9 Months with GPT, Can I Fire My Devs Now?
ฝัง
- เผยแพร่เมื่อ 17 ธ.ค. 2023
- Recorded live on twitch, GET IN
/ theprimeagen
Help Prime: www.boot.dev/?promo=PRIMEYT
Boot.dev article: blog.boot.dev/computer-scienc...
By: Lane Wagner | / wagslane
Discount Code: PRIMEYT
MY MAIN YT CHANNEL: Has well edited engineering videos
/ theprimeagen
Discord
/ discord
Have something for me to read or react to?: / theprimeagenreact
Kinesis Advantage 360: bit.ly/Prime-Kinesis
Hey I am sponsored by Turso, an edge database. I think they are pretty neet. Give them a try for free and if you want you can get a decent amount off (the free tier is the best (better than planetscale or any other))
turso.tech/deeznuts - วิทยาศาสตร์และเทคโนโลยี
The fact your camera crashed and then you have this animated cutout of yourself as a backup then you take it a step further at add angry # when you scream. Simply iconic .
Yeah that was so cool. I think his voice alone is really entertaining!
That's why the need for developers won't go away. We can't even get software right now because companies are cutting cost, we need more developers. But I think machine-learning will just make the barrier of entry even higher because people will think machines are able to do it and not put up with the high curve of learning.
Google vtuber
@@nailbomb420 Self-proclaimed cartoon character voice.
PrimeGPT
Devs are safe. People can't restart their computer let alone install python at a minimum.
"hi chatgpt, plz install the snake"
Sometimes devs can't reboot their computer if they are pure developers only after education, and not because the computer is their life. There're also a lot of people who takes discipline and dedication wrong. They learn only things THEY think they need, without seeing a whole picture or even taking a wrong direction. Since you experienced - you'll not be trashed.
My favourite meme is an FB chat screenshot "Hey dude your job is doomed! I made this website with GPT!!"
recipient asks for the link and the user sends an address that starts with "C:\"
Laziness and stupidity, always on the left side of the bell curve, many people don’t own computers or if they do it’s a Mac, but they couldn’t pull up a Unix terminal to save their life. Most people use phones. They can’t read or write without autocorrection, and they most certainly cannot think for themselves
Just imagine how confused people who aren't developers get when they encounter an AI hallucinating code solutions that don't have any basis in reality
Wait.. so lawyers are just prompt engineers for judges? 😳
..... omg.....new lawyer definition unlocked
If I ever have to explain why I didn't follow the law, I'll just tell them that their natural-language program is too ambiguous and buggy.
😳
😳
Always have been.
If you're already really good in the language you're writing, copilot is not a great help. If you're diving into new territory though, it can be awesome. Same with GPT-4. If you're learning a whole new thing -- a framework, a big pattern, a new stack -- you'll have all kinds of questions to ask it, and if you know how to formulate prompts well, you'll get awesome, to-the-point answers that are vastly superior to what you get from hours of scouring docs and the rest of the internet. Not that you should't get a handle on those docs, mind you.
i was learning express js, golang, and rust from scratch from chatgpt 😅. now i understand express js and golang, and able to create backend using axum, actix, pg pool, still in proccess learning rust
Exactly my experience. I am a data scientist and used it to learn frontend/react for a project and also some c and now kubernetes. For beginner topics hallucinations are not a real problem
Completely agree. If you know what to ask and how to ask it, it saves hours upon hours of time that would have been spent googling.
Yep, I asked how to convert DVD to mp4 using CLI and it told me to use ffmpeg. I then specified what exactly I want (original quality, multiple DVD files to 1 mp4 file) and got the exact command to use. Finally, I asked if the DVD file names for are ordered sequentially and it explained to me that they are.
There’s no way I could have overcome all those uncertainties without AI. I had no general idea of how to approach the problem, no idea of how use the tool and no idea of details and managed to get the thing done in like half an hour. Without ever touching the ffmpeg docs or DVD/mp4 specs
Yep what I did when learning in my bootcamp. But I started to notice it's limits.
You write more code, and your product becomes more complex - compounded non-linearly with the added amount. You'll need more attention to your product not less.
Yeah but sometimes that complexity is good because of the functionality it affords.
Example: Graph API call to webhook triggers flask listener on prod server to start Selenium webdriver to fetch a web application to generate a .csv to be inserted into html template to send via email to some other business that still runs on humans.
Tom could never be made obsolete by an AI because he's too much of a genius.
Tom will write the first AGI in JDSL.
