IT DOES NOT MATTER IF MODELS GET BETTER

NOT AI SLOP , WRITING WITH ERRORS TO PRESERVE AUTHENTICITY OF THE POST

I have been using AI ever since the launch of GPT, and its interesting to see how things are changing not going to say its good or bad, but here are some observations I have made.

Back in 2016, I just started my junior college and wrote my first line of code , I had no idea how it worked but it was still exciting to see, I wanted to make a game so that is why i started to deep dive and learn coding, but I realised it was not very easy writing a game in c, I needed to understand a lot of concepts about game engine, garbage collector , so I stopped, I made the smallest CLI tictactoe game with the help of gfg and stackoverflow and got all dopamine hit i could from it.

The next few years I spent some time getting a cs batchelors degree to write a lot of web dev code, made a lot of automation bots for freelancing clients, made java software, I made a better spaceshooter and headhunter pygame all by hand again, but this Time I had better understanding of the fundamentals of what I was doing and why I was doing it, I had a lot of fun typing all night writing code and listening to music. But I wont lie , I did not write every single line of code myself, I reused opensourced libraries, I copy pasted stack overflow and frustrated senior devs by asking them dumb questions, I remeber using something like tabnine, which used to autocomplete my code.

I was intrested in coding so i decided to purse masters to make my understanding even better for computer science , and I did , I also contributed to a lot of opensource at this time and prepared for interviews (I did not enjoy DSA or cp), but i did it anyway. At the same time gpt3 had launched , I was still writing code, new software but something changed fundamentally, I stopped using tabnine , I used to now prompt chatgpt to give it a full file of jsx, so that i could just copy paste it and use it instead of typing it by hand. It was frustrating intially because it did not have the context of my codebase , I realised it was writing a lot of slop code which I ended up debugging later.

But LLMS improved, the more data they had the better they became, And soon enough I shifted to copilot , atleast the models had the context to my codebase, now I could generate better code, there was still a lot of slop but it was much much better than what i had been doing previously , infact I almost stopped typing code , Now I wasnt declaring any variable or function by hand , I also stopped going to stackoveflow for dumb doubts because most of these dumb doubts were now solved by AI.

At my first FTE job at browserstack , I got access to frontier AI models which not only generated better code but were also giving be better software design suggestions, they could optimise and refactor code much easily. After using Fable and opus which i think was an important turning point, for the first time I closed vscode and shifted to CLI just to see if I could code and I got my first pr merged without even opening vscode. I was in awe, I could never think this could even happen, this was also the time where a lot of layoffs were happening in the industry. I was scared too , But our company being a startup we did what any company would do in such a time ,produce more code consume more tokens, which we are still doing to this date, every day.

Now in these last few months I went through more experiments with ai and changes than i did for the rest of my coding career and here are my few observations, predictions and the real post:

observations:

