Nvidia Promotes Its OWN LLM After Telling Kids Not To Code
ฝัง
- เผยแพร่เมื่อ 5 มี.ค. 2024
- Recorded live on twitch, GET IN
/ theprimeagen
Become a backend engineer. Its my favorite site
boot.dev/?promo=PRIMEYT
This is also the best way to support me is to support yourself becoming a better backend engineer.
Article link: developer.nvidia.com/blog/unl...
By: Chia-Chih Chen, Ao Tang, Bobby Chen, Chintan Patel and Shashank Verma
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
TLDR, It's just an attempt at creating a monopoly by nvidia. They make the gold and sell the pickaxe.
Just like Cisco before them
as nvidia investor i see that as an absolute win.
Yupp. listen if you want to make an ad, make an ad.
But saying that kids shouldn't learn to code "just use our product instead" is just wrong
Make consumers too stupid to do anything, make them pay to do anything.
You will own nothing and be happy.
@@pandoorapirat8644 take some earnings and buy a map and a translator: 國立故宮博物院
"No, don't learn math! Just buy my calculator!"
yeah ... try our calculator which will cost you only 10000$ for one month subscription .... just trust us ...
exactly
262,682 tech workers across 1,186 companies were laid off in 2023 and 41,000 as of March 1, 2024 only a fool would get into software engineering unless they’re a genius or plan to be elite level
@NPCSpotter yes, when the market suddenly shifts and less workers are required, they're usually laid off. What's your point? You think it's because of AI? No business is laying people off because of AI right now. Will that happen in the future? Probably. But it 100% isn't happening right now. This is a correction from the massive leap in infrastructure usage during COVID, among a million other things.
lmaoo fr, speaking of i gotta do some math 🤓
Nvidia is the most investor friendly and anti-consumer company in the world and I hate it.
For now. They will run out of people to screw in a few years.
Ever heard of CUDA?
You are saying Apple wrong
@@alondjeckto trash vs garbage
How they keep pulling their shit and being successful is beyond me.
The thing is the open source licenses were not written with LLMs in mind. No one expected their code to be hovered and reproduced without attribution when they licensed their code. Most Opensource licenses explicitly require attribution so all of these LLMs that use the code but do not provide attribute are breaking the license agreements.
I think all model makers should be liable for the end use. Such a load of crap what they get away with.
It's a good time to be a lawyer in copyright law is all I'm going to say. These are all just plagiarism machines who only get the right answer by accident. (so useful /s)
@@chinogambino9375 That'll never happen for the same reason gun makers are generally not held liable for their guns being used to do the single job they're designed to do - there's too much money in the industry and too many politicians willing to look the other way as long as a little bit of it "accidentally" falls into their pockets.
Maybe the EU will do something sooner or later, though how effective it will be is anyone's guess. I wouldn't be holding my breath for an AI equivalent of what GDPR did for global security and privacy. Not in the near future, anyway.
Perhaps if one or more of the constant fearmongering ideas starts looking like it might actually come true rather than just short-sighted people panicking (or intentionally spreading doubt for their own gain), we'll start seeing more explicit and strengthened legislation coming to fruition. As long as the "harmed" parties are just randos on the internet though you'll see at best a token response. (Especially when a good portion - or at least a loud portion - of those randos have spent the past few decades ranting about copyright and trademark law stifling innovation.)
You're making the usual fundamental mistake. Your code is NOT being reproduced. The pattern your code fitted in is. It's the reason LLM's are SO GOOD at making template/scaffolding code. There's S**TLOADS of it and it all looks the same... Very specific code/patterns that it only saw once in a single place? If you're not over fitting, you'll be lucky to even see remnants of it in the latent representation.
Funny enough, this "weakness" is the reason these LLM's won't be a "terminal threat" in their current state. If they need Tom, Dick and Harry's "same'ish code" to learn from, John, who writes very very specific code will be "secure". It's everyone that's doing the same as Tom, Dick and Harry that isn't. Webdevs are ten a penny, embed systems devs aren't. Just an example.
Maybe I'm in a bubble as this is completely anecdotal, but I feel like most OSS doesn't require attribution.
What I worry about is not AI replacing junior programmers and creating "forever juniors" it's actually far worse. My fear is that companies will recklessly replace their juniors with AI, putting more work onto the plates of seniors who will then leave for obvious reasons of not wanting to do the work of 5 people. Even if AI was as good as they say it is, doing the work of a whole team by yourself sucks and I don't see any senior devs putting up with it for very long.
