[warrioraks]

Originally Posted by Tensor

My logical and educated guess in this whole GenAI and software engineers loosing job.
...
All things considered it likely takes 36-48 months more for most of the entry and mid level software devs to become redundant and makes AI more cheaper and more intelligent, most of tier 2,3 B.Tech CSE grads will have a hard time getting jobs and potentially their degree might even be useless since software engineering as a whole gets commoditised to everyone and with a few prompts building large scale apps will be available to all.
...
Obviously there won’t be a sudden layoff situation where everyone gets laid off simultaneously rather it’ll be a slow decline with less hiring and more periodic firing of people in chunks over the decade, the current trend of slump in IT jobs would very likely continue.

Great assessment. I agree with nearly everything thing you mentioned including the timeline forecasts. Your scenario of of reasoning & planning through one model and actual programming through another model is addressed through AI agent frameworks. These are in very nascent stages and will take 3-5 years to achieve the stability required to make a tangible impact. Most average software developers of today will resemble the factory workers in a highly automed manufacturing plant. We can expect unicorns which are way leaner than what we have today. AI will be a significant productivity boost to software engineering teams.

[Tensor]

Originally Posted by warrioraks

Great assessment. I agree with nearly everything thing you mentioned including the timeline forecasts. Your scenario of of reasoning & planning through one model and actual programming through another model is addressed through AI agent frameworks. These are in very nascent stages and will take 3-5 years to achieve the stability required to make a tangible impact. Most average software developers of today will resemble the factory workers in a highly automed manufacturing plant. We can expect unicorns which are way leaner than what we have today. AI will be a significant productivity boost to software engineering teams.

Yep only a few people actually understand the amount of development that’s happening in background, ask any software engineer about this and almost 90% of them get triggered and will start to cope and say it’s impossible to replace them and they have xyz skills but what they don’t realise is how dumb current LLMs are compared to the future LLMs, I’m pretty confident that even the current gen LLMs like sonnet 3.5 will steamroll any software engineer with less than 10 yoe in terms of architecture design, scaling methodologies, system design or even just pure coding.

The amount of errors generated by these LLMs are almost negligible right now compared to 20 months back.

The o1 model is such a beast in reasoning, most people in the thread claiming it’s about understanding client requirements etc just use o1 for an hour and I’m pretty sure you’ll start seeing the future of LLMs in terms of thinking and reasoning.

Here o1 is solving unseen PhD problems that are not even there on the internet in 2 minutes which otherwise takes 10 days for any smartest human to solve.

https://youtu.be/scOb0XCkWho?si=mN3htSZVROu7M4U0

https://youtu.be/a8QvnIAGjPA?si=JIf-fda01sMIJ7u-

How long until it builds 100% accurate software and not 99% as people think?
(My two cents here: Use cursor as IDE and hook up you’re anthropic/openai API or use pro plan and start coding and then you’ll start realising 99% of whatever project you build, English will be just be enough to make it a reality use replit for easy hosting and cloud flare for free CDN)

And there are still people saying AI won’t reduce the number of software engineers (delusion is literally peaking)

All I’m saying is Downsizing of teams if you’re team had 12 members it’ll reduce to 5 and remaining 7 will be redundant with no where to go even with whatever skills they have it won’t matter because someone with similar skill set will work for cheaper it’s as simple as that. I agree you’ll up skill but so does every single one of your peers and some of them don’t mind working 70+ hours a week with reduced pay.

The software industry will now ship a ton of more products, unprecedented amounts of software will be shipped compared to before. It’s now not who builds apps or better software it’s just who can get users and retain them which in turn will kill a lot of startup’s since most of them make fancy MVPs and run behind valuations without any meaningful revenue.

TLDR; No the software industry won’t collapse, just people working there will reduce by 40-50% in like 4-5 years, and No the economy won’t collapse just because IT employees are reduced there are a ton more occupations/business that spend more in an economy.

Let’s be grateful for having a magic intelligence in the sky that enables us to solve meaningful problems through our phones in a few minutes.

What a time to be alive.

[Turbohead]

Originally Posted by Tensor

TLDR; No the software industry won’t collapse, just people working there will reduce by 40-50% in like 4-5 years, and No the economy won’t collapse just because IT employees are reduced there are a ton more occupations/business that spend more in an economy.

