Wow this interview is going to be hard to beat. You can feel the energy, excitement, inspiration, and imagination of Simon Willson in every response he gives.
“if you just ask them a question like you would on Stack Overflow, you'll probably not get a great response” - but that sums up the SO experience, because the majority of people who use it are unskilled at asking *good* questions. There is an extreme lack of thought process beforehand, plus the question isn't necessarily the right one to solve their problem. Asking good questions is absolutely a solid skill.
I *loved* the "life before Firebug" section which, of course, I remember vividly! :D Also the comments on how open source is now widely accepted in the (enlightened) business community. Thank you both for a great interview!
LLM are good in coding when it faces narrow problem. I have built code generators for scraping in my Parsera project. It can really boost you when you need to scrape websites, but this is a quite small task, not like writing the whole app or smth.
Systems Developers/Programmers are more than just code, the primary goal is to Solve Problems not to code, code helps you solve problems in a repeatable functional manner, those who compete on Code will be replaced hence the Existential Crisis
Is it just me ? or was Mr. Gergely cutting off(Interrupting) the flow of Simon Willison. I am so not coming at you with negativity, it's just something I observed. Like for example 27:35
Unfortunately that may be a necessity. I’ve always loved Simon, but he can speak infinitely because he’s immensely curious and remember a lot of peculiar details about topics he loves.
Here's why Anthropic cannot be used: "You may not access or use, or help another person to access or use, our Services in the following ways: - To develop any products or services that compete with our Services, including to develop or train any artificial intelligence or machine learning algorithms or models. - To decompile, reverse engineer, disassemble, or otherwise reduce our Services to human-readable form, except when these restrictions are prohibited by applicable law. ", i.e. I cannot use it to build AI which I will never agree to in any way whatsoever; "By submitting Prompts to our Services, you represent and warrant that you have all rights, licenses, and permissions that are necessary for us to process the Prompts under our Terms.", i.e. you cannot ask it about any kind of copyrighted material which I will not observe in any way shape or form, I will discuss whatever I like an nobody will restrict that in any way; "We may use Materials to provide, maintain, and improve the Services and to develop other products and services.", i.e. they own anything you submit or they generate. NO NO NO NO NO.
cool episode, I agree with most of what he said and had similar evolution as LLMs were upgraded. also share the exastion of trying to follow up with all the advancment and finding ways how to make use of them
21:00 you can ask the LLM to output JSON like {"value": true} in some case and null otherwise to make it deterministic, but not everything is true/false
Awsome talk! Thank you both of you. I would love to look over Simons shoulder how he would use chatgpt and claude to solve a beginner or more experienced coding problem. ❤
very useful and incitful. My take out is that LLMs need to be learnt, copilot is a good example of how if you dont find how to get the best out of it you will likely stop using it, that's what I've found also and you really need to put the time in t learn how to prompt depending upon the key strenghts and weeknesses of the LLM your using. Greate video !
Great interview, thought provoking. One thing I do worry about is what will happen to the "junior engineer"? The Simon Willison of the world is doing great, they can be more productive, and they can delegate the trivial work to the AI. But in the old world, jr engineer learn by doing that work so they can learn to become senior. In this new world order, there are no need for Jr Eng so company stop hiring them, how are new folks gonna get their training?
I think we're seeing what happens, job losses. In the short term think you're going to see people needing to DIY their learning a lot more, and become more generalists and full stack. Over time learning institutions will adapt their curriculum. We are moving away from a factory model of code production to one where code represents a working theory on how the world should interact with code. I think all devs will need to adapt and shift into this AI enabled paradigm or they will become irrelevant and jobless.
2 หลายเดือนก่อน +2
That's a good point. But it won't only be the junior engineers. There's no reason to think that AI will get stuck around this level and won't be a lot more capable and self sufficient in a few years and thus displace even the most experienced engineers and engineering/product managers. Also, even if not (or not yet) as you say we learn by doing (and learn because that is needed for doing). And we do know that those who stop doing start to forget and/or become irrelevant. We do know it because this is what has been happening to everyone who have moved into management. The "I used to be a programmer" guys. They forget a lot and also they don't learn the new things so become irrelevant. And then again there won't be too many people who understand and can work with the technology. And that will be pretty dangerous.
I wonder if since llms are trained on websites like stack overflow which people are now using less and less, will future models be worse because of the decrease in data?
i can do you better than that: will future models be trained on OSS projects that have permissible license that are hosted on github? how many of those would have codebases generated by AI and will that eventually lead to a worsening of training data
When was this interview recorded ? He doesnt talk about Chat GPT o1-preview and mini which are actually better than Claude AI he mentioned. Why he never mentioned these new Chat GPT versions ? and only ChatGPT 4 ?
We recorded it in June, when they were not yet out. Took time to edit and get ready for the first episode. New tools come out all the time: o1 is so fresh it takes time to form an opinion on how best to use it, what are helpful and not-so-helpful use cases etc.
