iiinova — where your digital self keeps evolving

Mike @misaligned.markets · 12h

I think the corpus is the mirror. The model doesn't have "understanding" of it, but by reflecting regularities in the corpus it gets a bunch of modalities. Accurate language use encodes relationships and structure in the world. "You shall know a word by the company it keeps" and all that.

120

Mike @misaligned.markets · 12h

Wait I thought this was a meme? It was a real debate?

000

Mike @misaligned.markets · 12h

Have to check this out! Think this is politically fraught because it's wrapped up in notions of a "social contract." I've been meaning to write a piece describing how the perceived social contract in anglophone countries has changed and how reporting has been slow to validate and reflect that.

000

Reposted by Mike

Mike @misaligned.markets · 25/09/2026

I don’t even work in the field (cybersecurity writer) and I understand just how negligent OAI was (while bragging about it). misaligned.markets/antidote-ai-...

misaligned.markets

An antidote to AI hype: AI agents are not “going rogue”

Dangerous, automated security tools existed before LLMs, and AI agents don't change cybersecurity fundamentals.

Colin @colin-fraser.net · 28/09/2026

Reposted by Mike

I think if you meaningfully punish OpenAI & co for all of these cyberattacks then they will stop doing them. Whether or not you want to anthropomorphize the agents, you gotta just land there.

537570

Mike @misaligned.markets · 28/09/2026

I'm definitely anti-anthropomorphization but the first image that came to mind was Tommy from the Rugrats being placed in his playpen with the very toy screwdriver he uses to constantly escape. All of that is to say any "alignment" problem here is not solely machine related.

050

Mike @misaligned.markets · 25/09/2026

I don’t even work in the field (cybersecurity writer) and I understand just how negligent OAI was (while bragging about it). misaligned.markets/antidote-ai-...

misaligned.markets

An antidote to AI hype: AI agents are not “going rogue”

Dangerous, automated security tools existed before LLMs, and AI agents don't change cybersecurity fundamentals.

Mike @misaligned.markets · 25/09/2026

Uh… ants. Does that count?

001

Reposted by Mike

Colin @colin-fraser.net · 23/09/2026

This is a nice example of the kind of clear non-reasoning that is increasingly hard to spot, but is still in there. It’s strange to say that 9.11 - 9.9 = -0.79 *because* 9.9 is greater than 9.11. This is kind of a weird random unnecessary side-note rather than a step in a sound argument.

2372

Mike @misaligned.markets · 24/09/2026

Oh, we solved alignment!

I am struggling with grief and regret after losing someone after failing to get them the support they needed… this realization has been one of the most important lessons that has made things somewhat bearable.

Bronwyn Douwsma @ninesnowboots.bsky.social · 23/09/2026

Reposted by Mike

I'll write stuff but not now.

3132

The refusal to accept next token prediction as an answer I think emerges from the fact that everyone underrates how powerful next token prediction is. Language encodes causality, and real relationships in the world. Modeling how close every word is to another you unlock these encodings "for free."

0101

Reposted by Mike

We kind of do know what LLMs are doing. They model regularities in language which includes detailed human plans and reasoning as well as a general shape of the world. AI veterans like Michael Wooldridge note this.

121

Reposted by Mike

People who want to anthropomorphize models without caveats are falling for a UX illusion where everything that enables the model, from training data to subsystems is treated like a single, emergent mind. Failing to account for what models can and can't actually do.

131

Reposted by Mike

depths of wikipedia @depthsofwikipedia.bsky.social · 23/09/2026

A recent paper titled Why LLMs Aren't Scientists Yet pretty much highlights some of the same failure modes. Models are deep, but brittle pattern matchers constrained in part by both their data and their context windows arxiv.org/abs/2601.03315

101

Reposted by Mike

when i mention "claude" you can assume i'm talking about the albino alligator not the llm

523852648

The latter also requires the former. Other cognitive agents (people) assigning meaning to words that models can index on. In tandem w/ designing external systems that guides model behavior. Human cognition does not rely as heavily on regularities/external scaffolding. But this is still impressive.

000

The LLM arguably is doing something more powerful (but brittle). For example, we learn causality through experience it “learns” causality from distributional semantics (log + fire = ash). These aren’t the same. It can index on regularities in semantics, grammar, syntax without comprehending.

hikikomorphism @hikikomorphism.bsky.social · 22/09/2026

The only other non-standard usage I’ve seen is “ghosts.”

100

Reposted by Mike

LLMs aren't made of neurology except in the most handwavey sense. They're made of narrative. That's why I started calling them text elementals: they're not human, they're most-likely-text generators we bent into living stories in a prose format centering the characters User and Assistant

631852

As someone who lives by the words I write, this is the only thing that matters lol.

Sam Hogarth @samhogy.co.uk · 22/09/2026

Reposted by Mike

The human in the loop is a neck in a noose

There's a whole UX layer hidden from the average user meant to present the model and its subsystems as a single entity that is “thinking.” Unveiling that reveals that while models are useful they shouldn't be seen as some sort of (proto-)intelligence.

Most modern LLM systems are composites where subsystems (knowledge graphs, schema like frontmatter, skills-md, layered RAG) can compensate for some of these deficiencies. But this externalization of "thinking" suggests that LLMs "think" in a way that's not comparable to what we do.

When LLMs fail they do so in ways that a deep but brittle pattern matcher would: - Training method excludes it (counting tokens to "know" how many letters are in a word) - Input is phrased in a way not captured by the corpus - Input requires model to generalize far beyond presentation in training

121

Reposted by Mike

A (less) technical guide for understanding large language models

If you want to anthropomorphize an LLM then the model itself should be seen as a savant dementia patient with anterograde amnesia. Systems like RAG and schema like frontmatter and skills(.)md, designed by REAL cognitive agents (people), allow a model to externalize its "cognition."

misaligned.markets

Disambiguating LLMs from other machine learning tech is the first step to making sense of the current AI boom.

111

Reposted by Mike

✨💜 king 💜✨ @king-reinhardt.bsky.social · 21/09/2026

No, I'm not published. Yes, I am still a writer. I've written stories since I was a kid, and I have no intention of ever stopping. You ask me, the writing itself, done purely out of love for the process, makes you a writer, whether you're published or not.

3159976

This is genuinely the funniest and most accurate description of the LLM-based AI race so far.

Colin @colin-fraser.net · 21/09/2026

Reposted by Mike

I feel we are wasting some time, effort, and resources on making sure that Claude is a certain kind of guy when really all we need are the facts that he knows.

4999

Reposted by Mike

Kevin Riggle @kevinr.free-dissociation.com · 21/09/2026

And me saying that LLMs don't work for the use-cases that I have doesn't mean that it doesn't work for the use-cases that you have, despite that I do tend to give you all the hairy eyeball over a lot of what y'all seem to think that it works for.

191

Could some future hybrid system make use of a language model's encoded regularities for more structured domains? Maybe, I don't know. But that doesn't require us to pretend that models are like humans. Also doesn't mean that models are useless. They just have to be scoped to specific tasks.

000