Let’s be grateful for having a magic intelligence in the sky that enables us to solve meaningful problems through our phones in a few minutes.

What a time to be alive.

I don't think you understand the implications and effects of what you're saying. Reducing 40% of workforce in an industry of a country like India where the service industry is one of the top employers with relatively higher wages will cause massive economic downturn. I'm making up numbers but I'm guessing atleast 20% of income tax would be collected from individuals working in the service industry. It would affect so many industries not limited to real estate, consumer goods and even the food and beverage industry.

One of the videos shared above is about o1 trying to solve a physics problem which suggests that LLMs will affect not just IT/ software development but many complex professions.

Which also makes me think, if everyone is a superhero then no one is a superhero. Wouldn't the competition rise heavily to provide services and experiences that is better than the competition and what is already generated by LLMs? So will the number of jobs really reduce? I think it's still debatable. The advent of computers have only brought more work.

Let me ask one more thing. How much is this thing going to cost? It won't be cheap for long right? All these are introductory prices right? How is it going to be priced in the future since every answer generated produces varying degree of value?

You can say I'm coping but if the real cost of generating such answers is very expensive and exceeds human costs, I don't think it's viable.

[Jeroen]

Originally Posted by Tensor

.

Would love to hear from more experienced folks serving in product and service sectors what they think about this

Thanks for sharing your thoughts. Very interesting. I really hope you notice of Samurai’s comment of what is lacking in your analysis and prediction, the customer view.

I would like to expand on that a bit. Technology development never happens in a vacuum. It is part of a much larger ecosystem, our whole society, globally. Yours is a prediction based on expected technological advancement. Which is always an interesting exercise. However, history has shown that it rarely is an accurate way to predict the future.

In the end, technology must serve some (business) purpose. Without oversimplifying your predictions it seems to be mostly about cost/efficiency. Very few businesses are built on just being the cheapest.

I think you need to take into account a number of other elements, all from a global perspective, rather than just from an Indian perspective.

- as per Samurai, the customer perspective
- what market can be addressed, what size, who are or will be the players doing what
- legislation (current and to be)
- political interference (usually leading to legislation
- general sentiment in markets to AI from the public at large.

Personally, my experience with AI has been on various applications in the Telecom space. We use(d) AI/ML extensively in managing (mobile) networks. Although cost/efficiency was a factor, we mostly focused on features that would be able to bring extra value for our customers. So increasing uptime of the network. It’s much more interesting selling high value than low cost stuff.

Also, AI let us do stuff that is simply beyond the capabilities of humans. Again, that would be a value, rather than a cost related approach.

I have said it before, I think India needs to move away from being and producing low cost services/products and focus more on value.

Jeroen

[warrioraks]

Originally Posted by Turbohead

I don't think you understand the implications and effects of what you're saying. Reducing 40% of workforce in an industry of a country like India where the service industry is one of the top employers with relatively higher wages will cause massive economic downturn. I'm making up numbers but I'm guessing atleast 20% of income tax would be collected from individuals working in the service industry. It would affect so many industries not limited to real estate, consumer goods and even the food and beverage industry.

...

Let me ask one more thing. How much is this thing going to cost? It won't be cheap for long right? All these are introductory prices right? How is it going to be priced in the future since every answer generated produces varying degree of value?

You can say I'm coping but if the real cost of generating such answers is very expensive and exceeds human costs, I don't think it's viable.

40% reduction in workforce in a sector is not unheard of. This is what happens when bubbles burst and massive recessions kick-in. Having said that, I do not believe that we are at such a dooms-day scenario yet. If AI evolves as good as what I believe, this will lead to a massive transformation in how we operate as humanity. Just like the internet of 2000s - there will be a host lot of new business and jobs that we cannot imagine at this point. Technology has been a net job creator over last 3 decades and will continue to be in the long term future as well.

On the commercial viability aspect of AI, human/labor is one of the biggest costs tech organizations are struggling with today. Covid led to high salaries all around and there are layoffs left, right center because of this over-hiring and overpaying. On the other hand, the beauty with tech is that the cost almost always goes down with time. And this will be the case with AI models and hardware as well. Infact, I have had so many people come up to me and express surprise on how cheap these pioneer models are for the right use-case.

