Page 8 of 8
Re: The AI Thread
Posted: Fri Mar 27, 2026 9:42 pm
by Gavster
I was kinda referring to Mito’s comment about not getting great outputs, and how the prompt (regardless of whether you write it or get AI to write it), has a massive impact on the quality of the output.
Re: The AI Thread
Posted: Fri Mar 27, 2026 9:44 pm
by DeskJockey
Ah. Then I agree with you.
Re: The AI Thread
Posted: Fri Apr 03, 2026 1:44 pm
by DeskJockey
https://www.forbes.com/sites/amirhusain ... g-systems/
This is one of those good and bad at once situations, with the decision only resting on who found the bug and what they did with it.
Re: The AI Thread
Posted: Thu Apr 09, 2026 1:08 pm
by Jobbo
Analysis Finds That Google’s AI Overviews Are Providing Misinformation at a Scale Possibly Unprecedented in the History of Human Civilization
https://futurism.com/artificial-intelli ... nformation
Large language models adopt an authoritative tone and can confidently present fabricated information as fact
That will come from them being trained using online sources - probably quite a lot from the forum posts which gave rise to this:
https://sniffpetrol.com/author/tpfaif/
Re: The AI Thread
Posted: Thu Apr 09, 2026 1:39 pm
by DeskJockey
Jobbo wrote: Thu Apr 09, 2026 1:08 pm
Analysis Finds That Google’s AI Overviews Are Providing Misinformation at a Scale Possibly Unprecedented in the History of Human Civilization
https://futurism.com/artificial-intelli ... nformation
Large language models adopt an authoritative tone and can confidently present fabricated information as fact
That will come from them being trained using online sources - probably quite a lot from the forum posts which gave rise to this:
https://sniffpetrol.com/author/tpfaif/
They've pulled the "medical" advice, because it was downright dangerous at times.
This is making the rounds in security groups
https://www.eweek.com/news/anthropic-my ... ity-risks/
It is being widely reported. At best it'll be a brief reprieve before someone less scrupulous releases their version of it. Some hyperbole, but not entirely.
Re: The AI Thread
Posted: Sun May 10, 2026 10:43 am
by Zonda_
Are there any reliable AI checkers? I currently deliver courses online and a lot of the ‘less able’ students try to get away with using it for their answers. The problem is I’ve run the same answer through different checkers, namely Grammerly and Zerogpt and one gives a percentage of 100% AI use and the other 0!!
Re: The AI Thread
Posted: Sun May 10, 2026 11:10 am
by DeskJockey
I don't think there are any very accurate ones.
One possible approach (if allowed) could be to permit it, but it has to be explicitly called out and justified. E.g. don't just use AI to write the answer, explain why you used it and what value it adds. I know that's been discussed in other educational contexts.
Re: The AI Thread
Posted: Tue May 26, 2026 7:40 pm
by Beany
Nice bit fun to help you remember: Is AI profitable?
If you can't be arsed clicking, the answer is no, unless you're Nvidia.
https://isaiprofitable.com/
Re: The AI Thread
Posted: Tue May 26, 2026 8:01 pm
by DeskJockey
Latest fun with AI.
Rovo (Atlassian's AI, Gemini in drag I think): asked it to export some data for me. It says it has done it. There's no file. I ask where the file is and it apologises promising it is definitely there now. It isn't. Third attempt. Still insists the file is ready, except it isn't there.
ChatGPT: generate an image with this data. Be accurate with data points. It generates an image where 3.35 is nearer to 4 and 3.78 is at 4.5. Ask why. It apologises and explains it was using the wrong ability, and should have used a different one. Ask it to do that, it doesn't. Numbers are still on the wrong end of the scale it itself generated. Offers the same explanation when asked.
Copilot in Excel: can you merge the data from these cells together. No, but here's how you can do it. Yes thank you, I know how, I wanted you to do it for me. What's the point of "integrating" in the app, if it can't actually manipulate the data?
As impressive as some of it can be, it is still so very unintelligent.
(Note: all versions used were the paid for corporate ones).
Re: The AI Thread
Posted: Wed May 27, 2026 11:10 pm
by Shlergen
Good channel on the toob if you're interested to get more into the workings - 3brown1blue. Has a series on Neural networks, certainly an eye opener for me.
I am pretty impressed with some of the stuff I have been able to generate quite quickly with Gemini Flash and Agent dev kit (ADK). Very little coding effort.
Re: The AI Thread
Posted: Thu May 28, 2026 10:05 am
by GG.
DeskJockey wrote: Tue May 26, 2026 8:01 pm
Latest fun with AI.
Rovo (Atlassian's AI, Gemini in drag I think): asked it to export some data for me. It says it has done it. There's no file. I ask where the file is and it apologises promising it is definitely there now. It isn't. Third attempt. Still insists the file is ready, except it isn't there.
ChatGPT: generate an image with this data. Be accurate with data points. It generates an image where 3.35 is nearer to 4 and 3.78 is at 4.5. Ask why. It apologises and explains it was using the wrong ability, and should have used a different one. Ask it to do that, it doesn't. Numbers are still on the wrong end of the scale it itself generated. Offers the same explanation when asked.
Copilot in Excel: can you merge the data from these cells together. No, but here's how you can do it. Yes thank you, I know how, I wanted you to do it for me. What's the point of "integrating" in the app, if it can't actually manipulate the data?
As impressive as some of it can be, it is still so very unintelligent.
(Note: all versions used were the paid for corporate ones).
Yes - the more I use it (including industry specific programmes) the more impressed I am that people use it and get intelligible outputs. Some of it is incomprehensible, i.e. it would summarise a document saying some language / a concept has been included when clearly it has been deleted compared to the base document.
It has aspects that are akin to a skilled / trained adult professional and then deficiencies / errors that you would be concerned about if a 5/10 year old made them.
ETA - I also tried to get Chat GPT to reorientate a plan that was drawn with north as a reference point by rotating c. 45 degrees. It literally could not do it and was insistent on re-drawing it with all the lines / objects distorted compared to the original.
Re: The AI Thread
Posted: Thu May 28, 2026 10:13 am
by Rich B
Chatting to a mate over the weekend, it seems to be very good/useful for programming - he has used it to make a trading app successfully with limited knowledge of actual coding.
Re: The AI Thread
Posted: Thu May 28, 2026 10:24 am
by Gavster
A friend of mine used Claude to make his website, and as someone with no knowledge of coding it did a brilliant job for his budget, i.e. free!
When it comes to my work, I use Gemini, Claude and NotebookLM for various things and find a lot of little errors. It's still useful as a 'junior researcher' kind of role, however have to verify and check citations before commiting anything to a video. Nonetheless, it's way quicker than me doing the research from a blank sheet. Also Claude is incredibly useful for bringing together data and building tables, graphics etc.
Also, one use case which I've found AI to be exceptionally useful is analysing my running training sessions. I set up a Gemini chat with all my running history, injurys, goals etc and then am feeding all CSVs from training data and letting it crunch the numbers. That's one thing it's really good at, simple number crunching when presented with hundreds of data points.
Re: The AI Thread
Posted: Thu May 28, 2026 10:29 am
by DeskJockey
Vibe coding is the new kid on the block, being used loads. Whether the code is good quality and secure is a another matter.
Re: The AI Thread
Posted: Fri May 29, 2026 5:20 pm
by Explosive Newt
Likewise I use Claude to adapt existing code in Python or Matlab to my own data and needs. Works well. Not as good at writing code de novo, and I think you do need to have a little bit of code knowledge to bug catch and a little bit of domain knowledge to know if it is actually doing what it says.
Huge time saver though.