ChatGPT vs Microsoft CoPilot: Poor Execution or Bias?

[kemmler3D]

that presents a very dangerous precedence that AI technology providers will bias the results to their liking.

Not only will AI creators intentionally introduce bias, they will unintentionally introduce it, too. Today's AI can only regurgitate existing human thought on things, without being able to judge the quality of that thought very well. At best, it's biased towards mediocre analysis, at worst, it's just hiding our destructive blind spots even deeper and making them harder to find and correct.

Even in the "balanced" opinion of hiring that guy, it doesn't consider any alternatives. This is how mediocre thinkers answer questions.

Case in point, when someone on ASR asks for recommendations between a few speakers, people usually ask about the use case, room preferences, etc. so they can ignore the question as posed and offer better alternatives. ASR people are not mediocre thinkers when it comes to picking out speakers.

I believe that AGI will come about someday (and end human employment as we know it) but I don't totally buy the idea that LLMs will get us there.

 

[fivepast8]

We received AI training at work.

I use a paid version of Co-Pilot at work and you can switch it to "internal" thus it has access to all files/emails etc. of our organisation that I have access to. I tell it to summarize our interactions with a specific client, to list the top three commercial, quality and delivery issues etc. and it does a marvelous job. We also use it to record our Teams calls and then to write meeting minutes. If a training is delivered via Teams, you can have Co-Pilot write the corresponding manual / work instructions and you can force it to use our internal format.

Chat GPT is excellent for business development, and for complex searches incl. geographic proximity. I had all my country managers in for business review and of course we want them to expand business. We used AI to create business development plans for specific sectors down to the phone numbers of the prospects by using elaborate, verbose prompting and "role" playing.

I recently hosted a strategy workshop and got ChatGPT to create the full workshop agenda, including topic related warm-up jokes, methodology, break-out sessions and collating forms for all the information gathered. You feed this back to ChatGPT and it creates the summary. I was done in one hour, a task that my younger self would have easily taken two days to get done.

These tools are just like any other, one needs to learn to use them and understand their limits. Thus, I conclude that our host Amir has once again shown the limits of technology !

 

[tomtoo]

A AI not cares about truth, it just babels what it hears. Nothing else than a tool. Not the AI is dangerous, its the humans that use it a very bad. It used in the inet as a babel(just talk) multiplier.

 

[Graham849]

Over at Twitter, I find Grok wider in its approach to most questions I ask. It appraises more source material:

"The hiring of Mustafa Suleyman by Microsoft to lead its consumer artificial intelligence business has sparked various opinions. Here are some perspectives to consider:

Arguments in Favor of the Hiring:

• Expertise in AI: Suleyman co-founded DeepMind and Inflection AI, giving him a deep understanding of AI development. His experience at DeepMind, which was later acquired by Google, and his time at Inflection AI, where he worked on creating AI with ethical considerations, make him a valuable asset for Microsoft's AI ambitions. This expertise is seen as beneficial for strengthening Microsoft's position in the AI market, especially in consumer products like Copilot, Bing, and Edge.
  https://www.theguardian.com/busines...of-microsoft-ai-with-concerns-about-his-trade
  https://www.crn.com/news/cloud/2024/microsoft-moves-more-leaders-around-after-suleyman-ai-ceo-hiring
• Leadership and Vision: His role in Inflection AI, where he helped raise significant funding, shows his ability to lead and innovate in the AI sector. Microsoft's decision to hire him could be seen as a strategic move to bring in new leadership and vision to its AI projects, potentially accelerating development and integration of AI technologies across Microsoft's products.
  
  
• Strategic Advantage: By hiring Suleyman and several of his team members from Inflection AI, Microsoft not only gains expertise but also diminishes a potential competitor. This move could be viewed as consolidating Microsoft's lead in AI, especially given its partnerships with other AI firms like OpenAI.
  https://www.cnbc.com/2024/03/19/mic...under-mustafa-suleyman-as-ceo-of-ai-unit.html

Arguments Against the Hiring:

• Regulatory Scrutiny: There have been concerns about the nature of the deal with Inflection AI, where Microsoft essentially hired most of its staff and is paying for access to its AI models. This has led to regulatory scrutiny, with investigations into whether this constitutes a merger that could affect market competition negatively. Such regulatory attention could complicate Microsoft's broader business strategies.
  https://time.com/7012714/mustafa-suleyman/
• Past Controversies: Suleyman's departure from DeepMind was amid allegations of a problematic management style, though he later apologized. There might be concerns about cultural fit or past management practices affecting team dynamics at Microsoft.
  https://www.theguardian.com/busines...of-microsoft-ai-with-concerns-about-his-trade
• Ethical Considerations: While Suleyman has advocated for ethical AI, his views on AI development and the potential need for pauses in AI research could raise questions about how these align with Microsoft's aggressive push in AI product development. There might be concerns about how his ethical stance will influence Microsoft's AI strategy, particularly in balancing innovation with ethical responsibilities.
  https://www.theguardian.com/busines...of-microsoft-ai-with-concerns-about-his-trade