I also like to think of it this way. The current automotive eco-system would not have been able to build cars and vehicles at scale if humans were still handmaking it. The industrial revolution and factory automation brought in an unheard level of productivity and quality to manufacturing output. Despite all this automation, we still have millions of folks employed in the auto sector. The same is going to happen in software sector with the advent of AI agents.

[PreludeSH]

Similar to software, are there any such tools for hardware ? Something that knows about every kind of hardware chip that is available, data sheets etc and can come up with a PCB design meeting the requirements ?

Talking about TCS reducing its staff, why would you need a company like TCS for services if companies can solve such issues themselves using AI ? Or even more very less number of such companies themselves are needed.

[Tensor]

Originally Posted by Turbohead

I don't think you understand the implications and effects of what you're saying. Reducing 40% of workforce.

Well it’s already happening in your IT industry just see the hiring scenario in your own company or among your peers, try switching to a different company with higher package like the past few years, Many people working are periodically being weeded out from the company (or I must say forced resignation and not outright firing that’s why there’s lesser news coverage) all these are ground realities currently in IT sector and not some imagination of the future(be it service or product)

Coming to service specific data:
TCS has slowed down its hiring for the first time they did not even slow down 2008 during the financial crisis or Covid.

Originally Posted by Turbohead

One of the videos shared above is about o1 trying to solve a physics problem which suggests that LLMs will affect not just IT/ software development but many complex professions.

Yes not just IT, o1 is steamrolling doctors, lawyers essentially every white collar job which involves cognitive abilities.
Here’s the data to back it :





But guess what IT employees are the lowest hanging fruits to get reduced,No Government will let AI to replace Doctors be it partially or completely in the next 5 years or lawyers in court.

But Hey in our country there is no labour laws forget a law to protect IT jobs, employees often end up working 10 hours on an average per day.

Originally Posted by Turbohead

Which also makes me think, if everyone is a superhero then no one is a superhero. Wouldn't the competition rise heavily to provide services and experiences that is better than the competition and what is already generated by LLMs? So will the number of jobs really reduce?

Exactly correct, building software now is the easiest a fresh graduate with SOTA LLM can be at par with any mid level IT employee right now. You observed a perfect pattern if everyone can build software then the inherent use of software engineers as a whole plummets and value of them reduces hence lower pay and longer working hours.

Originally Posted by Turbohead

Let me ask one more thing. How much is this thing going to cost? It won't be cheap for long right? All these are introductory prices right? How is it going to be priced in the future since every answer generated produces varying degree of value?
You can say I'm coping but if the real cost of generating such answers is very expensive and exceeds human costs, I don't think it's viable.

It’s not like car manufacturers where they have introductory pricing and jack up with time in tech it doesn’t work like that it’s the complete opposite take any tech product it always has become more cheaper and better as time progressed.

Again I can’t keep answering to these questions please refer to my post here I have explained in detail how efficiency increases with scale and the current state of chips running the LLM even if you don’t agree with my timeline it’s worth a read.

https://www.team-bhp.com/forum/shift...ml#post5841641 (Is the golden era of the software engineer over?)

Also by being an IT professional I assume you must be very well informed on how fab designs get smaller with lesser power draw and more performance per watt.

Originally Posted by Jeroen

In the end, technology must serve some (business) purpose. Without oversimplifying your predictions it seems to be mostly about cost/efficiency. Very few businesses are built on just being the cheapest.

It’s not just cost reduction it’s a small part the main thing is the accuracy, use sonnet and you’ll quickly realise how shockingly accurate it gets your query and does your task its rate of error is way less than most people think (most people in Indian tech space used gpt3.5 a year ago and thought its bad and won’t improve further) and it’s now for free just see where things go in the next 20 months.

[Tensor]

Originally Posted by PreludeSH

Similar to software, are there any such tools for hardware ? Something that knows about every kind of hardware chip that is available, data sheets etc and can come up with a PCB design meeting the requirements ?

