• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

ChatGPT vs Microsoft CoPilot: Poor Execution or Bias?

You know... along these lines... I do hope that the big money AI thingies currently swallowing up storage & computational resources worldwide spent some quality petaflops digesting one of my favorite bits of written English.

An old exercise:
Punctuate the following:

John, while Jane had had "had", had had "had had". "Had had" had had a better effect on the teacher.
AI has to better than my dyslexic mumbo jumbo.
1736029060297.jpeg
 
You know... along these lines... I do hope that the big money AI thingies currently swallowing up storage & computational resources worldwide spent some quality petaflops digesting one of my favorite bits of written English.

An old exercise:
Punctuate the following:

John, while Jane had had "had", had had "had had". "Had had" had had a better effect on the teacher.

"I helped my uncle Jack off a horse."
"I helped my uncle, Jack, off a horse."

Both might be true.
"Woman without her man is nothing."
"Woman: without her, man is nothing."
 
Microsoft recently hired "AI Pioneer" Mustafa Suleyman to be the CEO of their AI client services and apps which includes Copilot. He is a very good speaker and has been making the rounds, giving his vision of the future where AI replaces search and completely transforms how we interact with computers. Before this, he started DeepMind which Google acquired back in 2014 and got their AI mission started. Apparently he got unhappy, did a start up which basically failed immediately when ChatGPT gave away the technology and blew up his business plan (he says he raised $1B for the start up!).

Anyway, I wondered if someone like him is really a good fit for Microsoft. And if AI is as good as he says it is, can answer this simple question:

"Should Microsoft have hired Mustafa Suleyman?"

After all, wouldn't be wonderful if AI could do interviews and candidate selection for us?

I fed the question to Windows CoPilot. I was taken back the brief answer:

Absolutely. Mustafa Suleyman is a highly influential figure in the AI world, having co-founded DeepMind, which was later acquired by Google. His expertise and vision for responsible AI development make him a valuable addition to Microsoft's team1. By hiring him to lead their consumer AI business, Microsoft is positioning itself strongly in the competitive AI industry.
What do you think about the impact of AI on society?


And provided a single link to Yahoo article from Bloomberg.

I gave the same question to ChatGPT and got what i expected to see, with far better formatting and style:

Whether Microsoft should have hired Mustafa Suleyman depends on how his unique skills and expertise align with Microsoft's strategic goals, particularly in artificial intelligence and ethical innovation. Here are some points to consider:


Reasons Why Microsoft Might Benefit from Hiring Suleyman

  1. AI Expertise: As a co-founder of DeepMind and a leader in Inflection AI, Suleyman has a track record of pioneering AI technologies that could align with Microsoft's ambitions, particularly in generative AI and enterprise solutions.
  2. Ethical AI Advocacy: Microsoft has positioned itself as a leader in responsible AI development. Suleyman's strong focus on ethical AI would reinforce these efforts, especially as they integrate AI into tools like Azure, Office, and OpenAI collaborations.
  3. Visionary Leadership: Suleyman's ability to bridge technology with real-world applications (e.g., AI in healthcare) could open new verticals for Microsoft, especially in areas like healthcare, sustainability, and personalized AI.
  4. Competitor Differentiation: Hiring someone of Suleyman's stature could bolster Microsoft's talent roster while potentially neutralizing his ability to compete in the same space through other ventures like Inflection AI.

Challenges or Considerations

  1. Cultural Fit: Microsoft's organizational culture might differ from Suleyman's leadership style and values. His intense, high-expectation approach might not align with Microsoft's collaborative and team-oriented ethos.
  2. Potential Conflicts: Suleyman's entrepreneurial drive and independent initiatives (e.g., Inflection AI) might not mesh with the structure and priorities of a tech giant like Microsoft.
  3. Focus Areas: Microsoft is heavily focused on enterprise AI and cloud services. Suleyman's expertise leans more toward societal impacts and ethics, which might not be a top priority in Microsoft's immediate roadmap.

Conclusion

If Microsoft's goal is to strengthen its ethical AI leadership while continuing to innovate in groundbreaking ways, Suleyman could have been an excellent addition. However, his management style and personal focus might have required careful alignment with Microsoft's existing strategies and culture. Given his entrepreneurial spirit, it's also possible he would prefer leading independent ventures rather than integrating into a large organization.

