When I was at school I remember finding an advanced physics text book full of really interesting problems. I used a calculator because the arithmetic was not the point and was a distraction from engaging with harder content and higher level reasoning.

This sounds very familiar to me. I spend a lot more time reading, writing, speaking and listening to natural language than before and much less time coding. Dialogic free writing is a good phrase.

really good from @grimalkina.bsky.social. I’ve been thinking about writing something about a habit or maybe skill I have, which is that I’m good at treating things I hear as tentative and holding onto them for later confirmation (or not) www.fightforthehuman.com/cognitive-he...

guys, you’re missing the point. AI doesn’t have a purpose like parking lots, which help me fit fossil fuel burning more conveniently into my day

oil companies watching "environmentalists" focus all of their attention on data centers

To clarify this isn't to say it's bad if organisations pay for canopy solar. Just that the framing of it as a major solution as if space were the main constraining factor rather than being a consideration along side cost.

Small reminder that parking lot canopy solar is MUCH more expensive than utility scale field solar. For the same investment you'll get several times less capacity. It mostly solves for not having to think too hard about broader land use policy and economics.

This is not "whataboutism". If you are serious about these issues you need to know what the levers are, all of them, how big they are, and the costs/constraints of pulling them. If you're only thinking about one lever, you're thinking about the lever, not the system.

There are a lot of people who by their actions and blue whale sized omissions in their concerns show they don't actually functionally care about the environmental issues they say they care about.

I guess we're going to need to learn the hard way that you can't eat cotton... nor the varieties of corn grown for ethanol.

Everyone was harping on about model collapse, but as far as I can tell this is the clearest example of the proposed failure mode and while weird, hardly counts as collapse!

The question of LLMs is not "what can it do for me" but "what kind of cognitive ecology can I build with this?"

In parallel, I'd expect anyone in a position of authority to be able to articulate *why* on this matter. If your argument for changing a scientific classification is "because i have the feelies", you are too immature or feeble minded for public office.

The data center and renewable energy discourse reminded me of a thought experiment a explored a few months ago:

Rage as national newspaper uses manipulative framing for political ends.

Can we have a moral panic about golf courses please? Such a panic, if even marginally successful would have significant positive environmental and social impacts.

This is an excellent conversation. Now that the LLMs have gotten really good, it’s interesting to see (at least) two types of outcomes. If you can’t think, you’ll get much better and more stylish boilerplate. If you can, LLMs can facilitate interesting writing. I see a lot of the former around.

I think we should happily stop trying to determine if LLMs have "consciousness" or "sentience" and recognize they have some kind of completely novel cognition and just try to understand the consequences of that

The appropriate response for a colour blind person being repeatedly asked to distinguish colours they are blind to is "no, fuck off". The appropriate "intelligent" response to these questions to LLMs should be the same.

Also very much this.

100% this. Using LLMs well requires developing a "theory of LLM cognition". That structure exists whether or not you accept the language I used. It applies at a more limited but rich scale for all expert tool use.

There are two very different epistemic posture on show here. If you care about understanding reality, you'll engage with one of them while checking sources, thinking critically and pushing back if you find mistakes. The other one: you should flee from and never look back.

I don't think it is remotely 1:1 but it's one of the more clear skill transfers at least.

"Argued into existence" New essay on authorship when writing with LLMs: marenonnostrum.com/essay-on-aut...

I'm inclined to disagree. Using LLMs is ultimately a kind of collaboration and communication activity. We widely accept that good communication is a skill or a set of skills relating to different communication contexts.

Unpaintable — a quality of effectiveness that resists celebration. "Boy, we're unpaintable, aren't we." — Dockworkers finishing a hull repair at midnight, in time for the morning tide. One of 16 idioms from the world of Mare Non Nostrum.

if the presence of hype makes you completely unable to think about reality independent of the hype, that is not resisting hype, it is surrendering to it totally

I've been messing with creative writing projects with LLMs for several years. It was the recent harness improvements of Cowork/Code plus recent models better at long context generation which enabled this project to work.

