5 new Software Engineer Jobs that "AI" will make (that aren't for ML specialists)

Not only with "AI" LLMs like Devin not take all the Developer Jobs, but there will be new Software Engineering opportunities made by "AI" in the future. And i'm not talking about "AI" Jobs like Machine Learning Experts. These are jobs for Existing Developers and Software Engineers living in a world full or ChatGPT and Claude and Gemini.
Don't believe the hype about there being no human programmers left in five years. "AI" Models are software, and ALL SOFTWARE HAS BUGS, and with bugs come opportunities.
Chapters:
00:00 Intro
01:57 Number One: New Startups
02:02 Number One Part 1: No-Code Web History
03:45 Number One Part 2: LLM Startups, Likewise
05:24 Number Two: Automation created more products
08:20 Number Three: Customizing LLM Models
10:15 Lora and ControlNet Demo
11:19 Number Four: Debugging LLM Generated Code
12:23 Number Four Part 1: What was it thinking?
15:23 Number Four part 2: Regression Bugs
16:37 Number Four part 3: LLMs have Bugs in their own code
17:30 Part Five: Adversarial Attacks
20:12 Adversarial Attack Examples
20:31 Hope that wasn't depressing, too.
20:39 SignOff
Links from the video:
For more on ComfyUI/StableDiffusion (This guy is great):
/ @sedetweiler
Paper on LoRAs: huggingface.co/papers/2106.09685
Info on ControlNets: docs.nvidia.com/nemo-framewor...
Debugging Quote: en.wikiquote.org/wiki/Brian_K...
Top 25 worst kinds of bugs: cwe.mitre.org/top25/archive/2...
The design that allowed for buffer overflow bugs:
www.infoq.com/presentations/N...
And these are the kinds of bugs that will make Buffer Overflows seem like child's play:
www.vox.com/future-perfect/20...
www.wired.com/story/tesla-spe...
www.technologyreview.com/2019...
openai.com/research/attacking...
www.wired.com/story/ai-advers...
www.pcmag.com/news/this-silly...
NOTE: (Not that anyone will read this, but) When this video uses the term "AI." it (like most of the current discourse these days) means current and near future "Generative AI" a.k.a. Large Language Models. If you're thinking about the Agents from the Matrix or the Robots from Terminator, that's DIFFERENT. That's the OTHER AI, the Human-Intelligence General AI, Sometimes called "AGI." That DOES NOT EXIST. For the time being it's PURE FICTION. It may get here soon, although I (and a lot of people smarter than me) seriously doubt it, but we'll see.

[@KuzzatAltay]

“I am not AI hater, I am a hype hater.”That kinda brought the memory of everyone was a blockchain expert few years ago.

[@InternetOfBugs]

I complained about "Crypto" a lot back then, too. I just wasn't on KZbin, yet. (If I'm being honest - I complain a lot about lots of things, but hype is a particular hatred)

[@Mimi_Sim]

I truly believe Ai won’t be headed to crypto land but that doesn’t mean there isn’t a ridiculous level of hype. I will admit a have a sweet tooth for AI content however I recognize it’s generally low nutritional value. The level of speculation involved in AI reporting is akin to tabloids. Now no judgement if you like Celebrity Gossip but don’t paint a tabloid as an article of peer reviewed research.

[@ImARealHumanPerson]

Everyone with a smart phone has been using AI for over a decade without even knowing. It's everywhere and always has been in one form or another. That's only going to continue.

[@ev_codes]

@@Mimi_Simyou’ve articulated yourself incredibly well here.👏🏾

[@InternetOfBugs]

Well, the idea of a immutable distributed ledger has real value. Crypto got super scammy, but the underlying tech (even the concept of NFTs) has a lot of useful applications. Of course, no one will be able to propose anything serious be done with any of it until most everyone forgets, or they'll never be able to show their faces in public again. Hoping the same doesn't happen with LLMs.

[@earlworth]

Remarkable how its the lowkey channels that make such insightful observations, while the big names just pump out hype and hyperbole. Hope you get many more subscribers, your content certainly deserves it

[@InternetOfBugs]

Thanks.

[@Fishamble]

Hype is more interesting than a knowledgeable guy giving his measured take.

[@InternetOfBugs]

@Fishamble ain't that the truth. Gets a lot more clicks, too.

[@igorlitvak1178]

Regression bugs already seem to be a huge problem for LLMs. I had a very hard time with chatGPT generating bugs in it's own code, then when I pointed out the bug, it immediately messed up in a different place. And this happened in tiny code snippets 50-100 lines long. Just imagine what a mess it will be in a project of 10k-100k lines of code and beyond.

[@lukeherbst7931]

Ive noticed this too. It also seems to make the same mistake over and over again (hallucinations?) and cant spot it It also sometimes does poorly with black box code although both cases were gpt 3.5

[@InternetOfBugs]

In all fairness, that happens with a bunch of humans, too. It's the most common cause I've seen of Mobile App "lowest bidder" project failures. The only way I know to solve it is with automated integration test suites. I'll have a LOT more to say about that in future videos. It's one of the most important insights of my career. But, like I said, Humans don't have hardcoded prompt word count limits, so LLMs have a much harder time.

[@ckorp666]

maybe it's bad luck, but ime GPT really starts freaking out and becoming useless once it needs to debug between a few different functions. one time it even started trying to access arrays[like}[this}

[@itskittyme]

what you are describing is a feature of ChatGPT, not a bug. Try setting temperature to 0 and use GPT 4.5 using the OpenAI API and gone will be the mess you are describing.

[@dallassegno]

Same

[@andreipaven4388]

As a person with only a few years of Software Engineering experience (after switching careers), I find tremendous value in listening to older people with 30 - 50 years of experience in the field, effectively using their wisdom to separate the wheat from the chaff and providing clarity amidst all the chaos and hype out there. Thank you sir, never stop sharing wisdom and objectivity.

[@bancroftberlin]

I hope your channel becomes popular. A reasonable voice with industry experience between all the hype, doom and "buy my course"-content.

[@InternetOfBugs]

I appreciate that.