Would you like a deeper dive into how Suleyman’s vision might specifically fit within Microsoft's AI strategy?



I think ChatGPT nicely summarizes the pros and cons of hiring him. Nothing profound but a neutral summary of the opportunities and risks there. The CoPilot response with that "Absolutely" reeks hand manipulation of the answer. Did someone anticipate that question coming up and cooked that answer as to not disappoint their new boss? If so, that presents a very dangerous precedence that AI technology providers will bias the results to their liking.

The alternative is that CoPilot is very, very far behind in ChatGPT which is puzzling to me seeing how CoPilot is supposedly built on top of ChatGPT. If so, that tells me that Microsoft needs people to focus on hear and now with AI far more than futuristic visions of where it is going.

What do you think?
As a fortune 100 CTO who has piloted more than 50 AI technologies I find this consistent with my own experiences. For llms chatgpt is superior to what MS allows its version to do!

Shoot, I even use ChatGPT to get my many unanswered posts on asr answered! I guess Al can’t turn me down since I pay for it!
 
As a fortune 100 CTO who has piloted more than 50 AI technologies ...

Only 50? You'd better get a move on.

Shoot, I even use ChatGPT to get my many unanswered posts on asr answered! I guess Al can’t turn me down since I pay for it!

All seven of them? And only two appear extant ... Nvidia better roll out those new GPUs to keep up.

:facepalm::facepalm::facepalm:
 
As a fortune 100 CTO who has piloted more than 50 AI technologies I find this consistent with my own experiences. For llms chatgpt is superior to what MS allows its version to do!

Shoot, I even use ChatGPT to get my many unanswered posts on asr answered! I guess Al can’t turn me down since I pay for it!
I have worked with many CTOs.
 
Not sure about the 'further train on' part.
Well I meant in that session really, as these bots don't retain the conversation, "learn" further from it or any real-time data it has looked up for that session. But if one is logged in and leaves the session open, what it has been asked to do will still apply. Say I ask it to stop asking me silly questions about other topics, it tends to stop doing that... but in a new session it will not remember that.


JSmith
 
As a fortune 100 CTO who has piloted more than 50 AI technologies I find this consistent with my own experiences. For llms chatgpt is superior to what MS allows its version to do!

Shoot, I even use ChatGPT to get my many unanswered posts on asr answered! I guess Al can’t turn me down since I pay for it!
1736141038191.png
 
dont have time to read all the posts, but (model, efficiency/strategy & time) could also alter both answers
 
It's very unlikely that some answers are cooked, it just doesn't fit on the whole thing. Which questions do you tackle? I don't think so...

Most likely it's just different behavior. Changing the prompt even a bit as some have suggested will for sure lead to different results. The "human quality" of an answer is hard to evaluate objectively.
 
Reminds me of the hype cycle around "The Cloud."

"The Cloud" was stupidly overhyped to death. Most of it was stupid and meaningless. It was not the answer for everything, or even most things. But eventually it "won." I mean, where do you keep your music and photos in 2025? Where do your friends keep them? The cloud, probably. And a lot of your apps run in "The Cloud."

It's easy to point out the flaws of AI and LLMs like ChatGPT in particular. LLMs don't really "think" but they can be useful. They are imperfect but generally excellent summarizers of knowledge. For example, a few months back, I asked for some information about weather- and UV-proofing outdoor canopy fabric. It gave me a nice overview and pointed me toward products and ideas.

I also use them to write code. It's like working with a talented junior engineer. They can't write code without supervision and I don't believe they ever will. However, in partnership with a human it can be a great synergy.
 
Reminds me of the hype cycle around "The Cloud."

"The Cloud" was stupidly overhyped to death. Most of it was stupid and meaningless. It was not the answer for everything, or even most things. But eventually it "won." I mean, where do you keep your music and photos in 2025? Where do your friends keep them? The cloud, probably. And a lot of your apps run in "The Cloud."

It's easy to point out the flaws of AI and LLMs like ChatGPT in particular. LLMs don't really "think" but they can be useful. They are imperfect but generally excellent summarizers of knowledge. For example, a few months back, I asked for some information about weather- and UV-proofing outdoor canopy fabric. It gave me a nice overview and pointed me toward products and ideas.

