AI makes the easy tasks easier and the hard tasks harder. Not only does it make a hard task harder, it pisses you off in the process. The 2nd half is a perfect example of my problems with these tools. You're just fighting a stupid AI
Those are one of the nuances when ai work on project tasks *without* human input and direction. Also, "stupid" ai. 'Stupid' is the keyword. Such ais are as competent. "Smart" ai should do the heavy lifting, while human coders/programmers not only oversee and guide/direct their progress, but to ensure that the steps needed to reach the project's goal and objective(s) are met. As far as task difficulty is connect, the _more complex_ the task are, the higher chance you'll run into caveats and nuances. This would require human reasoning, intuition, and creativity to address.
Think you might be right. Inexperienced dev here, basic coding, current project is just fairly straightforward text processing with a Streamlit UI. With a nice big local LLM (like Mistral Large V2 or Codestral 22bn), I can really fly. What would’ve likely taken about a week of Stackoverflow, head-scratching, and counting on fingers before, can be smashed out in an hour or few. And I learn a lot along the way. (“Btw, what does that part mean? What’s a good way to remember this? Can you explain why you used that library, rather than..” etc) But before I found a good one, I’d often end up in massive, circular, frustrating arguments with it: “Wait, why have you changed this other unrelated part though? I didn’t ask you to??”, “It’s still not working. New error message: …”, “are you sure? because.. don’t you need to..” etc etc I imagine more experienced folks, working on more complex projects, are hitting that wall a lot faster, and even with the best models? 🤷♂️
The problem is we usually don't feel dumb while being dumb. We must learn to become one with dumb and let it breathe through us. Then we shall learn. 😅
this entire video is a showcase of why i will never use copilot and stuff like it. it just wastes your time because you're trying to tell it what you want it to do in a way it can properly understand it, instead of taking that time to just make the feature yourself
nobody uses ai tools to do stuff like this tho, its used as new stackoverflow and boilerplate autocomplete and the whole story about trying to tell ai assistant what to do is pretty similar to situation with new/unexperienced devs - sure you can do this task faster, but you waste time for the hope of bigger returns (experience for another dev in second situation and experience in understanding of ai capabilities in first)
The code assistant tools aren’t quite ready for real work and they actually choke harder the more context you try to give them … but they do reasonable boilerplate code and simple stuff in a tiny context window. Inline code autocompletion is where I find AI assistance makes me most grateful. Second to that is when AI helps me write supporting scripts to do straightforward tasks that have a clear workflow. AI coding assistants simply cannot discern the logic of a multi-threaded application with many disconnected components without first going through a masters-level analysis to write out and explain everything to itself in an orderly way. Until you see your coding assistant doing a deep dive and cogitating for an hour about your codebase, don’t expect coherent results in any large project.
The real problem with these AI tools is the constant outdated APIs, lack of any connection of semver, and the annoying frequency to which they just straight up get things wrong. I’ve been an engineer for 25 years and I do find it helpful, I also find it equally frustrating.
He has to or the "gotcha" crowd will cry foul when they point out how he is a biased investor in what he's talking about if he doesn't mention it preemptively.
Watching this makes me even more sure of my decision to not use these ai tools. So much time fiddling with getting the ai to work, that could have been spent just doing it
Rounding up the current total population to 8.2 billion people to have 1 billion unique devs at a time would translate to approximately 12.2% of the total global population being 'developers', which if this were to occur there should be serious concerns about programming becoming such a large sector of global economy as any major shifts to the field would have the potential to see hundreds of millions of people lose their livelihoods which is a terrifying concept especially considering the AI the Microsoft says will help them reach that 1 billion developer number has already seen some developers replaced by AI
The demand isn't there, so they are trying to build up an artificial demand base... They are just putting a free sign at that fork in the road.. and once they get you going down their road: they'll eventually get you to pay for the full version.
Exactly! They hope that if enough future devs spend the time we spent learning to code, learning to wrestle AIs, they won't be able to do this themselves anymore and depend on the AIs.
11:35 I disagree, the fact that they support Claude is a big selling point for me because I know it's generally better at code completion. Lots of devs interested in AI tools are probably aware of this too
The point is not that Claude is not better, it's that having to manually select a model is not a feature, instead they should know what works best for what and auto assign it for the current task.
I've found, that generally Claude is better, but in translations GPT-4o gives better results. Claude can sometimes use weird constructs or even invent new words in Polish.
@@marcola80 different models are better at different coding tasks, in addition the pace of how quickly new models flagship come out , it's probably easier to give the choice to the user than abstract that away
I personally want to keep AI out of my editor. I like using AI for chat and solving problems, but I want to be in charge of what goes into my application and what doesn't.
You still are, lol. You don't have to accept the changes im confused. Why do people make this excuse? Are you sure you just don't want to use it in general?
For a task like adding that text, by the time you fight the model, go look for the file to let it know, then run it again, you could have just added the text. A senior dev without AI would still be able to run circles around a junior that heavily relies on it.
7:50 it never stops surprising me how much differently people approach learning new things in terms of having fun - personally I find these beginnings of learning something new to be very fun, and possibly where my curiosity peaks - it's once that the problem is solved that it's not as fun anymore. I learn human languages too and it's similar, I'm having a lot of fun in the beginning when everything's new and I'm trying to figure everything out, and while many people hate this stage for me this novelty provides enough of a kick to keep me motivated and engaged for a long time on its own.
