Trust me this version of me can't write code better than an LLM. I think even the LLMs from 2 years ago could do as much.
Only thing we have solved with LLMs is how to generate code, as in literally converting what you type in as text into something somewhat working. Fixing the details is impossible for AI.
Today and beyond.
Now Open AI and Anthropic are launching own Service firm/wing for maximizing ROI. Direct Competition. Huge Loss for Indian IT Service Firms.
They paid AI providers to Train own automated competition end to end so much so that they are learnt the gaps and how-to own the market by corrections, private IP Source code access and embedded expertise extraction (stopping short of calling it literal corporate espionage, they got all the know-how and especially for modernizing legacy tech and integrations).
They are untrustworthy especially where IP is involved.
Indian and other IT Firms who did not use private self-hosted AI for important things nor monitored the usage and did not train their employees to think what ( or how much) of our institutional knowledge goes out to third-party are already facing trouble.
You can also find this in science, I feel like software is about finding the Gauss’s law for the dozens of electrostatic principles the masses keep coming up with.
AI strips away the need for simplicity by making brute forcing details via verbosity that ruins the need for "good" software, and makes all software by definition unmaintainable since maintainable product as a unit has to be simple enough for a LLM or a human mind to understand.
Currently a human context window is in 100s of billions of tokens, while LLMs seem to max out at 1M.
For anyone thinking but humans can't produce all of the text verbatim, that's the difference b/w humans and machines we can parse and take the most essential parts from 100B token context it's just are speed of prefil and recall is slower. And we memoize learning so if we know X then all variants of X are learnt on the same foundation.
Human learning is so very different from LLMs is so very fascinating as a trad "AI engineer", well some person who has worked and built/shipped ai agents and models at scale.
Fun analogy Kinda like power tools like are we tightening a screw using a screwdriver or doing it with a power drill. Case in point there is chance of ease as well as more risk while doing it with power drill.
You are absolutely right, it can be guided to produce code but needs corrections like its too easy to create unmaintainable solutions.
Also to untrained eye without good architecture, it does the job but when one who doesn't have architectural skill will think that's enough and there's no more room for improvement. This cause troubles in their production.
Like if it solves their pain points and makes them think systematically and critically. That is better.
What i don't like is giving much information to 3rd party service providers of AI developer tooling LLMs + Copilot and such, which is causing lot of real market level Risk for Indian IT Service Firms and has started costing jobs and literally erosion of market share because the service providers like Open AI launched their own firms like Deploy.Co.
Smaller companies who didn't know any better using tooling and developing stuff and on growth bringing professionals to fix mistakes has been happening long. There are literally companies who made solutions by Access and Sold it to clients, we literally told them even as competitors to compile the things as programs to avoid from future cases of fraud by their Users who were doing shady things.
A proxy for it would be we see a huge wave of lay offs.
This is not happening.
We can now simply buy an artificial form of intelligence to solve intellectual problems. That doesn't mean we can buy our way into solving every problem in a week.
Building Photoshop or Excel still has a huge intellectual cost. We can use AI to help us get there, but it won't be free, it's just getting marginally cheaper.
That, too, we're not seeing that at all.
I think of "tokens" as units of intelligence spent on a problem.
Producing software like Photoshop, Linux, or Excel literally cost billions of tokens (human intelligence back then, and now both human and artificial intelligence combined). A lot of deep intellectual problems were solved along the way, and the token cost of solving them accumulated. The true cost of Photoshop today should be measured in billions, maybe trillions of tokens.
AI can definitely generate some tokens, and it has accelerated everything involving intelligence. But it still can't produce 100 billion tokens worth of intelligence in a week. When that day comes, that's superintelligence. We'll go to Mars, we'll make fusion work. That's the singularity.