They already don't hire them cause they don't want to train them (and because good mentors are a rarity in general). LLM or not, junior are currently getting fucked.
@@nekogami87 Yes fucked BIG time!
What's obvious here? The AI will do the work of five people. You just have to manage 5 more projects. That is all. You have to work for 8 hours. And you will, senior or no senior. And okay, maybe it will be only 3 or 4 projects rather than 5, but you WILL DO more projects thanks to AI. Your power as a senior isn't as great as you think! You didn't want to put time in a junior, well now you don't have to anymore hahahaha.
@@paulholsters7932 I am a junior dev you idiot. Take your sour grapes somewhere else.
Any dev, senior or junior is on the chopping block. You dont seem to understand where this is all headed. Youre a straight goon if you think youre too smart or essential to be replaced.
I feel like training this LLM on Github is like Google training Gemini on Reddit.
hahahaha
sorry if I misunderstand this, but isnt that what already happened ?
@@erieschlif youre not aware Geminihad a big pr event recently… ill let you look it up.
The implication is that theyre goofy af for using github when most of the code on there is junk student projects and broken repos
I mean.. it could be trained on StackOverflow..
True story!
My Documents folder is a better training dataset than Github or StackOverflow…
We now need permissive licenses like MIT and such that explicitly prohibit use by major LLMs.
Quick shot - you can use GPL and then sue CoPilot into oblivion ;)
GPL, LGPL.
Might be able to sue on the attribution clause of MIT/BSD license.
@@georgerogers1166 AGPL too, which is GPL but designed for networked services
We needed those licenses 10 years ago. Everything is fine. Ish.
Saying that one should not learn to code because there are LLMs that code is the same as saying you should not learn mental arithmetic because calculators exist.
Or not to learn to write because printers exist
More like don't learn calculus(or any prerequisites) because you can use a scientific calculator
More like don't learn calculus because you have a pocket calculator...
you will still have to code albeit in the very high level language of english lol. there will still be things to learn in order to better conduct the automation
@@finnaplow you still must learn all the basics despite many things being automated. You won't be able to reason wirhout an interdisciplinary fundamental knowledge, and AI will not replace this necessity.
10:51 But that is the dream of every single company, sell the pickaxe, rent the mine, and mine aswell, and on top of everything if it is able to sell the gold back to you is even better. Companies care about one thing and one thing only, profit.
It is called vertical integration.
Not only did he say not to learn to code but stop learning computer science. He also said don't learn to code because "WE" will just do it for you. Sounds like a monopoly to me!
I've always argued that at this stage basic computer science should be taught in school starting in junior high or high school. Its not necessarily to learn to be future software engineers or anything like that. I just think it would be good for a society that relies so much on computer technology to function like ours for all citizens to have a basic understanding or introduction of how computers and software works under the hood. This knowledge should not be held behind sophomore undergrad courses for CS/Engineering majors anymore. When I took my first embedded programing course and learned the basics of how Instructions get pulled from memory and get processed by a CPU I didn't think it was too difficult of a concept to teach at a high school level. Same with some of the concepts we learn in an operating systems course. At the very least this should be provided at the AP level for highschool students.
@@MrIndiemusic101Given you already understand something, it is difficult to put yourself in the shoes of somebody who has no clue about it yet.
I would also tend to say that learning about second quantization in quantum mechanics and operator theory is simple and everybody should be able to understand it. But also this is probably not true.
@@tybaltmercutio Good point. Ironically, the people taking classes on quantum mechanics, or embedded programming, are already a biased sample of the students in the education system, due to filters like grading, college admissions, and natural inclinations and interests. While I definitely think society could benefit from a base level of knowledge about computer logic concepts (mainly to prevent exploitative marketing IMO), providing such education at scale tends to fail as you end up teaching things students just don't want to learn. Of course, some students also don't exhibit the "flavor of intelligence" required to excel in these subjects either. For example, it wouldn't make sense to teach Jacob Collier how to program an operating system or learn quantum mechanics.