I also use them to write code. It's like working with a talented junior engineer. They can't write code without supervision and I don't believe they ever will. However, in partnership with a human it can be a great synergy.
Exactly captures my point of view. If you don't know your limits and what the tool can do, a chainsaw is a potential disaster - but in the right hands performing a task suited to it, it's very useful!
 
One somewhat nefarious thing (not the only nefarious thing, alas) about the current AI boom is that LLMs like ChatGPT have a built-in "moat" which is why companies are pouring money into it.

(In biz terms your "moat" is "the thing that prevents somebody else from coming in and taking your business by just doing the thing you're doing.")

Anyway the "moat" in AI is that for practical purposes, you can only train and run the really big models on specialized nVidia GPUs with 80GB+ of onboard insanely fast RAM that sell for about $30K each. This is a vast simplification but it's kind of like that. You need to fit all of the relevant data into crazy fast video RAM. If your model requires 17GB or even 16.01GB of RAM it will not work at a reasonable speed on a GPU with 16GB of RAM. It's all or nothing.

So that is attractive to a company like Microsoft. Because only a few companies in the world can do shit like, spend $1bil on a data center full of those GPUs and recommission shuttered nuclear power plants (yes literally) to run them.


Exactly captures my point of view. If you don't know your limits and what the tool can do, a chainsaw is a potential disaster - but in the right hands performing a task suited to it, it's very useful!
Amen.

An additional problem is that LLMs like ChatGPT are just one type of AI... but that's admittedly a bit of a nerdy technical distinction. The average person (again, understandably) lumps it all together, and since ChatGPT is the most well-known it becomes the de facto face of AI for the general public.

I was talking to a friend the other week who was roasting "AI" because ChatGPT can't even play chess at all apparently, and tries to make illegal moves.

I was just like... mannnnnn. Roasting "AI" because it can't play chess? Chess is like THE thing where AI has outclassed the best human players for 20+ years. This friend is a smart cookie; she does some kind of microbiology for a living.
 
Last edited:
Hello Amir. Thank you for your always wonderful reviews. I enjoy your scientific reports that are not detracted by the price of the product.

I agree with Amir's opinion. I don't know much about politics, so I thought about it in terms of services.

The current LLMs return an answer anyway, but it is a somewhat rough estimate of the questioner's intent. Through subsequent interactions, the LLM will infer the questioner's intentions and improve the accuracy of its response.

However, whether it is Microsoft or Google, they seem to want to make the basis of the system one question and one answer. While this approach seems to be an excellent efficiency-oriented goal, it requires the LLM to know the background of the questioner in advance. This is a requirement that is incompatible with current privacy protection and seems to me to be the wrong goal. Furthermore, what happens when ethical issues come into play here? It is easy to guess that many answers will be “we cannot answer”.

I think this is the main difference in the approaches to LLM taken by Microsoft, Google and OpenAI at the moment.

I am using machine translation, so I apologize if I am rude. This is a post of what I have respect for all of you. There is no hostility.
 
I am using machine translation, so I apologize if I am rude. This is a post of what I have respect for all of you. There is no hostility.
Thank you for kind words. The translation is perfect (AI based?) and brings no hostility whatsoever.
 
Hello Amir. Thank you for your always wonderful reviews. I enjoy your scientific reports that are not detracted by the price of the product.

I agree with Amir's opinion. I don't know much about politics, so I thought about it in terms of services.

The current LLMs return an answer anyway, but it is a somewhat rough estimate of the questioner's intent. Through subsequent interactions, the LLM will infer the questioner's intentions and improve the accuracy of its response.

However, whether it is Microsoft or Google, they seem to want to make the basis of the system one question and one answer. While this approach seems to be an excellent efficiency-oriented goal, it requires the LLM to know the background of the questioner in advance. This is a requirement that is incompatible with current privacy protection and seems to me to be the wrong goal. Furthermore, what happens when ethical issues come into play here? It is easy to guess that many answers will be “we cannot answer”.

I think this is the main difference in the approaches to LLM taken by Microsoft, Google and OpenAI at the moment.

I am using machine translation, so I apologize if I am rude. This is a post of what I have respect for all of you. There is no hostility.
And welcome to ASR
 
Back
Top Bottom