JDSLLM is going to crush GPT4.5
an AI that could generate hot takes on other TH-camr hottakes?
Yeah, that could be done - probably is already someone's business plan for starting a new TH-cam channel
there are rumors he already made it, it just takes years to startup@@rj7250a
J-DIESEL 🔥🔥
Why aren't there more job openings for junior devs then? Oh, right, it's because companies would rather force senior devs to do more work to save money, rather than hire more devs...
I think that it might be that a lot of work that used to be handed off to jr devs can now be automated. I use AI to generate the more tedious stuff that I've done a thousand times, and it's sped up my workflow a TON. This is stuff I used to give to someone lower on the ladder than me. It's given me a lot more time to focus on more important parts of my job.
It’s not that guys. We’re literally in a recession . There isn’t a lot of investment right now. We have two major wars going on, uncertainty over energy future , current housing and car situation. Once stuff gets easier and interest rates get cut and money starts being printed again we’ll be fine .
Right now companies are forced to only pursue projects that are sure bets .
It would be NOOO different if you were a roofer or plumber . When the economy goes down ppl will wait before fixing that thing
Just look at any other industry. It's basically the same for the majority of the non-essentials. I've noticed even pharmaceuticals are down.
high interest rates basically
@@ci6516I wish people online knew what a recession is.(easy to look up) The Internet was supposed to make us smarter. 5 percent gdp growth and 3.7 percent unemployment is not a recession. Just because some people have it tough does not equal a recession most industries just go through natural business cycles and tech us no different
We should appreciate that ThePrimeTime is not screaming hard at us anymore in recent videos 😅
I love your videos!
Anybody who tried to use Cat-GPT for some time stops being worried about AI taking his job in the nearest future.
Cat-GPT - the AI for cats. ;)
The issue right now in short term isn't really being worried about ChatGPT doing worse but about non technical people axing devs *thinking* ChatGPT will replace everyone. Because of that I've invested in getting some free ai/ml certs and it's helped the perception issues with gatekeepers but essentially was 80% a waste of time other than to satisfy my curiosity.
If things keep improving at the rate they have been in 5 years, not so sure about that.
Well, as a code monkey, I'm sad to hear this prediction. Apparently, I don't see any improvement in my productivity from chatGPT, because I am just a meatGPT. Rewriting same questionable site ideas into newer frameworks is what I do. It's always just some sort of CRUDe wrapper over a database with some fancy predefined ifs and filters to lessen the cognitive load of meatbag end users. My value is only in allowing bureaucrats of different sorts to pretend that they are doing something when they spend money.
If only the true engineers who design really new things are going to survive, only dark days ahead await me. Oh no, I shouldn't write 'await me' because 'me' is not promising at all.
That last sentence gave me asyncing feeling :(
Look a bank tellers. There are ATM everywhere and banks are still hiring loads of staff, plus there are lots of jobs in maintenance
I'm in the process of learning python and I needed this. Thank you for motivation!
The vanishing gradient problem is the result of trying to tell how much each node of the previous layer influenced the error amount of a given node. It's obviously a fraction of the total error amount. With very deep neural networks, the multiplication of these fractions leads to the error of the first layer to be effectively zero. This is a problem because the error is what determines how much each node changes during training, so zero error means zero learning.
Gpt says it’s been solved this way:
The vanishing gradient problem in neural networks, particularly in deep learning, was a significant challenge that affected the training process. It was primarily addressed through several strategies:
Activation Functions: Switching from sigmoid and tanh to ReLU (Rectified Linear Unit) and its variants (like Leaky ReLU, ELU, etc.) helped alleviate the problem. ReLU and its variants do not saturate in the positive domain, which helps in preventing gradients from vanishing during backpropagation.
Weight Initialization Techniques: Improved methods for initializing the weights of neural networks, such as Xavier/Glorot and He initialization, were developed. These techniques help in keeping the variance of outputs across layers consistent, which mitigates the vanishing gradient issue.
Batch Normalization: This technique normalizes the inputs of each layer in such a way that they have a mean of zero and a variance of one. This stabilization of the distribution of the inputs to different layers helps in combating the vanishing gradient problem.
Skip Connections / Residual Networks: In architectures like ResNets, skip connections allow gradients to flow through a network without passing through deep layers of transformations. This approach effectively allows for training much deeper networks by providing a direct path for the gradients.
Use of LSTM/GRU in RNNs: For recurrent neural networks (RNNs), Long Short-Term Memory (LSTM) units and Gated Recurrent Units (GRU) were designed to have mechanisms (like gates) to control the flow of information. These structures help in mitigating the vanishing gradient problem in sequences.
