karitham.dev website! atproto & fancy ssg + csr & blog articles
karitham.dev
atproto
gleam
blog
nix
post: new i like llms post
+132
-2
+110
priv/posts/i-like-llms/index.md
+110
priv/posts/i-like-llms/index.md
···
1
+
---
2
+
title: I like llms
3
+
description: >
4
+
llms are fun! research is fun! copyright is fake and ethics are hard.
5
+
date: 2026-07-19
6
+
tags: [llm, opinion]
7
+
draft: false
8
+
image: ./perceptron.png
9
+
---
10
+
11
+
I like LLMs. I also think they're doing real harm. To knowledge workers, to the environment, to our shared sense of truth. This post is about how I deal with both.
12
+
13
+
I don't want to reiterate what others before me have said better than I could.
14
+
Every major American (and European? hello, you up?) tech market is fake.
15
+
AGI is a marketing buzzword used to gather investor money.
16
+
17
+
LLMs are real, they're here to stay, and we're far from seeing the end of them, or the end of their advances. They're incredible pieces of technology, and both the research behind them and their _outputs_ could advance human civilization in ways comparable to the printing press or the industrial revolution.
18
+
19
+
LLMs allow generating fake news and muddying truth at unprecedented scale.
20
+
Trained on the entire corpus of _books_, _articles_, and an ever-growing flood of _LLM-generated_ text, all of our accumulated knowledge is in there.
21
+
22
+
They also require large amounts of _electricity_ and _water_ to run at scale.
23
+
24
+
They're already a commodity, and they're actively pushing artists, programmers, writers, and other knowledge workers out of work.
25
+
26
+
So what should we do? Are we to revolt at the thought of using these tools?
27
+
Are we to cower under the _"No ethical consumption under capitalism"_ maxim, bow our heads and use these tools nonetheless?
28
+
29
+
I don't claim to be a good enough writer to convince anyone, neither the AI fanatics nor the skeptics. I don't plan to say things like "both extremes are bad". I'm not a centrist...
30
+
31
+
Let me just lay out my interpretation of what exists now, what's to come, and how to deal with it, opinionated, so I can come back in a few years and see how wrong I was.
32
+
33
+
## Ethics
34
+
35
+
AI enjoyers seem to love skipping this part, I'll try not to. I want to make clear that it's important to me, that I think about it, even as someone who likes the tech.
36
+
37
+
Consider the ecological impact, paying an American corporation hundreds of dollars for access to the sum of human knowledge (that it _stole_), the human impact of these tools taking away jobs, the side effects of their takeover of the computer hardware market...
38
+
39
+
I don't plan to dismantle these points, because that would require research that both enjoyers and haters have already done better than I could. I'll sidestep the question instead.
40
+
41
+
Is it worth it? **I think so**.
42
+
43
+
Should we ignore these ethical issues then? **Obviously not**. You think I'm a Sam Altman clone or something?
44
+
45
+
## Case by case
46
+
47
+
Just like with other ethical problems, we need to take them seriously, stay informed, and fight them case by case, in proportion to the harm they cause.
48
+
49
+
### Should you fight the people trying to build a datacenter in your backyard?
50
+
51
+
- How would such a datacenter be powered?
52
+
- What kind of cooling technology does it use?
53
+
- Would it create local jobs and strengthen the local economy?
54
+
- Could it replace _heavier_ industries that are worse?
55
+
56
+
With a clear view of the situation, organize and fight. Where I live, building large numbers of datacenters would actually be a pretty good idea: with sane, durable, well-thought-through construction plans, the local infrastructure could handle a lot of them.
57
+
58
+
### Should you use LLMs yourself?
59
+
60
+
Just like before, we can weigh the ins and outs. Should I use Claude, paid for out of my own pocket, running in the [Colossus datacenter](https://en.wikipedia.org/wiki/Colossus_%28data_center%29)?
61
+
62
+
Should I use open-weight models hosted by third-party inference providers, running in datacenters with a _reasonable_ ecological impact?
63
+
64
+
The question is once again one of scale and nuance, and I can live with that.