Entire section of where Theo was fighting with the AI to make it copy an images background color was more frustrating than actually trying to solve the problem yourself, and when the problem was "solved" (code was quickly skimmed through without even checking what kind of logic it used to do this) there was a SIGH of relief, instead of having a dopamine rush for solving a problem yourself, not only do you learn from the experience, have a REFERENCE to come back to AND learnt in the process, that the next time you have to do something remotely similar, it'll be 100x easier.
He's "super energized" about this? Idk, sounds like copilot wrote its own promo there. Who says that. Anyway, I mean, it's cool, I'm sure many devs will appreciate having access to it for free, as limited as it might be. But I agree, it's dangerous if new devs will rely on it too much instead of learning the details themselves.
great i feel bad for all us open source devs that are going to be inundated with crappy pull requests made by ai. it s already getting annoying just going through human pull requests. Sooo many ai slop code , and why cant ai write short function? have to read so much
We can battle AI pull requests with AI PR reviews but it needs to be fine-tuned to your repository style and commits first. I think in the end all things combined (Copilot + generic AI reviewing your code + fine-tuned AI giving feedback on PR) should increase an average PR quality, not decrease it.
Ai is not the problem haha it's new devs thinking the are gods cause of ai I'm pretty experienced dev game make my own game engine in c++, c#, etc without ai and I got to say that it's amazing how fast i can get a setup going with actual Clean, modular, robust code!
@@Aves_1 I'm not talking about AI approving code. I'm talking about helping to review it, meaning it can make suggestions. Approval and actual PR updates is still has to be done by humans. The point is to make improvements to submitted code that human would have to waste their time on. Things like refactoring, variable naming, etc. Maybe if AI is smart enough it can notice some security issues too. And then human makes a final review to see if PR actually makes sense and there are no problems.
Sr engineer here, I don't know about y'all, but I tend to write code faster without AI. Maybe there's a use cases for helping with finding documentation/examples and writing simple algos
I find it weird that you wouldn't want developers to have the option to change out the model. Even if you have a model that runs other models, you want to have variations of that to select from. There will never be a one size fits all AI that is the best at everything, and you should be able to swap AI models like you can swap game cartridges in a game console.
It's crazy how both tools use external AI models and Copilot keeps managing to shoot itself in the foot. Cursor seems interesting but it's a bit expensive and I feel like soon (hopefully) Github will catch up in speed and UX.
When I tried cursor I figured it couldn't be much different from vscode w/ copilot, but I have had basically the same experience as you show here. Vscode is just so much more frustrating to use, especially with a complex-ish problem. Cursor's ux is miles ahead and even using the same models it tends to give tons better solutions. I wonder how many devs will try the free copilot and have their first experience be super annoying and full of bugs. Can't imagine they'd spring for the paid tier after that.
This is obviously a great tool to have in a dev's belt, and it's definitely already a part of the world; I am still really concerned about the sudden massive increase in electricity usage of all of these AI queries.
Many devs don’t write system code, yet we got Zig and Ghostty. The abstraction lets you be useful and then deep dive / specialty in what you want to do down the line. This hasn’t changed, this filter / level of abstraction has always existed. It’s how I got started. I didn’t seek to understand anything at the start 20 years ago, I just wanted to bash together code segments (from textbooks). Being a broke college student, I was sitting in book stores trying to figure out how to put Java together. Haha. It’s only after a large exposure and 10 years before I started to care and choose what to love, currently that’s tensors.
I've recently ditched VS Code for Cursor, because Copilot was really annoying and dumb even on the paid tier, so I am definitely not coming back because of this.
Having a free tier is great because it's encouraging competition, that said, you've completely put me off trying copilot, my experiences with claude and cursor have been pretty much the same as yours here. It just seems to get context so much better even if it is pretty expensive when doing it. Also I have rarely actually been interested in a sponsor, have never encountered Sevella before but I think they could solve an issue for me in a way I actually hadn't considered before, so thanks for that!
AI is great for writing the kind of boilerplate that I used to have hundreds of snippets for (Bootstrapping components, writing imports, setting up database connections, looping syntaxes, etc) or for replacing intellisense in reminding me of the syntax for certain functions. But I definitely wouldn't trust it for actual logic. It's not nearly reliable enough for that yet.
I think it's really useful, that you can choose the models, you want to use, because in my opinion some models are better in some specific stuff, than others. For example if I want to generate a modern frontend I definitely choose Claude, because their model is way better than openais model... A person who doesn't know or want to change that stuff probably doesn't care, that there is a dropdown for it.
Ahh i paid for it and now its actually free(actually i didnt i got it with the student email lol). But anyways copilot isnt that helpful normally for learning, you will just be spamming tabs. Only time i have used it was in one rust 3rd party component library where documentation was obscure and i couldnt read the whole code as easy, so i just fed the entire module to copilot and had it suggest me what traits and options were available for certain conditions i required. Anyway i use zed, it already has a free AI sidebar, though i only use it when stuck on how to do something like syntax. Though i always have inline completion turned off cause it feels like its hogging screen and too much happening in screen with the hints wjen writing code like it is showing me my function as i write but incorrect one so its kinda distracting.
Maybe you should go back to basics if you have issues with syntax (assuming that it isn’t a new language you’re using; but even then you aren’t going to learn it properly if it just tells you)
@@aquilafasciata5781 no i mean about some obscure external libraries, like in rust i use different crates for different purpose and they have different custom syntax for colors like for some its just color while others may have viewcolor, some have support for rgb while others only support 8 basic colors and so on. Finding documentation for them is also hard cause many of these are new crates and their wiki isnt properly updated or is as thorough. So i just scour through their library to see how structs and enums and traits are defined to understand how they work but some times they r quite huge so that where i use the AI mostly to list out all those available functions and syntax
I code with nano as my editor (I haven’t used any other terminal editor), over ssh. Don’t question it, but it means I can’t use Copilot and I probably wont ever, as long as I code like this
Related to his last question. Tested windsurf within the first 5 days after launch, and it was great, unlimited everything. Then did not used it for a month and now it is useless and everyone I've read on blogs says is not worth it and many going back to cursor. Is paying the $20 or $60 worth it or still will reach the limits too quick? I've read plenty bad experiences, has any heavy user had a good experience to tell with any of the paid tiers?
