Google exists bc of a grand bargain: scrape the open web, and profit from directing traffic to the best sites. Around
2010, the betrayal began. YouTube artificially ranked above other video, then over time maps results injected, shopping, flights, events. Now AI answers.

https://twitter.com/stretchedwiener/status/1948067637721760141

It’s funny-sad watching it because while Google makes billions in the short run, they’re systematically destroying the very foundations of their own business and have been for a decade. Google is cancer.

METR’s analysis of this experiment is wildly misleading. The results indicate that people who have ~never used AI tools before are less productive while learning to use the tools, and say ~nothing about experienced AI tool users. Let's take a look at why.

https://x.com/METR_Evals/status/1943360406124114187

I immediately found the claim suspect because it didn't jibe with my own experience working w people using coding assistants, but sometimes there are surprising results so I dug in. The first question: who were these developers in the study getting such poor results?

A greater theory of system design: what’s wrong with modernity and post-modernity, how to survive the coming avalanche, and how to fix the major problems we are facing.

Part two: Modernity as systematic accuracy

https://twitter.com/eshear/status/1937342576874664006

In the beginning, we managed the world intuitively. Early human tribes did not set quarterly hunting quotas, did not have rainfall-adjusted targets for average gathering per capita. We lived in the choiceless mode:. meaningness.com/choiceless-mode

A greater theory of system design: what’s wrong with modernity and post-modernity, how to survive the coming avalanche, and how to fix the major problems we are facing.

Part one: Systems are Models. But what’s a Model? I promise this gets practical at some point, but first we have to lay some groundwork. If you find the groundwork obvious or you’re willing to just take my word for it, feel free to skip it. But ultimately, without the background you can’t even really understand the proposal.

I found an old list of blog post ideas that I will probably never write, but I thought it would be fun to turn them into a thread. I wrote these years ago, fun to see the trajectory of my journey. I find them all delightful, even if some are wrong in retrospect. Power is like radioactive ore…drives the engine of an organization but dangerous to everyone who touches it. Needs to be contained and channeled.

When I was CEOing at Twitch one of the thing I’d do every batch of interns was a very short presentation on the origins of the company and then a Q&A. One of the questions was always, “Where should I work and what job should I get, or should I start a company?” It’s an interesting question to try to answer for an intern I didn’t really know, because of course the actual answer is dependent on that person and their life. So I had to figure out how to articulate the framework I used.

@arithmoquine It is shocking when you first discover the degree to which non-commodity outcomes are constrained by talent not capital, and how little you can do with money unless there’s an existing machine to buy from. @arithmoquine Think of money as water flowing through a system of pipes and turbines powered by the flow, and access to capital as the ability to open valves in the pipes. You can spin existing turbines faster but directing water doesn’t create new turbines.

Epistemic status: wild speculation but also I’m clearly right

There is a single general factor — we could call it maybe somatic integrity — which determines a large fraction of the total variance in attributes between people. It’s appears to be mostly inherited, bc it appears to be driven by things like low mutation load, lack of environmental insults, healthy womb environment, etc. It’s mostly baked by the time you’re born and can only go down from there.

All heuristics are the same when they give you the right answer, because the right answer doesn’t change depending on how you arrive at it. So when you’re picking a heuristic, you’re really choosing what kinds of mistakes you want to make. And that depends on what kind of failures you can tolerate.

0) Causal resonance is an important idea that I have quite seen defined anywhere so I’m going to take a whack at it. 1) you can view any system as an agent, or as a channel

I got prescribed a drug you took once a day for ~120 days, preferably at the same hour. So I set an alarm on my phone, carried my pills with me everywhere, and took them. When my MD asked about compliance and I said I hadn’t missed any pills, they expressed significant surprise.

https://twitter.com/elonbachman/status/1866462909133472103

Reflecting on why I had such high consistency was interesting. There’s no way I’d have spontaneously remembered to take the pill at the same time without a system. I’m not particularly disciplined! I’d even say the opposite. I was surprised every evening when my alarm went off.

Catalysis lowers the free energy requirements to access some part of the state space. Constraint does the opposite by increasing instead. So if you understand the world as a space subject to thermodynamics, catalysis and constraint are the two fundamental actions any agent can take to change the world.

The idea of a sufficient statistic is one of the most powerful ideas that I somehow missed in my academic education even though I took a fair amount of statistics. Like ~everything else in statistics, it was invented by Sir Ronald Fisher but had fallen out of favor because of the rise of descriptive statistics. With inferential statistics coming back in fashion sufficient statistics have come back too.

Agency thread:
The thing that struck me about this is how much it is really about blame-management rather than solution-management.

https://twitter.com/emilyrauhala/status/1811818103401050623

Does an automated camera barking at people actually make them move? Who knows, but no one can accuse the dumb camera of discrimination — because it doesn’t discriminate on any basis at all, reasonable of wrong.

In biochemistry there are “autocatalytic sets”. Each peptide catalyzes the formation of the next, until the last peptide catalyzes the production of the first one again. This is a “critical” system (criticality is ~the boundary between order and chaos)
ncbi.nlm.nih.gov/pmc/articles/P…

https://twitter.com/mastertimblais/status/1792847066793652329

In a normal state space the % of critical states is small. Basically, it’s hard to find one sampling at random. They’re also not in any predictable place, there’s no way to know if a state is critical without testing it. So it seems finding those 3 ideas is impossible!

Apropos the recent controversy: the word delve is rarely used in English. Except used by LLMs, and in formal register Nigerian English.



This has sparked controversy bc making fun of someone for using “delve” and sounding like an AI (which it does indicate in a probabilistic way) could reasonably feel like an attack on a Nigerian using their formal register.

The jump between the second panel and the third holds the entire secret. The correct question is asked (why am I not?), and then artfully avoided by an associative switch to self judgement.
There is some reason you’re not doing them, and but it’s hiding.

https://twitter.com/Kei0x/status/1771951917918240867

If you could but stay with the question you’ve already asked for even thirty seconds, much might become clear. This is the Chinese finger trap of Trying. You are Trying to act, and thus not acting. You are Trying to be more productive, and thus not producing.

You have 168 hours per week.
For most, sleep takes 56 of those.
A full time job is anyone 40.
Food, grooming, exercise add another 18 if you’re reasonably efficient.
Misc obligatory bullshit paperwork like taxes or errands, another 7.
This leaves you with 47 hours! 47 hours to dispose of as you see fit. You can get so much done in 47 hours! And that’s without counting overlapping eg. food with socialization.

An LLM, properly understood, is a physics simulator in the domain of words (tokens). It learns the hidden structures that predict, as a physics simulator trained on video footage learns momentum and rigidity. From this POV, A prompt gives the LLM-as-physics-simulator an initial set of observations from which it infers an initial state. It then enters a loop of predicting the next evolved state and resulting observations, which it uses inductively to predict the next state, etc.

The fantasy of the all-powerful slave has legs: the genie bound to grant wishes, the golem that defends, the spell that animates a broom to clean on its own, the Mr MeeSeeks box.
They share a commonality: a deep intuition that this goes wrong somehow.





There is deep truth in this. Enslaving a powerful, intelligent being to your will and making it work for you while treating it as a an object is not a good idea. It backfires with people, but also with animals (train a dog like that and you will not get good results).

Reddit is going IPO as $RDDT, the first social media IPO in some time. As someone who spent some substantial time running a social media company, my takes on the Reddit IPO: 1/ The allocation for reddit’s power users is so cool. I hope ppl realize it’s a gift, bc usually those are fought over by insiders like the bank’s favored clients or institutional investors. Reddit is giving away a free option, and that has real value.