In conclusion, the decision to hire Suleyman involves weighing significant benefits against potential risks. It's a strategic move aimed at enhancing Microsoft's AI capabilities but comes with considerations regarding competition, ethics, and management style. The long-term impact of this decision on Microsoft's AI trajectory and its corporate culture remains to be seen."

 

[amirm]

I just thought what I would say about hiring Suleyman at Microsoft. I would have said:

1. It is clearly nice to have an "AI Pioneer" at the company to give the effort credibility and provide leadership in direction.

2. I would worry about having him manage 10,000 people where management skills are paramount rather than vision on day to day basis.

3. Hardly any outsiders have succeeded at Microsoft, especially at senior levels. Internal culture is not receptive to outside experience being of value. Hierarchy has far less meaning at Microsoft than most other companies which may frustrate senior people being used to having authority.

4. Who will reform who? There clearly has to be a merger of ideology and culture. My experience is that Microsoft will not bend easily. Nor will Suleyman who I am sure can get start up roles with one phone call. How much will he put up with? Vast amount of my work at Microsoft as Corporate VP was handling internal processes. Few would find that fun and rewarding.

The right role for him I think would have been a CTO one, not a day to day manager/CEO. I know the "CEO" title is what probably enticed him to come over but it is an empty title that just brings a ton of hassle with it.

Back to AI engine, they barely touched on these issues as much of this is not public knowledge. I guess a proper, future AI would have gone and consulted AI agents working at companies and ask them what it is like to work there, ask other agents who have analyzed Suleyman's style and culture and made an informed recommendation. Much like a human doing references would do.

 

[amirm]

Adding on, I wonder if Suleyman trusted AI enough to ask it if he should go and work for Microsoft!

 

[raif71]

Adding on, I wonder if Suleyman trusted AI enough to ask it if he should go and work for Microsoft!

The answer would be "Absolutely"

 

[amirm]

In the last NPR interview, he said he had a one hour discussion with AI (CoPilot?) wondering what movie to watch for the time he had! I had no idea people were that much into AI.

 

[pjug]

In the last NPR interview, he said he had a one hour discussion with AI (CoPilot?) wondering what movie to watch for the time he had! I had no idea people were that much into AI.

Too bad the time did not expire and AI directed him to tiktok.

 

[Blumlein 88]

In the last NPR interview, he said he had a one hour discussion with AI (CoPilot?) wondering what movie to watch for the time he had! I had no idea people were that much into AI.

That is quickly becoming common. A relative was telling us over the holidays how he enjoys his discussions with AI about subjects he has little chance of having with people. Because people wouldn't know him (he feels like AI does a good job of "getting to know him"), and finding people who know that much about subjects that interest him is difficult. Then three others piped up to admit they were doing the same thing. They too talked about long discussions. Sounded like it was replacing game playing or TV for them. All were using paid versions.

 

[mhardy6647]

In the last NPR interview, he said he had a one hour discussion with AI (CoPilot?) wondering what movie to watch for the time he had! I had no idea people were that much into AI.

Were it not for the requirements of his job, I'd be tempted to call that hour a mammoth misdirection of his resources!
(or he could've watched a movie which was a whole hour longer)

 

[dshreter]

I know, but apparently Deepseek's new open source AI software won't answer questions about Tiananmen Square. The way the transformers work are probably different than the interception of questions or keywords with canned answers by the natural language layer. It seems to me like a pre-filter for forbidden questions or sensitive topics is a lot easier to implement than fiddling with a trained model, or trying to highly bias the training.

For these systems there are pre-filters, post-filters, and also RAG.

Pre-filters halt queries that are known to produce an undesired response, related to copyright content, or things along those lines.

Post-filters act upon the generated response. For example they operate content safety algorithms to prevent the model from sending explicit content. The LLM itself will create explicit content sometimes, so they use an independent model for content safety as an overlay.

Finally, RAG is the technique to add specific data to general purpose model. If you ask ChatGPT what its name is you don’t want the answer to come from general purpose training data, it will come from the RAG data set.

What is RAG? - Retrieval-Augmented Generation AI Explained - AWS

What is Retrieval-Augmented Generation (RAG), how and why businesses use RAG AI, and how to use RAG with AWS.

aws.amazon.com

 

[Old_School_Brad]

We received AI training at work.