@@tybaltmercutio For some background I have tutored high school students in coding as a side job so I have a bit of hands on experience teaching children about CS. Regarding your point At a university undergrad level of understanding you are probably right its not appropriate for junior high or high school students to learn. However I believe there's enough simpler introductory ground level concepts that a high school or junior high level of understanding of how computers and software work that can be made into a curriculum. And there are definitely some more complex topics that can be dumbed down as some level of detail can wait for university and that is okay as that's all high school level courses are in the grand scheme an introduction to certain specialized topics that you'll end up re learning anyways in your undergrad. Ideally it shouldn't be any different than grade school level Physics, Biology, and Chemistry. Those are subjects introduced to you at a particular level of understanding that you end up having to re-learn should you peruse an undergrad in STEM. For example memorizing and applying basic physics formulas in high school vs at the appropriate undergrad level of understanding where instead of memorizing formulas you relearn everything again except this time you apply calculus instead of the plug and play formulas you learned in grade school.
I honestly don't think the introductory courses I took at age 19 during my undergrad were all that Arcane that someone 2 years younger and just as little real world experience in the topic wouldn't be able to grasp also. I'm not expecting high school students to be able pickup an instruction set booklet and begin programming a chip on a board. But my prior example of the very basics of how a CPU instruction cycle works with some visuals and diagrams can be made digestible enough for kids to understand. And on top of it its a branch of science that's a lot more visible as far as tangible things that students interact with every day which makes it an easier topic to engage students in. Blowing my nephews little kid brain when I was building a desktop pc and explaining to him what all the pieces do and how they are essentially in his ipad, and other devices comes to mind. Things like that is why I think its important. People these days are too disconnected from the very tools they rely on to work, and keep society afloat which is why I'm an advocate for atleast illuminating some bits of CS earlier.
It's mind-blowing to me that a company that derives value from programmers would want to nerf their future talent pipeline by telling kids that.
Also, one thing I think that's important to remeber is that learning to progam isn't isolated to the benefits of producing applications, automation, etc. Learning to code has shown other benefits in the context of learning to think critically, learning concepts of logic, being a rewarding activity to people of all ages, a way to express creativity, and help people from a mental health perspecitve. To limit this scope is like saying, "don't learn English or how to write because an LLM can make a mediocre blog post sometimes."
It's also insane.
Human language cannot convey logic in the same way I can with code. It's just physically impossible.
You could get close, but you would probably have to make it.... look like code.
Psuedo code could maybe work... but then... you're still coding!
Agree, but "mind-blowing"? Come on. We are talking about business people. If it produces money, there are no scrupules. Ever. Unless the law forbids them.
"derives value from programmers" , what are you talking about?
This is money!, programmers are an unwanted expenditure in business, whose only value is in being exploited as much as possible and then some.
Trying to argue against that, that somehow a publicly traded company should have ethics towards programmers is spitting in the wind.
@@TheNewton
My primary point was that companies like Nvidia heavily rely on software and hardware engineers to create their products and derive their primary revenue from them. It seems contradictory for such companies to discourage kids from learning to code, as they are the ones who likely benefit the most from a large pool of talented programmers. These programmers are essential for developing and potentially profiting from powerful AI tools like Starcoder 2.
This contrasts with companies like Kellogg's, where coding might be a secondary function and a cost center rather than a core part of their business model.
While I would personally prefer to live and work in a world where companies consider the ethics and morality of decisions involving their employees, my main focus was on the practical implications of discouraging kids from coding. By doing so, companies like Nvidia could be directly impacting the size and quality of the future talent pool they have to draw from.
However, it's worth noting that by potentially limiting the available talent pool, companies may see AI development tools as a more critical requirement. This could factor into their decision to discourage kids from pursuing coding. Nonetheless, my point specifically relates to companies who produce software, hardware, and other technologies as their primary source of revenue, not those where software development is simply an overhead cost or a secondary revenue stream.
Unlearning LLMs is pretty cool, essentially the model is split into tiny shards that vote on output, if data needs to be removed, the shard can be retrained without the removed data.
Proposed test question for StarCoder:-
request:Generate code to produce a clone of Minecraft written in ANS FORTH.
Why make it so difficult? Wouldn't it be more interesting to just ask it to produce a moderate complex webapplication with one prompt? Because as far as I can tell, I haven't seen AI do anything like that yet...
@@paulholsters7932 It was a joke.
Skeptical take: LLMs are like search results turned into condensed content… a fancy computer-assisted way to plagiarize the work of others.
Just in case you still weren't convinced nVidia was a garbage company...
Media, hype, and hysteria. Really dragging me down lately :( tech bros going crazy
The ride won't end until we hit the bottom of the gorge
15:45 The AGI utopia will not happen, indeed. But the AGI dystopia will...
Seeing as More likely just went satire on the concept a perfect society with his coined term...