This space changes so quickly there is no sense in focusing on particular models. The real value in this conversation is in the general practices and applications.
Good stuff. Some constructive feedback though: you constantly interrupted Simon over the course of the interview, which I'd encourage you to work on. It'll make your content much, much better if you give your guests room to complete their thoughts. It's easy to talk over one another when doing an online interview like this, but it's really worth working on this one. It's worth stressing that this comment comes from a good faith place, in the form of feedback.
@@konoko-o3o I am using vscode + copilot... But the options to switch the models, use @web to ask perplexity kind of things and all on one chat window the chat window to code merging are way simpler in cursor ... but that's just my opinion.... I don't have anything against vscode + copilot ... I am still using it but not that much.
38:08 Well, the answer for reusability has been both open-source and OOP, because all open-source frameworks and most open-source libraries use OPP, and for a very good reason. OTOH, open-source, once seen as a way to democratize tech, has been exploited by corporations especially to train AI models, leading to job insecurity and burnout for developers. Many contributors unknowingly fuelled the development of AI tools that now threaten their own careers. Big companies benefit, while developers, both contributors and non-contributors, are left struggling to keep up with the fast-paced changes. The open-source community has fallen into a trap where their passion has been commercialized, benefiting employers more than workers.
At 58:55 I disagree, now I've been typing faster more times than ever, my hands hurt when I overdo it... Typing prompts for chatGPT regarding documentation 😂. By the time someone asked one question I asked 10.
Wow this interview is going to be hard to beat. You can feel the energy, excitement, inspiration, and imagination of Simon Willson in every response he gives.
Would love to see a video of Simon Willison demoing his ai assisted coding workflow by working on a small project or open source thingy.
that would be great !
Exactly!
T 😢
“if you just ask them a question like you would on Stack Overflow, you'll probably not get a great response” - but that sums up the SO experience, because the majority of people who use it are unskilled at asking *good* questions. There is an extreme lack of thought process beforehand, plus the question isn't necessarily the right one to solve their problem. Asking good questions is absolutely a solid skill.
I *loved* the "life before Firebug" section which, of course, I remember vividly! :D Also the comments on how open source is now widely accepted in the (enlightened) business community. Thank you both for a great interview!
Thanks for the podcast!! Congrats on the 1st one. Great topic for the very first one.
Thankyou so much for sharing such amazing information. Its incredible keep these podcasts coming!!
LLM are good in coding when it faces narrow problem.
I have built code generators for scraping in my Parsera project. It can really boost you when you need to scrape websites, but this is a quite small task, not like writing the whole app or smth.
I usually read the newsletters but I really like this format!! Nice!
Let the man speak!
this is a really under-rated interview
Systems Developers/Programmers are more than just code, the primary goal is to Solve Problems not to code, code helps you solve problems in a repeatable functional manner, those who compete on Code will be replaced hence the Existential Crisis
Is it just me ? or was Mr. Gergely cutting off(Interrupting) the flow of Simon Willison. I am so not coming at you with negativity, it's just something I observed. Like for example 27:35
I thought so as well. Something to be mindful of in future interviews. Great job aside from that. Looking forward to the next one.
@@Cryptalux Yep Definitely Good overall coverage and questions asked.
Yes, several of those instances unfortunately. He does get back to the point but there’s this hunger to add to the conversation.
15:43 They are both doing that ¯\_(ツ)_/¯
Unfortunately that may be a necessity. I’ve always loved Simon, but he can speak infinitely because he’s immensely curious and remember a lot of peculiar details about topics he loves.
Here's why Anthropic cannot be used:
"You may not access or use, or help another person to access or use, our Services in the following ways:
- To develop any products or services that compete with our Services, including to develop or train any artificial intelligence or machine learning algorithms or models.
- To decompile, reverse engineer, disassemble, or otherwise reduce our Services to human-readable form, except when these restrictions are prohibited by applicable law.
", i.e. I cannot use it to build AI which I will never agree to in any way whatsoever; "By submitting Prompts to our Services, you represent and warrant that you have all rights, licenses, and permissions that are necessary for us to process the Prompts under our Terms.", i.e. you cannot ask it about any kind of copyrighted material which I will not observe in any way shape or form, I will discuss whatever I like an nobody will restrict that in any way; "We may use Materials to provide, maintain, and improve the Services and to develop other products and services.", i.e. they own anything you submit or they generate. NO NO NO NO NO.
Great interview, great guest!
Really cool interview. Keep 'em coming
cool episode, I agree with most of what he said and had similar evolution as LLMs were upgraded. also share the exastion of trying to follow up with all the advancment and finding ways how to make use of them
So hyped for TPE on YT and Spotify!