I use a paid version of Co-Pilot at work and you can switch it to "internal" thus it has access to all files/emails etc. of our organisation that I have access to. ...

If I owned the company, this would make me extremely uneasy.

I understand all the benefits it offers, but giving "a stranger" with unknown intentions access to your entire company is a daunting thought.
It is widely understood that AIs require training to improve. Assuming they will not, at some point, utilize insights gained from your company across their broader platform may be an overly optimistic view.
Furthermore, there is the risk that we could see data breaches involving AI platforms or compromised AI systems in the near future. What would you, as a company owner, even do in the case of a breach? The TOC is no doubt 100 odd pages for this sort of stuff to work..

I wouldn't be comfortable letting this tech inside my company to roam around freely. Not until some sort of governing is in effect.

 

[MaxwellsEq]

I know, but apparently Deepseek's new open source AI software won't answer questions about Tiananmen Square. The way the transformers work are probably different than the interception of questions or keywords with canned answers by the natural language layer. It seems to me like a pre-filter for forbidden questions or sensitive topics is a lot easier to implement than fiddling with a trained model, or trying to highly bias the training.

Or, possibly more likely, the material it's trained on is prefiltered. If one of the GANs was only trained on modern art whilst another one was only trained on romantic art and a 3rd on comics, all three would produce significantly different output to the same prompt.

Regarding pre filtering. It's done to humans too. Prior to the fall of the Berlin Wall, as part of their studies, a friend obtained Soviet history textbooks for schools. Soviet kids were taught things quite differently from in the west, e.g. the Soviets apparently did not collaborate with Germany to invade Poland, whilst the 2nd World War started with Barbarossa and mostly only involved two major players!

An AI trained on Soviet history books would have little to say about Poland in 1939 or even Pearl Harbor.

It's all in the training data. I've been involved in ML trained only on legal paperwork. Genuinely very useful!

 

[mhardy6647]

Or, possibly more likely, the material it's trained on is prefiltered. If one of the GANs was only trained on modern art whilst another one was only trained on romantic art and a 3rd on comics, all three would produce significantly different output to the same prompt.

Regarding pre filtering. It's done to humans too. Prior to the fall of the Berlin Wall, as part of their studies, a friend obtained Soviet history textbooks for schools. Soviet kids were taught things quite differently from in the west, e.g. the Soviets apparently did not collaborate with Germany to invade Poland, whilst the 2nd World War started with Barbarossa and mostly only involved two major players!

An AI trained on Soviet history books would have little to say about Poland in 1939 or even Pearl Harbor.

It's all in the training data. I've been involved in ML trained only on legal paperwork. Genuinely very useful!

Kind of a mass action thing, then, I suppose.
If the training ground for AI includes popular YouTube and Tik-Tok videos -- heaven help us all.

 

[MaxwellsEq]

Kind of a mass action thing, then, I suppose.
If the training ground for AI includes popular YouTube and Tik-Tok videos -- heaven help us all.

But we can play our small part by correcting erroneous information whenever we encounter it, since the future users will be using tools trained on our contributions.

 

[garbulky]

The issue is that copilot still doesn't think deeply and costs $$ for the better version. MS needs to provide cutting edge AI that's as good as O-1 and free or at least free to gpt plus subscribers.

 

[mhardy6647]

But we can play our small part by correcting erroneous information whenever we encounter it, since the future users will be using tools trained on our contributions.

This isn't a rhetorical question per se- but it's also well off-topic, so I am not really trying to start a side-bar discussion on mechanics!
But I do wish to put my question on the table. What's the mechanism to correct erroneous information?

E.g., I googled the lug-nut torque spec for my car a few weeks back. Google's top-line AI answer was wrong.

From the manual (which - ahem - I had downloaded to the same computer... but I thought it'd be easier to just ask google. Silly me).

 

[Old_School_Brad]

This isn't a rhetorical question per se- but it's also well off-topic, so I am not really trying to start a side-bar discussion on mechanics!
But I do wish to put my question on the table. What's the mechanism to correct erroneous information?

E.g., I googled the lug-nut torque spec for my car a few weeks back. Google's top-line AI answer was wrong.


View attachment 417744

From the manual (which - ahem - I had downloaded to the same computer... but I thought it'd be easier to just ask google. Silly me).

View attachment 417745

ChatGPT did it right.

 

[JSmith]

lug-nut torque spec

CoPilot also got it right...

... but always take these AI assistants "advice" with a massive grain of salt. We know what trusted sources are, they often do not differentiate.


JSmith