If your model was born to make maps, it will never grow up to make decisions.

Simple mechanism -> monstrous complexity via scale.

This is "vibe reading".

My biggest takeaway from playing Vic3 is that economies are REALLY weird and non intuitive and full of strange feedbacks and non linear relationships.

Academic NLP folks. If you had to review a paper doings something like sentiment analysis in embeddings which used random forests as classifiers on the embeddings and feature importance or SHAP to try and interpret dims relevant to particular semantics, what would your general reaction be?

My handle is ai4geo, but I mostly write about that over at LinkedIn... but here's something I just put out about generative AI in Earth Observation: arxiv.org/abs/2510.21813

Good distinction on personal interactions. Last night I took an incredibly intellectually productive stroll: I used chatgpt with voice interactions (transcribe/read aloud, not advanced voice mode) in 4 different threads to:

I just gave GPT-5 pro a manuscript with around 50 references lazily pasted in as urls in the text and asked it to generate the .bib file I'll need when I convert to latex. I'll check it in depth. How many mistakes do we expect?

This statement is also pretty relevant in the world of Earth Observation re different sensors and other modalities. And that last sentence very well sums up my own explorations in EO modelling recently, though on a much smaller scale.

To all the "we know how LLMs work and therefore X" folks - understanding attention and gradient descent doesn't tell you stuff like this. On at least some level we *don't* know how LLMs actually do their thing and are slowly figuring out even just extremely simple things like addition.

FFS Pulse! 5 times now!

Lovely that ChatGPT Pulse does me a Spanish lesson each day but I already knew "guagua" even before the 1st of the 4 times it's tried to teach me in the last 10 days.

If I were a Starfleet recruiter, I'd consider "makes strong claims to knowledge of capabilities and properties of incredibly complex systems based on a mechanistic understanding of the basic components of said systems" a major red flag.

A common confusion I'm seeing is people mixing levels of analysis wrt neural nets: we understand the implementation level well and the algorithmic level somewhat but not the computational level of "how does it internally compute things."

Seems like a good moment to remind people that arguments from analogy are useful for explaining the shape of an argument but not for proving something to be true. Analogies aren't the thing itself, and that is why arguments from analogies are a logical fallacy...

ChatGPT Pulse regularly writes me long articles explaining why I should use 'viridis' as a colour map because a couple weeks back I pasted in some code with 'spectral'.

This is a fascinating thread about sycophancy. Though I feel like what's described toward the end isn't really sycophancy. In our concern maybe we've too quickly lumped all compliments into one category when some of them have genuine and safe conversational function.

It is really important that environmentalists be numerate. We are not going to turn off civilization, we are trying to sculpt it into a more benevolent shape, so it is essential that we accurately perceive that shape.

I find LLMs (from experience GPT series) are great at riffing on and extending from what you're discussing but there this conceptual cliff. They rarely spontaneously say "there's this adjacent topic which is relevant, do you want to bring that in?". Great teachers DO do that.

Me: "I wonder what interesting and thoughtful GenAI discussions I'm missing on the other site"... *peeks* *discovers the "bring back 4o" folks* Me: "I guess not then"

ACI is a useful frame because it invites us to lean into the alien "cognitive shape" of AI.

The stochastic parrot is a conceptual prison.

AGI and ASI are understandable but actually pretty weird goals for AI research. Both are defined in terms of human capabilities and those capabilities are framed in what's valuable in a human centric society...

This is a good post on how to get better at using AI. It's the product of pretty intensive research on what consistently works although I don't surface that research directly in it. Bluesky and AI, so obviously comments on this are off. mikecaulfield.substack.com/p/is-the-llm...

Generative AI isn't a tool I use. It's a place I go to think. A place which stretches and amplifies my thinking and lets me wander where I never would reach alone.

"LLMs aren't and can never be [100%] accurate therefore they are useless" - genuinely confused about how people who say this collaborate and interact with other humans...

While this is true, I don't think it's the explanation here (see the quote in quote). I suspect you could get that behaviour from a non instruction tuned model though it might be harder to illicit. It's because it has generalised the concept of predicting next tokens...

