I always appreciate these personal approaches and descriptions. It's good to know that you've been around the block several times, and you've settled on these particular tools. I myself have centered most of my research around ChatGPT Plus, and as a bonus, I get really good image generation. I use Gemini (or Google's Experimental Model) to read and "grade" my work - ChatGPT is still better at reviewing writing, but Gemini will notice some errors ChatGPT will not.
Perplexity has become a favorite too. It's amazing for quick research, probably better than the other 2 I mentioned. I've used the paid versions of GPT and Gemini, but only the free Perplexity model, and it is nearly as good as the paid models for my needs.
I might need a crash course on Perplexity one of these days. You're one of many people on Substack who seem to use it regularly, but while I've tried it out multiple times, I could never get it to "stick" in my routine.
Same. I have now started using it, especially the "Pro" mode for deeper, slightly more complex searchers. The "Pro" mode appears to be somewhere between traditional AI search and the "Gemini Deep Research" you mentioned in a separate comment. It makes an action/research plan and follows it, although it doesn't give you the option to edit it and doesn't crawl nearly as many pages.
Yeah, Perplexity is growing on me too but not to the extend where it becomes my go-to search engine. It's definitely a nice supplement for those in-between, semi-complex queries.
Oh hey, I actually *enjoyed* using Google's 2.0 Flash Thinking Experimental with apps (the title rolls off of the keyboard just as smoothly as it rolls off the tongue) today. That's a first for talking to Google about anything, ever. Big step up!
For me, if I already have an idea about something but want to verify info, it's very fast and very effective, and very transparent insofar as where the info is coming from.
I do note that what we're using these tools for very much determines which tool is the best to use. I'm not sure that was the case a year ago.
Agreed. Perhaps the reason I never got "into" Perplexity is that my current writing and research routine isn't a good fit for it. I mean, it's clearly a solid product, but I just don't find myself naturally navigating to it.
I've started using Napkin ai for ai based dynamic graphic images. It's kind of hard to explain, but it's my new go to for creating dynamic colourful images using prompts. Amazing
Oh yeah, Napkin's supposed to be great at turning concepts from text into visuals. I played around with it back when it was launched but never really used it for my own work. I might have to revisit it, thanks for reminding me.
Count me in as someone who's well aware of Perplexity but isn't using it actively. I remember trying it way back in early 2023 and even showcasing it as the future of search to some friends. I also know many people here on Substack who absolutely swear by Perplexity.
I've made several attempts at incorporating Perplexity into my daily life, going as far as following specific instructions on YouTube about setting up special browser extensions for it and using it as my default search engine. At the time, I went back to Google because I didn't like how Perplexity handled navigational queries - it'd take me to a Perplexity answer for a company instead of just opening their website when I'd type the name into the search bar.
So, while I recognize that Perplexity is a very useful tool for many people, it'd feel disingenuous to put it on a list of my personal used tools. I briefly considered putting it on the list (under "Research," exactly as you mentioned) but then I'd have to add the disclaimer that I don't use it, which might've given the impression that I don't approve of it. But the simple story is that it's just not in my daily repertoire. Perhaps it will be at some later stage.
Solid list Daniel and great for beginners. I remember when I caved to pay for GPT+ last year, one of my favorite $20/mo. That broke open a productivity door for me, and since, I've been exploring business-level applications like Kore.ai agentic builder, Clay for sales prospecting, etc.
Focused on productivity gains via AI in the next 6-months.. excited for what OpenAI and Perplexity have rolling out soon for + users in terms of Operator Agents. Cool stuff ahead!
Thanks for putting Kore and Clay on my radar. I typically focus on generic, jack-of-all-trades tools here on Why Try AI to keep things relevant to wider audiences, but there are so many niche business products out there for almost any purpose imaginable.
Would love to hear more about the way you use Kore and Clay and what your thoughts on them are, if you feel like sharing.
Happy you found this useful, Vic! And yeah, Grok 3 came out after my last update and it's pretty solid in my tests. But it will likely soon be hidden behind an X Premium paywall after the public beta phase, so let's see if it'll remain as accessible.
Hey Arthur, the answer depends a bit on the exact purpose you want to use the LLM for.
If it's for proofreading/grammar/beta reader feedback, then almost any LLM will do well (Gemini is my preferred "beta reader" as I wrote, but most chatbots will be helpful here.)