Regularization Techniques: Techniques like dropout also indirectly help in reducing the vanishing gradient problem by preventing overfitting and ensuring that the network does not rely too heavily on a particular path.
Each of these solutions addresses specific aspects of the vanishing gradient problem, and often, a combination of these strategies is employed to effectively train deep neural networks.
@@qwzr7083 if this was pre 2017, this would be the most correct answer
Because every technique mentioned was used in every model training and they still had problems with vanishing gradients
@@qwzr7083 You're going to super hell
Pretty hard to avoid the sigmod activation function eventually becoming flat.
I feel exactly the same!
GPT makes me like 3-5% more productive, it definitely helps, but most of the times it’s just a waste of time - so much easier and faster to do the research and coding myself than try to separate bullshit from legit code.
It helps a lot with the stuff I don’t know / don’t wanna know! It shows people that make these confessions aren’t very good at their craft
I use it more to brainstorm or implement repetitive tests. I give one test I made and then the class, and ask it to implement the rest based on the test I gave, it works great for that. But really do not see how people can say it replaces even a Junior dev, you can do a lot of things, but the more complex the problem is, the more you have to take time to write a pretty specific prompt, which most of the times seems slower than just solving the "old" ways. Most times I brainstorm with it and then fill in the gaps with traditional search or previous knowledge.
Usually, I end up using it as sounding board for the ideas, not the code. Frankly, it's responses are pretty much expected, with some exceptions, which are either very useful or very useless. As such, I talk to the AI when I get bored talking to my rubber ducky. Outside of coding, it's useful for looking up the terms I've forgotten by explaining the concept, which happens with I'm on a person project involving some math or something. Once I know the term, pop over to search engine and lookup the meat and potatoes. I actually haven't tried co-pilot yet. When messing with ChatGPT4, I've found that the limited token span makes it difficult to build up methods by describing functionality (I did a few times for my Pico Pi for the kicks and giggles). As far as the code generation it concerned, it does okay with stubbing out things, which is fine when I'm too feeling too lazy to make boilerplate, but once any real complication gets introduced, things falls apart quickly. Might be solvable with an increased token storage, but that would drastically increase response times, in theory.
Exactly, I almost never have any use of it in my actual work, and we are not allowed to give it that information anway.
The only times I use GPT at work is for like errors and troubleshooting or very generic solutions.
I use it much more in private when I work in other areas of coding that I am not as knowledgeable in, so it's great for my hobby projects where I mostly do app dev which isn't my work.
@@Lucas-mk1gi Pretty much yeah, as experiences devs we basically already know how we want to implement something anyway, since we know the context of the projects we are working in. Most of that is basically solved beforehand at refinements anyway. Having to re-tell all that to gpt and then trying to fight with it to get a correct implementation with the rest of the code just ends up taking more time.
There is one guy at my work that I know use GPT much more, and he's the slowest one and always the one asking very basic questions even after years.
lol imagine taking a bootcamp to try and improve yourself. Two years later sitting here in the dark, and the jobs go out of reach like a fucking genie extending it's platform.
Pretty sad.
That's literally how every other field of employment has gone, and how everything you might want to buy has gone.
Yeah imagine haha🥲
The longer we walk, the more we have to walk yet
Not really that funny. Imagine no new devs coming through… what a woeful future.
A fantastic article. I think I've said some of the same things before but with a modicum of the eloquently, I blame the natural language xD
We'd need literal sentient AI to translate product, design, user opinions into code and it'd probably be harder to do it genuinely accurately to or from natural language.
I really like that AI generated prime lipsyncer. Almost like real Prime.
I’ve used GPT-4 when it first came out when I was starting to work on my own startup. It came in as a great help. Now the issue I encountered was that after a certain point it became more of a drawback than a help. There were lots of times where I just didn’t feel like dealing with chatGPT and instead, just write it myself. It almost always turned out to be more performant, less lines, and more readable if I just did it myself.
I work in a very datacentric environment (think elastic data go in, picture or table come out). I think the use of LLMs for that problem set crystallizes the question of "Why can't I just ask GPT to do it?" - Its not just the ability to state the end product clearly, but also the domain knowledge to ask relevant questions. The humans that do this have both the knowledge of the data and the languages to manipulate them because they have to. Not just because that's how the systems work, but also because we can't be effective with only the "what" or the "how", it requires both.
So, your claim is basically that your knowledge will keep you safe from an LLM? You do realize that the goal is to create an LLM that considers this (futher) knowledge, too, right?
