domain:doyourownresearch.substack.com
Perhaps it would've been more accurate of me to say "This is part of the reason why LLMs have such difficulty counting..."
But even if you configure your model to treat each individual character as its own token, it is still going to struggle with counting and other basic mathematical operations in large part for the reasons I describe.
Grant_us_eyes
20hr ago
I'm curious what you mean by 'low-level', here. I've heard Obsidian described many ways, but I don't think I've heard 'low-level' before.
kky
20hr ago
Good post. Interesting to see how your perspective intersects with the other critics of LLMs, like Gary Marcus’ consistently effective methods for getting the systems to spit out absurd output.
In my own experience, the actual current value of neural network systems (and thus LLMs) is fuzzy UIs or APIs. Traditional software relies on static algorithms that expect consistent and limited data which can be transformed in highly predictable ways. They don’t handle rougher data very well. LLMs, however, can make a stab at analyzing arbitrary human input and matching it to statistically likely output. It’s thus useful for querying for things where you don’t already know the keywords - like, say, asking which combination of shell utilities will perform as you desire. As people get more used to LLMs, I predict we will see them tuned more to specialized use cases in UI and less to “general” text, and suddenly become quite profitable for a focused little industry.
LLMs will be useful as a sort of image recognition for text. Image recognition is useful! But it is not especially intelligent.
OliveTapenade
20hr ago
But what RLHF does is create a meta-level reward landscape. The model learns that generating text which corresponds to verifiable facts gets a positive reward, and generating text that gets corrected by users gets a negative reward. It's not learning the "vector for truth." It's learning a phenomenally complex function that approximates the behavior of "being truthful." It is, in effect, learning a policy of truth-telling because it is rewarded for it.
I'm not sure how this makes sense? The model has no access to verifiable facts - it has no way to determine 'truth'. What it can do is try to generate text that users approve of, and to avoid text that will get corrected. But that's not optimising for truth, whatever that is. That's optimising for getting humans to pat it on the head.
From the LLM's perspective (which is an anthropomorphisation I don't like, but let's use it for convenience), there is no difference between a true statement and a false statement. There are only differences between statements that get rewarded and statements that get corrected.
100ProofTollBooth
Dumber than a man, but faster than a dog.
20hr ago
and I might claim what I see as a simpler explanation - they wanted cheap labor but felt bad about it.
Yes! So ask the next logical question; why is labor so expensive? Why is there the need to exploit interns? Why can't they just hire a guy or gal who wants a job for a reasonable price? Because labor laws make it too expensive to hire people cheaply!
The government should have an agency or committee dedicated periodic review of laws to see which laws can be retired, or if multiple overlapping laws can be combined for clarity and brevity.
And then what?
The only body who can actually change laws in Congress. Do you know how many congressionally mandated reports there are? Approximately eleventy billion. Do you know how many congress deeply reviews to implement their recommendations? Approximately negative eleventy billion. What would be the point of your hypothetical new agency? And, in order for it to be created, Congress would have to authorize and fund it. That's not going to happen. You're suggesting a "fix" that is obviously and demonstrably untenable.
100ProofTollBooth
Dumber than a man, but faster than a dog.
20hr ago
I don’t know, im personally of the opinion that there are good and bad ways to achieve any goal and that there are always trade-offs that come with any of it.
...
I just want a country where the bulk of people can live a reasonable lifestyle, and where a setback isn’t fatal.
Define "reasonable." Because what you define as reasonable may be reprehensible to me.
A lot of the tradeoffs we're implicitly talking about have to do with security over possibility. If the shot at a successful and independent life means that I could also, with equal or even higher probability, end up destitute, i'll take the deal so long as I am in control of myself. Sacrificing autonomy and independence so that the government can spoon feed me a "comfortable" (but dependent) life? No thank you.
I’m generally with at least tge idea that whatever the form a government takes, the most important thing is customer service.
For a lot of bedrock constitutional reasons, the American government can never be good at what you term "customer service." The only way is to let the customers help themselves - i.e. less government.
Joyful
20hr ago
I've been journaling off and on since childhood and digital journaling consistently since 2013.
TLDR: Obsidian. I use a template some of the time, and a journal review process I'll describe below.
