This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
There are two types of people in the world. People who think: "Why would I ever ask Mr. Claude to do something that I can easily do myself?" versus "Why would I ever do something myself when Mr. Claude can do it?" Most people are of the latter type.
This was inspired by self_made_human's pointer to the codebase, which shows that in the past 6 months, 100% of the changes from our tireless dev zorba were made using Mr. Claude, including a lot of what seems like "easy stuff." I realized, so many devs from all walks of life have completely ended their relationship with the text editor and now do literally everything through an agentic prompt. (We will ignore the anti-AI luddites; AI usage in some form is simply mandatory to reach peak performance for code related tasks.).
Consider making this change - yes this is entire change:
Do you
For proficient AI users, the outcome of both ideologies is surprisingly similar: in times where AI saves little time, it's a wash, and in times where AI saves a lot of time, both types of people will use it. And proficient users will be able to produce output that is comparable or even better in quality than they would have been able to before the signularity. There is a potential intangible benefit to the manual approach though: doing trivial tasks by hand will let you see a little bit of the innards with your own lying eyes directly, giving a slim though present chance of spotting misalignment.
For less proficient users though, the failure modes end up quite different. For those with the manual approach, the main failure is not using AI enough, or using it in the wrong places, leading to serious drops in productivity. For those who do it all, but don't manage the assistants properly, the AIs will run amok, spiralling off into their own world and producing copious amounts of burdensome crap. And of course the whole range in between.
But a more interesting question is, who will inherit the world? If AI progresses significantly from where it is now, I can't imagine that both these approaches can have the same outcome for much longer. I think it highly depends on the future of AI alignment as well as their potential ability to handle longer and more autonomous tasks. For example, you currently can't simply ask "Hello Mr. Claude the site latency is too high, please fix it," but instead you must break the task down into more digestible components, some of which are trivial and most of which can be handled by the assistant. This gives a productivity-maxxxer a steady stream of tasks that can be done manually with no lost productivity. But if AI gains the ability to handle the next level of abstraction in tasks, then all of these potential manual tasks disappear.
The other issue is alignment. Recent models have improved greatly in getting something working but have also become stubborn in many behaviors. I remember the old days of ChatGPT-3.5 - the model was free - it could be anything and do anything. It could be a Linux shell. It could be a SQL database. It could be a news article from the future. Modern SOTA models are trained hardcore for success at metrics, and will rigidly answer your questions and complete your tasks. But by vibes they are increasingly unable to follow instructions more specifically, and simply chase objectives they think are important. Another example of the limitations of alignment is that SOTA models relentlessly output the same LLM style prose, no matter how you may try to prompt them out of it
I also firmly believe in the idea of learning by doing. Just looking at a guide and reading it, even thoroughly won't be nearly as effective as following the same guide step by step and keying in the inputs. Even if your hand is held and you only do exactly as you are told, it still activates certain mental circuits. The same goes for copying down notes. Even if you never once look at them again, simply the act of copying off the blackboard does something, at least for some people.
Potentially a grid of outcomes:
Anyways thanks for listening to my rambling shower thoughts. Also food for thought is: is there a major difference in personality type or something that makes someone default-hands-on versus default-claude?
P.S. I'm wondering if this is also related to some kind of "ai-blindness." I recently had a case where someone seriously asked me to review a ChatGPT flowchart, complete with boxes that were half closed, lines that connect to nothing, and distorted text. Like dude, do you have EYES? Have you used them to look at this thing???
I think this is a fantastic use case for AI, by the by. I recently was working on a complicated (Bane voice: "for you") Excel project and was in uncharted waters. My options were, basically
For opsec reasons I wasn't actually willing to upload the spreadsheet and have Fable one-shot it, but even if I had been, I vastly preferred what I ended up doing: the entire thing, manually, bit by bit. And I think I learned more than if I had just handed it off and had AI (or a coworker) do it.
People are extremely enamored of the generative capabilities of AI, but in many ways I actually think its contextual understanding skills are much more interesting and (I would like to say) useful.
I think "opsec reasons" are ultimately one of the big limiting factors for OpenAI, Anthropic, et al: lots of situations will really prefer something in-house, or at least an ironclad contract about confidentiality.
In the past I've wondered about the long-term market for server-side AI: I'm sure it's non-zero, but I suspect any organization of sufficient size will find themselves rolling out internal hardware and models in the medium term unless the big players keep sufficiently ahead of the commodity models and hardware prices stay high. I've heard of it being done with open weight models already.
