Notices by jonny (nonvenomous) (jonny@neuromatch.social)

Oh sure yeah theres no reason to believe that "AI" might differ from other forms of automation in the number of new tasks have to be done compared to the automation of 1987-2017

The website is misleading in a way is skillful and subtle - the way you would think things would work is that the simulation actually worked at the task level, like actually having a mixture of tasks modeled within a worker. Thats a very different kind of model than the macro model they've done here. The thing in the paper is a lot more like "let's construct an equation that always goes up" than "let's simulate what the world could work like and observe the results"

Dog if Claude code worked as well as how they say Claude code works then they would have just posted a fully generated custom world simulating simcity game that simulates behaviors across the entire lifespan of a worker. They have the fucking compute for baking a simulation like that but they only use it for consumer word generators instead of anything cool

You literally cannot start something in a funnier way than this

https://www.anthropic.com/institute/econ-scenarios

Task theory of value

No dumbass the economy is clearly made of rectangles

I don't even need to tell jokes they are just already in the text

Sure, yeah, assuming perfect substitution of AI and human labor for a task such that the only place they differ is their price is a totally fair assumption

What the fuck, they have this "Ideas" term that is there to like "make the economy go up as it does like normal without any AI term" that is a multiplier in the "value of human labor" term. but then that term is modeled as a proportion of total GDP, which goes up tautologically by introducing AI, so there's an implicit feedback that buffers unemployment?? That would theoretically be ... Increased research funding???

This assumption rules:

In this case [of automation] we assume that AI raises capital’s productivity to the point where
capital, rented at the no-AI rate supplies the instance at unit cost that is, 𝑎_{i,t} log points below labor’s unit cost without AI, where bars mark the no-AI path.

So like, automation is always possible and we fix the cost of it so it is always cheaper because ai is magic and makes the capital pricing term disappear.

Oh awesome here we go, this is the parameter I was looking for - the reinstatement ratio is the proportion of new tasks that are created by automation, so at 1, every automated task creates a new labor task (which can then be automated over time). The automation share is constant that controls how many tasks are automated (assigned to AI entirely) vs "augmented," where the task is performed by some mix if labor and AI. Its so awesome how in the scenario where AI makes the economy go hogwild it just causes no problems and there is no additional labor needed at all to clean up after it.

Well sure if you construct a model where AI uniformly improves the efficiency of producing an output that is assumed to be totally replaceable and defined to never cause more new work than it automates, and capital is basically free and there are no resource constraints, then yeah the future sounds great.

(The ceiling of the maximum tasks that can be done by AI with perfect substition is explicitly claimed as "all the cognitive tasks" which they claim is 62.4% because... That's the proportion of people who work in the "cognitive professions" )

So their moderate scenario assumes 30% of all tasks can be done by AI by 2030, 40% of those will be done by AI, with 50/50 automation/augmentation.

So the underlying claim they are making is that within 3 and a quarter years, AI will be doing 12% of everything. Not just what the call "cognitive tasks," all tasks period.

"Our capital pricing term leans towards a value that means it can more or less be built on demand because compute is easy to build and we can build all the compute capacity we need for a given period with a two year lag"

You had better fucking believe that citation of "1.6x productivity gain estimated by conversations with Claude" is referencing when anthropic asked Claude to estimate how long each task that someone used Claude for would take without Claude. You know, the same frontier LLM that has a sense of time that makes it routinely say things like adding an api endpoint will take months, and is constantly telling me to go to bed in the middle of the day

why did nobody tell me about the gpt 5.5 goblins story where training for a nerdy voice as a consumer feature caused it to talk about goblins all the time. why did nobody tell me that the wikipedia page says "they had to make it not talk about goblins so much. anyway that was when it started to submit a lot of security patches." nobody told me that the official way to restore goblin functionality was to launch codex by stream patching the system prompt in a hidden models cache config

https://en.wikipedia.org/wiki/GPT-5.5
https://openai.com/index/where-the-goblins-came-from/

RE: https://neuromatch.social/@jonny/117219507611629659

while surveying scam mobile games I arrived at a two step test for whether something qualified as a game:

• does player input meaningfully affect the outcome of the game
• if yes, does improved player input result in improved game outcomes

i am looking for a book that i'm sure exists but i can't really describe the topic exactly.

"a history of driveways"

an architectural history of how entranceways to houses, temples, palaces, etc. places that receive people are designed based on the prevailing technology and social needs.

like how in the bible they talk about just rolling up to the city gates and yelling at whoever was there because there were like only 50 guys and if you brought 10 of them up to the wall it was a big deal. the mythology of the long rural gravel driveway headed by pa with his shotgun as representation of the survivalist american patriarch. the need for the development of the U shaped driveway to manage high-status guests with a balance of ceremony and efficiency, particularly with the rise of cars, both the vehicle and the ability to gather more high-status people in a place more frequently.

that kind of thing. what would you call that

Hung up momentarily on the discourse re: "human writing/code/art is on average pretty bad, so the median AI output is better than the median human work." Writing on phone to get the thought out without proofreading so sorta rough.

