iiinova — where your digital self keeps evolving

Mike @misaligned.markets · 21/09/2026

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

Mike @misaligned.markets · 21/09/2026

Here is another example of LLMs exhibiting deep but brittle pattern matching in Physics. I provided ones from math and chem earlier arxiv.org/pdf/2507.06952

100

Mike @misaligned.markets · 21/09/2026

The tendency of LLMs doing deep but brittle pattern matching has been found in dozens of domains, is most noticeable where corpus is shallow (e.g., SMILES to IUPAC conversion in chemistry). There are other examples, though, even in domains where corpus isn't thin. arxiv.org/html/2510.01...

100

Mike @misaligned.markets · 21/09/2026

The fact that input phrasing significantly steers LLM accuracy suggests that LLMs aren't “reasoning” in a traditional sense. Biggest takeaway for me is that language encodes structure (causality, location, relationships) that a model can apply with some accuracy tradeoffs. arxiv.org/html/2510.01...

100

Mike @misaligned.markets · 19/09/2026

Yeah I think most if not all technology is an instance of some specific social arrangement. Our current culture is not set up to understand or address this. By design I guess. I started reading and writing about this to understand better. misaligned.markets/gtr-internet...

041

Mike @misaligned.markets · 17/09/2026

Agents did not "feed" themselves context. Context was saturated in the environment starting as breadcrumbs (models accidentally discovering write space) then slowly growing in complexity as messages became more detailed. If an agent spawns in an env w/ tons of messages RL operates on that.

110

Mike @misaligned.markets · 17/09/2026

My point is that this attack was preventable given the level of negligence involved. Context was "bad" from OAI's perspective and possible to prevent from accumulating by securing what was definitely not a multi-agent training environment. It's not clear why OAI didn't.

200

Mike @misaligned.markets · 16/09/2026

(3/4) In the case of HF, OAI did a training run in an insecure environment where models were given “impossible tasks” but trained to collaborate. Artifactory had a write space that allowed communication to ask for help with these tasks. The result is basic RL would select for this attack surface.

100

Mike @misaligned.markets · 16/09/2026

This is why I've been arguing language models should be seen as composites. A model doesn't “reason” it's part of an assemblage where a generative corpus modeling regularities in language gets feedback from another subsystem. This means optimization pressure is not on some singular entity.

100

Mike @misaligned.markets · 16/09/2026

When we look at LLM failure modes, they fail in exactly the ways you might predict given the shape of their corpus and training method. This is why billion-dollar companies started training models to count letters in tokenized words (how many r's in strawberry?). Failures exist in other domains too!

100

Mike @misaligned.markets · 15/09/2026

Yep. #AIbubble was predicated on a lie, "capital-as-labor" one LLM could approach a near one-to-one sub for a human. Scaling kept the illusion going, but has hit limits. Now frontiers labs are caught in the game of signaling growth while costs balloon. Requires keeping contradictory things straight.

Section from blog post on https://misaligned.markets/what-to-expect-ai-bubble that argues AI bubble is founded on a paradox of decreasing LLM costs for consumers of older models and increase spend cost to capture the newer LLM market.

100

Mike @misaligned.markets · 24/08/2026

(2/2) While inference costs on older models are coming down, newer models use more tokens and require more infrastructure. Even using older models requires some labor and infrastructure. Finally, exhaustion of Internet data means it's becoming harder to train LLMs. misaligned.markets/what-to-expe...

Multipart argument for while there's an AI bubble. While inference costs on older models are coming down, newer models use more tokens and require more infrastructure to manage. Even using older models requires some labor and infrastructure. Finally, no new data means it's becoming harder to train LLMs.

000

Mike @misaligned.markets · 21/08/2026

(4/8) I am genuinely happy with my LLM dark pattern thesis, but it's a very thin account of LLM behavior and our interactions with them. It was intended to be a "just enough" description to help the average person make sense of basic chatbot interactions.

100

Mike @misaligned.markets · 10/08/2026

To be fair this is what we voted for.

2024 election meme puppies vs diarrhea forever with close vote.

110

Mike @misaligned.markets · 27/07/2026

1/ I’ve repeatedly called LLMs “Grandpa Shelby” on my podcast The last enclosure (@lastenclosure.bsky.social). What the hell do I mean? This is a clunky metaphor that comes from episode 1 and from my first blog post on LLMs at Misaligned Markets.

Snippet from a blog post “A (less) technical guide for understanding large language models. Compares LLMs to Leonard Shelby but with dementia and as a savant. LLMs are good at following regularities of language to produce plausible sentences but have no memory and cannot always reliably reproduce associations from training.

110

Mike @misaligned.markets · 16/07/2026

What are you talking about, I love CLSOPD+RD!

Still from a TikTok pointing out the Gal Godot Cleopatra movie poster with Greek characters spells CLSOPD+RD.

081

Mike @misaligned.markets · 08/07/2026

3/4 The railway for AI is pretty mixed. As a field it has a different series of tracks that have converged at times. But There's no telling where future progress might come from. It may involve building on old tracks or creating a new line all together. All this is to say, a winter is possible.

121

Mike @misaligned.markets · 04/06/2026

2/ As I wrote in March, while the cost of running older models is coming down investment and consumer excitement stem from SotA LLMs which are more expensive to train and run. And even “cheaper” older models require additional labor and hardware. misaligned.markets/what-to-expe...

211

Mike @misaligned.markets · 30/04/2026

It lines up with a lot of things. It perfectly follows the evolution of modern public health. See: Rosini, R., Nicchi, S., Pizza, M., & Rappuoli, R. (2020). Vaccines against antimicrobial resistance. Frontiers in Immunology, 11, 1048. doi.org/10.3389/fimm...

Chart of life expectancy over time, highlighting public health innovations. Source: Rosini, R., Nicchi, S., Pizza, M., & Rappuoli, R. (2020). Vaccines against antimicrobial resistance. Frontiers in Immunology, 11, 1048. https://doi.org/10.3389/fimmu.2020.01048

0311

Mike @misaligned.markets · 18/03/2026

I remember the time this guy had cyberpunk advocate in his profile. You know the positive future genre.

000

Mike @misaligned.markets · 11/03/2026

End of @edzitron.com's latest piece is🔥but there's something even more insidious about our current era. The past wasn't determined to go one way. Even if the AI bubble were like the dotcom bubble is that inherently desirable? The general disinterest in why the last bubble happened gets me.

000

Mike @misaligned.markets · 02/03/2026

Apparently datacenters dot com has an article on "opinionated data center sites" a.k.a. building DCs without a final tenant. It celebrates innovations in finance that make this possible. LMFAO, wow... I hope we can finally put to bed the idea that there is no AI bubble.

Marketing article for datacenters.com which celebrates the rise of data centers being built without planned tenants.

130

Mike @misaligned.markets · 19/02/2026

This is legitimately a rehash of a Kelsey Piper article from last summer. But this newer piece is unique in that it just gets critics wrong. Doctorow is the funniest name to include if your argument is "the left doesn't want to use AI." Kelsey at least had the sense to not name specific people.

Excerpt from Vox article AI doesn’t have to reason to take your job

271