I tried using chatgpt for doing an analysis of a scientific subject for me.
Oh boy.
It was quick, really quick. It located obscure references and summarized them for me in seconds.
But there were still major issues that I had to deal with.
I asked it to pull up a list of compounds and while it was correct sometimes it would mix branded products with scientific names.
Which was confusing.
But the worst…the WORST part was it started to hallucinate scientific papers with links that didn’t exist. It literally comes up with paper headings, summaries and authors and even provides a hyperlink…to nothing.
Imagine if you used that for a public commercial or scientific report, you could ruin your reputation or even be sued for fraud if you didn’t check the references were real or not.
I caught on to some of the reference list as they papers with ‘no authors’. hmmmm. But for some it hallucinated a bunch of authors as well.
I went back to it and said you’ve been naughty only give me real verified references and it said the usual ‘sorry’ and it seemed to do a better job then.
It also concerns me that if you use it to generate reports that it may base the conclusion of the report on fake data it itself , or other AI, have generated. I understand this is a concern already whereby AI feed on simulated content as it has has already devoured most of the worlds libraries.
Public.
All current bots have this tendencies to hallucinate. But it may be that future versions can check each other better.
Supposedly there’s a thing called agentic processes where you could get AI to check each other and take on different parts of the process more efficiently. One might operate as a fact or source checker?
My point for this one is don’t trust their data or references 100%, it’s for guidance only.
They are FAR more powerful than traditional search though. I asked it to find me references for specific products and their uses in the scientific literature’ and it aced what may have taken me hours in secs. Previously we had to have one guy to maintain a library manually part time but now I think AI could do this work much better and practically in real time.
But I still have to check one by one those references that it gives and be VERY hesitant to rely on it for statistical analysis.
You see it’s one thing to get it to rewrite essays but another for it to give you wrong conclusions based on false data that are then used for real world decisions.
It doesn’t work for those who knows what’s what, and can pick out the hallucination. The problem is chatgpt looks very well written for those who don’t know, or doesn’t know to check everything.
But the reason it does this is because it doesn’t understand anything. It bases everything on the data and how people statistically answers a question. There’s no real understanding at all. Someone called it auto complete on steroids.
So I noticed today that Google searches are AI powered, at least they are when the search query is phrased in question form.
I had two issues at work today. I asked the questions in the Google search bar, and the answers I received were spot on, with the websites used to generate the answers given. I’m impressed.
I’m still trying to figure out the implications of this. For now, I can see only the positive in this.
The TLDR is that it didn’t make any very surprising decisions, which has pros and cons. You get a flat median sort of perspective. I feel like the same might apply to AI answers to Google searches.
There’s a sort of Nash Equilibrium here where using AI is probably beneficial to any individual person, but if we’re all using AI then we risk losing the outliers that matter.
The multi-trillion-dollar artificial intelligence (AI) boom was built on certainty that generative models would keep getting exponentially better. Spoiler alert: they are not.
In simple terms, “scaling laws” said that if you threw more data and computing power at an AI model, its capabilities would continuously grow. However, a recent flurry of press reports suggests that is no longer the case, and AI’s leading developers are finding their models are not improving as dramatically as they used to.
This guy calls it the largest socialist wealth transfer. Basically ai trains on all data out there, even copyrighted or trademarked work. So if you asked an ai to draw you Disney characters, it does that. It says open ai has a few copyright or trademark suits because of that.
I just saw an ad on FB for AI language learning. I’m not sure how realistic it is, but a non-native speaker (student) was talking to an AI generated person (teacher), and the teacher was correcting mistakes made by the student. There was some good two-way exchange, very much like what you might have with an actual teacher during a private lesson.
Are there such programs available now? And how about for other languages? I would love to have that for Mandarin and Korean.
No idea about actual dedicated programs, but I think you could get a passable result just using ChatGPT.
I suggested something similar to a friend recently, using a prompt something like this (it’s probably not perfect and could be improved, it’s just what I quickly came up with on the spot):
I want you to explain the basic aspects of computer chip design to me in simple terms for professional purposes. I have recently participated in several meetings discussing chip design, where I didn’t understand many of the technical terms because my background is not scientific. Also, I’m Vietnamese and English is not my native language, and I’m working in Taiwan. What I would like is for you to gradually explain things to me, where you tell me the basics of semiconductors and chip design starting from the beginning, and I can ask you further questions about things that I don’t understand. For any technical terms, it would be helpful if after the English you put the Vietnamese and traditional Chinese (Taiwan) terms in brackets to help build my vocabulary. Then after you answer my question, you can check I understand before continuing with the discussion. Can you do this?
Okay, spoken language would be an issue. I’m not sure ChatGPT supports that natively yet — I thought not, but might be wrong. I’d be surprised if it couldn’t be accomplished through plugins though.
But you can quite easily have two-way written conversations with ChatGPT already. The prompt above does that to a degree, but it’s totally possible to make the interaction more conversational, e.g., by asking it to correct and explain any errors you make.
I tried something similar before, just to mess around. Though I didn’t go into that much detail, probably because I got bored of talking to a robot.
You said:
I would like you to help me improve my English. I want us to have a conversation where we talk to each other and you correct my answers, giving me the corrected answer then explaining my mistakes as bullet points. Can we do that?
ChatGPT said:
Of course! I’d be happy to help you improve your English. Let’s start with a simple conversation. Feel free to answer each question, and I’ll correct your responses and explain any mistakes.
Question 1: What did you do last weekend?
Please respond, and I’ll provide corrections and explanations.
You said:
I go to the park then eat dinner
ChatGPT said:
Great, here’s the corrected version of your response:
Corrected Answer: I went to the park and then had dinner.
Mistakes:
Verb Tense: “go” should be “went” to match the past tense context.
Conjunction: “then” should be preceded by “and” to properly connect the two actions.
Verb Choice: “eat” should be “had” to correctly convey the action of having dinner in the past.
Question 2: What is your favorite hobby and why do you enjoy it?
You said:
I like eat Vietnamese food and sleep
ChatGPT said:
Great effort! Here’s the corrected version of your response:
Corrected Answer: I like eating Vietnamese food and sleeping.
Mistakes:
Gerunds: “eat” should be “eating” to correctly follow “like” when describing activities.
Parallel Structure: “sleep” should be “sleeping” to maintain parallel structure in the sentence.
Question 3: What are your plans for the upcoming weekend?
It’s curious why text in images seems to be a major technical sticking point. It seems like that should be one of the easier things to get right (compared with the rest of the image, I mean).
Even if the initial models don’t do this very well, you would have thought that things like OCR are now developed more than enough to just be able to overlay the correct text if asked (and convincingly merge it into the graphic).
It’s troubling and indicates low standards that they are using fake photos for supposedly factual news.
What’s interesting is that AI will use this fake photo now as a future reference for photos representing Mandarin Airlines. It shows the need for clear labelling of AI generated content.