Does Microsoft pay for GPT and Claude on the user's behalf, or does the developer have to pay separately for the AI service? Because if it is the latter, then Copilot is largely the wrapper and is not the expensive part of the stack, so it being free isn't even that much of a dumping move.
i have been using codium and copilot, i mostly use it to write small chunks of code and never complete functionality because it almost never gets it write, instead i would myself break thhe problem down into subproblems and then use autocomplete feature so the AI can solve the smallest subproblem and then when all of them are solved my orignal problem gets solved, its like i am a conductor of an orchestra guiding different models and tools to solve the problem for me instead of letting them solve the problem thmeselves
I noticed that you mentioned that file hopping is something these tools are not good at, have you tried windsurf yet? The thing I noticed is that Windsurf is really good at that, somehow its more capable of making more sense of the context, While the Cursors DX is better when coding on your own.
Codeium, IMO, did a bit of a dick move a while back. I had just paid my first $10 when they decided to up that to $15, which was borderline ok, but they also hamstrung that tier badly with such a low token count that it's way less than Copilot's free tier. Especially as it looks like Codeiums model eats through tokens really fast.
I tried it yesterday and "/fix" broke my code. It took an undefined let and assigned it as an empty array. It didn't care it was used for an if condition, it decided it should return true always. Thanks copilot!
"I'm an investor in Microsoft", it's a trillion dollar company lmao, what you say wouldn't have an impact on the stock at all. Besides, 80% of Americans are investors in MSFT since they all have 401ks and it's part of the S&P 500.
@@AbstruseJoker No I don’t think so. More like “there’s a small conflict of interest because I have skin in the game. It might be affecting my opinions. And so it’s only fair that I should state that up front. So you can take it into account and slightly discount for example any positive spin or excuses I make, for things that aren’t very good.”
0:19 Yes you used it for advent of code, where you maybe write 100 locs max in probably JS or python which are the 2 most used languages for similar problems. Try it on a *real* problem and see it die. Try writing a Windows kernel driver. You'll bluescreen yourself before you can say "AI can only solve trivial problems in an empty code base".
HARD disagree on "eating your vegetables for you." If you want to be healthy you need to LOVE eating your vegetables. You need to develop a palate for vegetables. The same goes here for code. You need to find appreciation in the tedious things, especially for maintability purposes. Making a variable naming convention that you and your team understand at a glance, refactoring your code to make it yours and understand to your core exactly how it works, so that when the time comes to revisit, you have a holistic understanding how why you wrote that code in the first place. I'm not saying ditch the AI at all, hell no. I'm sayinf have the AI help you, but you do the implementation. I don't think people are ready for the technical debt they'll be hit with when every small thing would have been made by AI.
I'm wondering if there will be languages that "win" or "lose" as a result of tools like copilot. LLM's are not magic they need to be fed coding examples before they can be useful. Also I wonder if it be harder now to build a new language and get it adopted? Maybe not if you could make it LLM friendly.
The Satya era at the very least refocused Microsoft towards Azure as a profit center therefore taking some of the weight off of Windows. This meant Microsoft was no longer at war against Linux and at least up to some point embraced FLOSS. VSCodium (VSCode is only open source as source), .Net, parts of WSL are great and it is great to have them as open source. These should not be taken as granted and Microsoft can at any point return to the Embrace-Extend-Extinguish strategy if they are in a position of strength. The Satya era also meant they abandoned Windows on phones (even though Win10 Mobile was finally pretty good), Windows suffered from enshitification, more tracking, more subscriptions, a forcefull push for Azure dependence, more abandoned UI frameworks, selling the generally apreciated peripheral division, launching and shutting down WSA and of course all of the AI bullshit.
@RemotHuman rather than integrating Copilot I am more worried about the weird licensing of VSCode which bans the open source version from several things including the extension store, the control over .Net including the discontinuing of Mono and the contol they have over GitHub which so far has not had major detrimental effects.
I'd rather pay 60$ for unlimited chats and 3000 tool runs in Windsurf. It's been great. The base autocomplete model has unlimited uses and is good enough.
If the product is free, you are the product. This is a clear grab for as much code as possible to improve the models. Not that that’s an inherently bad thing if it’s moves us forward, but let’s be clear about what it is.
it's not only bash, the whole toolchain. Most people who heavily rely on IDEs (say Visual Studio or Rider) plus all DI and 'magic' and that's all they know, have no idea what's happening in the background. Anyhow, I personally have more fun with an editor like Code even when writing C#.
I myself dont use copilot for actual programing completion or writing but as an internal chat to ask for documentation and or recommendations for something specific i want to do I have learnt js like this and its great for asking how to syntax what it means and where to use it usually Its great, now for code completion only if i really need "code that works now" and then to alter it for my specific use case because a lot of times the code just doesnt work (It solved the what and why problem but not the will you just f*cking work problem)
i like the thumbnail, the co-pilot tapeworm… the ai tapeworm sums up ai, the largest ‘theft’ if intellectual property in human history. all if our work gas been used to train these ai tools, enriching their owners… who amongst has been asked to make this donation… and i don’t count permission gained by small print in click over t&c’s
Meanwhile Cursor keeps getting worse. The queue'd slow requests gets longer and slower as it grows. It's code suggestions have been steadily getting less accurate, requiring ever increasingly more requests to get what you want, at your cost no less. And Cursor is showing they don't care, just look at the forum. They never directly address this problem instead either don't answer at all or ask irrelevant questions
Truthfully just using AI for templating and concepts, then I manually go do all the other things. Everything has seemingly gotten worse over time, but that may be just me.