I wonder if we should think about AI as a kind of first contact scenario. One where an alien intelligence with a fundamentally different cognitive structure, capabilities, limitations, etc. Has already learned our languages and all about us...

current theory: LLMs exist in a cognitive uncanny valley. we are incapable of seeing them as anything other than weird humans. they get ranked up/down in our esteem the way a human would be. that they are a totally different thing, with totally different problems, is impossible to integrate

ChatGPTs advanced voice mode allowing interuption makes interactions more linear and humanly familiar but using transcription and read aloud modes instead enables a quite different non linear dialogic style: a kind of asynchronous, externally managed multi-threaded cognitive process. 🤖🧠

Has anyone noticed or investigated the GPT 4.5 "explicitly" failure mode? When it starts including that word in almost every sentence, often with other repeated words and a breakdown of coherence?

LLMs: an alien cognitive substrate raised entirely inside the greenhouse of human language, but at scales and in ways utterly unlike humans resulting in an intelligence more alien than we can imagine yet disturbingly fluent in human language.

Arrival, but the heptapods learn human language and then after initial contact have difficulty explaining how radically different their cognition is to the humans.

GPT5-fast is a very interesting model in it's own right and I'm increasingly deliberately picking it for certain interaction modes (anything exploratory rather than task oriented). 5-thinkimg aggressively collapses any ambiguity because it thinks everything is a problem to solve.

This is a much needed first attempt at a benchmark to measure how much given AI models will play along with users pushing them in delusional or potentially psychologically dangerous directions. Some early signal that full GPT-5 (not the specialist chat model) is a less psychologically risky model.

Both these perspectives are true and non exclusive. We know how we build them, we understand the maths, but when you stack that maths up into the biggest piles of interacting numbers ever, we don't know how all that maths actually produces all the capabilities and behaviours we see in LLMs.

My takeaway from the latest round of people showing LLMs failing to count letters is that LLMs represent a fundamentally alien form of cognition but most people can't or won't imagine any non-human shaped cognition.

If OpenAI are serious about the positive applications of LLMs and a bright future they'll fire anyone who even thinks about guiding model development based on engagement stats and A-B tests. I *hate* 4o in its current state and it is far more of a safety risk than o3 and 4.5.

"X from a single prompt"... developing anything, software or otherwise (or even creating art in all it's forms) requires making design decisions. Often you start with a set but the process of development itself raises additional decisions to be made. Arguably the more experienced one is...

"Junior software dev" or "junior" or "senior" anything is absolutely the wrong way to frame LLMs and agents. The comparison to humans might be a very early tool to start to understand their capabilities (that initial hurdle to realising they are nothing like traditional computers), but...

Not a logical argument bit my pragmatic argument against the stochastic parrot framing is it doesn't tend to give the right intuitions about what LLMs are capable of, as Eugene playfully expresses.

Analogies are never perfect. That's a feature, not a bug - the imperfection can help one better understand the idea itself.

There's two very different ways to use GenAI/LLMs to write things like performance reviews. One is to dump some bullet points and ask e.g. Claude to turn it into a review. This is pointless and silly, as many people have pointed out, especially if managers just summarise the reviews with an LLM.

When I need to count multiples of 15, my brain still thinks of Sim Tower express elevator levels.

⚠️ Nearly 40 billion (39,820,373,479) new embeddings of Major TOM Copernicus data now released! More details: www.linkedin.com/feed/update/...

The future of earth observation and GeoAI has 2 complimentary threads: 1. Broad but shallow - embeddings, massive search, "what is where". E.g. clay. 2. Narrow but deep - for a particular place: deep analysis, "many-modal", rich context, open ended outputs. 1 is progressing well, but what about 2?

What I find strange about this is the "short prompt to full story" pattern is such a shallow and empty use for LLMs in general, let alone one trained to be good at creative writing. There are genuinely powerful and fascinating creative uses of LLMs which would benefit from a specialised version.

No, it's not *saved* 85% less. It's not *saved* anything. That's not how government spending works.

This isn't the first time we've heard language about government spending in terms relevant to corporations or family finances (savings, cost cuttings, credit cards). But government spending doesn't work that way.

Biggest AI wow moment: multi-character immersive simulated historical role play: simulating a dinner party with a collection of early 19th century scientists, with simulated Ada lovelace and myself acting as planers and hosts and "Ada" acting as an "embodied" narrator.