How can you opt out if they've already acquired the data, and then telling you, you could have been opted out? It's the situation from Hitchhiker's Galaxy - where the plans of the road going through the road were accessible for 50 years at Alpha Century, it's your fault you haven't checked
Does anyone really believe that developers will be useless in the near future? I don't… i think we are so far from that. Or do we shift our expretise for developing better AI?
In college, so no experience here. I see a minor threat to people that refuse to accept that LLMs are here, but to those that work with the technology and have an understanding of how it works as well as computer science fundamentals I don’t think it’s too bad
If AI could replace developers it could also replace many other jobs: accounting, marketing, HR, sales, CEOs, etc.
@@user-mg9hi5ln8nAs If only senior devs could use AI 😂 Bro what stuff are you on? xD
@@user-mg9hi5ln8n If junior developers are replaced now, there won't be any senior developers in a few years and companies will be in a real world of shit. And I'd argue that anyone who takes this overhyped bubble of a "threat" too seriously drinks too much Kool-Aid
Personally I'm willing to jump into programming still. There is way too much uncertainty with this stuff.
we already have self-driving vehicles, the problem is they only work in controlled areas, IE; mule yards/warehouses etc
11:50 when you have issue going back then click and hold the back arrow and it will open dropdown menu with your history of that specific tab and you can just go directly back to any page you opened in that tab
or just right click
i can buy the "you don't need to learn to code"-thing .. if you only think about languages on the surface ..
in reality learning to code is quite language agnostic .. what you really learn is the ability to solve problems .. and that will never be obsolete.
i wonder if they checked if the repos in 'permissively licensed data from Github' actually compile and build. On the other side - we need to have new license that disallow this sh&t.
Currently running starcoder 2 on my AMD GPU with ROCm thanks NVIDIA! :troll_face:
Outside of making an LLM produce assembly and the only abstraction is prompting the LLM, I don't understand how we replace programmers. If a new version of C++ comes out with additional syntax, no LLM will know what that syntax is and won't offer it for a while since there will be no code to feed it. Maybe I don't understand LLMs enough to understand how they can replace people but it seems like the bottle neck is the requirement to train them. LLMs seem like they can never been on the bleeding edge of development because they can't come up with new ideas.
I'm having that exact problem right now.
I'm trying to make a blender 4.0 interface, it came out after all these LLMs were made...
It can't do it.
And LLMs if they get trained on data produced by llms start to degrade (like a copy of a copy it messes with their statistical model and makes them less useful)
So what do we do? just no longer develop software and let ourselves be frozen in 2023 forever?
Maybe they can be fine tuned to learn the new syntax from the docs and extrapolate? Anyway, looks like we are not at this level. I wonder when NVidia will start to layoff their developers and launch a GPU driver fully written by a LLM.
I worry that we just stop improving software. Peg to cpp11 or something.
Which I mean cpp is just assembly with a mask on, but still.
Jetbrains does this, but also provides a lifetime licence pegged to last month you payed your last month subscription.
IF you dont need anything new, you can keep with old versions without any issue.
You can click-and-drag down on the back button to quickly escape history traps
So instead of writing bad code and refactoring, I'll be inputting bad prompts and reprompting? The more things change...
the DougDOug Primeagen crossover was not something I was aware I needed
Called it. We went from "you own nothing and will be happy" to "you will own nothing you MAKE and be happy".
Vegan Code -- LOL
Hahahaha! 😂
Could all humanity's "necessary code" be already written, or written before all current developers retire?
The longer and thicker the bar is, Prime, the better. We all know that.
gotta watch out for Big Code, all they care about is profits!
Prime knowing about DougDoug is the greatest crossover since Avengers Endgame
btw, they seem to also use licenceless repos, which are full copyright (all rights reserved and all) by default
good lord, the caption for the bar chart is horrendously funny
Actually you can train negatives in AI models with a few different techniques, though it doesn't necessarily remove a specific point of training data, but rather, you can remove a specific concept if needed.
Generally you would just remove the data from the dataset for future models, though. It does create a weird situation where someone might train a V1 of a model with data that later was opted out, produce data for a V2 model with test-time compute or something in the V1 model, and so the benefit of training on the opted out data might remain in the V2 model in certain training schemes, so it can get kind of weird.
@thePrimeTimeagen,
my thoughts go to not only the permanent Jr's, but an insidious strategy to FORCE the Sr++ people, as you phrase it, to also become permanent Jr's as well.
