Hacker News new | ask | show | jobs
by ilaksh 614 days ago
I've been programming since approximately age 8, so around 40 years. For the past few years have been focused on generative AI, agents, etc. I have built multiple agent frameworks in Node.js, Rust and now Python.

I use AI with my agent system to help me write code for projects and my agent system. The leading edge models are incredible, but still somewhat brittle. They can make some weird mistakes. And I definitely have to go back over things.

But actually given all of the necessary context they can already complete a surprising number of programming tasks with good supervision.

It is night and day beyond the capabilities that it had a couple of years ago when I built my first agentic coding tests.

The thing that people strangely still seem to not get is that these systems will continue to improve rapidly. We continue to break through barrier after barrier with new paradigms. Now we are on the inference time ramp which is going to lead to large gains in intelligence.

There are more significant gains to be had also with things like multimodality and just plain efficiency increases from ASICs.

Beyond that, new machine learning approaches such a JEPA, new hardware paradigms built on memristors, etc.

Robotics and integration of new general purpose machine learning into robots continues to improve.

In less than five years, human software engineers will not be able to compete with AI in most contexts. Not only that, they generally will hold projects back if they are involved.

Before we get to the fully general purpose AI software engineers, platform specific tools with integrated agents can already with current models be very competitive in producing customized software for specific domains.

But it's not just software engineering. It's literally every single job. Including plumbers and all of the physical labor which currently is safe but improved robotics are just a few years away.

It seems obvious to me that normal careers are just not a good plan.

Instead, think of AI and robotics as incredibly cheap labor that will be available to just about everyone. So you become an entrepreneur, or the manager of a sophisticated network of teams of robots/AI running on a blockchain. Think about how to leverage these systems to provide goods and services.

7 comments

The thing that people strangely still seem to not get is that these systems will continue to improve rapidly.

This baffles me as well. So many low hanging fruits to pick, we literally just started with these large models. One obvious direction is predicting the next video frame. Infinite training data is available.

> One obvious direction is predicting the next video frame.

This is being done. There are incredible generative video models...

I would not call them “incredible”. Video generation is still at gpt-1 level. It will need several orders of magnitude scaling to get to gpt-2 level. We’ve just started making very first baby steps in this direction.
Given the only real mass market "robots" that are popular currently are stuff like hoovers or lawnmowers I can't possible see how trades such as plumbing are at risk any time in the next 20+ years. This is from the perspective of UK / European housing where a huge amount of houses are over 50 years old and so each set up is completely different. Different heating systems, thermostats, piping, connectors. Maybe in a new build property where every house is essentially the same there might be some chance of automation, but in traditional housing I can't see how that would work until robots are essentially "humans".
It's very obvious that humanoid robots will be quite capable and mass-produced in well under 20 years. Just do some searching on YouTube for humanoid robot and combine that with the progress in AI. The robots are improving quickly. I would be surprised if we didn't have something quite useful for the price of a car in less than 10 or maybe even 5 years. Just look at the state of the art and project improvements every year.
Instead, think of AI and robotics as incredibly cheap labor that will be available to just about everyone. So you become an entrepreneur, or the manager of a sophisticated network of teams of robots/AI running on a blockchain. Think about how to leverage these systems to provide goods and services.

This doesn't make sense to me... It is a genuine doubt: if you, and everyone else by extent, have the same ability to create a product or service, that can be easily replicated by AI, what value you are creating? What will you sell? Who will want to buy what you're selling if everyone can make the same thing?

Have you seen the average person’s Googling skills. How are they going to be able to produce anything complex with an LLM?
> So you become an entrepreneur, or the manager of a sophisticated network of teams of robots/AI running on a blockchain.

Wont managers and entrepreneurs be automated as well? Their jobs are not neccessarily harder to automate than software engineering.

If AI agents do get better than human SWEs, they'll also get better than their human counterparts at other jobs - including being an entrepreneur.

When people call out the doom of SWE careers at the hands of AI, they seem to think that only SWE careers will be affected, which seems weirdly biased to me. If AI comes for our jobs, it'll come for everyone's jobs. And if that is the case, economics/capitalism as it exists today doesn't make any sense anymore.

I'm not sure I agree. LLMs are great coders, because we have a lot of training data to make next token prediction work well for writing software programs.

A mistake to make here is to conflate 'AI' with some non sensual dream of 'AGI'.

LLMs are able to do a lot of what a swe does and if nothing else that just means we need less SWEs.

It absolutely still makes sense. It just won’t be so great for the humans.
In which case there will be massive social upheaval, which won't be good for the newly minted god-CEOs. They are incentivized to share the spoils of infinite productivity. What good is being a god of a world in flames?
> The thing that people strangely still seem to not get is that these systems will continue to improve rapidly

Where is the plateau though? What we have so far is definitely impressive and useful in some cases, but I can't help but eye-roll every time I hear someone seriously talk about making nuclear reactors JUST to power the millions of GPUs needed to train these models.

We have already seen multiple plateaus as we reached barriers in terms of scaling the models the hardware and software. We have continued to innovate and break through them. The latest barrier is scaling training via inference time.

New hardware and even learning paradigms are being researched to get us over the next wall. This has been shown time and time again we keep innovating.

The power requirements with force investment to scale out the new memory-based compute paradigms.

I would love to hear your experience around building agents

And how did you even got started with building agents