5:32 - "[The Copilot-authored PR] didn't do what it was asked, and it added a new license to the repo..." An AI company, _lying_ to exaggerate its usefulness? _Gasp!_ Say it isn't so!
Microsoft can do free stuff because globally they have extra AI capacity they have to have to address paying users peak demand. AWS and Google ditto. All the other players in general have to pay by the token for someone else's capacity, or have premium cost owned instances.
50 isn't much especially when AI gives so many errors. I'm in AI hell with Windsurf/Claude right now and I have to wait 2 weeks for my quota to reset. Wouldn't say it's a ton of fun!
Some of you will run situation in your project where the I doesn't do as it should, likely being not as competent or not performing well as it should. Its good to have around for mundane or assistive task (that aren't to overly complex). But do be _too dependent_ on it. You stumble upon a problem where the ai takes on a project task *without* human input and direction. "Smart" ai should do the heavy lifting, while human coders/programmers not only oversee and guide/direct their progress, but to ensure that the steps needed to reach the project's goal and objective(s) are met. As far as task difficulty is connect, the _more complex_ the task are, the higher chance you'll run into caveats and nuances. This would require human reasoning, intuition, and creativity to address. Nothing wrong with using tools, but don't expect it to do everything for you, lest you stay stagnant in your skills as a coder and programmer.
Do yourself a favour and try use the VS code extension called Cline. It is way better than copilot. You can choose whatever model you want even locally hosted models like deepseek. I use it with Claude Sonnet and it is so good. I prefer it over any of the other AI editors..
"I have been very surprised with how much more fun coding has been since [AI]" cue 21 minutes of unmitigated stress (and a sponsor slot).
Hey, making content profiting off of AI shit show is fun!
LMAOOOO
lololololol so true. Theo groaned so many times and eventually said how he would do it himself by the end. How fun!
It's fun if you use it for auto-completion, and even then it can be overly helpful and thus annoying.
2,000 completions and 50 chats per month isn't free, it's a trial
yep. opted out of the paid version on 22nd last month. less than a week i had reached the autocomplete cap
@@titato I might be using tabnine extension for now, just getting tired of writing the redundant line of code
Obviously the end goal of any free tier is conversion to a paying customer. What's even the alternative? Running the service costs money on their end.
him calling that reasonable is wild lmao
Isn’t it per month
AI makes the easy tasks easier and the hard tasks harder. Not only does it make a hard task harder, it pisses you off in the process. The 2nd half is a perfect example of my problems with these tools. You're just fighting a stupid AI
So don't use AI for hard tasks
Those are one of the nuances when ai work on project tasks *without* human input and direction.
Also, "stupid" ai. 'Stupid' is the keyword. Such ais are as competent. "Smart" ai should do the heavy lifting, while human coders/programmers not only oversee and guide/direct their progress, but to ensure that the steps needed to reach the project's goal and objective(s) are met.
As far as task difficulty is connect, the _more complex_ the task are, the higher chance you'll run into caveats and nuances. This would require human reasoning, intuition, and creativity to address.
The moment it does hard tasks is the moment we become obsolete
Think you might be right. Inexperienced dev here, basic coding, current project is just fairly straightforward text processing with a Streamlit UI. With a nice big local LLM (like Mistral Large V2 or Codestral 22bn), I can really fly. What would’ve likely taken about a week of Stackoverflow, head-scratching, and counting on fingers before, can be smashed out in an hour or few.
And I learn a lot along the way. (“Btw, what does that part mean? What’s a good way to remember this? Can you explain why you used that library, rather than..” etc)
But before I found a good one, I’d often end up in massive, circular, frustrating arguments with it: “Wait, why have you changed this other unrelated part though? I didn’t ask you to??”, “It’s still not working. New error message: …”, “are you sure? because.. don’t you need to..” etc etc
I imagine more experienced folks, working on more complex projects, are hitting that wall a lot faster, and even with the best models? 🤷♂️
@@ilikenika There's going to be industry/military cps on it. There's no way any tech is being 'allowed' without the strictest ordinances.
It's free and I still want my money back.
yes
its not free, its a trial of 50 messages
It is free, I think you have a problem with the definition of a "trial"...
@@flrn84791 It's a free trial that resets every month. lol
Feeling dumb is the fun part. Learning and finally getting that "ah ha!" moment is the dopamine rush that makes me want to feel dumb more
The problem is we usually don't feel dumb while being dumb. We must learn to become one with dumb and let it breathe through us. Then we shall learn. 😅
this entire video is a showcase of why i will never use copilot and stuff like it.
it just wastes your time because you're trying to tell it what you want it to do in a way it can properly understand it, instead of taking that time to just make the feature yourself
nobody uses ai tools to do stuff like this tho, its used as new stackoverflow and boilerplate autocomplete
and the whole story about trying to tell ai assistant what to do is pretty similar to situation with new/unexperienced devs - sure you can do this task faster, but you waste time for the hope of bigger returns (experience for another dev in second situation and experience in understanding of ai capabilities in first)
Right? If only there was a system of precise language we could use to tell computer exactly what we want them to do.
