Why are some other... LARGE Youtube personalities so wrong about how "Stupid" AI is, and how it's always wrong?

Dev Zone Thoughts on AI
1

So - I know AI is kind of a hot-button topic - and, I’m sure my perception is a bit skewed because I’m in the AI space - but - I just can’t help but think some of the larger tech YouTubers (I’m not going to name names) have come together as a group and decided all collectivelly run an ongoing campeign about how LLMs are always wrong and provide wrong infomation.

Like - it literally makes me question how accurate some of their content is.

Yes - I’ll be the first to agree, that Gemini has not been so gear (Even 3.1 Extended Pro isn’t great) for quite sometime - and - if their core expierence with AI is the google search overview, free Copilot without selecting a model, or ChatGPT Instant - I may could understand it - but - these are some of the largest computer hardware / tech / gaming channels - and I respect many of them, and honestly think a lot of them.

But when I hear them cover things like news stories, or bring it up on a podcast or video - I just cringe - so many of the things they’re saying to millions of viewers is either due to a misconception, misunderstanding, a bad expierence, or just factually wrong. I’m sure if they were pushed by someome they respect like Windell, many of them would play the “I only use what most people use - which is the free version” card - which maybe true - but, that’s not what they say time again. I can’t remember who I was watching last night - but he said “AI is wrong at least 80% of the time” (And this isn’t some rando - if you’re on this forum, you’re probably subbed to them on YT) and I was just like wtf. Words matter bro. Another channel covered the Anthropic shared chat’s showing up in google story a few days ago, and again, what they said was VERY skewed from the perspective of “This is why AI is bad” - and not the facts.

So - maybe I’m just jaded because I’ve been doing this stuff every day since 2022 - but - I can tell you right now (Maybe it’s my ADHD talking) - if I was going to build a custom loop for a PC - I would spend HOURS - hell, DAYS researching things to the n’th degree until I felt like I could confidently do it, and carry on a conversation about it without sounding like an idiot. I can’t believe so many tech-youtubers aren’t the same way. We’re (they’re, myself, likely most of you) are all in tech - it isn’t THAT hard to learn. We LOVE to learn things. We love to build homelabs and expierement.

I can’t tell you how much I’ve learned over the last 3 years - things that would of taken me days, weeks, or months to learn (or never, lol), I’ve been able to teach myself thanks to AI.

So… am I that far off base? Am I the only one?

2 Likes
2

A lot of tech Youtuber are entertainer with a production team and a lot of shallow knowledge, a Linus for example reads scripts for a lot of projects, period, the deep dive is done by his team members. Secondly, nuanced discussion is not encouraged by the platforms- you either are tech bro who likes that NVIDIA puts mini data centers in everybodys front lawn, or you have to hate everything LLM related.

I think LLM can be a good tool, if you leave your brain on. Another reason might be that Internet search and most websites are shit nowadays (and a lot of knowledge is hidden on unsearchable plattforms like Discord) and that you unfortunately need a tool that can regurgitate preprocessed information.

6 Likes
3

The reality is that, for most people, AI is not a must.

Regular people use it for fun, and they’ll keep using it, as long as it’s free.

Few users would pay actual money, to use it.

9 Likes
4

Most “tech” youtubers are just flavored entertainment. They dont know tech, their viewers dont know it either. They just follow public sentiment about AI and pander to their base, as usual

6 Likes
5

AI aren’t the only sources of false-, and mis-information. Humans also make mistakes and draw incorrect conclusions (as well as out right lie.)

AI are blindly trusted by many people. Content creators are blindly trusted by many people. Celebrities are blindly trusted by many people (see crypro scams and pump and dumps, as well as advertising)

It is only sensible, to verify, vet and check any source.

I guess people presume AI has some magical built in fact checker system, and that there are repercussions to the fallacies and innacuracies that seep in?

You’ll see on the board, that often people deploying AI, need to include tests and checking agents, to verify and bugfix the output they generate, but it can be effective to have AI agents, check other AI, to reduce innacuracies, and provide better code and output

Like, with several iterations, AI is perfectly capable of creating working programs, with just some guidance of people completely devoid of programming skills. Just needs iterations, checking and refining, not running with first results every time.