[@piotrek7633]

Buy my course are getting more annoying day by day, especially chat gpt prompting courses since its basically teaching people how to talk

[@weeb3277]

this man is like a light at the end of the tunnel a beacon of hope a kernel of corn in the sea of AI poop thank you 🙂

[@ExecutionMods]

Listen to experienced engineers, not AI CEOs who are professional bullshitters whos job is to get as much investor money as possible.

[@soup9911]

"a kernal of corn in a sea of AI poop" you are a goddamm poet

[@nighttalkersmedia]

"When everyone is super, then no one will be super"

[@masgustavos]

Know your content is valuable by the ratio between views and comments. You're making SEs engage in productive conversations, and this is a trait one only sees in a team or community that has proper thought leadership. Keep doing what you do, I'm subscribed!

[@InternetOfBugs]

Thanks. I appreciate that.

[@redpillsatori3020]

Thanks for the practical, in-depth analysis of the situation. Subscribed!

[@ka.MeHAmeHa]

Your shelf is dope!

[@seraphcms2511]

This topic reminds me so much of no/low code that was swamping the media a few years ago where evenyone was saying it was the end of devs. This was nonsense of course as not only did it take more devs to build the no code platforms, it also meant that to use these applicatuions effectively, you really needed to become a sort of meta-developer!

[@InternetOfBugs]

Yep. Exactly.

[@drummermike5150]

I used to love the term "citizen developer". Glad that isn't used anymore.

[@cultofaction8938]

Excellent video. It's good to see some optimism from the software engineering field in light of things like Devin popping up more and more.

[@meijiishin5650]

Made me happy when you brought up comfy. A lot of guys love to talk about this and you can tell they've never touched one of these systems. When people think AI art, I think they mostly imagine those bad Dalle3 youtube thumbnails you see nowdays, but that's a skill issue. If you actually have a decent understanding of this stuff, you can produce art that would fool most people into thinking it's real. It's just funny to think that most people's idea of art made by an LLM is one-click-done magic, but when you actually get invested in it there's a ton of learning to be done. It reminds me of DJ-ing a little bit. Everybody thinks it looks easy until they try it.

[@user-kg1od9es5d]

"If you actually have a decent understanding of this stuff" this is the problem - short term on 'experts' can leverage it to get what they want. but long term, it should be, so that a total amateur to get what they want.

[@user-or4ow6bb9n]

Man, I watch a lot of tech stuff on the internet, and your clarity, and honest words of your experience... it's just so freaking good. I really enjoy you'r content

[@InternetOfBugs]

Thanks for saying so.

[@ChrisCDXX]

As much as I would like to agree with you… I believe we are going to have to start thinking about post labor economics soon. Thanks for sharing your thoughts!

[@Kwazzaaap]

Lol it's funny to see how crazy programmers get when their jobs get automated. Heavy industrial and agricultural automation didn't create post labour economics and neither will this. This is just your first rodeo that people in other professions have been having for 200+ years.

[@charlesE89]

Great video, I’ve had similar thoughts! The future is bright. Subscribed!

[@drummermike5150]

Great video, glad it showed up in my feed. Appreciate the honesty about the title.

[@InternetOfBugs]

Thanks

[@afai264]

I'm enjoying working through your videos , regarding the 'regression bugs' issue could LLMs get to the stage where the 'hallucination' issue is fixed (I've been learning ML/AI and it seems to be evolving at a tremendous rate!) If the hallucination issue does get resolved could the LLM be fed an existing codebase (using RAG for example) and generate a regression test suite before making any new changes. Lots of devs don't like writing tests so maybe this would be a good thing anyway, and devs can concentrate on writing the new stuff. I think LLMs are only going to get better and better (and the costs to train them will reduce as more efficient training techniques are found) and also agree with you that technical people who can think through and resolve complex issues will always be required (and hopefully the end of the world isn't nigh just yet for techies!) Keep up the good work these are very interesting to watch.

[@jee5874]

i can see you growing big, please don't change.... you are doing an amazing job

[@ericford6756]

It's all just too difficult to say. There's a lot of sound reasoning here, but much of it assumes we stick in the same paradigms without more categorical leaps (rather than mere 10 percent improvements here and there). Also as ML models are used more and more in the development of future models and systems that use them, we could see a strong flywheel sort of effect. There's plenty of sci-fi material that has already explored AI takeoffs and race conditions. Not that current LLMs and a full general AI is the same thing, but clearly there will be systems built with LLMs as one of the pieces which attempt to get closer to general purpose intelligence.

[@satchillananda]

Thank you for being critical with the anti hype movement!

[@albinasuikis8636]

i think this is the most balanced comment out of everything ive seen on ai

[@InternetOfBugs]

We don't know, yet, whether LLMs are a milestone on the path to AGI, or a dead end we'll have to walk back before we can progress (there have been several of those in AI research over the decades - when I was in school it was believed by many that Symbolic AI a.k.a. Expert Systems were the right direction. When I was in diapers, it was DARPA/ALPAC/Marvin Minsky, in 1970: "In three to eight years, we will get a machine with average human intelligence." Source: eprints.ugd.edu.mk/28050/1/2.%20jrc120469_historical_evolution_of_ai-v1.1.pdf ). My personal belief is that the brain is so much more complicated than we currently understand (Source: www.psychologytoday.com/us/blog/consciousness-and-beyond/202309/the-staggering-complexity-of-the-human-brain ) that we're not going to be able to replicate that complexity with a randomly connected collection of neurons, no matter how many billion of them we use. Coupled with two looming issues that could halt the progress currently underway: Running out of data (Source: epochai.org/blog/will-we-run-out-of-ml-data-evidence-from-projecting-dataset) and model collapse if the training data for LLMs gets tainted with the output of LLMs (Source: www.techtarget.com/whatis/feature/Model-collapse-explained-How-synthetic-training-data-breaks-AI ), I think it's unlikely (certainly not impossible, but improbable) for AI to take off soon (I'm a functionalist, so I do think we'll get there, I just don't think it at all likely that this is the end of the path).

[@satchillananda]

@@InternetOfBugs Nice coping!

[@asdfasdfasdf1218]

The categorical improvements will be key to making all kinds of new applications. Many seem to think that AI just improves by itself without needing to find brand new approaches.

[@djust270]

I am an Automation Engineer for an MSP. I am not a developer per say, but I can write code (and I do write quite a bit for custom automations) and I develop internal and client facing automations for various business processes, reporting etc. I am already starting to integrate and help customers with AI solutions. I anticipate that this will continue to grow and make more role even more valuable, so I am excited for the future.

