I seem to be on a "defend LLMs for personal use" kick. It seemed like fun to see how I've used them in the past month. In most cases, I first did a search that proved useless (often, spending a long time following links that looked like they *could* be helpful…but weren't). ↵ - ThreadSky

shriram.bsky.social • 59 days ago

I seem to be on a "defend LLMs for personal use" kick. It seemed like fun to see how I've used them in the past month. In most cases, I first did a search that proved useless (often, spending a long time following links that looked like they *could* be helpful…but weren't). ↵

Comments

krismicinski.bsky.social•59 days ago

like I've mentioned to other folks, I notice the predominant attitude on Bsky surrounding is the antithesis of the annoying grifting on Twitter. In many ways, I feel it's become a microcosm of the politics surrounding the technocratic right vs. the Twitter exodus.

krismicinski.bsky.social•59 days ago

I don't think it takes a genius to realize that LLMs have lots of issues (power use, market bubble, speculation, folks using them to replace humans when they absolutely shouldn't) but most skilled folks who use these tools recognize they are useful in spite of these issues in some circumstances.

shriram.bsky.social•59 days ago

It's the lack of good-faith on both sites/sides that is so frustrating. Like just a default assumption that you must be both a terrible person AND an extremely stupid one just because you (based on knowledge, experience, and expertise) dared to not take their absolutist position.

shoebum.bsky.social•59 days ago

I think, it might be wrong and I don't have any evidence to qualify, both sides in this discussion are not equally weighted when it comes to power dynamics.

And it's mostly because people are actually out of a job because any LLM based company is using the above stated legitimate reasons

shoebum.bsky.social•59 days ago

To perpetuate this notion - "These LLM based tools are at the inception of something that is to come and we can let go of all/most of the white collar workers, so give us money to invest in this endeavor"

I don't think anyone wants to support the above statement and want people who legitimise this

shoebum.bsky.social•59 days ago

To realise that hey we aren't saying AI is bad, it's good but it's essentially a tool that requires not only people reskilling but an endless looming threat to ones who don't have any insight into the tech that their days are numbered.

I don't think people are receptive to a fearful environment

krismicinski.bsky.social•59 days ago

here's an example of how I might use an LLM in daily non-programming life. It's a jumping-off point for ideas. After this I'd probably google about things for recipes written by humans, or throw them away entirely and do my own thing--it augments brainstorming.

krismicinski.bsky.social•59 days ago

but one proximate reason I think people will use LLMs to replace online recipes is that 95% of recipe sites are filled to the brim with ads and filler...

shriram.bsky.social•59 days ago

Came here to say exactly this.

felipe.rs•58 days ago

shriram.bsky.social•59 days ago

It really is *exactly* that. Very hard to miss, it's not even remotely subtle.

shriram.bsky.social•59 days ago

PROGRAMMING/TECH:

shriram.bsky.social•59 days ago

MATH:

shriram.bsky.social•59 days ago

LANGUAGE:

shriram.bsky.social•59 days ago

GENERAL KNOWLEDGE:

shriram.bsky.social•59 days ago

FOOD + MISC/RANDOM:

shriram.bsky.social•59 days ago

THE REAL TIME SAVERS:

sp.degabrielle.name•58 days ago

No _code review_?

Or is this something you tried but found it unhelpful?

shriram.bsky.social•58 days ago

1. I read my own code, why do I need another reviewer? (-:

2. As an experiment I spent a month last year using Copilot with Racket and gave up in exasperation, so I didn't think it would do much good reviewing my code.