As a society, we should be very careful about the developments, but your comment reads like an old man desperately trying to hold onto straws.
The thing that learns will learn what you know sooner or later, don't worry. You won't stay relevant for long with this mindset.
My point is that LLMs won't be a "Do what I mean" button, probably ever. For that you really need a general AI. For problems that are more like "Find the significant data in this set of logs, and determine how to parse and normalize them" or "How can I model abnormal activity in this data set?".
I went to a talk by Dr. Craig Martell where he made the point that language fluency messes with our ability to judge the actual intelligence of the responses. I don't think
I was thinking about it academically, I think I'll be ok with a combination of that fact that I'm getting old and the fact that I'm in security. Those who can't do secure, and those who can't secure audit. I'm in the middle, but I think I can run out the clock just filing paperwork before the AI boogeyman comes to get me. And if I'm wrong, well I guess you get an I told you so.@@lukasz96
@@petespyder6576 You sure as hell will be okay, it wasn't meant as an attack, pardon. I just think that people are neglecting many facts about their own design. A human is nothing more than a "(not A)I-driven" processor. We *will* get to the point where a Boston Dynamics robot powered by a general AI will replace a human in every given, work-related, context.
I was born in Poland, and the Poles have a saying that proves to be true pretty often: Nadzieja matka glupich - "Hope is the mother of the dumb"
I don't think that counting on hope, in regard of one, better CPU not replacing another, will be enough. It never is in this world.
I think that's a good point about hope, I just see there being a difference between where our hopes and worries lie. Your hope is that AI will more or less live up to the hype, and your worry is that it will effect the employment of the legions of technical people that rely on the good living that tech currently affords us. My worry is that not only are we atop the Dunning Kruger mount stupid with AI, but it's going to introduce even more complexity to the security and operations of the large organizations I tend to work for, and my hope is that it doesn't negatively effect my lifestyle. I have my reasons for feeling this way, from my experiences with new technologies in the past, to my research in the security issues related to LLMs that exist now.@@lukasz96
Yeah, based on GPT and other LLMs, I'd say we are still a far way out from general AI. And you're precisely right on needing to give the LLM precise direction. It doesn't have an "understanding" of what you said. It's a sophisticated word calculator. You can't ask "how well it knows x" or "do you understand?" and expect it to interpret meaning.
My company fired 99% of its engineers for AI (only keeping 2 SREs), along with its entire QA team, the entire UI team, all the managers, and then most of the remaining employees jumped ship - including the CTO.
wtf ? xd didnt see that happen ANYWHERE
utterly insane
Putting your company in the hands of a glorified auto complete what could go wrong?
Did they explicitly give that reason or is it just your assumption
"Ideas guy" is only cool if he is willing to make the ideas practical, research the possibilities and do the ground work.
And talk to users 😁
I've been trying this AI programming for while now but I'm not that convinced about it. I spend more time writing prompts and fixing and correcting the results than it would take me time to write it myself. It's very rare that the model just "gets it" if the problem is at all more complicated than some hello world. And it's really hard to explain your problem than just think what you want. On top of that the models seem to just invent libraries and own functions on random intervals. I'm often like amazed how novel solution the AI is giving me, like "why haven't I done that before", just to learn that yeah, that's not possible. I don't know but I feel like GPT-4 isn't quite there yet. It's neat for more generalized things but the moment you go into more detail it chokes. I like it, but I'm not super impressed. Sometimes I even find myself arguing with the AI because it's giving me wrong information many times in row.
concatenated and abbreviated duality is often an optimal function of how humans choose to delineate and evolve language and this follows for the most popular of computer languages today too. Given the template is as unreasonably effective as it is obvious the overwhelming possibilities it proposes inherently precludes other possibilities, the conclusions which are inevitably consuming almost everyone even peripherally interested in the future in the sphere are almost frozen by what is almost completely unbelievable.
But I almost bought it.
25 yr long senior software engineering dev here. I don't write self indulgent articles so I'll spew here. There will be devs, but only at FAANG (aka the CIA). For the rest of us who don't work at fantasy land companies deal with the "we're not a tech company" sentiment backed by with razor thing margins will see a reduction in force across all industries simultaneously. We're about to displace 80% of everybody who works at a desk. Such a disruption won't end well for anybody, even FAANG since the average person won't be able to afford even a Netflix subscription.