I started in a program called Liquid Story Binder X, a locally installed program for writing, where you could attach entries to a calendar and it was very rewarding to look at the calendar and see all the dates lit up for the dates you had done them. This was abandonware, however, and did not autosave (at least by default) so I eventually moved to Evernote Remember before one note or google keep or anything when Evernote was a sexy startup unicorn that you could access all of your personal files? And then there were some unpopular changes, they stopped doing their own storage (hence no advantage in avoiding google/amazon) and pared down the capabilities of free accounts so that you could no longer install it on all of your devices for free. I had a paid account some of the time but I just wasn't getting enough usage out of it. So since my journal was on google servers anyway I jumped to Google Docs I brought over all of my files and linked them in a spreadsheet. I also set up my template so that I was creating the journal entry in a google form, when I wanted to. This was when I started doing journal reviews some days. If I feel like doing it, I reread journal entries from the same day. As I stated before I started journaling in 2012, so I have between 4 and 12 entries for any given day. Sometimes I review all of the entries, sometimes only the even or odd years. On review I occasionally delete old entries (if they were sparse) or add a + to the end of the name if I know it is a good one. I became really disillusioned and untrusting of Google, I think it was when they started scanning your personal files for copyrighted material circa 2020 so I jumped to Obsidian I have a dataview table that shows me all of the entries for today (so I can do the review process above) and I store new entries in one big folder in the format 22JUL2025. I have a template with a few brain-dumpy questions (what would make today great? dreams?--logging your dreams helps you remember them) and some memory focused ones(Did anyone say anything funny? What was the best thing about yesterday? Who did you talk to? What are you reading/watching?).
100ProofTollBooth
Dumber than a man, but faster than a dog.
20hr ago
I didn't know these things.
Can you provide any good rec's for a "State of the SSPX vis-a-vis the Vatican" at present?
FarmReadyElephants
20hr ago
The US didn't really put all that much effort into taking out Assad. Turkey did.
Nah, the US oil sanctions were pretty decisive. You can see GDP per capita dive as they go into effect. US support for various jihadists also didn't help. But it was the sanctions that were the killer. I was hopeful that Russian aid could Assad hold on, but it wasn't enough.
IMO the post-Cold War USA record in the Middle East is a giant humanitarian tragedy, and it has wiped out most of the remnants of a Christian culture going back 1900 years. Many other people have made the case, so I'll spare both of us repeating it.
Amadan
Enjoying my short-lived victory
20hr ago
I lean more towards @TequilaMockingbird's take than yours but I agree that his explanation of why LLMs can't count threw me off. (If you ask ChatGPT why it has trouble doing simple math problems or counting r's in "strawberry," it will actually give you a pretty detailed and accurate answer!)
That said, a lot of your objections boil down to a philosophical debate about what "counts" as intelligence, and as far as that goes, I found your fish/bird metaphor profoundly unconvincing. If you define "intelligence" as "able to perform well in a specific domain" (which is what the fish judging birds to be unintelligent is doing) then we'd have to call calculators intelligent! After all, they clearly do math much better than humans.
FoundViaTwitter
20hr ago
There's too much overthinking in this thread.
Nazism is reviled because of the inherent implications for multiracial, multicultural societies. The main thrust of Nazism and Hitler - the enactment of an ethnonationalist society through violence on a country-wide scale - is incompatible in a nation where "less than half of US children under 15 are white".
Communism has an offramp because it's in principal an economic ideology.
Im not sure if it's fair to say it "destroys" anything, but it certainly fails to capture certain sorts of things and in the end the result is the same.
A lot of the frustration I've experienced, stems from these sorts issues where some guy who spends more time writing for thier substack than they do writing code dismisses issues such as those described in the section on Lorem Epsom as trivialities that will soon be rendered moot by Moore's Law. No bro they wont, If you're serious about "AI Alignment" solving those sort of issues is going to be something like 90% of the actual work.
As for the "foom" scenario, i am extremely skeptical but i could also be wrong.
SubstantialFrivolity
I'm not even supposed to be here today
21hr ago
If you can't get past the ridiculous "one unit per type per hex" limit, that's understandable
That change is the best change in the game! Warfare is so boring in Civ 4 because there's no gameplay to it, if you have a stack that counters their stack you win. I am sympathetic to the argument that doom stacks were better because the AI was more competent with them, but can't really understand preferring them as a game mechanic.
I have no reason to disbelieve that the sections you've quoted from the Red Terror entry on Wikipedia actually happened. In fact, I'm inclined to believe all of it.
However, I do have to say that some of it - (1) the Voronezh Cheka rolled naked people around in barrels studded internally with nails, (2) Chinese Cheka detachments placed rats in iron tubes sealed at one end with wire netting and the other placed against the body of a prisoner, with the tubes being heated until the rats gnawed through the victim's body in an effort to escape, (3) the Cheka in Kislovodsk, "for lack of a better idea", killed all the patients in the hospital - reads like the more fantastical and debunked stories of the Holocaust that deniers always trot out to muddy the waters.
I don't want to dwell too much on this topic, but could it be that these more horrific types of tortures were limited to just a handful of people and the rest were summarily executed?
hydroacetylene
21hr ago
I remember WWII history, but far less of it than Holocaust.
sammysoccer411
21hr ago
One can argue 'monetary value is not use-value'. Sure. But name a better predictor.
