I used AI and I like it

[loccdor]

I prefer using it for mathematical/physics/engineering types of problems where everything is experimentally verifiable and opinions don't really come into it.

 

[Alan Edward Klein]

A lot of what you cite has already started without AI. People's dependency on the internet has severely limited independent thinking, rote tasks such as spelling and arithmetic, knowledge of history and so much more.

Actually, letting AI deal with spelling, grammar, arithmetic, and other rote tasks will allow humans more time to think.

 

[Pieter12]

Actually, letting AI deal with spelling, grammar, arithmetic, and other rote tasks will allow humans more time to think.

Maybe, but it reduces the quality of their thinking. Disciplines like math and grammar require precision and critical thinking, aspects that are quickly diminishing among younger generations.

 

[filsogh]

So I was skeptical in my first attempts using ChatGPT. Someone suggested I try Claude.ai.

My use case was oriignally for language learning... I am learning a language that is not widely used and has about 8 million speakers. I have found that claude is VERY good at producing native fluent constructions of sentences and texts. By this I mean written in the manner in which a native would phrase it including idioms. These are cases where normal translators like google translate completely miss the mark.

I am in. I continue to experiment with other use cases. One big thing with these ai is to not just accept the first answer to a complicated question. Probing questions and dialog create much better results. It is sort like just accepting the google search result versus really probing, following link to other sources, etc.

 

[Sean]

I was going to echo some other comments that free tier models are way behind compared to the higher tier paid models. It can be a bit like looking at a 27" CRT television set from the 80s and thinking "bah, I don't see the big deal", while others are using 100" 8K OLED TVs eyes bulging.

 

[Alan Edward Klein]

Maybe, but it reduces the quality of their thinking. Disciplines like math and grammar require precision and critical thinking, aspects that are quickly diminishing among younger generations.

Adding 2+2 isn't critical thinking. If it is, we should get rid of calculators and Spellcheck. Critical thinking is understanding how inflation works and how it affects your paycheck and what to do to protect against it. Critical thinking is examining what photographic art is, not forgetting a comma.

 

[Alan Edward Klein]

So I was skeptical in my first attempts using ChatGPT. Someone suggested I try Claude.ai.

My use case was oriignally for language learning... I am learning a language that is not widely used and has about 8 million speakers. I have found that claude is VERY good at producing native fluent constructions of sentences and texts. By this I mean written in the manner in which a native would phrase it including idioms. These are cases where normal translators like google translate completely miss the mark.

I am in. I continue to experiment with other use cases. One big thing with these ai is to not just accept the first answer to a complicated question. Probing questions and dialog create much better results. It is sort like just accepting the google search result versus really probing, following link to other sources, etc.

I've been using Chrome's Gemini for getting quick answers instead of searching using Google. Yesterday, my driver's motorized seats and steering wheel stopped working. I asked AI while in the car how to fix, giving the model and year of the car. Gemini said to check the fuse giving me the fuse number and check the connector under the seat. Turned out it was the connector that opened (cause by my snow removal brush that slide under the seat and opened the connector when I slid it out.). Real-time help and action.

 

[Sean]

Might interest some here: https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf

 

[Pieter12]

Adding 2+2 isn't critical thinking. If it is, we should get rid of calculators and Spellcheck. Critical thinking is understanding how inflation works and how it affects your paycheck and what to do to protect against it. Critical thinking is examining what photographic art is, not forgetting a comma.

No, critical thinking is actively analyzing and evaluating rather than passively accepting information.

 

[mshchem]

AI requires human beings to copy.
Soon to be out of a job
Models for swimsuit catalogs
Annoying customer service phone operators
Porno movie actors
Etc.

Eventually the Silconoids will kill us all. But not before we are enslaved to clean up all the single use plastic yogurt containers and Kruerig pods everywhere.

 

[Pieter12]

AI requires human beings to copy.
Soon to be out of a job
Models for swimsuit catalogs
Annoying customer service phone operators
Porno movie actors
Etc.

Eventually the Silconoids will kill us all. But not before we are enslaved to clean up all the single use plastic yogurt containers and Kruerig pods everywhere.

Humans require humans to copy, too. But AI cannot as yet come up with original ideas. I know you're trying to be funny, but why should the robots care about our trash? Try to be more original, less like a machine.

 

[gbroadbridge]

Might interest some here: https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf

Many here may not know of the works of Donald Knuth, but no surprise that he would find the latest iterations of AI models to be thought provoking.

AI is a wonderful tool that is going to enormously accelerate human achievement and learning - for those that embrace it rather than fear it.

My son, who is a final year university engineering student is being encouraged to use AI tools, it is in fact a mandatory part of the syllabus.

 

[Alan Edward Klein]

No, critical thinking is actively analyzing and evaluating rather than passively accepting information.