@@ryannathanwilson😂
so you're saying you just don't understand any of it
bro free tier is like 15mins of use a month
Lol did he actually cancel
The code assistant tools aren’t quite ready for real work and they actually choke harder the more context you try to give them … but they do reasonable boilerplate code and simple stuff in a tiny context window. Inline code autocompletion is where I find AI assistance makes me most grateful. Second to that is when AI helps me write supporting scripts to do straightforward tasks that have a clear workflow. AI coding assistants simply cannot discern the logic of a multi-threaded application with many disconnected components without first going through a masters-level analysis to write out and explain everything to itself in an orderly way. Until you see your coding assistant doing a deep dive and cogitating for an hour about your codebase, don’t expect coherent results in any large project.
The real problem with these AI tools is the constant outdated APIs, lack of any connection of semver, and the annoying frequency to which they just straight up get things wrong. I’ve been an engineer for 25 years and I do find it helpful, I also find it equally frustrating.
"yes sorry, this functionality indeed does not exist"
Outdated APIs. From the last 3 years. Are you serious?
@@bootmii98 yes, JS/TS and modern Python libraries change frequently. Even the best AI gets these things wrong all the time.
bro loves flexing his investments on us
If you see that as a flex, says more about you than him 🤷🏽♂️
He has to or the "gotcha" crowd will cry foul when they point out how he is a biased investor in what he's talking about if he doesn't mention it preemptively.
Was thinking the same thing
Usually public figures have to share their investments if they are going to talk about information that can alter the stock price
@ bingo 💯
Watching this makes me even more sure of my decision to not use these ai tools. So much time fiddling with getting the ai to work, that could have been spent just doing it
They are useful to have around. Not to depend on. When you need boiler plate they are a godsend.
Rather have a competent ai to assist you, not entirely depend on.
Then you misunderstood how they can be useful 🙈
@28:30 "Enable one billion developers..." ??? WTF? Since when did we (developers) exceed 10% of the World's TOTAL POPULATION? What is smoking?
Rounding up the current total population to 8.2 billion people to have 1 billion unique devs at a time would translate to approximately 12.2% of the total global population being 'developers', which if this were to occur there should be serious concerns about programming becoming such a large sector of global economy as any major shifts to the field would have the potential to see hundreds of millions of people lose their livelihoods which is a terrifying concept especially considering the AI the Microsoft says will help them reach that 1 billion developer number has already seen some developers replaced by AI
The demand isn't there, so they are trying to build up an artificial demand base... They are just putting a free sign at that fork in the road.. and once they get you going down their road: they'll eventually get you to pay for the full version.
Exactly! They hope that if enough future devs spend the time we spent learning to code, learning to wrestle AIs, they won't be able to do this themselves anymore and depend on the AIs.
"There's a Smoking Hot Babe at the Top!"
11:35 I disagree, the fact that they support Claude is a big selling point for me because I know it's generally better at code completion. Lots of devs interested in AI tools are probably aware of this too
just use cursor
I agree with you I use claude it's better
The point is not that Claude is not better, it's that having to manually select a model is not a feature, instead they should know what works best for what and auto assign it for the current task.
I've found, that generally Claude is better, but in translations GPT-4o gives better results. Claude can sometimes use weird constructs or even invent new words in Polish.
@@marcola80 different models are better at different coding tasks, in addition the pace of how quickly new models flagship come out , it's probably easier to give the choice to the user than abstract that away
I personally want to keep AI out of my editor. I like using AI for chat and solving problems, but I want to be in charge of what goes into my application and what doesn't.
You still are, lol. You don't have to accept the changes im confused. Why do people make this excuse? Are you sure you just don't want to use it in general?
For a task like adding that text, by the time you fight the model, go look for the file to let it know, then run it again, you could have just added the text. A senior dev without AI would still be able to run circles around a junior that heavily relies on it.
One of the best features is searching your codebase with AI. I think this is underrated.
I still don’t understand this concept man. Is it like a folder with files in it or everything in the ide editor
14:04 copilot did get it correct. he just needed to hit apply
In other words, horrible UX/UI. Not surprising as it's Microsoft and they're atrocious with UX/UI.
Our planet is warming because Theo kept forgetting to select the right tab. RIP 🌎.
7:50 it never stops surprising me how much differently people approach learning new things in terms of having fun - personally I find these beginnings of learning something new to be very fun, and possibly where my curiosity peaks - it's once that the problem is solved that it's not as fun anymore. I learn human languages too and it's similar, I'm having a lot of fun in the beginning when everything's new and I'm trying to figure everything out, and while many people hate this stage for me this novelty provides enough of a kick to keep me motivated and engaged for a long time on its own.
Entire section of where Theo was fighting with the AI to make it copy an images background color was more frustrating than actually trying to solve the problem yourself, and when the problem was "solved" (code was quickly skimmed through without even checking what kind of logic it used to do this) there was a SIGH of relief, instead of having a dopamine rush for solving a problem yourself, not only do you learn from the experience, have a REFERENCE to come back to AND learnt in the process, that the next time you have to do something remotely similar, it'll be 100x easier.
He's "super energized" about this? Idk, sounds like copilot wrote its own promo there. Who says that.
Anyway, I mean, it's cool, I'm sure many devs will appreciate having access to it for free, as limited as it might be. But I agree, it's dangerous if new devs will rely on it too much instead of learning the details themselves.
great i feel bad for all us open source devs that are going to be inundated with crappy pull requests made by ai. it s already getting annoying just going through human pull requests. Sooo many ai slop code , and why cant ai write short function? have to read so much
We can battle AI pull requests with AI PR reviews but it needs to be fine-tuned to your repository style and commits first.