What if they (LLM's in general) are integrated into IDE's that make it so easy they many flock to them, but the code generated from the prompts is virtually encrypted via extreme obfuscation in a proprietary language where the IDE supplier legally owns all the IP of this proprietary language, which is only readable by their LLM itself, and permanently abstracted. so you could be that Sr C or rs, etc. dev, but the wave of Jr's will drown you out of ever being able to practically apply any of the human-readable languages into any serious project that pays. Basically invalidating an entire lifetime of someone's expertise, thereby eliminating the need to compensate them more, because you force their hand into using your tool instead.
there are many more ways than this just how ugly it can get.
Prime is right. The problem is not charging money for services with AI. I'm well aware that that hardware is very expensive to both buy and run. The problem is Nvidia's addiction to engaging in mutually detrimental behaviour; that is, behaviour which is bad for both themselves as a corporation, and us as consumers.
stock holders seem to like it, and that's their real customer right now.
"I have first hand experience with hand crafted my little ponies" - prime
Any opinion about when to start shorting nvidia?
I find AI code assistants helpful especially when they're wrong. I get these 2nd winds in motivation when I have to figure out why the ai is such an idiot and doesn't know what its doing. Nothing gives me more pleasure to understanding the exact specific malfunctions in someone's thought process.
You can run the model locally and it doesn't have to be on an Nvidia card - there's no monopoly being created by this
Ohh ffs dont be naive
First, make it free and available to anyone
Make people dependent on you
"oopss" now if you want the pro version, gotta pay up!
Been done a million times before
We need an “MIT v2” license that prohibits use for AI training:
Coding teaches you how to think. This severely undermines my trust in anything NVidia says anymore...
As long as we value learning, we will keep progressing. Mistakes will be made that is how we learn.
I love the person who draw 0 - 80% bar chart
Do most coders find cording “tedious”?
I don’t watch primagean to understand computers, I watch to see him mangle the social sciences and business
How bad it is compared to DeepSeek coder?
God, I'm so looking forward to the first lawsuit alleging code theft and the defense will be "well it's just what the AI spat out, _we_ didn't copy from anybody"
china of all places is the first place to fine someone for stealing ip. Ultraman apparently. Only 1200 dollars.... but multiply that by how many people they've taken data from and every country that has citizens they've taken data from.
People who are opposed to AI have kind of painted themselves into a weird corner - they're trying to argue that AI generated images for instance "aren't made by a human, so it isn't subject to IP laws".. but if they're arguing that it's not made by a human, then it also seems like it should also be impossible for anyone to break copyright using AI generated work, after all, if it supposedly wasn't created by a human then it can't really break copyright.
To me it feels like the obvious solution should've been that whoever uses the AI generated stuff is still treated as the person that created it.. but for some reason people have decided to go down the weird path where they try to argue that nobody has IP rights to it because it wasn't created by a human and I think it's going to set some really awful precedents.
3:48 always something strange tweet when he opens twitter
This code was responsibly written by a grass-fed vegan.
The thing about all of this is that no one actually needs this.
Its BigCode's fault that my app has no users
I think that the CEO of NVIDIA never tried to use the code from the LLMs to build something and/or his staff is trolling him.
But if everyone quits learning how to program then how would they be able to maintain their LLM? lol
in 30 years we all live in the scorch mad max style 🤩🤩
How is it possible these stupid excel bar graphs make it into an official post?
Just keep convo out of the stonks plz
So real talk: what application can you build with it?
A hello world calculator
@@-Engineering01- Amazing!
I'm about to create some new encryption, that doesn't just encrypt the software. It encrypts the binary code into smaller chunks that self-encrypts and binds itself to lru-memory in ram. This is the physical hardware. We then sent OTA updates that push the current build, which is also "b³-Encrytion". This then meets with the "b³-E" binary code on the physical hardware, and decodes each chunk with its corresponding couter-parts on both sides, and fully executes the code. Lol watch how this new tech goes over everybody's head. It's ultimately what is needed in next-gen Cyber-Security
"Invented by Steven Casimir"
This sounds like something i said 4 days ago 🤔
Well I'm probably in the last ten years of my career, so I don't care that much. But people starting now should probably consider it.
15:10 But he's gone on for the least 15 minutes saying how the big corpos decide everything and not "the market" 😂
16k context length is so far behind these days. Their entire article also reads like "I think my product is the perfect thing in the world, so invest in me!" and I seriously heavily doubt they've got any chance against Claude or the many other superior models from true research organizations out there.