"Rome's problems produced violence, which at least had the virtue of clarity. Carthage's problems produced meetings." Mare Non Nostrum

Carthago capta est — a grand declaration that means nothing because there is no one to hear it. "I win!" "Carthago capta est, little general — your sister stopped playing five minutes ago." One of 16 idioms from the world of Mare Non Nostrum.

(On the quoted article), the existence of bikes or taxis does not invalidate the value of participation or entertainment of the sports of running, cycling or motorsport. It's simply worth knowing which sport you're actually participating in.

Having written a 125k word book with AI, I can attest that if you have standards, AI certainly doesn't make writing anything substantial "easy".

When writing Mare Non Nostrum with AI, challenging as it was, the driving force was always "how can I make this interesting, entertaining and thought provoking to read". While it was hard to write (even with AI!), that's not a message I care for the reader to hear...

"Argued into existence" - the kind of AI involved writing where it takes well over a million words of dialogue to get 125k words in high quality book form. Full essay on the process here: marenonnostrum.com/essay.html

When Claude Cowork came out, I set myself the challenge of doing something substantial with it. Something high quality, not just a weekend project, and not something I could have attempted alone. What I ended up doing is something I'd never have put on my bucket list. I wrote a book.

My evidence-free guess is that the students who use LLMs as careful exploration and dialectical machines are running laps around others right now -- including the answer machine kids AND the refuse-to-touch-em kids.

I've seen a number of people who can't entertain the idea LLM use isn't cognitively damaging but I think the real issue is up stream: they don't entertain it as anything other than an answer machine.

Concern about the cognitive effects of LMs is based on the premise that they're an answer machine. And they can definitely be used that way. But I think ppl who are using models to solve actual problems (not homework problems) experience the tech as a free-writing, dialogue, and revision machine. +

There is going to be a big wave of "omg what if you didn't treat this as an answer machine" "what if we rediscover active learning" commentary from CS in about six months, especially the voices currently describing everything as "brain rot and cognitive decline," and I'm already a bit tired

If you’re an American should watch the Anthony Bourdain Parts Unknown ep on Iran. A deeply humanizing look at a country full of people just like you and me. I watched it after Anthony died and it moved in ways that are hard to relay. I also immediately ran out to try kabob koobideh and tahdig.

Almost everything you do on your phone is backed by a data centre. Your phone is just how you access data centres without knowing what you're doing.

A latent argument of most anti-AI, anti-datacenter positions is that the ~75% of all datacenters devoted to non-AI use is totally rational while the remaining 25% used by AI is unpopular, irrational, and hype-based waste. It says "*My* use of datacenters is justified. Yours must be destroyed."

this is why overgeneralized stigmatization of AI is bad. not everything AI helps is AI slop. not everything written without help of AI is a work of art. it’s not easy, but we need to be more nuanced about these things.

It breaks some of them, but we read a lot on the basis of recommendation and review/curation mechanisms (like academic journals). Yes LLM writing breaks easy heuristics from the text itself but there are and will be other ways.

There's no avoiding it — LLM written code and content *will* be all over the place. What we need is not to tell *if* something was influenced by an LLM, but better ways to tell if something was thoughtfully created and thoroughly tested.

I suspect a lot of people not only don't know what generalisation means in this context but have strong intuition that nothing beyond verbatim repetition is possible or would come from "next token prediction".

My favourite genre is "let's pretend this high dimensional space is just standard tabular data and apply trees to it so we can avoid engaging with a fundamentally different and brain melting reality".

The reaction to data centres is bizarre. It's like people saying roads are bad, they cause local noise, air and run off pollution. All true, but then they conclude we should destroy all roads...

Using origin as a proxy for quality or even using a proxy for origin as a chained proxy for quality... is a pattern which concerns me. I sure hope those who aggressively practice it about llms don't import that reasoning pattern into other parts of life.