In addition to your excellent points, is the issue of affordability for many school systems and colleges. These tools aren’t free or the free versions are limited in scope. Many smaller schools, like the one I recently retired from, simply don’t have the budgets to provide anything but Copilot, which is built into their existing campus Microsoft subscription. This feels like yet another area where a whole lot of people will be left behind.
Claude was best for the kind of analytical and synthesizing work I do, but they changed their pro plan to make it far less useful. Last I looked, their pro plan was about as limited as a free plan; so I switched to chatGPT. I use it primarily, going to perplexity or claude for specific small tasks- citations, reviewing. Why did Claude decide not to support small users?
You can go pretty far with the free plan. My problem with the pro plan is that shortly thereafter, they put you on a weekly limit as well as the 5-hour.
I write fiction (micro- and otherwise), have done so for decades, and am very familiar with the frustration of writing myself into a corner I can see no way out of... as I recently did with a novel in progress.
Having read this post, I finally decided to give AI (specifically, Claude) a shot. I asked it the simple question: "Can you help me resolve a plot hole in a story?"
To say that the experience both exhilarated and flabbergasted me -- after maybe 15 minutes of back-and-forth Q&A -- would hugely understate my reaction. I can't believe I didn't think of the resolution on my own.
Going forward, I don't know that I'll make *frequent* use of the tool; I like writing -- and problem-solving -- far too much. But I was dazzled by this one experience.
I read your other article on reducing AI hallucinations and liked that very much. Didn’t like this though. For one thing, the tone is very authoritative, as if you have done an extraordinary amount of comparison research between these four models. I doubted that that was the case, and after some very preliminary research, realize that most of these points are outdated. Google Gemini, for instance, has a much larger context window than Claude. Similarly, all of the models can do web search so that’s not a differentiator either. I could go on, but the point is perhaps just, maybe stick to the higher level library science material, which is timeless, rather than pointing to specific models, which are all changing rapidly.
You’re right that model capabilities shift quickly, and any comparison risks aging faster than the underlying argument. My aim with that piece was to give readers a way to think about matching questions to tools. Still, if the tone came across as more authoritative than the evidence warranted, that is useful for me to hear.
Your point about the library science layer being more durable is well taken. That is the spine of the project, and I will keep refining how much technical specificity serves the reader versus distracts from the larger frame.
Thank you for reading closely and pushing on it. Conversations like this make the work sharper.
The thing is, comparison is how or one of the ways we learn. I didn't feel like you gave a bunch of rules to remember- you demonstrated how to think about the situation. And identified for people that different LLMs have different focuses. Perhaps there could have been more about this being structural than technical, but at least for me your message came through loud and clear. I do wish people would call AI LLMs more often. They may become artificial intelligence over time, but at the moment they're not. And I don't have a problem calling them AI, I just like to refer to them as LLMs now and then to remind people. Or at least give them the concept of what they really are. Working with an LLM is like dealing with an autistic 3-year-old that has extraordinary language abilities. It takes a certain skill set.
They are changing rapidly, but there is a real value to "learning the domain", I guess you would call it. Regardless of whether they're changing or not, we need to be developing our tools and understanding. We can't wait until all the weekly updates for every tool are completely completely done. We'd never get anywhere.
Yes, exactly! If we wait for the tools to stop shifting before we learn how to use them, we will wait forever. Learning the domain means understanding patterns, behaviors, strengths, and failure modes, even as the surface details change. The habit of asking which tool fits which question is a skill that carries forward.
Perplexity internet search is reliable. You have to stay on top of the other chatbots, as they offer responses and information but when you click on the reference links, verify yourself or if you already know you'll see they hallucinate.
In addition to your excellent points, is the issue of affordability for many school systems and colleges. These tools aren’t free or the free versions are limited in scope. Many smaller schools, like the one I recently retired from, simply don’t have the budgets to provide anything but Copilot, which is built into their existing campus Microsoft subscription. This feels like yet another area where a whole lot of people will be left behind.
What an outstanding guide. It reflects my trial-and-error experience very well. I'm may want to save this for future reference.
Claude was best for the kind of analytical and synthesizing work I do, but they changed their pro plan to make it far less useful. Last I looked, their pro plan was about as limited as a free plan; so I switched to chatGPT. I use it primarily, going to perplexity or claude for specific small tasks- citations, reviewing. Why did Claude decide not to support small users?
I haven't used Claude much but have used it *some*, with a free account. At what point did you hit a limit with the pro plan?
You can go pretty far with the free plan. My problem with the pro plan is that shortly thereafter, they put you on a weekly limit as well as the 5-hour.
this librarian is happy to read this post!
Me too!
I write fiction (micro- and otherwise), have done so for decades, and am very familiar with the frustration of writing myself into a corner I can see no way out of... as I recently did with a novel in progress.
Having read this post, I finally decided to give AI (specifically, Claude) a shot. I asked it the simple question: "Can you help me resolve a plot hole in a story?"
To say that the experience both exhilarated and flabbergasted me -- after maybe 15 minutes of back-and-forth Q&A -- would hugely understate my reaction. I can't believe I didn't think of the resolution on my own.
Going forward, I don't know that I'll make *frequent* use of the tool; I like writing -- and problem-solving -- far too much. But I was dazzled by this one experience.
Thank you as always, Hana!
Excellent article, thanks.
I read your other article on reducing AI hallucinations and liked that very much. Didn’t like this though. For one thing, the tone is very authoritative, as if you have done an extraordinary amount of comparison research between these four models. I doubted that that was the case, and after some very preliminary research, realize that most of these points are outdated. Google Gemini, for instance, has a much larger context window than Claude. Similarly, all of the models can do web search so that’s not a differentiator either. I could go on, but the point is perhaps just, maybe stick to the higher level library science material, which is timeless, rather than pointing to specific models, which are all changing rapidly.
You’re right that model capabilities shift quickly, and any comparison risks aging faster than the underlying argument. My aim with that piece was to give readers a way to think about matching questions to tools. Still, if the tone came across as more authoritative than the evidence warranted, that is useful for me to hear.
Your point about the library science layer being more durable is well taken. That is the spine of the project, and I will keep refining how much technical specificity serves the reader versus distracts from the larger frame.
Thank you for reading closely and pushing on it. Conversations like this make the work sharper.
The thing is, comparison is how or one of the ways we learn. I didn't feel like you gave a bunch of rules to remember- you demonstrated how to think about the situation. And identified for people that different LLMs have different focuses. Perhaps there could have been more about this being structural than technical, but at least for me your message came through loud and clear. I do wish people would call AI LLMs more often. They may become artificial intelligence over time, but at the moment they're not. And I don't have a problem calling them AI, I just like to refer to them as LLMs now and then to remind people. Or at least give them the concept of what they really are. Working with an LLM is like dealing with an autistic 3-year-old that has extraordinary language abilities. It takes a certain skill set.
They are changing rapidly, but there is a real value to "learning the domain", I guess you would call it. Regardless of whether they're changing or not, we need to be developing our tools and understanding. We can't wait until all the weekly updates for every tool are completely completely done. We'd never get anywhere.
Yes, exactly! If we wait for the tools to stop shifting before we learn how to use them, we will wait forever. Learning the domain means understanding patterns, behaviors, strengths, and failure modes, even as the surface details change. The habit of asking which tool fits which question is a skill that carries forward.
Perplexity internet search is reliable. You have to stay on top of the other chatbots, as they offer responses and information but when you click on the reference links, verify yourself or if you already know you'll see they hallucinate.