I agree it covers that, pretty much my point.

 

[cliveh]

AI could copy the style of any photographer, but it would be made and not real.

 

[filsogh]

Many here may not know of the works of Donald Knuth, but no surprise that he would find the latest iterations of AI models to be thought provoking.

AI is a wonderful tool that is going to enormously accelerate human achievement and learning - for those that embrace it rather than fear it.

My son, who is a final year university engineering student is being encouraged to use AI tools, it is in fact a mandatory part of the syllabus.

Knuth (deep bow)! Back in another life before I retired I was a factory automation specialist - vison guided robotics and stuff like that. I can only imagine how much more effective I would have been had ai been around at even todays level.

I have been working on a web based automation project and using claude code cli as my assistant. The productivity is incredible! A task like refactoring code which can be a real vipers nest can be absolutely trivialised by claude. Debugging code? Once again trivial.

There are cavets. Claude needs guidance. Claude needs someone that understands good practices in the field to guide it. Claude will take shortcuts that result in poor practice or solutions that aren't universal and usable else where. But wow, with a little guidance in a technical domain it is outstanding.

 

[Pieter12]

AI could copy the style of any photographer, but it would be made and not real.

Sure. But AI has so many useful applications beyond that. Humans can copy the style of any photographer (or other artist, for that matter)--and not only is that real, it happens all the time.

 

[mshchem]

Knuth (deep bow)! Back in another life before I retired I was a factory automation specialist - vison guided robotics and stuff like that. I can only imagine how much more effective I would have been had ai been around at even todays level.

I have been working on a web based automation project and using claude code cli as my assistant. The productivity is incredible! A task like refactoring code which can be a real vipers nest can be absolutely trivialised by claude. Debugging code? Once again trivial.

There are cavets. Claude needs guidance. Claude needs someone that understands good practices in the field to guide it. Claude will take shortcuts that result in poor practice or solutions that aren't universal and usable else where. But wow, with a little guidance in a technical domain it is outstanding.

Ah the good side of technology. I worked for a year as an analytical chemist for General Mills, this was over 30 years ago. One of the products was a variation of Fruit Rollups. God awful fruit pectin and corn syrup sold as an alternative to "candy" . Early technology was using video to recognize the pattern of the product for qc.

I was amazed.

Unfortunately AI is going to be used to mislead and market. It's also going to revolutionize medical care, help people with every little thing. Drunk driving, car won't let it happen etc.

Good and bad.

 

[Alan Edward Klein]

Ah the good side of technology. I worked for a year as an analytical chemist for General Mills, this was over 30 years ago. One of the products was a variation of Fruit Rollups. God awful fruit pectin and corn syrup sold as an alternative to "candy" . Early technology was using video to recognize the pattern of the product for qc.

I was amazed.

Unfortunately AI is going to be used to mislead and market. It's also going to revolutionize medical care, help people with every little thing. Drunk driving, car won't let it happen etc.

Good and bad.

AI is a tool. Like a knife, it can be used for good and bad. You can cut up a chicken, or your neighbor.

 

[fotor]

Recently, I’ve been thinking about getting a short haircut, but I wasn’t sure which style would suit me. So I uploaded my photo and used AI to try on some trendy short hairstyles. In the end, I picked 4–5 styles that I think fit me best. This weekend, I plan to show these photos to my hairstylist and have them design the haircut that works for me. I really believe that using AI wisely can make things much easier and more efficient.

 

[loccdor]

It made the process of learning FreeCAD for 3d printing and 2d laser cutting much faster for me. The amount of effort it takes to learn certain pieces of software can definitely be a barrier for people like me with limited attention spans. In that way it's like being able to have an expert on hand to ask any dumb technical question - only you have to realize that maybe 1/3 of the answers will have some sort of mistake, but you can still at least piece something together from it.

@fotor Welcome to Photrio.

 

[tjwspm]

Despite all the criticism (which I share) and all the enthusiasm (which I also share), we should realize just how simple—not to say primitive—the algorithm behind this type of AI is. cirwin2010 described it perfectly:

AI/LLMs aren't true intelligence and work in "predictive" ways. At their core they predict what the next word, pixel, or result will be and chooses the "best" option.

ChatGPT ("Chat Generative Pre-trained Transformer") and all those LLMs ("Large Language Models") are exactly what their name implies: programs that have been pre-trained on massive amounts of text. They create databases containing statistical data on the frequency of individual words. The most readily available text sources today is the internet, and that is precisely what has been used.

The goal is to generate a response to a user’s prompt in the same way that a smartphone’s autocomplete feature works—only instead of a single word, it generates entire texts. The LLM simply adds the most likely next word based on its statistics, and so on.

What amazed me at first is that such a simple principle can generate such good texts without any understanding of the question nor the content of the answer.