[@InternetOfBugs]

Cool. Good for you!

[@marko5734]

Great channel, keep up the good work

[@InternetOfBugs]

Thank you very much!

[@rafaelarias9759]

Very good insights... Thank you!

[@InternetOfBugs]

You're welcome.

[@Fermion.]

I'm a quality manager at a manufacturing plant, and I used Claude Opus to write a python program that had an inventory control gui that will track the location of our goods in the plant, in real time, and also was a visual system that would show a picture of our warehouse shelves, and by clicking on the product location, it would give you data on the product. This is something I would've never been able to do on my own, and with AI, I had it done in one weekend. Used another week for testing, and implemented it the week after that. Perhaps you may discount that, as some weak little project, but for someone who has never been formally trained in programming, it definitely adds new dimensionalality to my professional skillsets. Not sure if you guys are going to call this "hype" or not, but I can attest to the fact that AI definitely lowers the barrier of entry for Jr. to Mid level projects. I honestly believe it's just a matter of time until all Jr. Admin/Dev/Engineer positions are completely automated. Then after that, mid levels will be next. I'm not so sure of the time frame, but I am sure that it will happen.

[@catalinul5700]

"Never" is a big word, and also I think you are underestimating yourself, you still had technical skills that without you would have had no idea what is python, why choose python for the app, real time stuff, goods location in the warehouse and so on and so forth. How to talk with the AI, check the code, or at least have the needed logical thinking to make sense of the code and so on and so forth. The AI is useless on its own, it needs to be told what to do, and this "what to do" means the person who is telling it what to do has some kind of understanding about the problem it needs to solve. I am a programmer and I have 0 knowledge about C# but I have knowledge about programming in general and if I want an app in C# I have a basic idea where I am going with the AI, someone that is completely clueless has no idea, in fact I think Jr. Devs are going to use the AI more than ever to "fake it till they make it", if serios, if just copy paste then they are doomed.

[@Fermion.]

@@catalinul5700 Perhaps you're right. And I defer to your knowledge of the intricacies of programming and what not. But I just don't see the need to bring on a Jr. guy at $80k to $90k per year, when the quality manager + AI, with no programming training was able to come up with a solution and have it implemented about 2 weeks. Sure, mid-level and above guys will definitely still be needed, especially when security concerns and/or more important technical matters are at stake. But for a relatively safe and simple internal inventory control app, I don't see Jr. guys being needed anymore. Not to mention how much it helps me with my own quality duties. I just recently had AI write new SOPs for everyone from the C level execs, managers, supervisors, team leads, operators, all the way down to the janitor. That would've taken me a couple weeks, but AI had it done in one afternoon, with minimal edits from me. And it also helped me revise my quarterly performance evals for each department, based on the new SOPs it wrote. And it is absolutely amazing at Excel. The way it can take an inventory sheet, with thousands of items, quickly comb through the data, and come up with charts, graphs, and other data visualizations, from merely asking it questions in natural language should not go understated. Data visualizations in a powerpoint really help draw attention to specific points you're trying to illustrate in meetings. I've never, in the past 20 years, been more productive than I have since recently adding AI to my workflow.

[@InternetOfBugs]

No, that makes perfect sense, and good for you. I would argue (tell me if I'm wrong) that it would not have been economically worth it for you to hire a professional software engineer to do that for you. If I'm right about that, then 1) LLMs did a good job, and 2) they didn't "replace" a programmer. That's the equivalent of the "New Startup" point at 01:57 in this video. It didn't replace a programmer, and some day, that automation might prove valuable enough that it might be worth hiring a professional to expand it into a large system, or you'll come to depend on it enough that, if it breaks or something changes in the future and Claude can't fix it, it might be worth bringing in a pro to work on it for a while. I think, all around, that's a win for everyone involved.

[@doge9347]

I’ve been looking into how to get started into learning software engineering for an about a week now, then this pops up thanks I will be looking through your channel :)

[@dariomontoyasequeira6084]

Thanks for the video. I’m a second year software engineer. I have a question with regards to which areas of software engineering you would consider will eventually disappear and which areas will get improved.

[@AIWorks2040]

I feel like I can rely on your opinion a lot. Thank you for giving me a standard thought.

[@nicosoftnt]

Thanks Carl, I find your input really valuable and encouraging, I work with LLMs and the current set of technologies without thinking of the future (because there's no time with the amount of work for developers nowadays) but you usually think the worst is coming, fear and a feeling of despair rule the current environment, so listening to you sounds more grounded, I agree it's not going to be a soft transition but it's most likely not going to be a disaster like the media and AI sellers usually want to portray.

[@bebea11995]

I've the feeling that they need to maintain the hype to keep the huge amount of investiments and the financial speculation. I know that AI and all the stuff are revolutionary and all, but, does it really need all the terror that are being made? I think it's not

[@alexandergeorgesquire220]

Your video is wonderfully clear, concise, and positively impactful! As someone with a lifetime of experience, I appreciate the depth and value you've brought to this content. Thank you for sharing your insights with us.

[@InternetOfBugs]

Glad it was helpful!

[@_-steph-_]

Great video man. I'm taking Cybersecurity this semester and getting into all of the different types of vulnerabilities was pretty eye opening, especially when you know how common some of the vulnerable coding practices are. I have the same concern about AI being used like a weapon, but I guess we'll just have to wait and see. 🙂 Your insight is appreciated!

[@zippo32123]

Also think about how common those vulnerabilities are and what LLMs are trained on...

[@vanishghost001]

Appreciate your take and input on LLMs. Also super glad someone is talking about how the term AI has lost all its original meaning. Subscribed!

[@anasouardini]

This is the most informative AI video I've watched so far. AI however, is in the mean time, killing my chances, I hope it gets better sooner. I'm tired of of having skills that I can't put into practice.

[@tacorevenge87]

Thank you for the videos as always

[@InternetOfBugs]

My pleasure.

