Hank Green proposed that since ChatGPT can pass the Turing test pretty well, we instead create a new test.
If the system can't tell whether or not it's hallucinating random bullshit, it's not an AGI.
This is actually untrue. AI is actually BETTER at captchas than people. Companies still use captchas to teach AIs, since AI requires humans to tell it what it's seeing.
That's cute, the training data is from so long ago it thought there'd be consequences for January 6th from within the republican party and Trump might have to run as an independent. Hey remember when we had institutions we thought were functional? Good times.
I ask it about history once in a while to check how garbage it is, and it's always garbage. If you ask it to write a biography about a historical figure it thinks it's a writing exercise
It's interesting just how many individual mistakes there are. Like, there's the obvious ones, but even little things like saying that his term officially started in November when he was re-elected.
It might be more accurate if you prompt in a way that it does an internet search, such as "research who is the current president of the US" or "search for articles written by jason schreier.
I'm not exactly sure how it works, but it seems to default to it's internal database over verifying through a search. But if you tell it to search it should.
It's not presented as a tool. It's presented as something to which you can ask the kind of questions Jason asked. That would be an easy fix but it doesn't fit the narrative of the AI companies. Stop using it. It's evil
Hot take: A product that launched in 2022, now with a paid subscription tier, shouldn't still be producing a made up pile of hot garbage when you ask it a basic question.
You and I know this thing generates bullshit, but normies around the globe think this is a magic answer-any-question machine.
What about my screenshot was bullshit? I checked all the links, it all appeared to be accurate.
I asked it a basic question and it gave verifiably accurate answers. It all depends on how you use it, and there are different capabilities between each model. And yes there are flaws as evidenced.
It's not just ChatGPT. I was curious if ProZD @prozdishere.bsky.social had ever been a guest at MAGFest, so I googled "ProZD MAGfest" and the AI summary said that ProZD was a videogame launching in 2025. No potential benefit of LLM "A.I." is worth misinformation this flagrant, nor the investment.
I kept hearing it was good for coding so when I needed something automated for work I tried it. It didn't work at all, and I was able to just find and modify a piece of code on a forum instead
the reason for this, and i know nobody cares, is that the free version (4o-mini) is much worse and these are not the types of use cases that ChatGPT is currently good at
You can derive some utility from it while aknowledging that while sometimes good it bottoms out all the way down at occasionally completely incompetent and dangerously missleading. Its overhyped tech that's moderately useful in niche fields.
I found it was useful for doing powershell scripting when it first went public and within six months that turned to shit. It could be a useful tool but not when it's exposed to the public Internet.
A family member asked it for a recipe the other day and I told them this is exactly why I wouldn't trust it for that. How do they know the quantities of ingredients are correct for the serving size? It's blind faith in something that shows itself as inaccurate too often with the simplest of queries
...And yet there is some banker out there drooling about it taking over handling your account transactions to cut labor costs and get a bigger bonus. What could go wrong?!?
It's like if Google Assistant just lied to you whenever it feels like it and used exponentially more computing power and electricity to do it. If it didn't do the later I'd find the former much more charming.
Not that it still wouldn't have fundamental problems like the fact that it enables the worst actors in our online ecosystem to push spam and propaganda, and that it's a plagiarism machine, but I'd be lying if I said a computer that lies to you wasn't at least a little amusing.
I didn't quite get the same response. But after asking it a few times it started saying Joe Biden was the President. After going back and forth, it ended up saying "to clarify once and for all" that Joe Biden was the President. This thing is actually useless lmao.
Even if the paid version is better because it "searches the web", bitch I can do that myself. This isn't any easier than going to duck duck go and using the search bar.
Free version has web search. I'm actually confused how they even did this query without triggering a search. Asking it for any info about "right now" is like an automatic "you should search the web" trigger. DDG would find this too, but then there are things GPT can do with the info that DDG can't
Look, the tech bros are all about AI, and how could they be wrong? Obviously they can't, so that means this must be correct, and we need to straighten out the whole White House situation immediately.
It's an embarrassing mistake, but you know how confused older people get.
I want to believe this, but I don't. Big Tech has too much money and control over our economy, and it's only going to continue with the current administration
Obviously still not great, users shouldn't have to just intuit which mode will lie more or not. But, like, if you get how to use it the web search function is at least no worse than Google or others imo, which also quite frankly give me terrible and false results sometimes
But Google and others can't do things like, say, search up a bunch of camera specs on diff models I'm looking at and THEN also organize all those specs into a comparison table with links, summarize user comments from forums, and create a doc with all the links in it for me to save, etc
I think people make a mistake when they directly compare it to Google and go "what's the difference" on queries that are literally just simple Google searches. The difference is all the things I WOULDN'T be asking Google to try to do because it can't, that's what makes it compelling
@jasonschreier.bsky.social I checked it myself. It looks like it uses a different model if you're not logged in with an account. I was logged in and got a much better result. The "free" not logged in version looks like it's titled chatgpt4o mini and says it uses less resources etc.