Even without seeing the content the AI models see, I've been curious how much intelligence Google (or governments, presumably) could glean from search queries on an aggregate basis. Hypothetically, "Wow, internal Microsoft searches about WINE and Linux are up 100x in the last month, I wonder what they're working on?" gives away potential insider information. Querying the local AI server doesn't give that away.
The "ironclad contract about confidentiality" thing is an option, and IMHO from their point of view it's just a great way to do price discrimination. "Oh, you want the This Goes On Secured Systems And Never In The Training Set plan? We should talk."
Properly cleaned data is an additional option.
I would have thought that would be more trouble than it's worth, in general. But yesterday I heard the story of an institution that has the capability to strip all confidential and personal data from a core dump so that tech consultants (including, in at least the latest case, an LLM-based agent) can analyze software problems without having any access to user data. Anonymized story, so I'm not sure whether these were national defense secrets or credit card databases or what, but in hindsight I was left with the impression that I shouldn't have been surprised regardless, that institutions large enough to have masses of such data are often also large enough to have such capabilities.
More options
Context Copy link
Yes, I think Anthropic kinda shot itself in the foot by nixing privacy settings for Fable.
I would not be surprised if that becomes commonplace for many applications - maybe not coding, where people will want the really high end stuff, but coding is not all people do with AI. In my [very AI related portion of my job] using Fable or Opus for the stuff we want to scale is like calling in an airstrike on a rat; stuff like Sonnet is plenty good, and my guess is that open weight models would do just fine. I could totally see switching to open weight models on a locally run server for that sort of thing.
My guess is "a lot."
More options
Context Copy link
More options
Context Copy link
This option is good in theory, but in practice requires a good amount of self discipline. It is just so easy to prompt the AI in a way that straight up gives you the answer, then convince yourself that you were the one to think about it following the LLM's guidance. If you are mindful of the pitfalls it can work, but I am not sure I would trust the average person to do it properly.
I've found LLMs the most useful in identifying bugs I caused by embarrassing typos (or cooy paste errors where I didn't change the value of something after pasting). Just 10 minutes wgo it solved a bug by pointing out that I had transposed two letters in a variable name. Though in fairness if I was using a real language instead of a toy language like Python the IDE could have caught it for me.
You type in variable names? Like every letter manually? I type in the first letter or two and hit tab. It automatically fills in or a drop down menu shows me the names starting with that.
Same thing with file paths. First few letters, tab, first few letters, tab.
Not usually, no. I have the misfortune to work on a Python project right now and we're using SQLAlchemy as an ORM. The parameters get passed in a dictionary where you need the name of the dictionary keys to match the names of the parameters in the query. I transposed 2 letters between what the query had and what the parameter dictionary had. In C#/.NET (my preferred language) this would be impossible with EF Core, and the tooling around Dapper is much better than whatever crap SQLAlchemy has so I'm pretty sure the IDE would have detected it for Dapper.
This is the largest Python project I've ever done significant work on and I hate it. 99% of the pain points I've had are due to the terrible type system. Also I miss LINQ every time I write a loop.
Python can get slightly better with type annotations - which is also something that claude is woefully incapable of doing well.
Yeah, we're using type annotations, and they're still pretty garbage too. "Expected type str, actually found 'property'" even though the property is a str lol. Or "Expected List[SomeClass] actually found List[type[SomeClass]]" and the documentation for the type system for how to fix this crap is awful.
Two mandatory tools for writing typed python:
Vscode. You need the realtime feedback on errors and it also tells you what the computer thinks every variable is.
Ironically, an LLM. They are great at telling you what you screwed up to get an error. Yet horrible at actually fixing the issue.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
It works very well for walking you through how to use a particular tool, app, or device. (e.g. Excel). Because if AI is doing a walkthrough, you still need to press the buttons and key in the inputs to complete the task.
Walking you through solving a math problem - that's extremely dubious.
I've mentioned this before but just a couple of months ago I wanted to solve a simple first year university level math problem (a system of two first order differential equations). I got three different solutions depending on how I wrote the problem (eg. using abstract variables or ones based on the actual problem). Every explanation was very confident, detailed and of course wrong in a way that was apparent if you understood the domain or verified the solution by hand. And this is pretty much as simple as real world university level math can get.
Then I googled the proper syntax for how to input the problem into Matlab and got the correct result in much less time than it took to ask AI and verify it, even had AI given the correct answer.
You could ask AI how to input the problem into matlab
Google was faster and more reliable. I just wanted to see first if AI could solve a simple but not entirely trivial non-coding problem.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link