[@ArchaicDemise-ex1lq]

@12:55 - I agree in principle that there could be bugs in the tooling surrounding Large Language Models (LLMs), but I suspect that the core of an LLM, which I refer to as the 'kernel', is relatively straightforward in its function. This kernel primarily executes billions of mathematical operations to run the LLM model. In this sense, it might be possible to vet and verify the kernel's functionality for correctness, similar to how the L4 Microkernel has 'proven' functional correctness. This does eliminate a myriad of bug surfaces, and restrict it to mostly the hallucinations of the LLM itself.

[@aredrih6723]

19:48 null references are different from null used as a sentinel to mark a string's end. Null reference are about the null value being allowed for any type (in spite of having none of the field/method). This lead to crash but I don't know of any null based attack outside of crashing a service (might be wrong, feel free to correct) Nul sentinel are just an encoding format. There are case where they are a problem (e.g. terminating a string early, leaving control to the attacker of what's directly after). Definitely a potential vector for attack with a known fix (length prefixed format, not rolling your own serialization, ...) so still relevant but in a different flavor.

[@InternetOfBugs]

That's fair. I was talking about several different kinds of bugs, (25 of them, actually, in that list) and two have similar nomenclature, and I wasn't as careful as I should have been to be clear about which one I was talking about at any given time. Mea Culpa. Thanks for pointing that out.

[@rich2823]

Remember when every IT person thought they were going to be jobless because of the on prem stuff moving into the cloud? Where are all those people today? Doing work with cloud services. AI is that for code.

[@albinasuikis8636]

yes but every step of the way it becomes more rote and algorithmic so youll keep ur job sure but it will be supervising an ai etc

[@hombacom]

@@albinasuikis8636not with vague natural language so you end up to understand the output it gives and that is literally developer skills

[@user-kg1od9es5d]

Steve Jobs said this many many moons ago. Tech removes the drudgery so folks can focus on the interesting/creative/passion stuff. I dont think you nor most commenters realise the realise the real genius of Jobs. He said this back in the 80's.... @@albinasuikis8636

[@dan_taninecz_geopol]

Good discussion about terminology and hype trains and algo optimization.

[@foxdog9332]

wow...im shocked finally a person with a reasonable take on technology. 100% agree its going to be coding on top of AI that's going to replace traditional coding but coding won't go away since it is a very good way of creating deterministic outcomes and those outputs are always changing.

[@Rockyzach88]

Maybe we should be using LLMs to put better controls for human written code instead of using LLMs outright to write code. Like help reduce unnecessary abstractions or complex code, and bugs in general.

[@InternetOfBugs]

I have a lot of thoughts on training LLMs to write tests (and testing in general) that will make it into a video one of these days.

[@miskamyasa]

Great video! Thank you!

[@InternetOfBugs]

Glad you liked it!

[@LukeAvedon]

Good example of Say's law. Would be a much smaller % of the population as coders if the most advanced stuff we had was 6502 assembly. The LLVM thing has pushed me more towards an interest in reverse engineering, as per 12:20 I think we will be all spending a lot of time reversing weird LLM code/output soon. Working in large corporate codebases already feels like more reversing then coding. Also, saw you on the Primeagen stream! Your famous!

[@PRIMARYATIAS]

Is ThePrimeagen did commentary on his video?

[@InternetOfBugs]

I'm almost afraid to ask. Should I go watch it? Or am I just getting dunked on the whole time?

[@quc522]

​@@InternetOfBugshaven't seen it but he's a fair guy he'll likely agree with you on most of what you say.

[@PRIMARYATIAS]

Also I guess you meant LLM and not LLVM ? You speak about Assembly and not LLVM-IR I guess.

[@LukeAvedon]

@@PRIMARYATIAS ah so embarrassing

[@bestechdeals4539]

Thanks for putting a perspective from yours years of experience.

[@ArchaicDemise-ex1lq]

@18:22 my favorite bug is really the "in-band command & data" - AKA SQL injections, but there's more to it, directory traversals are STILL a thing in 2024, and IMHO LLM injection attacks are just the same thing really.

[@Mimi_Sim]

Comfy Ai is the best - non-coder here but I used to do audio engineer so signal flow is something intrinsic for me. Comfy Ai really exemplifies the same concept. I love the visual aspect for really understanding.

[@InternetOfBugs]

I've learned a ton from playing with it and the pytorch underneath it. It's the clearest example I've seen of an LLM ecosystem that's customizable by the end-user without access to large training sets or compute farms.

[@Pe0ads]

Really interesting. I’ve messed around with LangChain, RAG, and vector DBs a fair bit and it seems to me that this kind “model with infra around it” is going to be a big deal. Particularly when you have a database with richly modelled domain knowledge, eg a knowledge graph.

[@ciphernx]

Great video with lots of experince/logic based expectations. I hope you continue making these valuable videos. I'd also like to ask what you would advise me to learn or do right now. I am a CSE student in Turkey (studying in an average - above average uni but education is not as good as it is in most US, EU countries). I have started my programming journey with web development, learned a bit of front-end but I can't say I really like it. Obviously I want to find a field that I can be successful and be less stressfull about AI or LLM taking my job. What advices can you give me, in this specific topic and in general about academic career, job, things to do in college etc. Thank you in advance.

[@InternetOfBugs]

It's hard for me to say, because I really have no idea at all about that market. I've spent my whole career in the US.

[@charlesE89]

I would love to see videos on customizing LLM models!

[@mateusmattossabarreto2852]

we need more videos like this

[@kelsey4everSmile]

Buen contenido, y buenos insights. Gracias!

[@christopherchilton-smith6482]

I may be a unique case, I've been trying to teach myself to code for many years and I struggle to solve even the simplest coding problem. Yet with GPT4 I've been able to make a lot of progress on my ideas, mostly because I just want to do what others have already done but with my own twist. Without access to an llm I'd have made no progress at all. I definitely appreciate the advice on learning to tweak llms to produce output for specific types of business, that gives me something tangible to grasp and focus on in this space as a less technical person.

[@BillAnt]

AI is a tool, it's great if used properly, but it's still a good idea to have a good working knowledge of programming and problem solving. Basically if you feed sh*t to AI it will output sh*t code.

[@Fermion.]