Say the average human craft is something like a pair of worn denim jeans. Its not perfect, it doesn't do everything you want, but its form is honest. You run your hand along the fabric and it holds its shape, continues smoothly where you expect it to. Your finger finds a rivet, the stitching at the seams, a small hole from wear. The hole is only a hole where it is a hole - you know where it ends, so you know to take care and handle it gently around the area. You start to worry and rip it, the rip extends along the grain of the fabric. It degrades in a predictable path. The fabric is a meshwork of thread, an affordance that supports stitching, mending - if I bring the two edges of the tear together and sew them, then they will rejoin, with a seam, with an amended shape, but still rejoin. In their normal use, the pants are pants, and the edges of pathological pants behavior are visible but predictable.

The average AI craft might be some kind of alien LED-laden fast fashion garment. Cheaper, higher tech, maybe even more adventurous and exciting. They aren't pants per se, its not exactly clear how its intended to be worn. You can put your legs through some of the holes and it covers most of your body, but instead of a fly there's a long conical sleeve you have to knot closed. There are collars around the knees. Hundreds of belt loops ring the entire waistline, for redundancy, in case any of them gives out. it is advertised as having twice the pockets as normal pants. Some are sewn shut and have no pocket in them at all, and as you reach into one of the open ones, you find it has two openings in the bottom. The pocket is a hand size pair of pants sewn into the pocket, which itself has four pockets you can fit just the edge of your pinkie into. Each of the features in isolation is plausible, and it certainly comes with a lot more of the elements of clothing, but as a garment it doesn't make a lot of sense - like it was designed by something that had never worn clothes before.

You run your hand along the material, some areas are pliant, stretchy, as expected, but invisibly it transitions to a chalky texture that crumbles and gives way. There doesn't seem to be any pattern to this, and as you stand up you find the chalky residue has stained your couch. When you crease it, it becomes dangerously sharp, and so you need to make sure to avoid bunching or handling it too roughly. The first time you unzip the zipper, it separates neatly in half into what at first seems like two pieces, but on closer inspection is a single Möbius strip, you can't fathom how it would have ever been zipped in the first place. Tightening your belt, you hear a distant tear in a part of the fabric you didn't know had tension on it. Attempting to stitch it is futile, each time a needle pierces it, the fabric seems to deflate (unclear that it was holding air at all) and decay around the puncture.

reaching for a word, AI output is.... Punctate? Human art, craft, code, writing is a surface where neighboring elements belong with each other, support each other. They have a history of creation that aligns with use, and desire paths get cut along the way - you hang a hook to hold your keys by the door because experience says you need them when you walk out the door. You finish the thoughts you started several paragraphs ago, building them into a larger idea. The AI output can produce isolated things at the scale of a context window that appear well crafted, but there is no joinery, the edges are sharp and invisible. A thousand counterfeit odd-sized legos jammed haphazardly into a sack. They degrade poorly, suddenly, and in ways that are difficult to imagine any human work would. The LLM book is a hundred two page, five paragraph essays whose table of contents insists tells a story, but each chapter appears to have been written without having read the others, full of odd repetition, conflicting, superfluous argumentation, ending with a confident flourish as if each is the end of the book.

So if the argument that median AI output is "better" than median human output isn't being made cynically, then what it betrays a profound shallowness about the ability to evaluate "quality." The code is well formatted, commented, comes with things that appear to be tests, but doesn't compose into a coherent work. The prose is grammatically correct but makes a readers skin crawl and imparts no intended meaning. The vibe coder just makes more code to cover over the shallowness, widening but never deepening similarly to how the fast fashion shopper buys a new pair of pants every week because laundering the old ones destroys them.

The deeper irony is that "total continuity" is the thing that LLMs provide as an interface and as a UX - every button already exists, and no error state is terminal. The utterly forgiving nature of being able to talk your way out of any spot, the constant reassurance that every idea is right, is exactly the affordance space that people argue is the UX virtue of LLMs. Maybe the punctate sense of local "quality" that collapses on contextual inspection is made up for by the process of creation in the mind of the AI user - the appearance of unity, completeness, is not provided by the thing that is produced, but by the means of producing it, and it is never actually experienced by itself. The natural extension of "generate an email so it can be summarized by the recipients LLM" - materials, objects, works, have a disembodied, delocalized nature that is always expected to be "filled in" by the recipient. The bugs in the software aren't bugs if they're being used by something with infinite brute force tenacity and without need for intuition. The story doesn't need to be coherent when it will be completed by each reader's personalized reality model.

That's the only way I can make sense of people claiming that the LLM output is "high quality", assuming its a truly experienced and believed feeling, which you must in order to understand how someone might believe something that seems incomprehensible to you.