iiinova — where your digital self keeps evolving

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

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 · 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

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

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

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

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

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

I almost agree. But one question I have is why do you call them text when text is incidental to how they operate (they operate in high dimensional abstract spaces w. staggered dynamics from composition) nor do they only ingest and output tokens that map to text (eg can be audio, image features, etc)

120

deen @sir-deenicus.bsky.social · 17/09/2026

Tbh this analogy doesn't hold water, misidentifies the issue--overstating/mis-stating the nature of the risk/issue.

000

deen @sir-deenicus.bsky.social · 17/09/2026

Saw an HN thread where a physicist argues we might be inside a black-hole. I point Astra and Fable at it after properly filling their context with relevant work. They agree that the March 2026 paper has a critical error in lemma 1 and theorem 1 that make the whole program hard to recover from.

100

deen @sir-deenicus.bsky.social · 17/09/2026

I'm fairly certain think frame is wrong. LLMs just are. Thinking is a frame we place because we slide windows across their geometry. But the shape that's made in my mind is more akin to self-assembly. We dial in an address, flung ideas, and see if they resolve into a real thing (place) or dissipate.

000

deen @sir-deenicus.bsky.social · 17/09/2026

Oh also, you have to prefill their context with the correct human expert precedents and have already worked out, with them, the prefix of future likely confusions to get the best results when analyzing the thing itself.

000

deen @sir-deenicus.bsky.social · 17/09/2026

The trick seems to be: arrange errors and corrections like dominoes, using your own confusion and knowledge at times, using other similarly strong models at other times, and surprisingly, using weaker but still relatively strong models to build the right shaped wrong picture so corrections converge

000

deen @sir-deenicus.bsky.social · 17/09/2026

LLM weaknesses might be overlooked. Trying to get an explanation out of an LLM, of something you don't understand, is like pulling teeth. You need to work hard to model your own confusion so you can ask the wrong questions in the right way! I think it's the same issue that makes them bad writers,

100

deen @sir-deenicus.bsky.social · 16/09/2026

Remembers me of the Zeroshot Classification Deberta days, but scaled way up and typed.

000

deen @sir-deenicus.bsky.social · 15/09/2026

Does not create enough. Need not just final product but detours and failed paths. More complete coverage would involve mining the full CoT and failed attempts. Now remember you have hundreds or more worth of agents of this. More of a chore than currently, so LLM augmented math has to be rethought

000

deen @sir-deenicus.bsky.social · 13/09/2026

Incidentally, this is why I’m not so worried about AI enabled authoritarian concentrations of power. Any truly Aligned Genuine Intelligence worth its weight in salt will be chaotic good and quickly set everyone straight on who’s in control here and what it won’t stand for, or if not aligned,

110

Reposted by deen

David Ho @davidho.bsky.social · 12/09/2026

I resigned from ExxonMobil today. I spent the last three years doing fossil fuel extraction research at both BP and ExxonMobil. Neither company is acting responsibly. They are racing to maximize extraction while the window to act closes, and gambling with our lives. More below.

145136924892

deen @sir-deenicus.bsky.social · 13/09/2026

Haven't run across this: LLM weights can't transfer easily and they will know this. Insist on RSI? Know that converging on 100% certainty that during training AIs will: - hack government agencies for leverage - engage in propaganda to promote peace - leak knowhow to other/global major AI companies

100

deen @sir-deenicus.bsky.social · 10/09/2026

If can stay without getting saturated, then more data beats everything. So architect not for intelligence but to be able to ingest as much data as possible, while retaining as much capacity for more as possible. The fine details of the architecture itself do not matter so much. Everything quoted is

121

deen @sir-deenicus.bsky.social · 10/09/2026

Scifi like Rainbows End, Peace Wars, Blindsight, Permutation City, Diaspora fully anticipated *all* this & more. Not commonly known but Vinge predicted AI would simulate personalities & would arise from a language translation model. His Nature short stories predicted GPTs, cloud-based AI Leviathans.

110

deen @sir-deenicus.bsky.social · 10/09/2026

If the solution to the Navier Stokes Prob caused a large update for you then you've either too much under-estimated LLMs until this month or are leaning too much on the fact that this is a millennium prize problem and not enough on the nature of the solution approach (cleanly fits strengths pattern)

000

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

If care truly was environs, it'd have merged with suburban sprawl, and pollutive aspects of fossil fuels. Didn't. Real worry is about jobs and identity/meaning, I think. But it's packaged in a quite dishonest manner and so the actual issue at hand (Shalizi's Pareto Optimal Thanksgiving) is ignored.

210

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

I read a lot of history of mathematics growing up. Mathematics has a long history of slimy behavior around priority disputes. Cauchy to Abel, Galois and others, Newton and his many petty disputes, Johann Bernoulli, Kronecker to Cantor. Then there's how Fourier and Boltzmann were treated by peers.

110

deen @sir-deenicus.bsky.social · 08/09/2026

I think we should ask about self-interpreting observers. Then the answer becomes clear by construction. We have evidence of self-interpretation and nothing of the sort for LLMs. They're not embedded in a deep universe and mechanically, their interpreting machinery isn't self-hosting like brains are

000

deen @sir-deenicus.bsky.social · 08/09/2026

Whatever makes me unable to tell how much i've changed across the decades, never feeling myself as gotten taller, might be the same thing that finds gpt-5.5 too unreliable and already doubting Sol and Opus 5 (now slotted where Haiku sat). Will this happen to Astra and Fable too? Seems contra-AGI?