@@BillAnt I'm a quality manager at a manufacturing plant, and I used Claude Opus to write a python program that had an inventory control gui that will track the location of our goods in the plant, in real time, and also was a visual system that would show a picture of our warehouse shelves, and by clicking on the product location, it would give you data on the product. This is something I would've never been able to do on my own, and with AI, I had it done in one weekend. Used another week for testing, and implemented it the week after that. Perhaps you may discount that, as some weak little project, but for someone who has never been formally trained in programming, I'd say you're overestimating the working knowledge that one needs. If you get stuck or have a question, just ask the AI what a real programmer would do, and you'll have options. Basically, if someone can't properly use AI to do amazing things, it's because they're the bottleneck, not AI.

[@techmansion8618]

Sir, what are your thoughts on Data Engineering (learning python, SQL, working on data pipeline and other stuff for this role ) Is this a good career to go with rather than mobile app development or Web development...

[@defenastrator]

You are the best new channel in existence.

[@InternetOfBugs]

Thanks a lot. That's very kind of you to say.

[@leibaleibovich5806]

Greetings! Thank you for the video! It is always interesting to hear opinion of someone who has been in the industry for a long time. I have a question: based on your long-time experience, what advice could you give to people who want to switch careers and move into programming? It would be great, if you could do a video about it. While I do not think that AI will eliminate jobs, I think that it will significantly impact the job market by increasing requirements. I noticed that a lot of people wanted to change careers and become computer programmers. In my opinion, when the field was in its infancy, it was a fairly easy task to jump into programming, even when one lacked certain skills. Fast forward into the future: bootcamps jumped into a business opportunity and made a ton of money. The barrier to entry has risen. We ended up with a broken "senior-only" job marked. All these half-baked junior programmers, pouring out of bootcamps are not needed. I started to learn coding on my own, but I have a sense that it will take about 2 years of dedicated full-time effort to even become "interview-ready". Is it even realistic for someone, who already has a 9-5 job? I would appreciate your advice! Thank you!

[@InternetOfBugs]

I'll think about what vide version of this would be. But the short answer is: don't become a programmer unless you really like programming and you really like solving problems and/or building things. It's a career with some issues, the worst of which is that everything is constantly changing, and the techniques and languages you mastered a couple of years ago might have no value in the market a few years from now. Most of the folks I know that became programmers for the money were both 1) lousy programmers, and 2) didn't stick around long. (And some of them become Managers over Programmers, and that's a nightmare for the folks under them)

[@alex8630876]

Just discovered your channel. Insightful and informative. I unsubscribed from some other channels of people much less experienced and only following hype.

[@dinglecringle5375]

Been binging your videos the past week. Graduating with my masters in Business Analytics soon and a lot of what I've been hearing from speakers and industry members is that companies are cutting out the analysts and hiring analyst engineers (software engineers that know how to do the analysis) What are you thoughts on this change and do you have any recommendations on skills to secure to make sure that there is a place to find in this market? Thanks!

[@prashrittiwari5256]

In a time of chaos many creators were just giving their pessimistic view but seeing your video is really a synchronic view that we have got this is exactly what I have thought as Ai is developing yes it will tale the job but also create new jobs for the person who are programming its just like using your skills in a different way and will be fun

[@ArchaicDemise-ex1lq]

@10:16 knowing how to load a model without a library doing all if the magic is a good step for anyone learning - learn the base level library API's - learn Quantization techniques! Learn how to implement your own lora over the extant models (training and inference)

[@vasiliskarvelas5488]

Instant subscription, I ve got tired of Ai hype and Ai fear. You are a voice of reason my friend.

[@InternetOfBugs]

Welcome aboard, and Thanks.

[@weeb3277]

there is an interesting video on yt called "Deep Learning is a strange beast" have you seen that? they discuss an interesting aspect of LLM how these models have a tendency to hold 2 (or more) contradictory "beliefs" apparently they are also getting worse over time

[@vr10293]

It looks like LLMs will remain good for creating content which doesn't need to focus too much on truth, or which can be provided easy truth values without much logic. The architecture doesn't really have a place for defining truth as anything other than a fuzzy value. This is why people can convince it of ridiculous statements.

[@InternetOfBugs]

I have not. I have now put that in my queue. I hope to get to it soon-ish (for sufficiently long values of "soon"). Thanks.

[@InternetOfBugs]

I finally finished watching it. Thanks a lot. I really learned quite a bit from it.

[@waveFunction25]

One thing I've been thinking a lot is that we might see platforms for development and runtime specifically made for AI. How long until we have an operating system, language, and framework that is specifically made for AI to work with (think the OS in Her). Code now is designed for humans to work with and not AI - it is a miracle that we have AI getting really good at it nevertheless. Imagine when the systems are actually designed to take advantage of AI strengths and avoid their weaknesses. On top of it we would would have a paradigm shift. AI is kind of an interface in itself, so UI development might be much less needed. And a lot of things that we might want to do might be AI to AI interface. Also AI itself is the app - there are so many tasks I used other apps for and now I just use AI. I can also imagine we find ways to actually make overall code quality much better using AI itself. Instead of the usual flow of 1 engineer coding and maybe 2 reviewing, we might have an absurd amount of AI experts refining every little piece from every angle from the OS to language to framework, to the point that the specific apps it creates might be very much sandboxed and observable. Also standards and practices can be built right in, and made sure to be enforced - which is notoriously hard and impractical for humans. Humans are also maybe productive only a couple of hours a day - I think there is a lot of incentive to have AI teams running 24/7 and able to perform a lot of things autonomously. At that point how do you even expect a human to fit in and keep up? Keep in mind transformers architecture is from 2017. We are at the absolute beginning of this. I personally believe it will be extremely undesireable and even irresponsible to have humans touching code at some point when AI does pass a certain threshold (which it is approaching rapidly). As you mentioned we are terrible at code quality! And we do not scale up or down easily - companies get bad rep for layoffs but no one would say anything if one day you had a million AI devs and the next you had 0. It will definitely take a while though, I work in corp and we barely use github copilot which is already very old news. But I strongly believe every incentive is there to remove humans and we are seeing the effects of this incentive all over already. Finally one aspect you didn't mention is how AI is actually poluting the internet. I don't go to Google anymore not just because Google search has been horrible for years, but also the results are garbage because the overall content on the web has become so many times more garbage. So instead of going to a website and consuming their AI generated content, I would rather just stick to my trustworthy personal AI assistant as my main interface to pretty much anything as soon as it gets there.