I think in the end all things combined (Copilot + generic AI reviewing your code + fine-tuned AI giving feedback on PR) should increase an average PR quality, not decrease it.
that's the progress everybody is talking about
@@alpuhagamewe do not need ai pr reviews. An AI approving whatever code it wants into a project is a terrible idea
Ai is not the problem haha it's new devs thinking the are gods cause of ai I'm pretty experienced dev game make my own game engine in c++, c#, etc without ai and I got to say that it's amazing how fast i can get a setup going with actual Clean, modular, robust code!
@@Aves_1 I'm not talking about AI approving code. I'm talking about helping to review it, meaning it can make suggestions. Approval and actual PR updates is still has to be done by humans.
The point is to make improvements to submitted code that human would have to waste their time on. Things like refactoring, variable naming, etc. Maybe if AI is smart enough it can notice some security issues too.
And then human makes a final review to see if PR actually makes sense and there are no problems.
Like price comparison sites in the 00's, we need an AI that picks the best AI model for the situation
We need an AI that figures out which AI to pass your code to.
Sr engineer here, I don't know about y'all, but I tend to write code faster without AI.
Maybe there's a use cases for helping with finding documentation/examples and writing simple algos
Good reminder to cancel my copilot plan.
I find it weird that you wouldn't want developers to have the option to change out the model. Even if you have a model that runs other models, you want to have variations of that to select from. There will never be a one size fits all AI that is the best at everything, and you should be able to swap AI models like you can swap game cartridges in a game console.
It's crazy how both tools use external AI models and Copilot keeps managing to shoot itself in the foot.
Cursor seems interesting but it's a bit expensive and I feel like soon (hopefully) Github will catch up in speed and UX.
When I tried cursor I figured it couldn't be much different from vscode w/ copilot, but I have had basically the same experience as you show here. Vscode is just so much more frustrating to use, especially with a complex-ish problem. Cursor's ux is miles ahead and even using the same models it tends to give tons better solutions. I wonder how many devs will try the free copilot and have their first experience be super annoying and full of bugs. Can't imagine they'd spring for the paid tier after that.
As a developer pumping data to feed this billionaire, possibly trillionaire industry, I think that having a free tier to use AI is just repayment.
This is obviously a great tool to have in a dev's belt, and it's definitely already a part of the world; I am still really concerned about the sudden massive increase in electricity usage of all of these AI queries.
Many devs don’t write system code, yet we got Zig and Ghostty. The abstraction lets you be useful and then deep dive / specialty in what you want to do down the line. This hasn’t changed, this filter / level of abstraction has always existed. It’s how I got started. I didn’t seek to understand anything at the start 20 years ago, I just wanted to bash together code segments (from textbooks). Being a broke college student, I was sitting in book stores trying to figure out how to put Java together. Haha. It’s only after a large exposure and 10 years before I started to care and choose what to love, currently that’s tensors.
I also had some issues with Prettier blocking saving. Not sure why. I disabled prettier.
That red square with the dot in the center means "non-veg" on food products in India ¯\_(ツ)_/¯
In my personal experience, Copilot is a really useful autocomplete tool
I've recently ditched VS Code for Cursor, because Copilot was really annoying and dumb even on the paid tier, so I am definitely not coming back because of this.
This is the best promo video for Cursor! I'm downloading Cursor now.
Having a free tier is great because it's encouraging competition, that said, you've completely put me off trying copilot, my experiences with claude and cursor have been pretty much the same as yours here. It just seems to get context so much better even if it is pretty expensive when doing it.
Also I have rarely actually been interested in a sponsor, have never encountered Sevella before but I think they could solve an issue for me in a way I actually hadn't considered before, so thanks for that!
This video is making the boring parts of using Copilot really come through, I don't think many people on RUclips manage to convey it this well.
I feel a little bit better about my job security after watching this video. Cursor: 10yrs. GHCP: 20yrs.
PhD grade models though
AI is great for writing the kind of boilerplate that I used to have hundreds of snippets for (Bootstrapping components, writing imports, setting up database connections, looping syntaxes, etc) or for replacing intellisense in reminding me of the syntax for certain functions. But I definitely wouldn't trust it for actual logic. It's not nearly reliable enough for that yet.
Me and my bro spent 3 hours for his final school project yesterday. Today i learned this new. It took 18 seconds to finish it
yes
I think it's really useful, that you can choose the models, you want to use, because in my opinion some models are better in some specific stuff, than others. For example if I want to generate a modern frontend I definitely choose Claude, because their model is way better than openais model...
A person who doesn't know or want to change that stuff probably doesn't care, that there is a dropdown for it.
Ahh i paid for it and now its actually free(actually i didnt i got it with the student email lol). But anyways copilot isnt that helpful normally for learning, you will just be spamming tabs. Only time i have used it was in one rust 3rd party component library where documentation was obscure and i couldnt read the whole code as easy, so i just fed the entire module to copilot and had it suggest me what traits and options were available for certain conditions i required.
Anyway i use zed, it already has a free AI sidebar, though i only use it when stuck on how to do something like syntax. Though i always have inline completion turned off cause it feels like its hogging screen and too much happening in screen with the hints wjen writing code like it is showing me my function as i write but incorrect one so its kinda distracting.
Zed is goated :D
Maybe you should go back to basics if you have issues with syntax (assuming that it isn’t a new language you’re using; but even then you aren’t going to learn it properly if it just tells you)
@@aquilafasciata5781 no i mean about some obscure external libraries, like in rust i use different crates for different purpose and they have different custom syntax for colors like for some its just color while others may have viewcolor, some have support for rgb while others only support 8 basic colors and so on.