But that is also the reason why it often struggles so surprisingly with math. If the calculation result is generated based on which digit most frequently follows the previous digits across the entire internet, then it probably won’t be the mathematically correct result.

You just have to use this software for what it’s intended for: summarizing or generating texts based on what the internet knows.

 

[gbroadbridge]

You just have to use this software for what it’s intended for: summarizing or generating texts based on what the internet knows.

You're describing the state of play 2-3 years ago. Not now.

 

[koraks]

You're describing the state of play 2-3 years ago. Not now.

Yup, it's clear @tjwspm hasn't had a lot of recent LLM runtime experience.
While it's technically true that it's predictive and that it doesn't really 'think' or have a concept of reality, if you aggregate the seemingly dumb 'summarizing and generating texts' to a higher level, it does things that rival and usually exceed the human ability to do things.

A good example is the generation of code. I've been using AI (Copilot and Claude) quite intensively for this (hobby projects). In the very recent past, it was kind of hit & miss whether a snippet of code would even compile, let alone do what you'd want it to do. Presently, we're at the point that the commercially available generation of e.g. Claude more or less autonomously can extend a project by adding modules, within an existing architecture, and make architectural changes as required to accommodate the extension of functionality. Likewise, it can build entire architectures from the ground up, and then implement them. For amusement's sake, the other day I had Copilot generate a specification and then asked Claude to implement it. The result just worked.

That's not to say I find the architectural decisions optimal or that I necessarily like them. I find that they generally deviate from how I had intuitively envisioned it, but I also recognize that this is mostly because I implicitly adhere to a couple of ground rules that I often fail to make explicit in my prompts (for the simple reason that I don't realize they're in the back of my head), and as a result, the LLM arrives at its own optimization of the problem. Currently, those optimizations and architectures vastly exceed what I could set up in a reasonable amount of time. AI just has a lot bigger repository of potential solutions that it can readily pick & choose from. It doesn't suffer from the same degree of mental limitation as my human brain.

This is just one specific example, and of course one that AI has been optimized for. That's also a matter of tactical choice on behalf of the AI companies: to make tools to help them develop the next generation of tools. It would be silly not to, after all.

Arguments involving "it's just this or that" are suspect by definition. The technology presently catches up with those arguments before they're even typed out.

 

[Cholentpot]

I used AI for nutrition and health routine.

I've lost 30lbs in 3 months, blood pressure is down and variety of small annoying issues have cleared up.

It took some tweaking and I'm due to tweak it again but people have paid more money than I can count to get to where I got and I did it all for free.

 

[tjwspm]

Yup, it's clear @tjwspm hasn't had a lot of recent LLM runtime experience.
While it's technically true that it's predictive and that it doesn't really 'think' or have a concept of reality, if you aggregate the seemingly dumb 'summarizing and generating texts' to a higher level, it does things that rival and usually exceed the human ability to do things.

A good example is the generation of code. I've been using AI (Copilot and Claude) quite intensively for this (hobby projects). In the very recent past, it was kind of hit & miss whether a snippet of code would even compile, let alone do what you'd want it to do. Presently, we're at the point that the commercially available generation of e.g. Claude more or less autonomously can extend a project by adding modules, within an existing architecture, and make architectural changes as required to accommodate the extension of functionality. Likewise, it can build entire architectures from the ground up, and then implement them. For amusement's sake, the other day I had Copilot generate a specification and then asked Claude to implement it. The result just worked.

That's not to say I find the architectural decisions optimal or that I necessarily like them. I find that they generally deviate from how I had intuitively envisioned it, but I also recognize that this is mostly because I implicitly adhere to a couple of ground rules that I often fail to make explicit in my prompts (for the simple reason that I don't realize they're in the back of my head), and as a result, the LLM arrives at its own optimization of the problem. Currently, those optimizations and architectures vastly exceed what I could set up in a reasonable amount of time. AI just has a lot bigger repository of potential solutions that it can readily pick & choose from. It doesn't suffer from the same degree of mental limitation as my human brain.

This is just one specific example, and of course one that AI has been optimized for. That's also a matter of tactical choice on behalf of the AI companies: to make tools to help them develop the next generation of tools. It would be silly not to, after all.

Arguments involving "it's just this or that" are suspect by definition. The technology presently catches up with those arguments before they're even typed out.

Thanks for the update on my 'experience,' @koraks. It’s fascinating to see that Claude is now making architectural decisions for you that exceed your own capabilities.

But be careful: just because the results (output) impress you doesn't change anything about the architecture under the hood (process). A flight simulator also flies incredibly realistically without ever leaving the ground. One can certainly acknowledge the impressive performance of the current generation without losing sight of the sober mathematical reality of the 'next-token predictor.

By the way, this response was AI-generated—proving my point that even a simple algorithm can deliver a decent punch.