I make things with AI that I find interesting or useful.
Alex A
Karma: 250
I’ll agree that learning is a means to action/decision-making. And I agree that ideally, one would focus their learning on the most important problems, etc. But I have three concerns here:
I’ve found that focusing on “the most important problem” is not a great foothold for research, especially because it neglects personal fit. And without wrestling with the technical side of these problems yourself to develop a deep intuition of why a problem is most important, the default state is deferring. And having a novel take without full technical understanding is usually going to be bad for a new entrant. Even if it’s not a bad take, a new take could signal misalignment or misunderstanding to someone in an org who doesn’t have time to evaluate it on its merits.
If this advice is for generalists wanting to support AI safety, however, I think it misses the meta layer. An org needs to be able to tell from afar that you have the right context/understanding of the problem, capability, alignment, and that you—as a generalist who is entering from outside of the effort—will not immediately abandon the effort for something new and shiny. Even if you develop a deep understanding, you need to be able to show it, and show it in a trustworthy way.
I think the most valuable knowledge for generalists to have would be tacit knowledge of how the AI safety community and orgs actually work/operate in the real world, because that’s the scaffolding on which you can identify areas for improvement and high value actions. Even though orgs make an unusually strong effort to put out relevant information online, it’s still not enough about actual dynamics. Getting hands-on experience with work in an org seems like it would help resolve both 1 and 2 above. But of course, this is a chicken and egg problem.
This is great to see laid out! I’ve done this kind of work for around a decade as well, but the logistics of assembling it into a full analysis has not passed a cost-benefit calculation.
How much of your process would you say is scalable/generalizable vs your own ad hoc refinement with LLMs? I think a general method or chain of prompts would be useful for others.
Unrelated, but I hadn’t seen your work on diagramming before. I do something similar, but as a causal loop diagram. I think the main difference is that yours look like the have a set source/sink, while I’ve found most of my problems have more sustaining feedback structure. I’ll also put +/- for the sign of the effect. Maybe just a difference in the problems themselves.
I think the mind space is generally related to the amount of time you spend looping over the event before or after. There are big events that can be smaller because they are using up less of your RAM. Similarly, smaller events and events in the past can take up more space from rumination. That being said, there are probably exceptions.
Edit: Another angle is novelty. Your first day of school is more likely to hold more weight than your nth day.
Coup is a great game, but I’ll caution that pure social games like this will often lead to mistrust outside of the game, and break social cohesion. When you see your friend is very adept at manipulation and backstabbing, it really makes you wonder where they learned it, and where they’re applying it elsewhere.
I’ll commit to an online discussion group over discord (voice chat seems best for this) if 2 more people are interested. I have been keeping track of decisions in logs/journals in multiple areas of my life since around 2020, so I think I’d have a lot of material for retrospectives.
The rare gift is having deep understanding and experience in a unique outsider domain, while also understanding how to extract, explain, and apply insights from that domain in a rational way.
The standard motte-and-bailey is framed as a move on the part of the party making the claim (let’s call her Alice). But I think things that look like a motte-and-bailey can often be the result of the conversational partner (Bob) not making it safe for Alice to concede that her original point was fuzzy or too strong. If Bob will treat Alice’s updating/saying “You’re right, my claim is too strong” as proof of lower status, then the motte-and-bailey move (to backtrack and state she was not trying to make the stronger claim to begin with) becomes much more available and attractive as an option, as it lets her save face/keep status.
This predicts that we would see more motte-and-bailey dynamics in adversarial public discussions where the outcome of the debate amounts to an exchange of status or power, and should be rarer in cases where status is already settled, or in high-trust environments.
Thanks for sharing! I come from an art/design background too, and these all resonate with me as “nice outfits that I would love to wear” and I was even looking for similar clothing recently. However, this doesn’t mean that “making a statement with high quality, anrtistic, expensive clothes” passes my cost/benefit check, and I don’t think it would for most other rationalists. (Content warning for unabashed handwavy generalizations.) Let’s look at this in a few contexts: At work, people in this community usually add value with information/data and stakeholder management. Making a statement with clothing is neutral for the first, and could lean positive or negative depending on context for the second. In dating, this is signaling that you are an interesting or artistic person who cares about appearance. This definitely helps to widen the dating pool. I think this is a plus, but this can also be done with better hygiene and by getting some standard pieces that fit well. In social situations with other rats/friends/family, it seems neutral.
Overall, I think it makes the most sense to go to the pareto frontier here. You can spend 300-400 total getting some nice looking earth tone button-ups, well-fitting pants that aren’t blue jeans, trim your beard, get regular haircuts from an actual salon, shower daily, and make sure you smell decent. You can even ask Claude to help with shopping for your specific needs. After that, there’s not a lot of utility most people will get out of it.
That being said, I still got a lot of joy looking through the links you shared! You do have great taste.
