One of my first experiences was to ask a chatbot to read 20+ points from a diagram and make a table of them. Then I found that some points were invented by the bot, together with their names. Then I asked how to remove picture's background in Mathematica. It returned several lines of code that did not work (and could not work, because the commands were invented). Later I found out that Mathematica has a RemoveBackground command which the bot did not recommend. I don't understand how these bots could present any danger to the humanity except making us all idiots if we begin using them massively.
Thanks Ugo. The models that regular people get are typically the worst, but you compared them at that same level.
I think the big thing with DeepSeek is that it is 50X thriftier, which breaks the business model for the huge server-farms with nuclear generation of electricity and all those Nvidia chip factories. Those stocks took a hit.
I have seen a report that 50,000 Nvidia chips were used to train DeepSeek, and stolen data, and so on. It is a dark world in the developing battlefield of AI.
War is how it is seen and spoken of in the halls of power, isn't it?
In my discipline of process safety management (PSM) I have found ChatGPT (the only one that I have tried in depth) to be good at opinions, but unreliable regarding facts.
For example, if I ask it which grade of stainless steel to use in a certain application, I have learned to double/triple-check the answer. It may be correct, but it could be badly off.
However, if I ask it ‘What is the most important management element in PSM?’ it will give an opinion that can be thought provoking and defensible answer. I feel as if I am talking to another PSM expert. The point in this case being that there is no right or wrong answer. There is a useful conversation.
The program can also be useful at creating checklists such as ‘What are the steps in starting up a centrifugal compressor?’
Overall, I have found ChatGPT to be helpful, but that’s all. It has not changed my PSM world.
The appreciation of AI greatly depends on what it's used for. The author of this article below only finds one flaw in a series of tests but is annoyed by other issues:
My only test for AI was to design a higher order compensation network so global negative feedback remains at maximum within 20Hz - 20 KHz and every AI even fails to "understand" the issue and outputs irrelevant search results from the web with which I am familiar.
One of my first experiences was to ask a chatbot to read 20+ points from a diagram and make a table of them. Then I found that some points were invented by the bot, together with their names. Then I asked how to remove picture's background in Mathematica. It returned several lines of code that did not work (and could not work, because the commands were invented). Later I found out that Mathematica has a RemoveBackground command which the bot did not recommend. I don't understand how these bots could present any danger to the humanity except making us all idiots if we begin using them massively.
That could happen...
;-(
Thanks Ugo. The models that regular people get are typically the worst, but you compared them at that same level.
I think the big thing with DeepSeek is that it is 50X thriftier, which breaks the business model for the huge server-farms with nuclear generation of electricity and all those Nvidia chip factories. Those stocks took a hit.
I have seen a report that 50,000 Nvidia chips were used to train DeepSeek, and stolen data, and so on. It is a dark world in the developing battlefield of AI.
War is how it is seen and spoken of in the halls of power, isn't it?
In my discipline of process safety management (PSM) I have found ChatGPT (the only one that I have tried in depth) to be good at opinions, but unreliable regarding facts.
For example, if I ask it which grade of stainless steel to use in a certain application, I have learned to double/triple-check the answer. It may be correct, but it could be badly off.
However, if I ask it ‘What is the most important management element in PSM?’ it will give an opinion that can be thought provoking and defensible answer. I feel as if I am talking to another PSM expert. The point in this case being that there is no right or wrong answer. There is a useful conversation.
The program can also be useful at creating checklists such as ‘What are the steps in starting up a centrifugal compressor?’
Overall, I have found ChatGPT to be helpful, but that’s all. It has not changed my PSM world.
The appreciation of AI greatly depends on what it's used for. The author of this article below only finds one flaw in a series of tests but is annoyed by other issues:
https://www.zdnet.com/article/i-put-deepseek-ais-coding-skills-to-the-test-heres-where-it-fell-apart/
My only test for AI was to design a higher order compensation network so global negative feedback remains at maximum within 20Hz - 20 KHz and every AI even fails to "understand" the issue and outputs irrelevant search results from the web with which I am familiar.
WOW!!😳