One more reason I think there's going to be even more jobs is if your ever imagined doing a startup you might've realised that there's so much to do to actually to ship at least MVP to a market that a lot of ideas are just had been thrown away. Like if you're a developer you probably don't know how to design or write copy for your ads, or don't know where get visuals for your landing. And now with things like ChatGPT you don't need that big team to finally try your idea (and might even don't need venture money). I personally expect even more start ups than before.
Great article. Clear, well written, engaging. On the topic itself, there is no way AI will replace developers. People will always be needed to check whatever code any model can spit out, especially when it comes to security.
it is not over yet 😢
Depends on what type of "developer", traditional mobile developers have a short time left on this world
@@Supersonicboom7 why mobile, mate? who's going to develop for los celulares?
@@Supersonicboom7 why mobile, mate? who's going to develop for los celulares?
As I was telling my fella this morning, I don’t see AI writing for Boeing or NASA any time soon. And if they do, I’ll just lean into agoraphobia more than I already do
The point about the vanishing gradient problem isn't quite right. It's true that the transformer architecture is nicely scalable, but we were training very deep large networks in the image space before transformers using residual connections, batch normalization and activation functions that were more resistant to this problem. The transformer did a lot of things but this isn't the thing that it "solved".
Yes. I would add that the transformer was the first architecture that does not lose information from beginning when the length of input sequence grows. Also, it processes its input at once, in parallel. These features allowed to scale the training datasets (and models), and it's a law in machine learning that the more data - the more capabilities.
However, these nice features come at a quadratic cost (with respect to sequence length) of memory and computation, so now everyone (and their mom... :) ) is trying to remove this constraint. See, eg., Mamba architecture.
When Visicalc came out apparently no one would need Accountants anymore - there's still Accountants, they're just expected to do that much more work.
18:36 you've heard of JDSL. Now get ready for LawDSL
tree rotation for insertion is such an important concept to understand..
This was a very interesting article to listen to. I'm very glad I discovered this TH-cam channel.
Pro tip. Don't use copilot to write dockerfiles. It grabs versions out its ass that aren't compatible with anything in the file xD
🎯 Key Takeaways for quick navigation:
00:00 🚀 *The video discusses the impact of GPT-4 on AI-assisted programming and the future of knowledge work.*
01:36 🧠 *GPT-4 has about 10x more parameters than GPT-3.5, leading to a more sophisticated understanding of the world.*
04:20 💼 *The future of AI-assisted programming suggests that while AI may eventually handle much of the work, learning fundamentals remains crucial for the foreseeable future.*
07:35 🤖 *The speaker shares experiences using GPT-4 in his business, emphasizing the potential increase in efficiency in writing educational content and software.*
09:57 🔄 *Increasing efficiency through AI doesn't necessarily lead to job reduction; instead, it can result in more projects and increased demand for developers.*
13:23 📈 *Despite automation in software development, the demand for developers has consistently risen due to companies seeking a competitive edge.*
16:11 🗣️ *Communicating with humans is challenging; writing code to describe programs is more effective than relying on natural language.*
18:57 🤖 *While AI may generate code, developers with programming knowledge will likely continue to play a crucial role in guiding and refining the AI's output.*
Made with HARPA AI
I've tried to use co-pilot chat in visual studio to optimize some large LINQ queries in our older C# app. None of the suggestions it came up with worked.
Jeez, I seriously thought that my yt was on 1.5 speed. You're on fire dude.
if the time to explain what you want (tProompt) plus the time to check the results is greater than the time the task takes it's not worth it to have the llm do it.
ChatGPT can replace chatty Kathys loooong before they replace high skill (doesn't mean high experience) engineers.
Just give it a random topic and it'll BS you for as long as you want it to 😅
First generation AI really about cross discipline. Jumping and connecting the knowledge silos between tech, accounting, marketing, legal, supply, manufacture etc. Finally, real world user acceptance testing throughout an organisation will be possible.
our production and demand will go way up and our wages won’t move let’s go!! go america
prime finally living up to that vtuber tag
I may be wrong... But I have been listening since the video started and now I looked at the screen for the 1st time at the 5 minute mark... I think Prime has been body snatched.
Prime... Use a strange unnatural expression if you're under duress. *Knew it*
The problem is automation paradox. In a line in a factory 0% automated until the line is ready. Then, when ready, suddenly 100% of automation is done. LLMs is the same thing.
While i agree that it's not doom and gloom, I'd hazard that the "job availability" (for code monkeys) will go down (quite a lot). Prime missed something because he's looking at it from his viewpoint, ie, a real developer. The skill needed to use a GPT-like thingy in the future won't be writing code, it will be "algorithmic thinking". And anyone thinking writing code implies you think in an algorithmic way (or close to), boy... do i have a bridge to sell you.