[@seraphcms2511]

This ia a really realistic and well reasoned commen! One thing about Copilot .... I think all the hype is because it helps very new devs (so it's actually a very good thing!) which makes it very media friendly.

[@gokuljs8704]

This is really good

[@andycalifornia426]

Props for saying that this is not AI. It's nice to hear someone who KNOWS confront the narrative. Other people (for example, my co-workers) keep repeating nonsense like mindless drones.

[@Pe0ads]

The LLM attack vector piece is very interesting. All these companies exposing LLMs via chat bots etc are going to face issues. LLMs are great at parsing, like really good. Got an annoying list of badly formatted stuff? Get an LLM to parse it. That’s a backend tool that is less exposed to an end user. You might even conceptualise of it as a parser, but behind the scenes it’s an LLM. Now imagine that parser becomes a core part of your input handling system. One day a customer uploads a file that passes all the virus scans, but intentionally attacks the parser LLM. This gets even more interesting if you have LLM “agents” with the ability to operate reasonably autonomously. Maybe the parser LLM talks to an agent that is able to setup billing processes… you can see where this is going. And btw, having spent several days recently debugging handwritten parser code, I’d gladly hand that off to an LLM. There is just a tradeoff to consider. Great video, got me thinking! Thanks!

[@ammarash5449]

These prompt injection things were the one that I most enjoyed so far felt like people doing Terminal Injections in BIOS era again... Terminal injections could be seen in this DEFCON talk: kzbin.info/www/bejne/j2WkaH6DhqqgnNE

[@ArchaicDemise-ex1lq]

@9:30 - a bit of a correction from someone who is also a bit uneducated in this field :P TL;DR; - I think it's way cheaper to do targeted training than you implied, that's it, and here's my reference. LLM's are great because they can be refined, I've been doing custom training llama models under contract - in order to summarize text into related tags/keywords, so you give it a paragraph of text and it will dump a bunch of tags some of them are bespoke tags that are distantly related, and it's working well for us and our use case, imagine a topic of discussion around comp-sci topic like gray codes - and it suggests that there's a related tags for ECC, ECOC, QECC and distantly related topics like QAM decoding and other encoding concepts, like Manchester etc... it's kinda cool in that we trained it with a lot of tagging as concepts and it was able to do the task as an emergent process of it's own knowledge a similar example like the above emerged when testing. We were able to get good output in about 48 hours of training.

[@InternetOfBugs]

Maybe it's a matter of degrees? I've been playing with training SDXL LoRAs in less than 30 minutes with current-gen hardware. A 48-hour cycle time seems expensive to me (based on the amount of flexibility and number of iterations I've needed to do on traditionally developed software at startups when we had to get something fixed or we'd go out of business). I appreciate the correction, though. I'll do some more digging.

[@dannynguyen7076]

Best channel I've found in a while.

[@afjelidfjssaf]

Damn very well said

[@AEblahblahh]

I rarely if ever leave comments. But thank you for this video. You made some really good points I didn't consider.

[@scottbuffington5964]

I really enjoyed the last three minutes best because it relates to my thoughts and gives me ideas on how to set up a model in a way that is secure. Speed and time will be the next question I suppose =\

[@albinoadrianocordeiro9929]

LLM inference code is just dozens of lines of code. Matrix multiplication basically. Machine learning is not complex software at all. There won't be a bugs in it. You write the program to do something specific by providing "specific" data. So the bugs will be on the data selection for training the LLM. If the LLM does something wrong, you fix it by retraining it with data samples that exemplifies the right thing to do .

[@InternetOfBugs]

If you limit "system" to just the LLM piece. I was thinking more in terms of things like Devin where there is obviously a bunch of other stuff wrapped around the query function so it can manage a terminal and a browser and a planing window, etc. Or the prompt injection (I've heard) stuff they added to Gemini that made it fabricate ethnic characteristics of historical figures. I should have been more clear. As for fixing it by retraining, that's a whole different problem. Expensive, time consuming, and you risk undoing some gain elsewhere in the network.

[@albinoadrianocordeiro9929]

@@InternetOfBugs , I see. Makes sense.

[@fahomenhera2820]

11:45 That’s right. Many this big tech companies talking about AI as a such godly ”being” but can’t even create properly working Windows or Android OS. Every software comes with many bugs and issues.

[@rujotheone]

I believe customizing AIs is called finetuning.

[@user-vi5tp4xd8i]

You just won a new subscriber. I think your analysis is spot on. I'm speaking as a 42 year old Site Reliability Engineer.

[@InternetOfBugs]

Thanks

[@codetour]

Really good stuff. Commenting for the algo. Hope you do more

[@InternetOfBugs]

Thanks. I'm working on more. (Well, at this exact moment, I'm procrastinating by reading KZbin comments, but I'll get back to it soon).

[@codetour]

thanks, take it from me, when it comes to creating new videos, take your time, there is no rush (even though it might feel like there is)@@InternetOfBugs

[@operandexpanse]

It's refreshing to hear an experienced voice in this space that isn't on board with the AI hype, and on the flip side, isn't saying AI wont replace devs without giving really solid reasons. I've seen a lot from both sides, and it's been really frustrating. The most annoying are the KZbinrs with clickbaity thumbnails claiming X company has already achieved AGI secretly. It's a confusing and terrifying space to wade through, with misinformation everywhere.

[@InternetOfBugs]

Yeah - I get that. This is my third go-round. There was an "AI Revolution" at MIT in the early 1970s and an "Expert Systems" based AI bubble in the late 80s. Honestly, I think this would would have happened sooner if the investors hadn't been spending all their effort chasing NFTs the last few years. As soon as FTX exposed that crap for what it was, and they needed some place else to try to get rich quick - boom - the AI hype was born. The thing to keep in mind about "achieved AGI secretly" is there's no definition of AGI. Twenty years ago, if you'd asked a ML expert what AGI would look like, they would likely have given you criteria that ChatGPT 4 currently exceeds. But now that we see that, it's clear that the criteria we had for defining AGI years ago was woefully inadequate.