1. Dunning - kruger effect(The Dunning–Kruger effect is a cognitive bias where individuals with low ability in a specific area overestimate their competence, while high performers often underestimate their skills.): the first people to get trapped in AI hype were not tech bros they were non tech bros , thinking AI is going to take away all tech jobs because anyone can do it now, which is partly true any one can wrote decent code now , make app in hours if not days. But only because ai is able to generate code within 10 minutes it does not mean it is the best code, neither does it mean you may be able to solve future issues and tech debt which comes with it. In corporate terms, code written ≠ dev done and dev done ≠ production ready. Problem arises when human in the loop will be required (because taste and accountability cannot be automated), What will you do when your aws throws you a 500$ monthly bill and you don’t know why , lets say you make a software and sell it but what happens when you start scaling. what happens if someone breaks into your system you cannot hold anyone accountable. what happens when you pitch clients an ETA of 2 days because you think ai can generate code easily and you can just ship it , but the actual effort is more than it and the solution itself may not be fundamentally possible. what if ai assumes problems in your code base and deletes half of it ( it is very good at assuming wrong things) while you keep pressing enter and sipping coffee. AI will only amplify the kind of developer you are , you may not need a developer to create a simple website anymore but you will need one to understand GTM metrics, AEO,SEO,GEO metrics, JS , lighthouse, code chunks, heatmaps, surveysensums, hosting ,cost optimizations, etc. yes you can learn these on your own ,you can use AI to understand what you are doing, but anyone more experienced will always be able to do better work than any random guy. This is another big problem AI has created, the leadership of any company overly estimates what can be done with AI while the reality on ground can differ, which causes a lot of distress, and as of now we dont have a line to draw, ie we dont know how to calculate the eta of a task if it is done using AI.
2. The Cobra Effect is a phenomenon where a well-intentioned solution to a problem inadvertently makes it worse, serving as a classic example of a perverse incentive. after the release of AI, many companies are measuring productivity by tokens spent per engineer or kloc , guess what happens when you measure productivity by tokens and kloc, you will see a lot of slop code being shipped anything that could be done in 5 lines will be deleberately written in 50 to increase kloc delivered and tokens consumed, employees will eventually find a way around.
3. Jevons effect is an economic phenomenon said to occur when technological improvements that increase the efficiency of a resource's use lead to a rise, rather than a fall, in total consumption of that resource. Greater efficiency reduces the amount of the resource needed per application, lowering its effective cost; if demand is sufficiently price elastic, this induces demand, frequently resulting in a net increase of total resource consumption. This simply means demands for sde will increase (but it will be still less, as an example a company might fire 100 mass code writers and hire 10 good engineers), , as things get more stable. So it might not be that bad.
4. Generalization is the new specialization , since companies will hire fewer people you will be expected to work across different codebases, technologies and tools. being a subject matter expert in one field may also be good, which means you are reliable and can be held accountable for something entirely , but the new hiring criteria may be more biased towards full stack generalists than specialists. A normal SSE will do what a tech EM did in pre AI era but with agents, we have just promoted ourselves without knowing it.
5. People who are most afraid about ai, are mostly script kiddies and people who have not used ai enough, have not even tried to embrace it. Infact people having access to frontier models are undergoing ai physcosis , I was one of them, talked to fable for 16 hours straight.
6. Most businesses have similar revenues pre and post AI eras, businesses are making money by cutting staff and replacing them with tokens instead, and counting the surplus as profit. but there will be a limit to what they can cut off their staff up to, they will need someone to spend the tokens they are buying.
7. AI companies are doing a lot of hype marketing and creating fear in the market so everyone buys up more tokens. so maybe dont trust altman if he says there are going to be no jobs by 2030, maybe let big corp fall for their marketing gimmik. It is an age old marketing tactic to create hype in the market to sell off something.
8. It is quite surprising to me that some big tech giants are not exploring this opportunity to create better models even though they have more resources, Infact all old players are pivoting more towards hardware or datacenters. Google, nvidia , aws , microsoft are just some examples.

Predictions:

1. IT DOES NOT MATTER IF MODELS GET BETTER: I think I have already reached a turning point where I have removed vscode and writing without it , which means my approach to coding has changed from keeping ai in the loop to autocomplete to keeping human in the loop to autocomplete the code. so It does not matter if the model gets better they are already generating pretty decent code, and maybe better models might be able to find more loopholes, vunerabilities, give better design, do it faster with less resources, but apart from that It does not matter at this point, I could almost achieve the same output in sonnet and opus . Im looking at it like new programming language releases we used to have back in 2016-2020, every new one was better in other in some way , and teams made decisions on which one to use for their projects.
2. Pure software will reach a plateau soon : Any piece of software which does not involve hardware ( like os companies ) , infra (browserstack), cloud provider/dependency ( vercel ), iot or embedded systems, will hit a plateau which means that they will stop shipping new features soon and will run in BAU mode, which means very few people can operate the product and manage entire codebase. This also means we might see more companies in unexplored niches like geospatial engineering, iot, biotech , embedded systems. Without ai it would take 10 decades, now it will take half a decade to reach this point.
3. Software development may not die out , but im not sure about other fields, what happens when everything on screen is automated , companies will hire people to automate stuff off screen, eg manufacturing, driving, cooking, all other menial jobs which involve labour will be cost optimized by automation but this is the far future assuming agi is in the timeline in this decade. However entry roles will shrink and only the people who are really good like will stay in the industry in these times, again im looking at next 5-10 years. how ever specialised coding jobs like python programmer, java developer, react developer are surely going to take a heavy hit.
4. Ai will create new problems, there will be problems like GEO, AEO, cybersec, model- poisioning, guardrails maintenance, training datasets, ai infra, harness engineering ,etc which might emerge as new fields.
5. Test driven development will be the new normal as people can now write a lot of code and create big prs with big changes, it is only wise to have a safeguard before hand to make sure nothing goes down in production
6. CS degree will still be worthy , AI often hallucinates writing solutions and that is the merely because it works on patterns, however in engineering there is no perfect solution but only tradeoffs, and to make any tradeoffs you need to understand how stuff works. To spot slop code you need to understand how slop code looks, for which you need to write slop code for years until your pr review is rejected by your seniors, which needs experience and fundamental knowledge. knowing what NOT to do makes you more worthy than AI.
7. Tech has always been built on opensource , look at the history , which will be repeated sooner or later , we will soon have an opensource model which will give better benchmark results than claude or any other ai. And when that happens many companies will definitely use those models , which means anthropic may not reach the valuation it predicted, atleast I would bet against the big giants in the long term. This is also my theory of dario asking to slow down frontier tech as they want to slow down open source / open weight models catching up to them.
8. a big prediction but devices may move from GUI to GUI + VUI, voice user interface , once we have AI os or something like that where we can look at screen but also operate systems handsfree
9. Licensed jobs like lawyers, doctors, CPAs will be safe , not because they are better than AI but because they are Licensed and are backed by the government. No random person can become a doctor or CA with the help of frontier ai model, even of they perform better than human doctors or CA’s, because without a license you cannot practice these things.