If you need help fleshing out the premise of a thesis and doing heavy research, then I'd consider a reasoning model like o1 or o3-mini in ChatGPT or DeepSeek-R1.
Well, DeepSeek-R1 is open-source. Technically, it's just "open-weights" as I described in my recent post: https://www.whytryai.com/p/open-source-vs-closed-source-ai, but so are Meta's Llama models and many other so-called "open-source" models.
But I can't say I use open-source models too much myself. They don't really have the same full-fledged interface as e.g. ChatGPT or Claude, so I don't find myself gravitating towards them for my needs. That might change in the future, though.
I use the paid Claude projects a lot but Inget annoyed a lot of times when it doe not adhere to intructions or prompts. Every single time I have to tell him not to include the examples but treat them as examples of my writing. And every time I ask for a, let say 300 word text, Claude gives me random word count usually way less than what I asked for. Never gives me the word count I ask for. Does anyone have solutions for this non-adherence problem?
Yes, getting LLMs to follow strict instructions is tricky. I've had similar experiences when building "Custom GPTs" - sometimes, including the same instruction even in multiple different ways and phrasings will still have GPT doing what you don't want it to do. For what it's worth, I learned to focus on what you WANT done instead of what you DO NOT want done. Focusing on negative instructions will sometimes have the opposite effect by anchoring the LLM to the thing you want to avoid.
The same goes for word counts and other very precise math-related instructions (e.g. help with writing a meta description of max 160 characters). Old-school LLMs aren't actually that solid when it comes to counting letters (the classic "how many r's in strawberry" example). The new reasoning models like DeepSeek-R1 and o1/o3 are definitely better on this front. Claude is still the old-school model, so I don't think you can expect that level of precision - it will always be "in the ballpark" rather than 100% on point.
It's a pity that Claude does not support web search at the moment. In terms of language and programming, it is, in my opinion, the best. I am also considering using Perplexity since it gives me the option to choose my model, even within the same chat. Grok has improved significantly in image creation as well, but I still prefer MidJourney.
Agreed. I think the lack of web access is one of the main "notches" against Claude right now. That, and the very limited bandwidth/rate limits for free accounts. I did try Grok's own "Aurora" image model and also found it pretty great. Midjourney is slowly losing ground to some newcomers, but on the other hand, they haven't had a new version out since late 2023 (!). I know they're working on V7 and I can't wait to see what that looks like!
My first hands-on experience with AI was through Canva. I started creating images for my templates. At the beginning they were weird but the improvement was increasing rapidly. Then I moved to ChatGPT (paid version) and now am using Adapta. I use ChatGPT to access Dall_e to create images; as I need to create content for my students, I upload images (photos) and ask AI to describe the scene. I gather both, the description (text) and and the image I created (JPG) and upload them with Canva to create a new template for a class. Those are the ones I use the most.
I encourage you to try other image models with the same prompts. DALL-E 3 is pretty good at instruction following, but I think it's falling a tiny bit behind in image quality. So try giving Ideogram or FLUX a shot!
I wrote about it and strongly considered pulling the trigger on Gemini Advanced free trial, but haven't gotten around to it yet. It looks really solid though!
I always appreciate these personal approaches and descriptions. It's good to know that you've been around the block several times, and you've settled on these particular tools. I myself have centered most of my research around ChatGPT Plus, and as a bonus, I get really good image generation. I use Gemini (or Google's Experimental Model) to read and "grade" my work - ChatGPT is still better at reviewing writing, but Gemini will notice some errors ChatGPT will not.
Perplexity has become a favorite too. It's amazing for quick research, probably better than the other 2 I mentioned. I've used the paid versions of GPT and Gemini, but only the free Perplexity model, and it is nearly as good as the paid models for my needs.
I might need a crash course on Perplexity one of these days. You're one of many people on Substack who seem to use it regularly, but while I've tried it out multiple times, I could never get it to "stick" in my routine.
For me it helps to think of Perplexity as a research tool, not a colleague that I’m chatting with. (Maybe this only makes sense to me though!)
Same. I have now started using it, especially the "Pro" mode for deeper, slightly more complex searchers. The "Pro" mode appears to be somewhere between traditional AI search and the "Gemini Deep Research" you mentioned in a separate comment. It makes an action/research plan and follows it, although it doesn't give you the option to edit it and doesn't crawl nearly as many pages.
