sort of thing, which Peirce called "hypothesis," or (later, and less felicitously) "abduction," is very useful to us bactra.org/reviews/hhnt...

bactra.org

Holland, Holyoak, Nisbett and Thagard, Induction

010

deen @sir-deenicus.bsky.social · 25/09/2026

I once landed at a seaport in a Turkish province; and, as I was walking up to the house which I was to visit, I met a man upon horseback, surrounded by four horsemen holding a canopy over his head. As the governor of the province was the only personage

100

deen @sir-deenicus.bsky.social · 25/09/2026

It's a form of abduction known as Pragmatic Reasoning Schemas. Shalizi covers it in his book review on Induction: "More provocatively, our authors suggest that rules implementing purely-formal deductive logic may actually be less accurate than pragmatic ones. Consider the following train of thought

110

deen @sir-deenicus.bsky.social · 25/09/2026

Also recommend Children of Memory (2022) by Adrian Tchaikovsky, the third book in his Children of Time series.

010

deen @sir-deenicus.bsky.social · 25/09/2026

View from nowhere non-thinking, non-understanding pattern-matchers this capable is every bit as fascinating maybe more! Both possibilities are incredibly unsettling if investigated beyond the surface level.

121

Reposted by deen

Cat Hicks @grimalkina.bsky.social · 24/09/2026

You think it's terrifying for a wet lab to exist. I think your average eighth grader should have access to one. We are not the same

59610

deen @sir-deenicus.bsky.social · 23/09/2026

The larger the model, the more likely such nudges will stick, because of increased memorization capacity. Also there are two pathways. Something is repeated enough and something is disconnected & high surprisal. Both boost memorization chance. While not certain,impact can be higher than you'd think!

000