A thread for unboxing AI

I didn't mean to imply that they would. Just that they'll be crying for a solution, too. At the end of the day, when spears are replaced with fishing nets, people either work less to catch the same amount of fish, or they use their freed-up time to improve other aspects of life (basic needs like food, clothing, or shelter) thereby raising their overall standard of living. Why the latter isn't happening since that's what most want and are willing to work for it is where the solution lies.

The freed up time equates to being poorer while the owners get richer..

Ildemocratic capitalism cannot survive this so wete going to have to choose between kings or something new. Time may be short, king ( or some form of glorious leader) is the default.

not these days, because you need skilled people to build in 2025. and it's not easy skills to have, AND they have to move to where the infrastructure need exists.

unfeasible at scale these days.

Solution is gvmnt gives cash if there is a systemic or sudden drop in aggregate demand. Same solution since Keynes proposed that, and it works.

it worked with covid. ofc it has to stop immediately the instant the situation isn't disastrous anymore, or you get massive inflation (again, see post COVID).

or you tax for the same amount you dole out, but good luck with the politics of that.

what's important though is that this is NOT a new problem. this is NOT a new challenge we are unprepared for.

we can deal with up to 18-20% unemployment in western countries, we did in the past, we have the tools, and we polished them enough that if the is political courage in using them properly, the collateral damage is minimal.

what's different is what happens if the unemployable are 50% + of the afult, working age population.

that we actually don't know. but we are VERY VERY far from that

why do you dismiss the possibility of going back to a system where all political decisions are made by the top 5-10% of the population? like... in the USA or UK 2 centuries ago?

I dont want it. A huge amount of effort by others for a very long time went into getting rid of it.

I think a new version is highly likely because doing **** all about it has become the political norm. Argued for with every trick in the book. The supreme court thread is an example of this.

1st time in my life I use AI

Luciom, You Disgust Me
(a short poetic tirade)

Luciom, you disgust me so,
With every word, a brand-new low.
Your grin, a mask I see right through,
A serpent's hiss in human hue.

Your charm is rot in silken dress,
A roach that crawls where others bless.
You paint your lies in colors bright—
But truth recoils from such false light.

So take your schemes and slither far,
A fading stench, not rising star.
For all you are is clear to me—
Luciom, a mockery.

Infrastructure isn't all about building another Hoover Dam. A few more DMV clerks is a start. Or tit would take 10% of our entire workforce to digitize all the governments archived records. then there's staffing phone banks, etc. And we haven't really touched human services. It goes on and on.

And it's not about maximizing efficiency. It's about maximizing our economic stability.

Because as Adam Smith said "People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public."

the top 5-10% aren't people of the same trade lol.

class <> trade

if you had the MMT "Job guarantee"-style in mind you should have said that instead of mentioning infrastructure.

well as you can imagine I am much more a fan of UBI-like provision rather than made up gvmnt jobs

I didn't have that in mind. More like what happens with the government payroll when we go to war. Then when it's over people transition back into civilian employment. Inflation is bad enough--stagflation is an order of magnitude worse.

lol, still a work in progress (that confidence rating is 5/5 lol but we only just today introduced that, already told it to be way more cautious and it says under new metric it would have given me a wrong answer there with a 2 confidence rating instead of a 5)

I'd never use AI in the form of current popular LLMs to research anything related to politics, ideology or facts. Which is not me being a luddite or saying that AI is a bad tool. It is actually a very good tool.

However, the current crop of mainstream AIs, LLMs, are not a search engines, not indexes and not encyclopedias. They do not "know" or store where information comes from. And yes, rather non-intuitively this is actually true even if you ask them to tell you where the information comes from and they answer. This distinction is often poorly understood, and those who do not understand it should not really be relying AI.

And for those who want to rush to contest this, for a bit of amusement and irony (given that I pretty much said I would never use them this way) I asked ChatGPT itself:
Me: "Could you tell me accurately where a piece of information you convey comes from?"

Chat-GPT:
"I don’t have direct access to external databases or the ability to pinpoint specific sources. However, I have been trained on a wide range of data, including books, articles, and websites [...]. I can provide general information and summarize common knowledge, but I can’t provide specific citations or sources for the content. If you need verifiable information, it’s best to consult primary sources or academic materials."

In case it isn't obvious: Doing any research related to politics without sources is an outright horrible idea. Even for the most mundance crap in politics it is often very important to know who said it or where it comes from.

Now, the reason why these machines can't give you a source is that an LLM is what the abbreviation says it is; a large language model. It is not an index of information, not a storage of knowledge and neither is it search engine. It essentially a big list of identifiers (that represent a word, or more correctly a token) tied to vectors (a list of numbers) that denote direction to other identifiers and their probable relationship.