In practice, prices are literally--like quite literally and exclusively--the result of billions of people voting with their dollars, based on how much utility they believe item x has. What could be a better predictor average of use-value than every persons' opinions on use-value, averaged out?
hydroacetylene
21hr ago
It's totally normal for people to describe shit they think as evil as 'kgb', though. You're correct that people call their political opponents Hitler more than Stalin, but there's always been a token of axiomatic evil in figurative speech- it used to be the biblical pharoah(like from Exodus). Hitler's portrayal during WWII was actually rather buffoonish more than outright evil; the Japs on the other hand...
Now why Hitler gets the title rather than Tojo, that might just be the dominance of Jews in Hollywood. I can remember old folks using terms like 'banzai' to refer to crazy evil, but that was more specific to the crazy part. I can definitely remember, quite recently and by younger people, Stalin used as a metaphor for totalitarian evil. But Hitler definitely takes the generic spot.
roystgnr
21hr ago
Half the reason TheMotte is here is that Scott went viral a decade ago discussing Social Justice And Words, Words, Words.
hydroacetylene
21hr ago
It is not surprising why Wokism had an evolutionary advantage on post-URSS marxism. All of this autism is pretty ick, it works on Reddit but not on real life, because every normal person can smell with a bullshit detector that this lines are actively trying to scam you as a North African reseller on an Italian beach. Wokism is better as an ideology because it refuses, partially, to play words.
This is just a baffling description of wokism. Wokism is all about political correctness, language policing, definitional games, etc. Changing the definition of 'racism' 'woman' etc.
5434a
21hr ago
Fair play, I can see the appeal of honesty over artifice.
I'm almost ready to start activating the glue on the replacement headphone band I've been making out of built up layers of edge banding I had leftover from my shelves that I've boiled and clamped on a form. It's only taken me... ohhh, four months. Lol. Gardening takes priority in the growing season.
If that doesn't work out I'll have to buy some solid oak stripwood and try plan B. Getting pretty tired of these dinky earbuds.
Dean
Flairless
21hr ago
Maybe I'm tired and not understanding correctly, but your use of the collective 'you' is reading to me as linking to both the perpetrator and victim of firebombing alike- or possibly both, as in someone deserving firebombing.
Might I ask you to reword this for clarity?
MaiqTheTrue
Renrijra Krin
21hr ago
I think they were both pretty equal in evil. The reason that Stalin gets a pass is that it makes an absolute mess of the moral certainty that the postwar order created. We were allies with Russia, and im not sure that the Allies would have won without Stalin and his war machine. If the war had remained a one front war, it’s possible that some form of Nazi German Empire would have survived. It was only because Russia was involved that we won, and thus talking about Holodomor and Gulag systems (which were absolutely as evil as any of the German labor camps) becomes a bit of a hagiographic problem. Stalin being known to be equally as ruthless would turn the story sideways. Which is a problem because the postwar mythological narrative of Liberal Western Globalist order is “we defeated the worst thing that had ever existed. Thus we have the moral right to rule over everything.” And furthermore it gives the new order a moral certainty— evil looks like Hitler, evil looks like straight armed salutes, arm bands, and speeches in big stadiums and big red flags.
Now they were obviously both evil and killed millions and committed genocide of people into the millions of people. But I don’t think the way the mythology works in th3 modern world works for a lot of reasons. For one thing, it turned things that used to be considered okay into evil simply because they’d been used to evil ends. Nationalism and patriotism are usually good things, they hold people together to build a country. It works in China. They think being Chinese is good and favor things that benefit China.
You could always just self-host a wiki with the same software that Wikipedia uses. (Other options include Tiddlywiki, Dokuwiki, and Wiki.js.)
ChickenOverlord
21hr ago
Yes lend-lease helped, but Soviet home industry did most of the heavy lifting.
That's dramatically underselling lend-lease. The US provided approximately 2/3 of the USSR's trucks, 60% of their aviation fuel, 10% of their planes, etc. And the US was also able to give all this to the Soviets while they also built the most powerful navy in human history and waged a war across North Africa, Europe, the Pacific, and Southeast Asia.
I've talked about this, not at great length, but I've mentioned it, before- cheap labour is cheap for a reason. Interns are a vast improvement over people who'll work for $15/hr in a corporate setting(and BTW, the tenth percentile wage is just above $14/hr. American labour is just really expensive).
There's plenty of working class people who happily make $15-$22 hr. There's a reason they don't get better jobs eventually. Corporate interns generally know things like 'how to keep themselves on track to hit deadlines without constant supervision' and 'how to follow directions correctly without asking for fifteen clarifications every sentence'. These may not be specific skills, but working in a white collar office environment requires abilities like this. Yes, requiring a college degree for this work is excessive, that's why students(who don't yet have one) are doing it as an internship. No, there's not really a solution here(go ahead, name it- no, things like 'flying pigs will provide character references' and 'we'll just kick everyone out of highschool who isn't college material so a diploma counts for the same thing' don't count).
More options
Context Copy link