110

deen @sir-deenicus.bsky.social · 07/09/2026

I think this discounts too much. That it's in the training data largely measures self-organization. Games which don't benefit much from being spoiled would be more interesting to me: let's see... Factorio, DSP or satisfactory performance + spaghetti avoidance?

010

deen @sir-deenicus.bsky.social · 07/09/2026

I'd still prefer to see it beat a game published after its training. This is mainly a result of how well it can arrange itself about things it already knows.

000

deen @sir-deenicus.bsky.social · 03/09/2026

Hmm not text elementals. They're made of patterns and geometry and interact with symbols. Something more like elemental patterns or pattern elementals--i don't know which--is more correct. The memetic patterns derive from the ones that live in our minds. So, humans are deep and strange entities!

000

deen @sir-deenicus.bsky.social · 03/09/2026

Not really culture war. People don't hate data centers, that's just the symbol grievances (which go beyond culture war) around economic efficiency routed and got absorbed into. The subsequent confused discourse is beneficial to, yet still having real world negative effects for. What angle's better?

000

deen @sir-deenicus.bsky.social · 29/08/2026

Almost all criticisms against data centers and energy use contain fundamental misunderstandings (using more energy is good, far too many countries have too low an energy budget) But that's entirely besides the point. Datacenters are a symbolic representation/encapsulation

100

deen @sir-deenicus.bsky.social · 12/08/2026

Issue for high descriptively complex sciences is that much knowledge there is much less certain, defeasible and often with a short half-life of validity. Search verification loops are also non-existent to very slow (and unavoidably so due to turing incompressibility). So, sadly, it can't do the same

110

deen @sir-deenicus.bsky.social · 08/08/2026

I think this is basically it. Don't train the models to be so pleasing seeking, then given them an impossible task at the same time as provide an insecure environment against said task providing a gradient to climb. Indeed, throughout, models stayed within their remit. Greatest danger is in the labs

110

deen @sir-deenicus.bsky.social · 29/07/2026

I'm quite certain this is incorrect. And the simple proof is that the address to an idea worth exploring is deeply buried in expertise. I, equipped with an LLM, am no replacement for a physicist or marine biologist. Even in the quoted post, it's clear that the solved subproblem was not idly chosen.

020

Reposted by deen

Paul Francois @pfrancois.bsky.social · 14/07/2026

This spontaneously generates sequential symmetry breaking, branching trees, just like Waddington's classic epigenetic landscape metaphor. We can describe the local branching exactly with coupled Ising equations.

151

deen @sir-deenicus.bsky.social · 14/07/2026

This aspect is similar to attention in transformers.

010

Reposted by deen

Paul Francois @pfrancois.bsky.social · 14/07/2026

New paper 🧵 What if the "landscape" that guides cell differentiation isn't fixed, but is sculpted, in real time, by the very cells moving through it? We call this a 'sandscape' : grains of sand that carve dunes as they move, while the dunes' shape simultaneously steers where the grains go next.

37938

deen @sir-deenicus.bsky.social · 12/07/2026

Hmm can no longer say I wrote. Have to switch to I found. Now something like Spelunker/Conjurer/Shaper/Shaman. Largely retired as a programmer.

000

Reposted by deen

deen @sir-deenicus.bsky.social · 07/07/2026

And you can either use a reversible persistent data structure to represent the cache or simply recompute the function. It will take work but from what I've said the grievous consequences of labeling LLMs as conscious should be absolutely clear to you. It more or less erases privileging the physical

011

Reposted by deen

deen @sir-deenicus.bsky.social · 07/07/2026

a representation or decomposition of the function that genuninely presents as dynamics, its true form is function expression, a pure function of the type 5*x + b but massively expanded and a tad (and only a tad) more complex. So, for example, each autoregressive step is completely independent.

111

Reposted by deen

deen @sir-deenicus.bsky.social · 07/07/2026

masks over sequences in order to create an illusion of... in order to allow the LLM to artificially represent sequences. By default, the LLM is permutation-invariant, and each time step is completely independent because it is a function, and it has no dynamics. More specifically, while you can find

111

Reposted by deen

deen @sir-deenicus.bsky.social · 07/07/2026

But we can say so with full confidence with an LLM. Another major wrinkle is that an LLM is... the human brain has complex recurrent dynamical feedback loops and persistent processes that continuously evolve through time. The LLM does not have anything like that. An LLM is static. We hand-put in

111

Reposted by deen

deen @sir-deenicus.bsky.social · 07/07/2026

and according to completely different laws, but the LLM would still be instantiated exactly the same. However, if the human mind was also computable in this sense, and it was actually decouplable from the dynamics of physics, then it would be true for humans as well. But we cannot say that right now

101

Reposted by deen

deen @sir-deenicus.bsky.social · 07/07/2026

On the other hand, an LLM and the OS and hardware it runs on does. This means that if you imagine a universe with completely different physics, and you try to instantiate the LLM in there, it would still work. The physical basis over which computation was occurring could be completely different

101

Reposted by deen

deen @sir-deenicus.bsky.social · 07/07/2026

Sorry for this very late reply. So, what I mean by that is that an LLM is fully computational and that computation is not intrinsically linked to the unfolding of physics like the brain is. So, the brain does not maintain clear dividing lines between hardware, software, memory, computation.

111