2 Likes
6

Here’s the problem - there’s a massive disconnect between the opinions of people who do the work (developers, office administrators etc) and the people who manage them. The former look at generative AI and see it as a tool to help - whether it’s as a sanity check on code, building scaffolds, copy-checking letters or anything else. There’s no problem with this; it’s a new tool, enhancing people’s capabilities and quality of life.

The latter, however, do generally see it as an absolute must - they want it in every product, they want everybody using it, and they think they themselves can suddenly do the jobs of the people who build/run their products without any of the knowledge which using those tools requires.

That’s where the problem lies; the people making the decisions lack the knowledge to understand when those decisions are bad, and the hype has made them immune to correction.

7 Likes
7

because they are not honest.

8

I’d say it’s best to look at the average audience for the respective youtube channel. For someone speaking to the average person, or even PC enthusiast, most of them will use tools with the lowest barrier to entry. I.e. free. The average person is not going to pay or even attempt to self host their own LLM.

Should also be a given to never take a single source at their word on a subject matter you know little about. People are imperfect and will make mistakes, lie, etc.

3 Likes
9
  1. People in the “stupid” camp tend to treat models like an extended autocomplete engine instead of interacting with it as a “project manager” and the model being an “individual contributor”. This doesn’t magically fix it on its own but the consequence of this mental shift tends to keep you from landing in OOD and fixes some of the more common prompting mistakes.
  2. Almost every time I have switched models, even in the same family but a newer release tag, I have had to run evals not only on my common language structure but on the fixed artifacts like skills.
  3. Prompting still matters, just more of it has been hidden in the harnesses you use today.

I think we will see a language standard like https://www.asd-ste100.org/ emerge for interacting with these things professionally OR prompts will be abstracted away into an IR and lowered into a model specific language before being passed to the model.

Just some random musing, not to be taken too seriously. <3

1 Like
10

Remember the video of that guy asking chatGPT to time his run to show it’s stupid? The whole time I was thinking…

…Why would you ask a back-and-forth chatbot to do that.

…Why not just use the timer app?

…I don’t think it’s ever said to be able to do that.

…That’s like asking an old magic 8 ball for live exchange rates, or something.

11

“Most” of the big ones do know a lot about tech.

Dr. Ian Cutress (TechTechPotato), Wendell (Level1Tech), Steve (Gamers Nexus), Steve (Hardware Unboxed), Linus Sebastian & his team (Linus Tech Tips, Short Circuit), etc.

The general hate on AI comes from the fact that, prices went up due to investments in AI data centers, and because companies start to offer less hardware, at the same money or more, but make-up by using AI features in their products.

2 Likes
12

Sir the “big ones” are LTT, Marques Brownlee, and whatnot. Multimillion sub channels mostly targeting normies. They are shill and hype factories, especially LTT and #LIEnus himself that just reads whatever his writers serve him that day with zero review but also Marques.

Ian Cutress is a sub-150k channel, Level1techs is 500k. Steve from Gamers Nexus is benchmarks and investigative journalism, they don’t personally know tech anywhere close to a Wendell or Ian Cutress. Steve had to summon a Wendell every time they had to rebuild their office server/network setup for example, they are open about what they don’t know.

That’s tangential, and does not justify misrepresenting the tech as said in the OP just to pander to their audience instead of trying to educate them about, you know, the technology and its actual uses and limitations.

Wendell is fighting the good fight here and getting flak in the comments from normies that just hate AI and can’t stand it, many others are just pandering to that sentiment for fun and profit.

2 Likes
13

Why are some other… LARGE Youtube personalities so wrong about how “Stupid” AI is, and how it’s always wrong?

Easy. You either work full time in tech, or full time on Youtube (or media in general).
Can’t have both.

Video every week means half week dives, unless a hired editor does the videos.

Video every day, read the script once, should do the trick!

2 Likes
14

Last time through the class I teach the big four LLMs managed about 25% on the homework questions I had time to feed them. Hoping to do a more complete evaluation next time through this fall. My guess is improvements to the course material will more than offset six months’ improvements to the models and they’ll get even lower Fs.

In my own use it’s my experience web search assists hallucinate answers more often than they provide anything related to actual sources. And, when sort of right, usually misrepresent the actual sources. In terms of correctly summarizing material in response to a search prompt, maybe 5-10% accuracy. So just about useless for anything besides returning another link or two in search results.