3. But seriously, I've never thought to use it for that. But also, #1.

daveandersen.bsky.social•59 days ago

Today's out of the box LLM success: Write a tiny python script that permuted a file with a single large JSON array into a new file with per-line JSON entries corresponding to each entry of the line. I'm sure there's some one-line jq invocation or something that could do it but it was easy w/copilot

daveandersen.bsky.social•59 days ago

actually there's a website that will do it. but it was just as fast to use copilot as to find the right search terms for the website. :) :)

shriram.bsky.social•59 days ago

Yeah, this is the perfect use-case because you *almost* don't even need to read the code; you know it's going to be right. And you can do "translation validation", ie, instead of verifying the code, look at the output data to confirm it's what you intended. Why even try doing this any other way.

arjunguha.bsky.social•59 days ago

What may be missing: there are certain things that you may not do without an LLM to help you blast through it. I wrote several one-off programs to help with in-class activities last year. I also wrote iPhone apps for myself, e.g., https://apps.apple.com/us/app/calendar-availability/id6667094800 (not worth my time to learn Swift)

shriram.bsky.social•59 days ago

Right. I think this is mainly a consequence of sampling during the busiest part of my semester, when I had no time for this. I have, for instance, built small apps that used LocalStorage for which I could never remember the APIs and wouldn't have bothered even trying but the LLM got me home in 5min.

atheorist.bsky.social•58 days ago

Keep in mind that the spectacular Fake Intelligence paydays now motivating investors have nothing to do with benign use cases like these. As the tech succeeds at putting creators out of business it will cease to be offered as a subsidized affordable tool for creators to noodle around with.

scottlindsey.bsky.social•58 days ago

Here's the thing -- AI slop is highly visible. AI being used well is not.

We're using AI to automate things and you don't see it when it is done well.

It's not going away.

025 will see price cuts as the cost of inference continues to decline.

https://finance.yahoo.com/news/alibabas-shocking-ai-price-cuts-163420743.html

atheorist.bsky.social•56 days ago

"AI being used well" might be a thing, sure, but putting people out of work and delivering crappier service is where the big bucks are and that is NOT what I would call using AI "well."

scottlindsey.bsky.social•58 days ago

2025, not “025”…

maxsnew.com•59 days ago

I'm pretty basic. I only use them for API/programming patterns in languages I don't use very often (Rust/bash), especially because Google has become useless for that. They seem to have very high signal to noise rate for that task compared to others I've tried

shriram.bsky.social•58 days ago

They are indeed especially good at those, and *especially* good relative to Google search, as you say. Like I cleaned up a Makefile that I struggled to fix for ages because I couldn't quite get the commands right, and it calmly got them right on first try.

dpiponi.bsky.social•59 days ago

I don't understand why there's a point to be made. It's trivially obvious that LLMs are useful. You just have to learn not to waste time in a loop trying to fix wrong results that are never going to be fixed. Not a hard skill to learn.

shriram.bsky.social•59 days ago

This thread is obviously written for people not like you.

dpiponi.bsky.social•58 days ago

Fair enough. Keep up the good work then!

vikramsaraph.com•58 days ago

I’ve found it helpful in reading/searching software documentation lately. Specifically Hugo, Rust, and Vue (all of which have been around long enough for there to be plenty of content for GPT models to have learned from).

ljones.xyz•58 days ago

It has been wickedly helpful for kernel coding in my case. Very very often it provides information that I am unable to easily find via Google (which has turned to shit) and feels like having a pair programmer kernel guru next to me

shriram.bsky.social•58 days ago

The rise of LLMs in parallel with the utter and total enshittification of Google search has really been a remarkable thing to watch.

Future generations will scarce comprehend when we tell them that Google search was one of the technological wonders of our time.

elfprince13.mumak.app•58 days ago

What’s funny of course is that the enshittification of Google has been largely LLM-driven

shriram.bsky.social•58 days ago

I don't think that's entirely true. SEO rubbish long predated them.

vikramsaraph.com•58 days ago

Yes, Google search results had been getting noticeably worse even before the LLM era. I wonder if there’s a way to measure how much LLMs have accelerated its decline.

vikramsaraph.com•58 days ago

Feels like there’s an Ouroboros joke here somewhere

ljones.xyz•58 days ago

Ask an LLM to create it

vikramsaraph.com•58 days ago

I’ll have a vague idea of a thing I want to do, ask a question, and it usually points me in the right direction towards the concept that I want to read more about in the actual document. After which I read the actual documentation.

vikramsaraph.com•58 days ago

Not so different than what RAG’s used for I suppose, except GPT models have read enough software documentation to not need augmentation

Comments

Posting Rules

Reply