Talking about TCS reducing its staff, why would you need a company like TCS for services if companies can solve such issues themselves using AI ? Or even more very less number of such companies themselves are needed.

Yep already businesses have started to ditch SaaS companies who used to sell software apparently it’s cheaper and better to build in house software as per need instead of out sourcing and having to pay recurring costs. Now it’s just the start expect more businesses to join this trend of in house software development

And coming to your initial question:
Yes, although I’m not very knowledgeable in hardware space but do checkout
NexusAI from Altium , PADS from Siemens, Ansys also has some tool.

[rkv_2401]

Originally Posted by Tensor

no LLM could ever do it before this quickly and not just that creating fully functioning clones of WhatsApp, YouTube etc are a breeze and can be made in 60 mins (I’m taking full stack/ fully functioning applications not a dummy front end copy)

Do you think making a MERN stack front-end, back end and database and hooking it all up together is all it takes to make a "fully-functioning clone" of WhatsApp or YouTube?

YouTube's bandwidth per second without extremely efficient video compression would be about 130x times the Global Internet Bandwidth - a convenient source for those who are looking to learn:

https://www.youtube.com/watch?v=swoIPD7EvEw&t=58s

Originally Posted by Tensor

My logical and educated guess in this whole GenAI and software engineers loosing job.

Unfortunately, most of your takes here don't come across as either logical or educated. Let me illustrate.

Originally Posted by Tensor

The LLM must be able to read into and write into large code bases like really large ones since most of the time we’ll be just tweaking/refactoring existing code bases and not writing from scratch for we need really large context windows like I’m taking 2M +

And what will you do if your magical tool in the sky hallucinates a grave error into the 2M+ token context window that it doesn't tell you is an error (because it thinks it did the right thing) and you don't notice because no human can sit through and read 2M+ words + you don't know enough about the codebase cuz you let the AI do your job for you? What'll you do then, get the AI to write test cases to prevent this for you as well?

Originally Posted by Tensor

Attachment 2654597

Ah yes, the mastery of solving cherry-picked CodeForces problems from old contests that have already been solved and the solutions to which are readily available: https://codeforces.com/blog/entry/211

In case you didn't know, GPT4 that released 1.5 years ago was already at the level of solving at the 85th-90th percentile of LeetCode problems with some clever prompting - refer "Arxiv: A systematic evaluation of large language models for generating programming code". But I have never seen it or Claude Opus solve a problem that I couldn't (easily) search for the solutions before.

Secondly we need AI to no just spit code out instantly we need it to architect the software first and then logically think further steps to take like which platform to use what programming paradigm and weather to containarize it or not etc. which is again being solved by another LLM o1 by OpenAI but the issue is it’s context window is 128k as of now which is a limiting factor.

Why do you need AI to tell you whether to "containarize" your code or not? And assuming this is the standard of development performance you are held to, would you even be able to tell whether the "architect"ed design is good or not?

Take a piece of paper draw a website layout it gives a fully functioning front end code which looks exactly like you wanted (obviously you must tweak some things which can be done in 1-2 shots later). These things were never before possible.

NeetCode dismissed this claim with his rebuild-NeetCode-challenge. Maybe o1 can do better with the reported $2000 a month plan.

Attachment 2654598

I can't decipher what the point you are trying to make with this chart. But even if I did know, I'd probably question the usage of this particular chart - the models are and have been trained on these benchmarks forever now. If your point is only to say that Claude 3.5 Sonnet is better than GPT 4o, I agree.

Some argue privacy of proprietary code well usage of any leading LLMs via enterprise plans or APIs won’t be used as training data or just use open source LLM like llama 3.1 405b and fine tune with codebase and run locally without internet

Oh my god. How do you plan to run a 405 billion parameter model LOCALLY without Internet? Do you even know how much VRAM it would take to run the 70b models, let alone FOUR HUNDRED AND FIVE billion parameters? And what kind of inference speed would you be looking at?

Why would you even do all of this? They have enterprise plans with the promise to not use your data for training. Some of the biggest companies in the world subscribe to such plans.

Attachment 2654599
look up hopper vs Blackwell and you can see Blackwell is 2000% more faster (most LLMs were trained on hopper H100s). Fab designs are still expected to have exponential improvements may still follow moore’s law.