I have little to add, but feel obligated since today's email triggered me like one of Pavlov's dogs.
I do use Perplexity from time to time, especially for a quick alternative "opinion" or incognito search where I don't want to sign in, etc.
Yeah, Perplexity is growing on me too but not to the extend where it becomes my go-to search engine. It's definitely a nice supplement for those in-between, semi-complex queries.
Oh hey, I actually *enjoyed* using Google's 2.0 Flash Thinking Experimental with apps (the title rolls off of the keyboard just as smoothly as it rolls off the tongue) today. That's a first for talking to Google about anything, ever. Big step up!
For me, if I already have an idea about something but want to verify info, it's very fast and very effective, and very transparent insofar as where the info is coming from.
I do note that what we're using these tools for very much determines which tool is the best to use. I'm not sure that was the case a year ago.
Agreed. Perhaps the reason I never got "into" Perplexity is that my current writing and research routine isn't a good fit for it. I mean, it's clearly a solid product, but I just don't find myself naturally navigating to it.
Yep. Same reason I don't use Midjourney every day.
I've started using Napkin ai for ai based dynamic graphic images. It's kind of hard to explain, but it's my new go to for creating dynamic colourful images using prompts. Amazing
Oh yeah, Napkin's supposed to be great at turning concepts from text into visuals. I played around with it back when it was launched but never really used it for my own work. I might have to revisit it, thanks for reminding me.
Did you miss to mention perplexitiy.ai?
I was expecting that to show up under research.
At this point, I'd be curious to hear from anyone who knows about perplexity and is _not_ using it. I guess that is not you, but just double-checking.
Hey Nico. Nope, haven't missed it.
Count me in as someone who's well aware of Perplexity but isn't using it actively. I remember trying it way back in early 2023 and even showcasing it as the future of search to some friends. I also know many people here on Substack who absolutely swear by Perplexity.
I've made several attempts at incorporating Perplexity into my daily life, going as far as following specific instructions on YouTube about setting up special browser extensions for it and using it as my default search engine. At the time, I went back to Google because I didn't like how Perplexity handled navigational queries - it'd take me to a Perplexity answer for a company instead of just opening their website when I'd type the name into the search bar.
So, while I recognize that Perplexity is a very useful tool for many people, it'd feel disingenuous to put it on a list of my personal used tools. I briefly considered putting it on the list (under "Research," exactly as you mentioned) but then I'd have to add the disclaimer that I don't use it, which might've given the impression that I don't approve of it. But the simple story is that it's just not in my daily repertoire. Perhaps it will be at some later stage.
Interesting
Solid list Daniel and great for beginners. I remember when I caved to pay for GPT+ last year, one of my favorite $20/mo. That broke open a productivity door for me, and since, I've been exploring business-level applications like Kore.ai agentic builder, Clay for sales prospecting, etc.
Focused on productivity gains via AI in the next 6-months.. excited for what OpenAI and Perplexity have rolling out soon for + users in terms of Operator Agents. Cool stuff ahead!
Glad you found it useful, Ian!
Thanks for putting Kore and Clay on my radar. I typically focus on generic, jack-of-all-trades tools here on Why Try AI to keep things relevant to wider audiences, but there are so many niche business products out there for almost any purpose imaginable.
Would love to hear more about the way you use Kore and Clay and what your thoughts on them are, if you feel like sharing.
Great list even for an old retired guy. I have added Grok 3. It is very handy for fact checking within X due to its accessibility.
Happy you found this useful, Vic! And yeah, Grok 3 came out after my last update and it's pretty solid in my tests. But it will likely soon be hidden behind an X Premium paywall after the public beta phase, so let's see if it'll remain as accessible.
This article is very informative! Thanks for sharing your thoughts on how AI tools work in different aspects of work.
Glad you found it useful Darius!
Excited to read it!!!
As a science professor helping PhD students write and edit their first paper or their thesis proposal what prompts work well and on which LLM?
Hey Arthur, the answer depends a bit on the exact purpose you want to use the LLM for.
If it's for proofreading/grammar/beta reader feedback, then almost any LLM will do well (Gemini is my preferred "beta reader" as I wrote, but most chatbots will be helpful here.)