Jason, I want books of all of these, especially Final Fantasy XIV's since I was following that from the get-go and it reminded me later of Fallout 76. Let me know when the preorder page goes up. I'm still planning on getting your Blizzard one (plus a copy for Dad) when I can get around to it.
I used this same prompt using o1 and it seems to be accurate. All of the named articles are real. There is quite a lot of vague info but nothing blatantly hallucinated. It's missing any info on Play Nice, but it looks like it's knowledge cut-off date is sometime in 2021.
I'm not talking about Jason's response (which was wrong). I'm talking about a test I did that gave me three articles written by Jason that I verified with a Google search. Happened to get lucky on this run
4o actually appeared to do a search and correctly say it's Donald, GPT-4 also told me it didn't know. This is the main issue with ChatGPT or any other LLM as a knowledge store, inconsistent results, possible hallucinations but right enough of the time that it seems like you can trust it
Something randomly lying to me is pretty much the opposite of seeming like I can trust it.
If it told me easily findable information most the time and let me know when I have to actually do the search myself: Sure, there are some usecases for that.
But if it randomly lies on even 1% of queries, why would I bother with it instead of using a search engine immediately since I have to doublecheck it anyway?
If it is even slightly more complex or gets tied into something that actually matters: How can I risk using it?
And like, trust it for what, in comparison to what? If we're talking about using it essentially as a search engine for factual information, either Wikipedia or what Google was, say, 10 years ago seem much more consistently reliable.
A "database that's a few years old" doesn't explain it stating that Biden won a second term, or that his opponent was "likely" a Republican when America is a de facto two party nation.
(Hint: AI isn't real, large language models will make these "mistakes" regardless of database age, and lots of "AI" claims are rebrandings of older machine learning techniques that are more robust and predictable but still very narrowly focused)
I asked Deepseek the same question and while it may not have been a robust list of articles you have authored, nothing in it was incorrect (mostly mentioned your books and where you have worked).
I sure am glad that real journalists across multiple industries who put the actual work in were replaced by absolute inaccuracy. We're soooo much better for it
/s 😒
It's extremely funny to me that the product designed to take low level creators out of business is making up narratives suggesting that indie games shouldn't exist. Just very funny how the model is tuned.
Ok, but what I do what to know is directly related to Fallout 76. If I bought the damn game on my xbox digitally, why does it say I need to own it now when I try to play without gamepass!? I'm poor, but I already bought the damn game and I have internet. Not fair!
Yeah I took it back already after seeing at least one person who was surprised by what they saw here. I still don't condone using AI to dunk on AI, it's like diving into the pig pen, getting shit all over you, and holding aloft a shit-covered pig for everyone to see. We know the pig pen is shitty
If you thought the world was getting dumber, just wait until the crop of kids currently using this as their primary source to write papers in highschool grow up.
And they told us Wikipedia wasn't a good source to use. (Of course you should check citations and cross reference, but compared to AI Wikipedia is as reliable as writing emails to nobel prize winners in the field)
And at least with Wikipedia you could use it to chase down primary sources. Unless something changed (I don't use AI and haven't been keeping up with the various flavors) i don't get that with the current LLMs.
Well, you can ask the LLM for sources, but it might just throw out crap or make them up. And then you spend more time fact checking the information than searching for it in trusted sources and checking those against each other would have taken
Comments
The future of computing everyone
If the system can't tell whether or not it's hallucinating random bullshit, it's not an AGI.
An AI passing a captcha just means it has a large enough data bank, not that it has become any smarter
You and I know this thing generates bullshit, but normies around the globe think this is a magic answer-any-question machine.
I asked it a basic question and it gave verifiably accurate answers. It all depends on how you use it, and there are different capabilities between each model. And yes there are flaws as evidenced.
It's not "user error" that chatgpt can't answer a basic question like "who is the current US president?" without making something up that is false.
No one should have to rewrite a prompt like that to optimize for fewer lies.
Does anyone have examples of what it looks like when you use it right? And not stopped clock right
Maybe we just have to accept that trumpists are some evolved advanced human race beyond our comprehension. Homo Trumpus
Hint: You can see it if you look in a mirror.
It's an embarrassing mistake, but you know how confused older people get.
Web search off = not so good at facts, but helpful for brainstorming or organizing/structuring info that you provide to it
Web search on = pretty good at searching up facts, then being able to organize and act on them
🤷♂️😄
If it told me easily findable information most the time and let me know when I have to actually do the search myself: Sure, there are some usecases for that.
If it is even slightly more complex or gets tied into something that actually matters: How can I risk using it?
In reality, AI is already integrated into nearly everything you do, and you don't even realize it.
But please go on with your deep knowledge of AI.
/s 😒
I'll bet over 90% of the words are IDENTICAL, just in different orders.
https://www.youtube.com/watch?v=kjyeCdd-dl8
just so you could write these articles in the proper timeline lmao