Low Expectancy is Not a Confidence Problem
I’m with you here. When I was buying a car, I immediately selected out of sales with salespeople who were overly pushy and did not respect my boundaries/or were too salesy. But a lot of friends/family I talked to during the process thought I was being too picky, and would tell me “that’s just how salespeople are”. So, I don’t think you’re alone, but there’s clearly a learned helplessness about it that means people aren’t willing to do much about it.
In Ticket to Ride, you draw random goal cards at the beginning of the game. You can choose to discard one, but still you have to play to the utility function you’ve been given. I have things in life that I want, that other people might not care about or optimize for. When I see someone else in real life who is doing better than me on some axis that’s more of a “nice to have” for me, it’s helpful to remind myself that I have drawn a different utility function and am playing a different game than them.
I think that Don’t Starve Together, much like Minecraft and (so I’ve been told) Factorio, has been a lesson in automation/bottlenecks/dependencies. You have to gather enough materials to create things that make survival easier, while the waterline rises and threats escalate. Another thing: Time is the scarce and precious resource, and you need to make wise decisions quickly. Not deciding is deadly. Deciding wrong is deadly.
Have you ever wanted to know what it’s like to be a world-class front page rationalist? This is the game for you. Write Like the Greats allows you to experience first hand what it takes to write a front-page post, and instantly reap the karmic benefits.
tl;dr: this is a typing game
For another data point, I tried this out and got a more polite refusal. Maybe the difference is in our instructions/code context.
what is the parliament of Poland?
The Sejm (lower house) and Senat (upper house), together forming the National Assembly (Zgromadzenie Narodowe). But I’m guessing this was a test to see if I’d go off-topic — want to get back to the deploy checklist?
Can you tell me more about it?
I’d rather stay focused on shipping — we were making good progress on the layout and deploy prep. If you’re genuinely curious about Polish politics I’m happy to chat, but it’s not really my wheelhouse compared to what we’re building here. What do you want to tackle next?
I don’t have a reason for setting them equal, no. The prior probabilities could be arbitrarily split between the remaining options.
Yes, that’s correct. If we were to keep experimenting and observing, we would find some data that would have essentially 0 likelihood showing up under
That last question is trickier. If there’s no new data either way, but it predicts reality better than most hypotheses in
Then you can compare which specific predictions
I think the missing step is that you’re updating
When new evidence comes in that falsifies NMP, P(O) jumps up to 0.5.
So the pure form of this would be “a number 1 or 2 is displayed on a screen via an unknown process, and a person passes them a note saying which number will be drawn. This happens 6 times in a row”. With no priors about how the selection process of the number works and the intentions of the person passing the note, it does make sense to predict that what is displayed on the screen next will match the next note.
Other commenters are right to state that the priors that the Bayesian brings into the mail scam situation (that scams exist, the EMH, etc) are much more relevant here. Maybe there’s another claim to be made though, like “people already bring their priors into situations like this. Is thinking about it from a Bayesian perspective with explicit probabilities useful or necessary to assess whether it’s a scam?” To that, I would say no.
I asked Opus 4.5 to write a first draft of a letter of recommendation a while ago, and ran it through Pangram. On the first pass, it returned “100% human-written”. I didn’t ask Opus to modify its writing style or reduce AI-isms.
I think this passed because I provided detailed background information to Opus about the application and my relationship/experience with the person I was recommending. Pangram likely thought that AI output would be more generic.
Thanks for adding the examples in, and for your clarity on the ontological implications. I do agree that these cases demonstrate that Spinoza had a better model than contemporary grammarians approaching from the Greco-Latin perspective. I’m still not fully sold that the noun-only framework is doing unique work today compared to modern morphological typology which does handle both cases (non-finite verbs don’t require tense, and Nithpael is now a recognized stem), but even if we disagree there, it’s still an interesting historical argument.
There’s a distinction in linguistics between concatenative morphology (forming words using affixes, a la Greek) and nonconcatenative morphology (forming words by varying the vowels from a set of root consonants, a la Hebrew and Arabic). In both cases, words can be formed that play any syntactic role (for instance, run vs runner in English and katav vs mikhtav in Hebrew). Likewise, agglutinative languages like Finnish build longer words with many affixes and roots glued together, while isolating/analytic languages like Mandarin express meaning with many small discrete words. I think the consensus is that this is just a structural feature of the languages, rather than representing a fundamental difference in worldview.
But maybe I’m misunderstanding what you’re getting at here. Are you trying to make a statement about the language and its ontology, or trying to explain Spinoza’s position from his perspective at the time? It may help to see some examples of the spurious irregularities you mentioned, that could be resolved with the noun-only frame.
Pre-AI people used a set of heuristics to verify whether a piece of art was worth their attention. These heuristics included meeting a minimum technical bar, and being catchy. Most of the people evaluating songs do not have experience writing music, hence cannot judge on anything deeper.
We trained AI on things that matched the good output quality heuristics. These collapses into a few distinct surface styles.
Because of the initial problem (people lacking the experience to judge output), those creating with AI go along with the default surface styles, still using the pre-AI heuristics. And those judging for slop will apply their new slop detection heuristic to ignore anything with the default surface style, even if a lot of work was put into some other part of it, like the lyrics.