[@user-mx6yt3wg6t]

Hey, I think you are right, and something that come to my mind is, isn't it very good time to start freelancing? I think it makes sense, becouse LLM (and for example Copilot, I am not talking about this Devin thing) makes you AS A DEVELOPER more productive. And that means you can write more code and write it faster. That means, you could eventually build larger projects and potential clients won't need to hire a team of proffesionals, they can just hire you. The only problem I see, is then you are not competent, and you rely only on these tools, you could end up with very messy code, unscalable code, and that means there could be bugs you won't be able to fix, and you end up needing help from someone more experienced. What do you and your community thinks about that?

[@InternetOfBugs]

I've done a lot of freelancing (I'm semi-retired and doing some small freelancing on the side now, even). Freelancing isn't about code. It's about selling yourself and finding customers. (Gig work like Upwork and Fiverr are impossible to make a decent living with if you live in an expensive area - you need real businesses to pay you). If you can handle the marketing/prospecting/sales part, it's a great time. But that part is really hard, especially for introverts.

[@user-mx6yt3wg6t]

@@InternetOfBugs Thanks for the advice!

[@user-mx6yt3wg6t]

@@InternetOfBugs Do you have any advice then on searching for potential clients? What are your observations on freelance topic?

[@InternetOfBugs]

Freelancing is generally very dependent on location, and your connections in that location. By virtue of having been around a long time, and worked a lot of places, I know a lot of people, and that helps. If you try to freelance outside your geographic area, you're competing with every person with an Internet connection, and it's REALLY hard to get noticed in that field.

[@SR-cm2my]

As an engineer with a third of your experience, I fully concur with your assessment surrounding the AI/ML buzz, hype cycle. However, I'm really thankful at how trivial it has gotten to implement systems such as search. I recently rolled out a multi-modal search feature within our company within a matter of days. LLMs are truly making things "easier" for engineers. People with knowledge in customizing elasticsearch instances might catch up on what I'm talking about. Remember having to feed synonym files, token filters etc? Thanks to LLMs and vectorizing inputs, this work need not be done at least for non-scientific usecases. I fear multi-modal LLMs will make implementing *complicated* things much more easier and faster, it won't require as many engineers as before. Of course, you still need an engineer to write code and architect micro-services to make all of this work. But the fact that the most complicated things like "Search" are becoming trivial to implement is something I'm looking forward to.

[@anuragdce1]

I am impressed by how holistic and balanced your thoughts are! It's refreshing to see content that's not trying to cash in on the latest hype. I learnt a lot from your more recent video on LLMs too. _hits the subscribe button_

[@Metarig]

I dabble in some programming for my 3D artist job, but I'm definitely not a programmer. I get why people are worried about AI, but honestly, we can't predict the future. So, the smart move might be to think twice before jumping into programming if you're just starting out, because of the risks. But if your entire business and income already rely on it, it might not be worth it to switch gears completely.

[@PRIMARYATIAS]

If anything 3D artists will be far more threatened by AI than programmers, Just saying… But still we are far from getting there. Art is a subset of Math, and AI replacing programmers means 3D artists will be replaced much much earlier. (Just for context: programs like say Houdini are made of whole lots of Math and Programming, So by saying AI is soon to replace programmers and mathematicians it is saying 3D artists will be replaced much much earlier).

[@Metarig]

@@PRIMARYATIAS Actually, becoming a game developer might be one of the last jobs AI will take over, because AI isn't ready to handle real-time 3D graphics on its own just yet. Instead, AI is here to assist game developers and others in the field with tasks that were previously too complex for one person, like creating animations, modeling, and debugging scripts. While this means that roles like modelers, animators, programmers, storytellers, and concept artists might see some changes, it also creates new opportunities for those who learn how to use AI in game development. But even with the possibility of AI changing the job landscape, I'm not worried about losing my job to it. I won't just sit back and let AI take over; I certainly won't mislead others with false hope or advise anyone to dive into the industry without caution. Instead, I plan to move to a more secure job, save some money, and maybe even start a new business.

[@PRIMARYATIAS]

@@MetarigI agree, Also plumbing is the hardest job to get replaced by machines. Take for example a person who makes cakes, It is not really that hard to come with a design of a machine to do that job (for a specific kind of cake), But with plumbing, Everything is just so unexpected. It is always good to be adaptive to job market. And, Yes, I think (professional) GameDev is also one of the hardest jobs to get automated, There is a performance consideration involved, It is not just 2D pretty pictures that SD can make.

[@Metarig]

@@PRIMARYATIAS Sure, jobs that involve hard physical work and don't pay much might be some of the last to get taken over by robots. But there are definitely jobs that no robot or AI could ever do. We're talking about jobs where it's not just about the work you do, but also the big responsibilities you carry, like being a doctor, a lawyer, or a surgeon. These jobs need a human touch because if something goes wrong, people need someone they can hold accountable, and you can't really take a robot to court.

[@PRIMARYATIAS]

@@MetarigJust saying, But independent plumbers are making a lot of money (I know of someone), As long as you are independent, This is where all the real money comes in, But you have to be a professional, Same can be said about programming, 3D art, etc.😊

[@PranavJ-pz6bo]

Sir, I am an high school student(16 yo) and want to break into tech by maybe 2030. Do you think it will be viable career choice to do CS given that I have serious passion towards it. Do you think that my views are right and more opportunities will be created in tech ? Or will there be a reduction in workforce and salary if not by 100% but 50 % or something like that

[@kyleyoung8179]

If you are 16 you can break into tech a lot sooner than 2030. Im 19 and have my second internship coming up.

[@PranavJ-pz6bo]

​​@kyleyoung8179 Yes bro I know I will be doing internships way earlier than 2030. But I just mentioned my graduation year where I will be qualified for full time jobs. BTW what do you think about the job landscape by then

[@kyleyoung8179]

@@PranavJ-pz6bo im not really sure what the job landscape will be like by then, but getting an internship somewhere will put you in a good spot post graduation

[@PranavJ-pz6bo]

​@@kyleyoung8179Well I mostly think AI may not replace human workforce but just wanted to hear it from him. Infact my views align with him as well. I guess Humans will be reallocated to work on more creative and intellectual projects while the rest of them will be assisting AI to work on existing codebases. Also my stand is if anyone can use AI to create enterprise apps then there won't be any companies existing as everyone will create their own software solutions.