But isn't gpt the best at writing algorithms? Just give it any leet interview question and it will crush it.
@@HyuLilium You just proved the point. If you can't even read properly how good, clear and concise do you expect your questioning to be?
It's still the old 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' joke because humans didn't change.
If i had $1 for every client and "developer" that can't properly express what they REALLY want (to do) I'd be retired by now. In my own island...
@@ErazerPT so basically you're implying soft skills are what matters the most now, got it.
@@HyuLilium While interpersonal communication, which is the soft skill you're thinking about, is paramount, i was actually talking more about reading comprehension. And that comes from reading A LOT. On paper. Whether it's Neuromancer or The art of war or SWEBOK or C Programming Language, 2nd Edition, read. It's hard to convey how deficient current reading skills are, but once you've read enough you'll be floored by how bad you were before. And reading A LOT wires your brain for "proper structure". Then you'll start "mentally restructuring" verbal communication, both yours and others. And things start being a lot less "muddy".
@@ErazerPT good point, I definitely put off reading more than I should, and I think that’s becoming very common. I just have a difficult time finding things I actually want to read.
Love your content. It’s the favorite part of my day sometimes.
Flip putting in the hours more than usual
Whenever I try to write informational paragraphs I can't do it. It always naturally turns into some more structured form like bullet points
tbh the talking mouth replacing your camera is hilarious and I don't mind it lmaoo
I will take this take and try to stop with all the AI doom wallow but given how openai revealed much about itself during that power struggle I still believe they are more likely to become public enemy than the new savior.
When it comes to using GPT-4 I mainly just use it for editing styles cause 1. I am not super artistic and 2. I find it extremely annoying and a waste of time. But with GPT-4 it usually has decent designs and I can tell it to change certain styles if I want to and it works great for simple designs that don't require a lot of advanced animations!
Its usefull for formatting and style until it forgets to set some bracket or changes ++a to a++ and hours later you wonder wtf is happening
@@dawre3124 trueeeee, I've also ran into that a few times, its a struggle but now I just look for those changes to double check if they were changed or not. If not then it saves me a decent amount of time. The worst one I ran into was it forgot a } at the end of one of my functions and it took me 2 days to find it.
Man what a great article. Towards the end , I found the discussion about English language fascinating. I'm thinking, the next huge Ai breakthrough might require changes not in the Ai itself, but in human language. As a native English speaker we tend to use English with GPT (obviously). But what about the 100+ other world languages? How about Esperanto, the modern invented language which was designed with all the knowledge we currently have. I'm imagining a future where we start interfacing with Ai using Esperanto or some other well formulated language.
It's a good point, but ultimately the reason that we use English to interface with LLMs is that they've broadly been trained on English training data. You can talk to GPT in other languages but the responses are nowhere near as good (and this is similar for code generation, languages with lots of usage and therefore examples are easier to use GPT for)
@@jacobbeaumont8716 true, it's all about the training data.
i'm brazillian and usually i use portuguese in chatgpt - and the results are ok
to be fair, i dont remember the last time that i was obligated to use english in chatgpt to reach some result that in portuguese i couldn't
but bare in mind that i don't consider myself even a jr - since i'm unemployed... my use of chatgpt is, usually, when i got stuck in something i paste a question like "what's the correct way of doing X" (in portuguese) and then i paste my code - and i go on from there... works quite well!!
It's available in almost all languages. Been using it in Romanian as well a lot. With some rare grammar mistakes, it is awesomely good
This video really helped my fear for not having a job in the future, thank you!
16:40 What the customer really wanted: a tire swing.
"Wilson" would make for a better sidekick.
"Wilsoooon!"
Lane and I are in sync on these ideas. Even the what is the point of english as a declaration of law I've long argued lawyer -> coder.
Wouldn't the next logical step be the following: How low can developers go? How low of a payment will they accept when the ChatGPT variants used on large scale will cost a company lots of money to rent and invest in hardware that can run the LLMs?
GPT can tell you how to construct an if statement or a while loop in a language you are not familiar with. That is useful.
Actually constructing the steps you need to take to get from input to output, you need a programmer for that.
Winston is willing to learn. Many non-technical new hires are unwilling to learn new skills.
Thanks for giving us some hope
Chat GPT: We give you a legacy code right at the beginning of your project to start with! 🎉
The anthropomorphisation of models will forever make me angry. It doesn’t think.. It doesn’t reason.. it doesn’t understand 😂.
