Thanks for sharing - really like the breakdown of how you integrated AI tools into a sprint. A few thoughts and questions came to mind:
- Claude vs. ChatGPT – Curious about your experience using Claude vs. ChatGPT in practice. From what I’ve read, aside from differences in training data and methodology, Claude seems slightly more nuanced and conversational but not dramatically different in output. Has it noticeably outperformed ChatGPT for you in certain cases?
- AI Tool Discovery – When it comes to discovering new AI tools, do you rely mostly on DeepResearch/ChatGPT (or other sources) to surface and assess tools? Would be great to hear if you have a go-to process for filtering.
- Customer Insights – DeepResearch and Societies.io seem like the most impactful tools from your list. I was wondering if you considered leveraging others in your sprint—were there any that didn’t make the cut, and if so, why?
Re: Claude vs ChatGPT. They have different personalities, but I use them largely interchangeably. Claude has projects that can have files, unlike ChatGPT, and I like its interface a bit better when it comes to building its "artifacts", but for most purposes they're similar. Sometimes, I run the same query in both to see which one I like better (e.g. summarising an article)
Re: discovery. It's a combination of Google, friends, LinkedIn, ChatGPT. There's no go-to process, I just see what others are using
Insights: Deep Research is definitely very important, whereas societies.io is simply promising at this stage. But I think they have potential. As for other tools, I don't use Perplexity that much, although others do, and maybe I should try again.
In my experience, they aren't a big problem. If perfect accuracy is a requirement, an LLM might not be the best tool: it's inherently probabilistic. Having said that, the rate of hallucinations seems to keep going down as models get better, including for the latest GPT4.5 released a few days ago. But in any case, everything very important is worth checking.
Obviously I'm shilling my own idea here, but I'm curious as to the level of trust that you place in Deep Research when it comes to such a big decision like stop/go on a start-up idea.
Did you share your deep research chat with someone with domain expertise to get their 15-minute input on a) your prompt and b) the highest signal part of its answer?
Imho, this would be v valuable and will become an increasingly common way of working
I wouldn't take Deep Research's answer at face value, but it's incredibly helpful at steering my thinking. But that's the same with people. If I'm looking to make a complex decision, I'll speak to many people, but I'll try to integrate their perspectives instead of blindly trusting one of them. So I find it very valuable, but as a thinking partner, not a source of final answers. And, on a couple of occasions, I did notice some minor mistakes (e.g. referencing a rule from FCA that isn't applied anymore but there's still a page online from 2005 describing it)
Hey Evgeny,
Thanks for sharing - really like the breakdown of how you integrated AI tools into a sprint. A few thoughts and questions came to mind:
- Claude vs. ChatGPT – Curious about your experience using Claude vs. ChatGPT in practice. From what I’ve read, aside from differences in training data and methodology, Claude seems slightly more nuanced and conversational but not dramatically different in output. Has it noticeably outperformed ChatGPT for you in certain cases?
- AI Tool Discovery – When it comes to discovering new AI tools, do you rely mostly on DeepResearch/ChatGPT (or other sources) to surface and assess tools? Would be great to hear if you have a go-to process for filtering.
- Customer Insights – DeepResearch and Societies.io seem like the most impactful tools from your list. I was wondering if you considered leveraging others in your sprint—were there any that didn’t make the cut, and if so, why?
Re: Claude vs ChatGPT. They have different personalities, but I use them largely interchangeably. Claude has projects that can have files, unlike ChatGPT, and I like its interface a bit better when it comes to building its "artifacts", but for most purposes they're similar. Sometimes, I run the same query in both to see which one I like better (e.g. summarising an article)
Re: discovery. It's a combination of Google, friends, LinkedIn, ChatGPT. There's no go-to process, I just see what others are using
Insights: Deep Research is definitely very important, whereas societies.io is simply promising at this stage. But I think they have potential. As for other tools, I don't use Perplexity that much, although others do, and maybe I should try again.
Also, the inaccuracies is a huge problem for me, did you find a tool to help fact check more efficiently at all?
In my experience, they aren't a big problem. If perfect accuracy is a requirement, an LLM might not be the best tool: it's inherently probabilistic. Having said that, the rate of hallucinations seems to keep going down as models get better, including for the latest GPT4.5 released a few days ago. But in any case, everything very important is worth checking.
Love this. Thanks for the detailed breakdown.
Obviously I'm shilling my own idea here, but I'm curious as to the level of trust that you place in Deep Research when it comes to such a big decision like stop/go on a start-up idea.
Did you share your deep research chat with someone with domain expertise to get their 15-minute input on a) your prompt and b) the highest signal part of its answer?
Imho, this would be v valuable and will become an increasingly common way of working
(more on why here: https://thefuturenormal.substack.com/p/chatgpts-deep-research-and-thinking)
I wouldn't take Deep Research's answer at face value, but it's incredibly helpful at steering my thinking. But that's the same with people. If I'm looking to make a complex decision, I'll speak to many people, but I'll try to integrate their perspectives instead of blindly trusting one of them. So I find it very valuable, but as a thinking partner, not a source of final answers. And, on a couple of occasions, I did notice some minor mistakes (e.g. referencing a rule from FCA that isn't applied anymore but there's still a page online from 2005 describing it)