You're probably also surprised at the shocking EXPONENTIAL performance gains of modern NPUs which can do 40+TOPS on the CPU/on an iPhone, aren't you?

Did you see that little "FP4" unit at the bottom of the Blackwell quoted figure? Did you notice that it's not the same unit of measurement as the Hopper and Ampere figures? Where did you get that 2000% figure from? Did you divide (20000 - 4000)/4, and then assume that 2 FP4 calculations = 1 FP8 calculation?

I'm not going to teach you a hardware design lesson because I'm not an expert in the field and I didn't study it either, but let me just say that that is a big, no, huge assumption to make. You also didn't check whether the FP4 and FP8 numbers are both sparse-to-dense or dense-to-spare matrix calculations. You didn't conduct an analysis of or study the upcoming chip design nearly as well as you need to to make a claim like that.

For the record - FP4 numbers are up 355% product-level to product-level and 254% normalized to power draw when comparing dense-to-dense matrix calculations, and FP16 numbers are up an impressive 153% - impressive, but nowhere near 2000% https://www.semianalysis.com/p/nvidi...f-tco-analysis

I'd also like to ask you if you think:
i) You can train and infer models like Claude Opus/GPT4 at the same level of accuracy they currently achieve running at 4 bit quantization (hint - they are currently trained and likely inferred at 16/32-bit precision: https://arxiv.org/pdf/2405.18710)
ii) You really thought 2000% performance improvement in hardware is not only possible from generation to generation, but is expected because of "fab designs that will have exponential improvements"?? I'd love to hear your take on what this "fab design" is - is fabrication yield an exponentially increasing factor, or did you mean chip design trends to exponential yields?

The more likely scenario is that Hopper came out just before/during the start of the current AI boom cycle and nVidia, like everyone else, dedicated more die space to the relevant floating point and maybe also Tensor compute units. Nothing about that improvement indicates that it is either possible to replicate or is the expectation for future generations of chips.

Add to that in less than 20 months we have sonnet 3.5 from ChatGPT (gpt 3.5 the dumb one) sonnet 3.5 steamrolls any college graduate with computer science degree who either spend the 4 years of college rote learning DSA or hop on the MERN stack trend

You seem to not understand a lot about the field. There is nothing special in building wrappers on top of GPT2 without understanding much about how the underlying neural networks work. You might end up being closer to this slightly-worse-than-mediocre college graduate than you think you are. FYI, neither the MERN stack nor DSA (by which you mean practicing LeetCode) takes 4 years to learn.

Anyone who is delusional that AI can’t replace them are just coping and won’t help them teams will shrink 5 people can achieve the work of 15 easily and those 10 gets displaced moreover most devs in India in service sector are not highly skilled.

And why should we take your opinion seriously, considering the level of knowledge and expertise in software development you've demonstrated with this post?

Software production will have massive boom with less people pumping out more software

This is the only thing in this post I agree with.

If I’m wrong please educate me with technical and logical arguments.

It's fine to not know things, look, if you can get the job done and go live the rest of your life, that's fine. But just like the AI FUD-ders before you, you have done a stellar job promoting misinformation and sometimes, sheer uneducated claims. I am biased, but not in the way you'd think - I am tired of misinformation and hype, therefore I am biased against that.

I actually use AI myself and have posted about it earlier in this thread, and there are things about o1 that I appreciate, but the hype and lack of critical thinking demonstrated in this particular cycle has been infuriating.

[acedriver_9000]

upvote for Tensor

Downvote for rkv_2401. You seem to be too angry at Tensor

Points raised by tensor look valid & convincing in my opinion. Came across people in govt offices say chat-gpt helps them in moving the files fast (atleast honest guys). Helps a lot with something like, say, legal work. I have experience of using it for coding, and helps well in making things faster.

[akhil_007]

All these chatgpt,Gemini,LLM's could be a "force multiplier" for a software engineer(broadly speaking) leading to faster development.