If you need help fleshing out the premise of a thesis and doing heavy research, then I'd consider a reasoning model like o1 or o3-mini in ChatGPT or DeepSeek-R1.
Another great list of AI Tools
Happy you find it useful!
What about the open source models?
Well, DeepSeek-R1 is open-source. Technically, it's just "open-weights" as I described in my recent post: https://www.whytryai.com/p/open-source-vs-closed-source-ai, but so are Meta's Llama models and many other so-called "open-source" models.
Llama 3 was also on my list of "free LLMs" here: https://www.whytryai.com/i/146200626/llama-b-instruct-meta
But I can't say I use open-source models too much myself. They don't really have the same full-fledged interface as e.g. ChatGPT or Claude, so I don't find myself gravitating towards them for my needs. That might change in the future, though.
Thanks for your reply.
I use the paid Claude projects a lot but Inget annoyed a lot of times when it doe not adhere to intructions or prompts. Every single time I have to tell him not to include the examples but treat them as examples of my writing. And every time I ask for a, let say 300 word text, Claude gives me random word count usually way less than what I asked for. Never gives me the word count I ask for. Does anyone have solutions for this non-adherence problem?
Yes, getting LLMs to follow strict instructions is tricky. I've had similar experiences when building "Custom GPTs" - sometimes, including the same instruction even in multiple different ways and phrasings will still have GPT doing what you don't want it to do. For what it's worth, I learned to focus on what you WANT done instead of what you DO NOT want done. Focusing on negative instructions will sometimes have the opposite effect by anchoring the LLM to the thing you want to avoid.
The same goes for word counts and other very precise math-related instructions (e.g. help with writing a meta description of max 160 characters). Old-school LLMs aren't actually that solid when it comes to counting letters (the classic "how many r's in strawberry" example). The new reasoning models like DeepSeek-R1 and o1/o3 are definitely better on this front. Claude is still the old-school model, so I don't think you can expect that level of precision - it will always be "in the ballpark" rather than 100% on point.
I have been planning to update myself on best image generation AIs and now you have done that for me. Thanks!
You got it! Let me know what you think of Imagen 3 and the others.
It's a pity that Claude does not support web search at the moment. In terms of language and programming, it is, in my opinion, the best. I am also considering using Perplexity since it gives me the option to choose my model, even within the same chat. Grok has improved significantly in image creation as well, but I still prefer MidJourney.
Agreed. I think the lack of web access is one of the main "notches" against Claude right now. That, and the very limited bandwidth/rate limits for free accounts. I did try Grok's own "Aurora" image model and also found it pretty great. Midjourney is slowly losing ground to some newcomers, but on the other hand, they haven't had a new version out since late 2023 (!). I know they're working on V7 and I can't wait to see what that looks like!
My first hands-on experience with AI was through Canva. I started creating images for my templates. At the beginning they were weird but the improvement was increasing rapidly. Then I moved to ChatGPT (paid version) and now am using Adapta. I use ChatGPT to access Dall_e to create images; as I need to create content for my students, I upload images (photos) and ask AI to describe the scene. I gather both, the description (text) and and the image I created (JPG) and upload them with Canva to create a new template for a class. Those are the ones I use the most.
That's a very solid workflow, Anna!
It's actually something I talked about early last year: Using chatbots to help you describe and make better prompts for image models: https://www.whytryai.com/p/ai-images-chatbots
I encourage you to try other image models with the same prompts. DALL-E 3 is pretty good at instruction following, but I think it's falling a tiny bit behind in image quality. So try giving Ideogram or FLUX a shot!
Have you check out Gemini’s Deep Research yet? It’s pretty amazing. https://open.substack.com/pub/aigoestocollege/p/gemini-deep-research-a-true-game
I wrote about it and strongly considered pulling the trigger on Gemini Advanced free trial, but haven't gotten around to it yet. It looks really solid though!
Recently, this open-source model came on my radar that does something vaguely similar: https://github.com/InternLM/MindSearch/blob/main/README.md
It's definitely on my list to take them for a spin!
Ideogram is a game changer if you need text in images.
Yup, it's top-tier in my latest test from November 2024: https://www.whytryai.com/p/ai-image-model-spelling-text
But don't sleep on FLUX and Recraft if spelling is your jam!