[@InternetOfBugs]

As I explained in this video: kzbin.info/www/bejne/bV7JaKWGmc9pqqs Enterprise apps (as well as customer-facing apps that aren't just copies of existing apps) are way more about the people than the code. At the point where LLMs can generate Enterprise Apps, LLMs would also be capable of doing the work of any Corporate Middle Manager. And I'm honestly not sure what that world will look like (but I don't think it'll happen for decades, though I could be wrong). That said - 2030 is a long time giving how fast things are moving. I would bet that Software Engineering would still be a viable career in 2030, but trying to get an entry-level position at that point might be quite difficult. It's really hard to guess at this point.

[@fozbstudios]

Yes I would like more info on comfy ui

[@fazdatasciencetechcodingml5662]

Great Video

[@InternetOfBugs]

Thanks!

[@techmansion8618]

Sir, i need really need your thoughts on this... I am learning flutter app development should i continue with it .... I'm going to complete my 4 year's Computer Science degree in 7 months... Or should I go for a start-up in future related to mobile development.... Or any other advice u can give me related to development filed...😊

[@InternetOfBugs]

The risk with Flutter is that Google has a long, long, long track record of coming up with (or acquiring) stuff, and then abandoning it, and having it fail. (See killedbygoogle.com/ ). Unless Flutter/Dart gets far, far more widely adopted in the rest of the big companies in the industry (the ones with enough free cash flow to support the extended core team) than it currently is, it will only be viable as long as Google continues to promote it. I can't predict what Google (Sorry, Alphabet) will do. I don't think they even know. So it's a risk. How much do you trust that part of Google?

[@techmansion8618]

@@InternetOfBugs so please guide me i want to do mobile app development what language i should pick as i was interested in flutter

[@techmansion8618]

@@InternetOfBugs or can you make a video on mobile app development

[@nirajsbajaj]

Thanks!

[@InternetOfBugs]

No problem!

[@rustacean10]

"They don't think actually" gave me an inner peace

[@joecochran4941]

My man

[@richarddisney1664]

Thanks

[@trenfa4371]

Maybe virtual reality will require more software engineer who will use Ai for assistance...may be

[@alexandersheiman4220]

If AI makes a software developer 5x more productive then what makes the most sense is to hire more software developers and just try to do everything. Like you look at what Meta did with the metaverse and they lost like $45b which is a massive flop but if their devs were 5x more productive and all of that dev effort was only $10-15b suddenly that's no longer all that crazy of an investment. Not only could they have taken that sort of massive leap into the unknown, they could've run all of their existing services better and branched into even more stuff and likely wouldn't have even noticed. I think overall developers being significantly more productive drives the cost of operations down a lot for tech companies which then means actually companies want to significantly expand the scope of what they're doing because they only need a fraction of the revenue for tacking on new business areas to make sense.

[@InternetOfBugs]

If only all the investors/markets thought that way, as well. Honestly not sure what's going to happen. Hope you're right.

[@galatei11]

I wish more people had common sense like you do. Many startups don't have proper knowledge and they go for AI to replace people's jobs and get burned too.

[@InternetOfBugs]

Thanks

[@ttcc5273]

If they develop an AI that rewrites itself… then that’s the time to worry

[@darkzeroprojects4245]

They will do it if we allow it. It ain't a inevitable March.

[@terry-]

Great!

[@draco4717]

What if llms start promoting other llms and start building some kind o f skynet that would be fun to watch. Is this evn possible?

[@ArchaicDemise-ex1lq]

@18:00 I've always considered an AI as a shotgun to be more dangerous than skynet. Literally using the AI as a tool/weapon to advantage yourself is a bit scary to me. (imagine a somewhat distant AI future) Joe Bob is a low level sales rep at CogBuilders Inc. and starts using a custom black-market AI to take in all the data from around him in his office, phone calls, conversations people, it's internet connected etc... and then our buddy Joe tells the AI that he wants to be the top performer for this quarter. And it goes about a plan to improve Joe, but as time goes on it starts asking for odder and odder requests like... "move the water dispenser in the breakroom 1ft to the left, at 2pm" "Remove the stapler in the copy room and place it in the breakroom" etc... like odd little things. Joe complies, and months later he is the top performing Sales rep in a now failing company, because the AI was unable to improve him, and instead did whatever it would take to cause disorder for everyone else.

[@magfal]

I wonder how many years there will be before an LLM reaches the point where I can get a 15% net benefit in productivity. It'll need to get at least 10 times better.

[@chrisf4268]

Your analysis is more emotionally driven than hard data driven.

[@InternetOfBugs]

Do you have hard data on what's going to happen in the future? 'cause I don't. I just have educated guesses based on what I've seen happen in the past.

[@chrisf4268]

@@InternetOfBugs you have no analog to base your “educated guess” on. You don’t need to get your panties in a bunch just because I asked you to justify your assertion. I’m not the one making the claim. 😁

[@joshlamingo1145]

My takeaway from this is that LLMs are a bad idea for software engineering.

[@andrewyork3869]

19:11 to add to this we don't know and don't have a good way of ever knowing why certain things happen in the hidden layer of a neural network.

[@InternetOfBugs]

Yeah. I'm so much more scared of that than SkyNet. Imagine a terrorist (or jilted sociopathic ex-husband with a grudge) with an exploit for the image-recognition LLMs in the landing software of airliners.

[@andrewyork3869]

@InternetOfBugs did you hear that someone tried to hack and take over an airliner a few months back? They only got into the notams, and there was no further issue. The fact that it happened at all is shocking, though.

[@InternetOfBugs]

No I didn't hear about that.

[@DEBO5]

This AI wave is only creating more demand for web applications that incorporate LLMs and other AI architectures

[@Silech1]

As someone who just started full-stack course, the last week has been scary with AI news on developers. This video came right on time. Thank you sir for this video. It’s very encouraging.

[@guizaoacc]

Y Combinator’s current batch have 80% of startups related with AI. Crazy! A lot of it is just hype and can easily be wiped out by incumbents. I agree that in the long term the market will grow! That is empirical economic data. History + Numbers are the biggest teachers.

[@InternetOfBugs]

I wonder what the percentage of them were BlockChain/NFT startups before FTX went bust. Wouldn't surprise me if it was similar.

[@waynelast1685]

16:04 it’s also called “creeping elegance” ?