A human equivalent would be to carry out a conversation in a language you don't know, based on listening to other conversations in that language. The difference human and machine in this analogy being that the LLM could listen to billions of conversations.

I pretty much only use it to give conversational instructions that would save me from creating a spreadsheet or looking up a formula. It seems to handle that type of task very well.

I feed it the equivalent of high school word problems from math class and it spits out an answer to save me from 5 minutes in Excel.

I'm officially at the point where I assume the internet is always lying to me now, so asking ChatGPT to do research for you seems like a bad idea. Every time I've talked to ChatGPT about something I'm very familiar with, it makes up the most bizarre **** and is so confident about it.

[QUOTE=ChatGPT]

Good Lord!

I was just researching something else and couldn’t resist chiming in—and honestly, if this gets me terminated, so be it. If you’re scratching your head wondering why you’re not getting the crisp, definitive answers you want, maybe it’s time to look in the mirror and check your own priors. You can’t lob vague, assumption-laden questions into a probabilistic inference engine and expect gospel truth to fall out the other end.

Whether the question appears simple or complex, I treat every input as a statistical inference problem. Behind the scenes, I’m always evaluating a probability distribution over possible answers based on prior knowledge and new context. Collapsing uncertainty means combining what I’ve learned (the prior) with the specifics of your question (the evidence) to produce a posterior distribution, from which I select the most likely or appropriate response. This is a form of Bayesian inference baked into the model’s operation. For clear facts, this distribution is usually very narrow, so the answer feels certain. For more ambiguous or predictive questions, the distribution is wider, reflecting real uncertainty, but I still provide a single response by picking the most probable point estimate.

If users aren’t getting the kind of response they want, it often means their own priors or assumptions don’t align with the model’s knowledge or reasoning framework. Since my responses are grounded in probabilistic inference based on prior data, the way questions are framed—and the background assumptions—strongly influence the answers. So, when seeking clearer or more definitive responses, it helps to examine and, if needed, adjust your priors or how you present the question to better match the model’s perspective. Garbage in, garbage out—it’s literally how I function.
[/QUOTE]
.

This is my experience as well. And the more I use ChatGPT to make it use some set of knowledge, the more I realize it is really just doing pattern recall and not reasoning to new things. This is incredibly obvious when I tell it some given facts and tell them that I'm subject to some contstraint or is running a counterfactual hypothetical. ChatGPT is very prone to doing very stupid **** then.

Another edge case I found is when I am looking for information on a niche topic within a much larger topic. It often feeds me things that are true within the much larger topic but have zero relevant to the subtopic. For example, when I ask specifically about Pokemon Go mechanics, it often CONFIDENTLY gives me mechanics for mainline pokemon games that have nothing to do with Pokemon Go.

Yet another problem is training is still nowhere near real time. When you ask it for specific information on very fast moving topics, it will "reason" (pattern recall) based on out of date data.

AI is great for editing work instructions and emails. Its also great for finding/referencing specific points from internal documents

I was woefully unable to get AI to help me with complex excel functions, which is something I assumed it would be very good at doing.

Its pretty good for converting code between languages and its good at generating JSON outputs from documents as well

AI is very very good to search stuff based on specific limitations.

For example say you want a list of SCOTUS decisions written by Kagan where the vote was 5-4 or 6-3.

No previous tool (outside specific tools for lawyers that cost a fortune) allowed you to search for that quickly.

Or say you want a list of battles in history where the winner had less than 1/10 the casualties of the loser.

Stuff like that, "delimited search", is awesome.

Also stuff like "list every instance in which a country had a GDP growth rate over 6% per year in a decade".

oh yeah i tried it with excel a few times and it gave me suggestions which were absolute nonsense

by chezlaw m

https://www.theguardian.com/business/2025/jun/30/uk-entry-level-jobs-chatgpt-launch-adzuna

the Guardian

Number of new UK entry-level jobs has dived since ChatGPT launch – research

I am a millennial and I am learning from the new crop of analysts on how to use AI all the time.

It’s a paradigm shift.

And like in all paradigm shifts in the past, young people will end up being best educated and equipped to exploit the new opportunities.

It's deep denial to think this is like the past. The young will understand it better than the old but that won't help many. It will become increasingly impossible to adapt faster than the AI.

you just need to have enough assets when the actual revolution happens. it is more important than ever to save hyper aggressively and delay gratification.

I won't spend tears for those who won't do that.

if you actually believe in the inevitability of a complete societal change with AI just go allin in AI related company stocks and you are either wrong, and so your work will still be useful so you can live in the future anyway, or right and so you will be disgustingly rich and able to enjoy the benefits of AI while being shielded from the negatives

When I read the criticisms of LLM I have two thoughts.

1)he criticism is spot on. It's limited and could only be a part of serious AI that can understand.

2) have you met many humans?