Finding documentation for them is also hard cause many of these are new crates and their wiki isnt properly updated or is as thorough. So i just scour through their library to see how structs and enums and traits are defined to understand how they work but some times they r quite huge so that where i use the AI mostly to list out all those available functions and syntax
I code with nano as my editor (I haven’t used any other terminal editor), over ssh. Don’t question it, but it means I can’t use Copilot and I probably wont ever, as long as I code like this
If a product is offered for free? You're the product.
Related to his last question. Tested windsurf within the first 5 days after launch, and it was great, unlimited everything. Then did not used it for a month and now it is useless and everyone I've read on blogs says is not worth it and many going back to cursor. Is paying the $20 or $60 worth it or still will reach the limits too quick? I've read plenty bad experiences, has any heavy user had a good experience to tell with any of the paid tiers?
Does Microsoft pay for GPT and Claude on the user's behalf, or does the developer have to pay separately for the AI service? Because if it is the latter, then Copilot is largely the wrapper and is not the expensive part of the stack, so it being free isn't even that much of a dumping move.
Really good video to highlight where AI tools excel and where they fall short.
i have been using codium and copilot, i mostly use it to write small chunks of code and never complete functionality because it almost never gets it write, instead i would myself break thhe problem down into subproblems and then use autocomplete feature so the AI can solve the smallest subproblem and then when all of them are solved my orignal problem gets solved,
its like i am a conductor of an orchestra guiding different models and tools to solve the problem for me instead of letting them solve the problem thmeselves
Watching from a small village of India at 6AM.
Sadly I pay for that pain you're feeling.
If only they've used that "lost" money to make VSCode better.
The could secretly flip the switch and let the AI train on all the free tier users. 😅
I noticed that you mentioned that file hopping is something these tools are not good at, have you tried windsurf yet? The thing I noticed is that Windsurf is really good at that, somehow its more capable of making more sense of the context, While the Cursors DX is better when coding on your own.
Codeium, IMO, did a bit of a dick move a while back. I had just paid my first $10 when they decided to up that to $15, which was borderline ok, but they also hamstrung that tier badly with such a low token count that it's way less than Copilot's free tier. Especially as it looks like Codeiums model eats through tokens really fast.
Aren't boths IDE workin on the same files and changing them Theo?
I tried it yesterday and "/fix" broke my code. It took an undefined let and assigned it as an empty array. It didn't care it was used for an if condition, it decided it should return true always.
Thanks copilot!
Is there a way to see your usage? Like how many completions are remaining.
"I'm an investor in Microsoft", it's a trillion dollar company lmao, what you say wouldn't have an impact on the stock at all.
Besides, 80% of Americans are investors in MSFT since they all have 401ks and it's part of the S&P 500.
It’s like saying, “I fund cancer research”, because you rounded up to the nearest dollar at the cash register to donate to a charity
he says it, because the following takes could then be biased
@@AbstruseJoker No I don’t think so. More like “there’s a small conflict of interest because I have skin in the game. It might be affecting my opinions. And so it’s only fair that I should state that up front. So you can take it into account and slightly discount for example any positive spin or excuses I make, for things that aren’t very good.”
Obviously he has no effect on the stock but he says it because he has said its a requirement to disclose stocks when talking about products
He says it because youre legally required to disclose your investments. Grow a brain
0:19 Yes you used it for advent of code, where you maybe write 100 locs max in probably JS or python which are the 2 most used languages for similar problems. Try it on a *real* problem and see it die. Try writing a Windows kernel driver. You'll bluescreen yourself before you can say "AI can only solve trivial problems in an empty code base".
This is incorrect i used ai to create a driver that bypassed easy anti cheat 😂
HARD disagree on "eating your vegetables for you." If you want to be healthy you need to LOVE eating your vegetables. You need to develop a palate for vegetables.
The same goes here for code. You need to find appreciation in the tedious things, especially for maintability purposes. Making a variable naming convention that you and your team understand at a glance, refactoring your code to make it yours and understand to your core exactly how it works, so that when the time comes to revisit, you have a holistic understanding how why you wrote that code in the first place. I'm not saying ditch the AI at all, hell no. I'm sayinf have the AI help you, but you do the implementation.
I don't think people are ready for the technical debt they'll be hit with when every small thing would have been made by AI.
Hi Theo, great video. But I am more interested in your theme. That is the theme you use ?
They are just buying the market. It's the first play they ever ran from their playbook since the inception of the company. It works.
0:00 Copilot's Evolution and Free Tier
1:10 Savala Automated Deployment Demo
3:10 GitHub Copilot Free in VS Code Overview
4:00 Copilot Free Feature Breakdown
8:00 Copilot vs Other Tools (Practical Comparison)
12:30 Hands-on VS Code Copilot Usage (and Problems)
21:00 Copilot Free vs Cursor Practical Usage
27:00 Summary and Conclusion on Copilot Free Tier
Generated by Snorvia AI chapter generator
Wtf why are those AI generated chapters so accurate
Are you a dude called Snorvia?
I'm wondering if there will be languages that "win" or "lose" as a result of tools like copilot. LLM's are not magic they need to be fed coding examples before they can be useful. Also I wonder if it be harder now to build a new language and get it adopted? Maybe not if you could make it LLM friendly.
I don't use any AI for work, it's nice.
Can you explain why you like Satya? Has he led Microsoft to do anything besides chasing trends (ineptly)?
The Satya era at the very least refocused Microsoft towards Azure as a profit center therefore taking some of the weight off of Windows. This meant Microsoft was no longer at war against Linux and at least up to some point embraced FLOSS. VSCodium (VSCode is only open source as source), .Net, parts of WSL are great and it is great to have them as open source. These should not be taken as granted and Microsoft can at any point return to the Embrace-Extend-Extinguish strategy if they are in a position of strength.