On a side note, gpt-copilot is a nice autocompleter; I dare you to look at the imports from the code completion LLMs generate. It’s good for small pieces that you can easily assess, and is generally terrible otherwise. It’s far off from generating production-grade code.
I agree.. but how do you know when it does?
There are times when emergent 'skills combination' emulate reasoning, though I would say it is probably not thinking, but I don't have knowledge to even say how one would be able to tell the difference. Maybe we should just get used to it now, I doubt it will go away.
Sanjev Arora gave a talk about LLMs developing new complex skills due to combination of simpler skills.
This has always been my take. Domt talk about it as if it is anything more than a sophisticated statistical model, and you will use it, rather than it use you
Von Neumann wrote about how the parts of a computer are similar to the human body. The entire CS community is based largely on anthropomorphism of machines.
Would it make you feel better if we said it "calculates", "computes" & "can logically apply it's dataset to" instead? Will that semantic shift calm you down while the models continue to accurately replicate our cognitive functions?
Why do you assume reasoning, thinking and understanding are limited to humans?
6:24 The issue is, if you put “generative” first then the acronym goes from AGI to GAI
10:37 the problem with this pov is that it assumes that demand for new features increases in the same or higher rate as productivity increases. It's not a linear relationship. You hire more developers if you have more demand for new features.
My opinion is that as AI will make us more productive we will be able to tackle more complex and exciting/cutting edge projects, but the market will become saturated and people won't value the product of our work as much. Hence, you won't be able to price the new feature at 100k and will have to drop to 50k.
At the end of the day it all comes down to money, not tech.
04:00 The juice ! The rest just normal Prime.
I have used GPT-4 it work but i am to lazy to login to it and ask five times to give me the right answer and wait for it. Turns out that is faster for me to search for answer or read documentation. When i read IT people saying cant love without GPT-4 it mean you are not good enough i am not a good programmer i want to get better so practice makes perfect.
Lol the Robin Hood fox got me. That's hilarious.
The DSL for law tends to be Latin since Latin still retains much greater grammatical specificity than English.
At University the picture with the tree was part of the curriculum
Each Fibonacci number is the sum of the two preceding ones, starting from zero and one.
You can write contracts in code with Solidity
We all need to start speaking in js to each other and answer in json when someone ask us about our day
Great article!
Why are they (prime and the writer) assuming AI engineers that can completely replace at least web devs are so far in the future? That's a fundamental mistake and just reflects our above-average driver bias. We are not really anything special for being able to take input, process it in our feeble brains and produce output, which include code or deployment commands on a Linux terminal.
Ilove the image from the Robin Hood cartoons
GPT-4 rarely gives the answer I’m looking for. I mean don’t get me wrong it’s helpful as a “guide” but it can’t think. It relies on your inputs.
I suspect we’ll be fixing a lot of GPT written code that people assume is going to be sustainable long term.
Yes that is a scary thought. Then again I fix a lot of non gpt code too
@crypticsailor on the upside it’ll keep you in employment for a lot while longer yet 😂
Idea I've thought about while watching this vid because why not (it's probably full of mistakes regardless I'll save it : D):
we need a way to generate responses slowly. and fast at the same time. then make a script. we ask it at step 1 "take your time and generate x" . step 1 of the script done. step 2 "fix a typo in the script (just a side note you don't need a model that does everything)" so the idea is to have a problem.. approach it in steps "slicing and dicing".. and let the appropriate tool fix it for you; If you don't like how a step is doing it's job well it's simple you make more steps within let's say step 1.... and now you've designed a sheet, script, a guide. to how problem X should Idealy be fixed. we will use a mixure of languge and script to approach the problem.
I love this, don’t be afraid of the future just get smarter and make more money
The DSL for legal is legaleze, in case you guys missed the memo...
I don't know what's a bigger W. The article or the camera
It seems the gap between knowledgeable expert and novices is only going to widen with LLMs. 🙃 Hopefully LLM can supply more and more context to help novices get to expert level of knowledge.
As if the jun devs are banned from using LLM to their advantage. Hell, the educational system is going to change massively in the next decade because right at this moment almost every university or high student is using LLM to help with the homework
100% agree...it's easier to write code or even pseudo-code than it is to write what you need a dev to do.
Being a software engineer for almost 20 years now, I absolutely agree.
And I share your opinion that it should be called GAI instead of AGI.