65
+
66
+
## Technology
67
+
68
+
The tech is real, it's incredible, and it's not stopping. The research is great, lots of it is open, and it's growing in ways that go beyond "more compute = more smarts".
69
+
In just a few years we went from tiny poem-generating models to ones that solve incredibly complex math problems by reasoning through dozens or hundreds of steps.
70
+
71
+
The space has real depth, and it's warranted. We went from basic transformer models that try to regurgitate human-sounding paragraphs, to large MoE models with fascinating regression properties, compressed context management, million-token-window computation in trillion-parameter models, and diffusion on tiny models for consumer hardware.
72
+
73
+
There's no telling when, or if, we'll hit a ceiling.
74
+
75
+
Reading about people experimenting with workflows or writing their own models, harnesses is extremely fun as well.
76
+
77
+
It's great to interact with and read about unhinged models roaming free on social media, and I like that we're actively developing tooling to allow humans to decide whether they want these interactions or not. There's still lots of research and tooling necessary to keep things sane.
78
+
79
+
## Misuse
80
+
81
+
LLMs are text-generation machines, so they can be used for knowledge work AND pure disinformation.
82
+
83
+
Mind you, we already have that much disinformation.
84
+
Our entire information system is in shambles. Until now, we've had to spend 10x the energy to disprove fake news as to produce it. LLMs can redraw this picture.
85
+
86
+
Sure, it's easy to generate slop, but now we can have autonomous claims verification.
87
+
88
+
It sounds stupid pitting LLMs against LLMs, but we've always just pitted humans against humans before. Why should truth require so much work? We now have a way to counter disinformation at scale, and I'm here for it.
89
+
90
+
## What I do with them
91
+
92
+
I use LLMs for brainstorming, writing code, thinking at higher levels, researching, verifying, understanding, rehashing. They are _so useful_ in my everyday life, and my life has changed, mostly for the better, over the past few years.
93
+
94
+
When used properly, they write high-quality code that is maintainable, correct, and well-tested.
95
+
96
+
I love being able to focus on the things that matter to me: software architecture, designing systems, products, integrations. I've never cared about how an individual function was written (I write Go, it was made to be ugly).
97
+
98
+
Compiling research for something? A problem? Shell out to a research agent, and get sourced results back within minutes. Search engines used to be decent for this, but Google is now so bad it's not even an option. I still pay for Kagi though, the product is good and a relevant search engine matters to me.
99
+
100
+
Generally, I commit more code than before, of similar quality, with equal or higher understanding of what it does.
101
+
102
+
From what I can measure, my code has fewer bugs than what I used to write by hand, with better regression properties (I can think more freely about that instead of syntax now).
103
+
104
+
## My guidelines
105
+
106
+
I try to use open models, and to pick the right size for the task.
107
+
108
+
I refuse to use AI to communicate with humans, and I clearly label AI-generated output. I try to educate the people close to me, and offer them the nuances above.
109
+
110
+
Using AI doesn't mean stopping learning, and it doesn't mean delegating your mind. It's here to empower me, and I love using it that way.
priv/posts/i-like-llms/perceptron.png
priv/posts/i-like-llms/perceptron.png
This is a binary file and will not be displayed.
+1
-1
priv/posts/minideck/index.md
+1
-1
priv/posts/minideck/index.md
+21
-1
priv/static/style.css
+21
-1
priv/static/style.css
···
114
114
}
115
115
116
116
strong {
117
-
color: var(--accent);
117
+
color: var(--blue);
118
118
}
119
119
120
120
i,
···
587
587
margin-top: 2.5rem;
588
588
}
589
589
590
+
.post-content h2::before {
591
+
content: "## ";
592
+
color: var(--accent);
593
+
}
594
+
590
595
.post-content h3 {
591
596
margin-top: 1.75rem;
597
+
}
598
+
599
+
.post-content h3::before {
600
+
content: "### ";
601
+
color: var(--green);
602
+
}
603
+
604
+
.post-content h4 {
605
+
margin-top: 1.25rem;
606
+
font-size: 1rem;
607
+
}
608
+
609
+
.post-content h4::before {
610
+
content: "#### ";
611
+
color: var(--peach);
592
612
}
593
613
594
614
.post-content pre {