In coding I have recently had pretty good results on basic autocompletion and such from Gemini and GitHub Copilot Pro and occasional good results on agentic tasks. But if it’s like plane trigonometry where understanding the differences between sin(), cos(), x, and y is important? Unmitigated disaster. Change the case on the first letter of 50 fields to finish promoting them to properties? Auger in and fail.

3 Likes
15

It seems like the majority doesn’t have a positive opinion about AI, being anti-AI probably gets you the most views. If you are a YouTuber you have to pick a side, the people who don’t love or hate AI is not very likely to watch videos about AI.

I feel like both sides tunnel vision on the positives or negatives of AI, and both side also overplay the effects of AI.

1 Like
16

I actually did a (very simple) trig test on my local agent - OpenCode, running on Qwen 3.8 27B. I gave it this image:

agent_test
agent_test811×689 67.9 KB

This was the response:

image
image669×302 32.9 KB

Admittedly pretty simple stuff (and admittedly 1b is wrong, unless I’ve missed something), but…from a scanned image? That’s simultaneously surprisingly good and fairly scary. I don’t envy your job, given the general unwillingness of kids to actually do their work.

1 Like
17

Here is the turnstone response. Used a slightly different prompt since I just attached the image to the message. Fully read and understand the provided image and then solve the two math problems providing the final answer in your response.

Solutions

1. Bearings of the ships

(a) Bearing of C from B

Because B is due west of A, the bearing of A from B is 090^\circ.

The angle between BA and BC is 22^\circ, with BC north of BA. Therefore,

090^\circ-22^\circ=068^\circ
\boxed{068^\circ}

(b) Bearing of A from C

Measuring clockwise from north at C to the line CA gives approximately 198^\circ.

\boxed{198^\circ}

2. Estimated height of the tree

The man is 50\text{ m} from the tree, and the angle of elevation from his eyes to the top is 30^\circ.

Let h be the height from his eye level to the top of the tree:

\tan 30^\circ=\frac{h}{50}

Using the supplied value \tan 30^\circ=0.577,

h=50(0.577)=28.85\text{ m}

Approximating the man’s eye height by his stated height of 1.8\text{ m}:

\text{Height of tree}=28.85+1.8=30.65\text{ m}

Thus, to a reasonable degree of accuracy,

\boxed{31\text{ m}}

(30.7\text{ m} to one decimal place.)

Answer Key

Question Final answer
1(a) \boxed{068^\circ}
1(b) \boxed{198^\circ} approximately
2 \boxed{31\text{ m}} approximately
3 Likes
18

I use various LLM’s daily in my sysadmin role. In 2026 I’ve found less hallucinations, but they will still get things wrong if the proper context isn’t provided. The truth is like with most technology, AI literacy is lacking. I would say the risks to critical thinking or its effects on our ability to determine what is true are of far more concern.

1 Like
19

Yes - this is a key piece I think so many people overlook - I see such a lack of critical thinking skills with a handful of my techs. A LLM should be a force multiplier, and not a crutch.

I think you nailed this. I’m SO glad I’m not the only one.

100%%% - I think this comes back to the critical thinking skills I mentioned above - if you can critically think, troubleshoot, and iterate…. AI is a gamechanger.

Oh, 100% - but sadly I think it’s MORE common for people to believe someone because they have some sort of following, somewhere.

No - You legit made all valid points. LLMs are getting better and better by the day (for the most part, ,lol) - but, just like anything, you have to figure it out. You have to learn the quirks. But - maybe because I’m just some super nerd - I have around 15 or so prompts that I use to test models - the same prompt - going back to GPT4. Every time a new frontier model drops, I run the prompt three times, and document it for historical reasons, and then figure out how to prompt it. Honestly - Opus 5 has been the biggest drift for me.

Of course - expecially when they’re in their lane - they’re great. But it seems like with the exception of Wendell, AI is not in their lane*. (*I’ve never watched TechTechPotato)

Well said.

This actually sounds like a fun test. I’d love to give it a shot lol

THIS - I think is nailed.

1 Like
20

One thing I’d like to call out - is this forum is a BREATH OF FRESH AIR. I feel like the antichrist expressing opinions about AI, ESPECIALY on YouTube or Reddit. So… Thanks everyone <3

3 Likes