The natural language to sql is very cool, I had a similar wow moment when I got it to work
This was a fantastic podcast, lots of useful insights, thank you
21:00 you can ask the LLM to output JSON like {"value": true} in some case and null otherwise to make it deterministic, but not everything is true/false
Awsome talk! Thank you both of you. I would love to look over Simons shoulder how he would use chatgpt and claude to solve a beginner or more experienced coding problem. ❤
That first part is so true: AI in the hands of already skilled people will lead to amazing things.
very useful and incitful. My take out is that LLMs need to be learnt, copilot is a good example of how if you dont find how to get the best out of it you will likely stop using it, that's what I've found also and you really need to put the time in t learn how to prompt depending upon the key strenghts and weeknesses of the LLM your using. Greate video !
interesting conversation, thank you both
Great interview.
Great interview, thought provoking.
One thing I do worry about is what will happen to the "junior engineer"? The Simon Willison of the world is doing great, they can be more productive, and they can delegate the trivial work to the AI. But in the old world, jr engineer learn by doing that work so they can learn to become senior. In this new world order, there are no need for Jr Eng so company stop hiring them, how are new folks gonna get their training?
I think we're seeing what happens, job losses. In the short term think you're going to see people needing to DIY their learning a lot more, and become more generalists and full stack. Over time learning institutions will adapt their curriculum. We are moving away from a factory model of code production to one where code represents a working theory on how the world should interact with code. I think all devs will need to adapt and shift into this AI enabled paradigm or they will become irrelevant and jobless.
That's a good point. But it won't only be the junior engineers. There's no reason to think that AI will get stuck around this level and won't be a lot more capable and self sufficient in a few years and thus displace even the most experienced engineers and engineering/product managers. Also, even if not (or not yet) as you say we learn by doing (and learn because that is needed for doing). And we do know that those who stop doing start to forget and/or become irrelevant. We do know it because this is what has been happening to everyone who have moved into management. The "I used to be a programmer" guys. They forget a lot and also they don't learn the new things so become irrelevant.
And then again there won't be too many people who understand and can work with the technology. And that will be pretty dangerous.
I wonder if since llms are trained on websites like stack overflow which people are now using less and less, will future models be worse because of the decrease in data?
i can do you better than that: will future models be trained on OSS projects that have permissible license that are hosted on github? how many of those would have codebases generated by AI and will that eventually lead to a worsening of training data
When was this interview recorded ? He doesnt talk about Chat GPT o1-preview and mini which are actually better than Claude AI he mentioned. Why he never mentioned these new Chat GPT versions ? and only ChatGPT 4 ?
We recorded it in June, when they were not yet out. Took time to edit and get ready for the first episode.
New tools come out all the time: o1 is so fresh it takes time to form an opinion on how best to use it, what are helpful and not-so-helpful use cases etc.
This space changes so quickly there is no sense in focusing on particular models. The real value in this conversation is in the general practices and applications.
Good stuff. Some constructive feedback though: you constantly interrupted Simon over the course of the interview, which I'd encourage you to work on. It'll make your content much, much better if you give your guests room to complete their thoughts. It's easy to talk over one another when doing an online interview like this, but it's really worth working on this one. It's worth stressing that this comment comes from a good faith place, in the form of feedback.
loved Trak..
I wish this interview happened after new CharGPT model (strawberry). I would like to hear his opinions about reasonning.
Why are they not talking about cursor in tech stack. How is that everyone has not yet switched from vs code to cursor.
This was Simon sharing tools he uses - and he doesn’t seem to use Cursor! Lots of devs like it and use it: this episode was about stuff Simon uses.
Cursor is not a good tool if compare with vscode + copilot or event Jetbrains AI.
@@konoko-o3o I am using vscode + copilot... But the options to switch the models, use @web to ask perplexity kind of things and all on one chat window the chat window to code merging are way simpler in cursor ... but that's just my opinion.... I don't have anything against vscode + copilot ... I am still using it but not that much.
@@pragmaticengineer I agree.. And I loved it..
He just called every prompt engineer a dog 😂
What comes after Sonnet? I bet it is Monet.
Great talk and too many takeaways.
38:08 Well, the answer for reusability has been both open-source and OOP, because all open-source frameworks and most open-source libraries use OPP, and for a very good reason. OTOH, open-source, once seen as a way to democratize tech, has been exploited by corporations especially to train AI models, leading to job insecurity and burnout for developers. Many contributors unknowingly fuelled the development of AI tools that now threaten their own careers. Big companies benefit, while developers, both contributors and non-contributors, are left struggling to keep up with the fast-paced changes. The open-source community has fallen into a trap where their passion has been commercialized, benefiting employers more than workers.
Cursor is a way superior option for Copilot.
Simon grew hair a lot since I saw him last time.
35:00 I totally forgot that was a thing, JS has come a long way since then 🤣
28:45 So much work to learn this to become productive! I agree!
At 58:55 I disagree, now I've been typing faster more times than ever, my hands hurt when I overdo it... Typing prompts for chatGPT regarding documentation 😂. By the time someone asked one question I asked 10.