Advices:

1. For anyone who is not gonna read the whole thing and wants a simple way out: Quit Tech, that is the safest bet you can make right now, its not because SDE is dead but because its saturated, we had 100X more people that we need , and being average is not going do you any good, Ai was just a cherry on top, making sure there is no sane reason for any company to hire junior developers any more. The market is extremely brutal , with extreme bloodshed ongoing, anyone who tells you otherwise is trying to sell you a course. If you are in a mid computer science degree , maybe try to switch carrers into an MBA or other profession , Im not sure which.

That being said , Im not quitting tech myself, and for one sole reason, I genuinely like being a Software engineer. Id rather be on the cutting edge of technology and see myself getting automated away than to change what I truly like doing, so easily, so yeah Im not gonna go it atleast in the Near future. And from my perspective , when AI gets better in doing stuff than you the only thing you get paid for is knowing what not to do, ie deciding the tradeoffs, which is what software engineering is under the hood anyways. If you feel the same, read ahead.

1. From what I can see everyone can now write code with AI, infact design systems , operate infra , pretty much do everything, but even with AI ,engineers who are not expertised in one field or do not know what they are doing will produce worse output than someone who is experienced (we tried this in our team and failed miserably). So that means freshers must not focus more on coding, syntaxes, outdated college syllabuses and focus more on core software engineering concepts and upcoming fields or more newer fields not included in old college syllabuses like cloud computing or finops. Atleast in college dont go chasing frontier models and try to write raw code by hand. Dont bet on specializing on one field.
2. Don’t do an AI course , AI is an ever changing field and you never know what is going to be the new norm to work with these models, Im not saying dont stay on the cutting edge of new developments, but definitley don’t invest too much , I went and spent 2 weeks learning openclaw when it was launched and became the fastest growing github repo, only to realise it was a waste of time and money months down the line.
3. If you have already graduated and not able to find a decent job, I understand its not entirely your fault , but for now, get whatever job you can in tech and stay there to gain experience, It will matter more in upcoming years. If you are not getting a job at all, get certifications/degrees , and do not sit idle.
4. If you are at a job , then expand your horizon. Most probably you were hired in pre ai era where you were hired because you were expert at a specific skill or could write decent code in a specific framework. but since now AI can write better code , most of your time at your job will go in orchestrating ai , you are not going to spend time in the time taking work of coding and debugging small issues atleast. This might be the perfect time to get your hands dirty in other fields, eg if you were hired as a Frontend dev, try diving into backend and infra, take on a learning curve and get more responsibilites in the team.

At the end of the day,
In every era, people who had access to frontier tech have been the beneficiaries and have always made the most money in the market, be it the green revolution, industrial revolution, web revolution, etc … being afraid of ai will not do you any good, it will only make you more tensed, you can either complaint or embrace it.
these are my own thoughts which I decided to share, feel free to disagree or follow.