Whereas for a Business, it can bring in extra value. It may be in terms of
-helping make business decisions easier (read as: pointers/parameters that might help take a decision. Not the decision itself
-increase operational efficiency
-help with faster growth

A well known firm has been creating AI models for its legal research and the team developing this model is continuously growing. Various organizations that uses this model have also increased their team size. Reason being more output,faster and accurate results

I think there will be more demand for software engineers once AI/ML/LLM's are adopted widely

[PreludeSH]

Originally Posted by akhil_007

A well known firm has been creating AI models for its legal research and the team developing this model is continuously growing. Various organizations that uses this model have also increased their team size. Reason being more output,faster and accurate results

I think there will be more demand for software engineers once AI/ML/LLM's are adopted widely

But it will also result in many firms going out of business. You cannot keep increasing output over demand. This is how we have seen new companies growing huge and older companies shutting down.

[mayukh42]

Originally Posted by akhil_007

I think there will be more demand for software engineers once AI/ML/LLM's are adopted widely

This is what happened in the past - when certain technologies become mainstream, they create new jobs that were unheard of. All it needs is for the individual to be flexible enough to learn and adapt if required, since the existing profiles can be achieved with less workforce.

Rather than a smarter job, could also be a competition - take the example of an earlier post where 7 out of 12 jobs become redundant. Now if 5 people can manage the product with AI help, what is stopping the remaining 7 to start a competition? So net jobs might not reduce, and with addition of new possibilities, only increase the count in the long term.

Only thing is that the median job profile changes. True to that, we can now achieve a lot more (even before the AI tools) individually, than the days of Borland or Turbo C++ (here just an IDE example, but the idea is same).

There are many unsolved problems in the tech industry which can actually help us if efficient solutions are found. We can also do well with competition rather than monopoly. If the overall cost to create software is reduced due to AI, the industry can get investments from smaller VCs who are waiting on sidelines.

Say, many hotels are forced to use Cleartrip or MakemyTrip for their online bookings, because they don't want to invest "so much" in having an own website with booking and payment facility. With reduced cost, each hotel can have their own efficient solution and we can get better prices. At a larger scale, Walmart can come up as a worthy competitor to Amazon in the e-commerce space, if the barrier to enter is reduced enough. Again, none of these are final solutions, instead we will always be in a cycle of innovation/ efficiency.

[beyondPika]

Okay guys, I would like you to think about it this way:
Assuming usage of AI results in 10X multiplication of work, so let's say a developer does x amount of work without usage of AI, they will be able to complete 10x amount of work with AI.

Now let's take 2 companies both with 100 developers doing 100 units of work, they discovered AI
Company A : Thinks that since 10 developers will be able to sustain 100 units of work, they fire the other 90 developers thereby maintaining 100 units work output.

Company B : Thinks that if they give all 100 developers AI, they will be able to work and generate 1,000 units of work. Meaning more features, more and better products and lesser product cost.

So company A competes with 100 units, company B brings 1000 units.
Which one do you think will be able to survive long term in the market?

Being a software developer, I really think of the age of AI to be as important as when humans learned how to use tools and fire. Companies have no choice but to adopt AI for the entirety of their workforce and FAST otherwise it won't take much time for them to be outcompeted.

[sirius]

Originally Posted by Tensor

If I’m wrong, please educate me with technical and logical arguments.

Appreciate you brought latest developments in AI space to the forum.

But Leave the future predictions to astrologers.

If you strongly think Software Engineering is going to be redundant and no future, and
1) If you are in SE field then better be prepared for alternate career .
2) If not in SE field, feel lucky.

For any reason convincing other folks will benefit you, please share me a video where any or combination of LLM/ai tool generating approach, design, code and documentation for below task (an actual task I did few years back and later tried to do that with chatGPT but what it generated is bullshit.). I will review and see how good are LLMs now.

- Build a multi Tenent data sync solution for workday entities (say candidates ) and our system (for simplicity sake assume we store, update, add and delete same candidates records in Mongo DB) where tenant will self configure i.e. Provides his workday api authentication credentials, etc. Utilise workday soap APIs. We should expect the data in workday to be reflected in our system within max delay of 2 hours. And any new data or updates at our end should be exported in max 10 minutes. Expected total records is 2 million in workday (need to import this initially to our system immediately after connecting the workday without blocking periodic sync). Daily updates at workday is 20k records and in our system 5k records.