Artificial General Intelligence -> AGI
General Artificial Intelligence -> GAI
I want to be GAI for pay. Yep I think we should change the acronym now! Being GAI just makes more sense.
Let's just hope that compensation also increases by the same proportions. Invest in your workers.
I'm a mis-educated power user who never learned how to code until later in life. GPT-4 to me is like strapping on a bionic mech power suit. It gets me over the threshold and teaches me in the process.
I'm an operations and logistics person. With GPT-4 I have set up VMs, automated my workflows, set up web APIs for data pipelines, and generally gotten tons of productivity increases.
For people like me GPT-4 is like pure magic because it lets me LEARN faster...
I feel similar. I‘m currently learning C++ via an online course. Whenever there is a topic which is not covered in sufficient depth for me I will go and have a chat with ChatGPT about it and usually helps to deepen my understanding of concepts.
However, when it comes to actually writing code (be it C++ or python which I am pretty familiar with) for a project I feel like debugging the code as well as writing it is more efficiently done by myself. Iterating with ChatGPT, giving it all the context, correcting it and so on often is a frustrating experience and takes more time than writing stuff myself.
For rather isolated problems which do not require much context (e.g. computing all k-th nearest neighbors on a n-dimensional hypercube) it can, however, be a huge time saver.
But also this has its caveats. After ChatGPT wrote that code for me I told it that the number of neighbors its code outputs is incorrect and it responded with something like „You are correct, I‘m sorry, Let‘s fix this […]“. After about an hour of debugging with ChatGPT and arguing it turned out that its initial code was actually correct and that I made a mistake in my test of the code. So, if ChatGPT actually stood its ground and said I was wrong in saying its code is wrong, that would have saved some time.
The calculation of income on feature is to simple. Yes if devs, become more effective they can producer more features in shorter amount of time, but this does not necessary translate into more customers. Also, all other companies devs are more effective, so it might end up just increase the quality and features of software in general across software, but don't necesssary lead to more customers income, since the competition will still be the same. Of course others perspective can be provided on this.
Love the pronunciation. Gipittyfour. Love it.
GPT4 and copilot have only made me a better developer. I'm no longer wasting time figuring out how to do some minor function or use some API. I can focus more on system level and architectural issues. Developing feels way more enjoyable.
i keep seeing this messages, people are you crazy? It's on the same level like revealing use of drugs. You're destroyed your reputation in industry with this. Do you also removed copyright from your code? You must, everything made with machines is public domain by worldwide courts position. That cat and mouse game is just waste of time, revealing in court the records use of ChatGPT immedialy destroying your defense and leaving fully to judge personal beliefs a decision.
Everyone must be fired in company by even mentioning Ai, it's the destruction of company intellectual property value.
17:34 #7 = thorough (uh). #8 Hiccough (up).
i understand what cost center means now in my company lmao
There's a simple reason for using "artificial general intelligence" not "general artificial intelligence": If you abbreviate it, you get AGI on the one hand and GAI on the other. So AI devs are obviously homophobic, spread the word!
6:20
Artificial General Intelligence is a general intelligence which is artificial. A general intelligence has the ability to learn any information and do any action instead of being specialized on specific tasks. So AGI is just an artificial version of a general intelligence.
General Artificial Intelligence would just be the general domain of artificial intelligence, which includes simple AI like videogame enemies
If I close my eyes, it sounds like I'm listening to Michael Scott from the office.
To 6:30
In German the term for an AGI is"Allgemeine Künstliche Intelligenz" which translates to General Artificial Intelligence, W German
but more often it's "Künstliche allgemeine Intelligenz" though
6:40 What is "Altron"?
Each time you add a dev you add coordination cost. That does not scale. You need to coordinate developers with users,, designers, product managers… it’s better to have 3 dudes super focused than 50 dudes having hobbies and side projects. If AI increases by 400% productivity, I would try to hire less devs but they have to be amazingly good and super highly paid.
there is so much demand, ai will help supply that demand
I think we overlooked the fact, we might not need to build CMS systems anymore? o.O
i think it’s kind of silly for people to assume AI is coming for programmers primarily. it’s definitely going to go for low-reliability type work first, work that requires testing is way out right now. i would say jobs like customer service, recruiters, jobs that basically have one off interactions with strangers and stuff like that is probably most under threat right now.
Um, massive increase in suppy always brings price down, unless you assume infinite demand. I don't think there is infinite demand for engineers.
Can I fire everyone at my work and use instead an AI that cannot tell which of two numbers is larger?
Yeah by the time software engineers are replaced- every other profession is even more screwed. Except for trade jobs of course