There is something called machine unlearning
that rbf sounds like taylor series, pretty neat.
IMO, Nvidia is doing the right thing to provide more to the client(personal and small group developer) and left less for its competitor such as AMD, Intel or other chip Co. ASAIK, they are also building their own software infra alongside with the hardware. It is a fair game, but good to clients.
We were promised jetpacks!
"big code" sounds like a monopoly
1:30 Prime not thinking big enough. The product won't charge a penny up front, it'll just cost you 2-5% of your equity
If it isn't a coincidence then it must be their position that "Blindly trusting AI code gen" is the future.
Yayyyy hes finialy talking about it
By the time it replaced seniors most people in trouble and it be a society issue.
isn't starcoder open source?
No way to have a monopoly, these LLMs are easily interchangeable, and a bunch of big companies are releasing new ones every month. Whenever one's better or cheaper you can just change your API config to point to another one. There's no cost of transition like you'd have with Databases or tech stack
They haven't tested it on Win API code as the accuracy of most LLM on Win API code is abysmal.
Did nvidia trained it on its own source code ?
No, They do not want us to see how “smart” their drivers and CUDA compiler.
Haha it doesn't pass the smell test
14:50 Yeah, the "permanent junior" and dying out of all humans capable of coding well is a real risk here.
sounds veeeeery suspicious! LMAO
2:57 They sell fool's gold actually.
Starcoder 1 was preceded by santacoder. And I used santacoder for a research project by finetuning it on shadercode.
I will benchmark starcoder2 on the new benchmark I am developing... Once it's upstreamed to transformers.
The model is fairly open source (some miracoulously read past that paragraph) and the dataset is much more open and important: the-stack.
Nvidia just provides engineers and compute (432 H100s) for this project. You can freely use this model on any hardware you want. Nothing is paid. In fact, I run it on Intel GPU and similar without issues.
There is a specific license issue. With Open Rail if you want to discuss this something...
Your whole argument is BS. Nothing that Jensen says matters for this model release. It's larger than Nvidia.
As for the benchmarks, Human Eval is really stupid and I can explain why it's busted and even give you research papers. Evaluating language models for code is a difficult problem and I am writing my thesis on the topic.
Please keep us posted here with your result. Thanks!
LMAO, the talk of responsibly sourced data does remind one of veganism. 🤣
“Don’t learn coding, you only need to learn typing words” smh…AI should be a tool that helps us leaning and experiencing new things, not prohibit us from them…
responsibly sourced data... 😄
U can't take out data from an llm
lol if i get my hands on starcoder i can build the next python ima call it starthon
Can't wait for AGI to takeover
the railroad did this after the steel barons proved you could pay half the population to oppress the other half
Next step: Make your own APIs and frameworks as complicated as possible, so no human wants to deal with it without relying on your LLM solution.
14:27 skills atrophying for juniors reminds me of how even fundamentals like view source (ctrl+u) in browsers has been undermined by "modern" web development practices in that they spew out this soup of divits, generated machine names for CSS class names, minified or outright obfuscated JS, etc.
Most anyone in web dev cut their teeth with view source, now newbies are left gumming inscrutable brick walls.
"Modern" web development has been pulling the ladder up behind itself for a some time now hobbling new learners in insidious non-obvious ways.
LLM code generation burns the ladder, and all the dangers that comes with.
responsibly sourced data
If these LLM’s are scanning git repos to train the AI. I wonder if they can determine if the code is a malicious code. Because if this malicious code passes to the security of the AI, oh boy...😀😀
"Imagine needing a subscription just to do your job" -- someone who doesnt use Visual Studio
Don’t forget Claude 3
Claude 3 Opus via API is too expensive for coding. Also, how come it can only generate 4096 tokens???
@@couldntfindafreename Interesting. I hear you can upscale it to 30000 or so tokens. On the other hand: what is "too" expensive? If it can create a complex webapplication on a day, than expensive is something different than when it just provides some code snippets right?
> I dont' know the an LLM works in that sense
Ok let me break it down. It's simpler than you imagine. You have a foundational model that's trained on language, and a fine tune dataset that's trained on code. Fine tuning runs happen at periodic intervals, and so opting out will mean your data gets removed from the fine tuning dataset, and when the model is next fine tuned, the models performance will change probably slightly, because you've removed some code from its dataset.
We'd have an easy win against Doug's soy chat