The Satya era also meant they abandoned Windows on phones (even though Win10 Mobile was finally pretty good), Windows suffered from enshitification, more tracking, more subscriptions, a forcefull push for Azure dependence, more abandoned UI frameworks, selling the generally apreciated peripheral division, launching and shutting down WSA and of course all of the AI bullshit.
Does adding copilot into vscode count as extending?
@RemotHuman rather than integrating Copilot I am more worried about the weird licensing of VSCode which bans the open source version from several things including the extension store, the control over .Net including the discontinuing of Mono and the contol they have over GitHub which so far has not had major detrimental effects.
Damn french bots
cursor is so desperate they made a free tier themselves from the beginning...
they have enough money to burn just like microsoft?
I'd rather pay 60$ for unlimited chats and 3000 tool runs in Windsurf. It's been great. The base autocomplete model has unlimited uses and is good enough.
If the product is free, you are the product. This is a clear grab for as much code as possible to improve the models. Not that that’s an inherently bad thing if it’s moves us forward, but let’s be clear about what it is.
He is super excited full of energy? Either this guy is the driest dude on earth or he is not happy to make this announcement
What Copilot is doing with their new free tier may be predatory pricing as described in competition law. If it is, it's actually illegal.
it's not only bash, the whole toolchain. Most people who heavily rely on IDEs (say Visual Studio or Rider) plus all DI and 'magic' and that's all they know, have no idea what's happening in the background. Anyhow, I personally have more fun with an editor like Code even when writing C#.
I would have been fine if it was just an extension. It being built into the editor mixed feelings about it.
The most spicy title I've seen in a while 😆
I myself dont use copilot for actual programing completion or writing but as an internal chat to ask for documentation and or recommendations for something specific i want to do
I have learnt js like this and its great for asking how to syntax what it means and where to use it usually
Its great, now for code completion only if i really need "code that works now" and then to alter it for my specific use case because a lot of times the code just doesnt work
(It solved the what and why problem but not the will you just f*cking work problem)
i like the thumbnail, the co-pilot tapeworm… the ai tapeworm
sums up ai, the largest ‘theft’ if intellectual property in human history.
all if our work gas been used to train these ai tools, enriching their owners… who amongst has been asked to make this donation… and i don’t count permission gained by small print in click over t&c’s
Meanwhile Cursor keeps getting worse. The queue'd slow requests gets longer and slower as it grows. It's code suggestions have been steadily getting less accurate, requiring ever increasingly more requests to get what you want, at your cost no less. And Cursor is showing they don't care, just look at the forum. They never directly address this problem instead either don't answer at all or ask irrelevant questions
Saving the tab bug disappears by switching tabs quickly
Cursor is so much better, the AI is much more integrated into the editor, unless copilot has changed since I last tried it.
Truthfully just using AI for templating and concepts, then I manually go do all the other things. Everything has seemingly gotten worse over time, but that may be just me.
You know it's sponsored when his voice changes lol
@4:35 Theo doesn't know what the term "leaky abstraction" means.
You didn’t apply the changes from Github chat
5:32 - "[The Copilot-authored PR] didn't do what it was asked, and it added a new license to the repo..."
An AI company, _lying_ to exaggerate its usefulness? _Gasp!_ Say it isn't so!
Did you ever try aider? will be interesting to hear you thoughts 🙌❤️
6:14 wait.... 1 billion developers?? This is every 8th human being..
As someone that moved away from VS Code to Cursor, I can definitely can say. I'll be sticking with Cursor lmao.
I wonder why using Copilot you can get high in leaderboard in advent of code. Maybe because Copilot takes data from it?
Maybe MS cooperating with cursor - making the ux in vsc/copilot so bad that you get a glimpse of how cursor might work ;)
Closed the video right away after I heard Laravel
Microsoft can do free stuff because globally they have extra AI capacity they have to have to address paying users peak demand. AWS and Google ditto. All the other players in general have to pay by the token for someone else's capacity, or have premium cost owned instances.
I got like 4-5 emials about my copilot license expiration … after another 2 reminder mails, they offered it for free lol
Never used it even though we had enterprise licenses for it 🗿
50 isn't much especially when AI gives so many errors. I'm in AI hell with Windsurf/Claude right now and I have to wait 2 weeks for my quota to reset. Wouldn't say it's a ton of fun!
Some of you will run situation in your project where the I doesn't do as it should, likely being not as competent or not performing well as it should. Its good to have around for mundane or assistive task (that aren't to overly complex). But do be _too dependent_ on it.
You stumble upon a problem where the ai takes on a project task *without* human input and direction.
"Smart" ai should do the heavy lifting, while human coders/programmers not only oversee and guide/direct their progress, but to ensure that the steps needed to reach the project's goal and objective(s) are met.
As far as task difficulty is connect, the _more complex_ the task are, the higher chance you'll run into caveats and nuances. This would require human reasoning, intuition, and creativity to address.
Nothing wrong with using tools, but don't expect it to do everything for you, lest you stay stagnant in your skills as a coder and programmer.
To be fair, it's up to them if they ever wanted to learn coding or just take advices from coding off from either online or the AI coder themselves.
I miss Theo's old hair style
Do yourself a favour and try use the VS code extension called Cline. It is way better than copilot. You can choose whatever model you want even locally hosted models like deepseek. I use it with Claude Sonnet and it is so good. I prefer it over any of the other AI editors..
you still use arc? :D
i dont. get back to ms edge.