Sounds like a hardware problem to me.
How long will it take the tool for the program to be finished? You mentioned that I should message next week, did you mean that you could only try it by next week?
Iâd love to get an update about it.
I was asking questions and your answer to it was âlol wtf, that wonât happen anytime soonâ. Then why do you think I was asking that stuff in the first place.
You think I wanted to talk about the theoretical possibility of a system being capable of writing code?
If you have special, letâs say innovative ideas you still have to explain how you want to solve it obviously.
I have some stuff to do to finish the system. Some stuff where I have to manually start scripts and adding capabilities to deploy model training to a provider of gpu for example. E.g. when it wants to create a diffusion model to create 3D meshes from 2D picsâŚ
Check this out! You can probably spin up a space on huggingface or your hoster of choice and have it generate 3D meshes from 2D pictures via an API. ![]()
Sure could, but training own models is a core feature. Should be able to do that.
Hey Jochen, cool topic, interesting perspective and indeed something to discuss.
I agree coding today is similar to âorange picker,â âcar mechanic,â carpet weaver⌠and we donât even know that many carpet weavers today. Machines do the job - period. Coding is no different. Of course, it is very hard to accept. But itâs inevitable. The good news is that candles were not 100% replaced by LED. So there will be coding jobs for decades.
MY EXPERIENCE
We are literally producing âenterprise-grade applicationsâ in no time. A compliance management solution takes us 4 weeks - to do. Not coding - this is done by AI, but putting it all together - the assembly job.
Not sure I should continue, but the world of SOFTWARE makes way to the world of artificial INTELLIGENCE. Software engineers must completely reimagine their future - itâs a mindset, unlearning and relearning. âŚ
Iâm a 70 year old former developer ![]()
I am also prompting more core functionality. Got some missing parts that need to be implemented (made PoC with quick n dirty code and wonât put dirty code into it).
I was building parts for it for nearly 5 years.
For example 2 years of a 400 hour per month journey to explore all kinds of OCR techniques and get a pdf pipeline that wonât be capable of extracting all kinds of data ever. Simply because you can put everything in a pdf.
Maybe a meshup of kafka injected neo4j where you press the graph into an optically connected quantum computer and use dwave could lead to 0.000001% hallucination rate⌠but deterministic pdf costs at least 50-80 million.
We need special hardware for that.
Which would be way better in terms of energy consumptions btw. Multimodal GPT most probably needs 80% more energy and takes hmm 30 minutes for 3k documents (against ~0.2 seconds with the quantum graph).
I feel that you might not have fully understood my main argument. What I was emphasizing is that individuals with coding skills are typically more capable of leveraging language models to generate superior code and software products compared to those without coding knowledge. This is because they possess a clearer understanding of code capabilities and limitations, allowing them to articulate their requirements more effectively. Without this expertise, many people might end up making vague requests to the models, such as asking for something âclassic, yet modern.â
While this difference will shrink over time, knowing to code will continue to convey curtain advantage in this regard for the foreseeable future.
AI is helping me not just in generating lines of code, but in algorithm development and implementation.
Also ideas on how to approach different problems, more efficiently.
Itâs not just all about code.
An example, that involves code, however, is when dealing with a DynamoDB table, and you want to query on two values at the same time. This would be more efficient than a query on one item, then looping over to align on the second item.
I wasnât sure how to do this, or if this were possible. But after I gave it enough assurance I knew the basics, it produced a composite secondary key approach and corresponding query syntax, and I got it to work in minutes!
Iâm learning new stuff all the time!
Now you would argue, âwho cares, in a couple days AI will write all the code!â, I think that is an unrealistic timeline, but maybe not an unrealistic expectation.
What I do think is humans do need to be in the loop as âoverwatchâ to make sure things are done right, and this, my friend, takes someone that knows how to code. ![]()
That is a question of how my system cuts down the user request and creates technical descriptions - which are made by me⌠a guy with nearly 35 years of sitting in front of code for 12+ hours a day on averageâŚ
It would be good, but is it neccesary? I would treat ai agents like interns and tell them how to reach my level. It may take some time..
Iâm intrigued by how your extensive expertise has enabled you to develop a heuristic that effectively comprehends the intentions behind such requests from any individual among the 8+ billion people worldwide.
While writing code once required manual human effort to accomplish something, the idea of delegating this task is not new.
The same thing that has always happened when things can be processed and delivered much cheaper as a unit: We build bigger, better and faster.
Statistics.
Do a thousand projects and youâll know as well.
This topic reminds me of when 3D animation and CGI arrived in the world of cinema.
At the time, there was a debate about whether live-action actors would completely disappear.
Well, many years have passed, and weâve seen that they havenât disappeared; instead, they coexist, bringing out the best of both worlds.
I imagine something similar will happen with highly skilled professionals like you and ignorant people like meâwe have the opportunity to develop basic code and gradually learn the intricacies of programming.
I guess that for more advanced users like you, there will come a point where you will stop writing code and simply oversee the machine.
Something like âcode supervisorsâ will emerge.
I want to do that too, but I donât want to blindly trust what the interns write. Maybe at some point, and the AI is stable. Itâs still a bit risky. It all depends on what you are writing and what the consequences of going wrong are.
Launching nukes ⌠probably need massive overwatch, or no AI at all.
A blog about your favorite hobbies ⌠let the AI go wild.
The middle ground, where most people are, are when reputations are at stake ⌠what happens when something goes out to the public and is messed up? How quickly can you save face and recover, is the speed up worth it? Is it too painful to clean up the mess? Etc.
worked on automating that task with AI agents based on rules in static code analysis and extensions to the iso standards made by me for ~a yearâŚ
You mean no LLM.. yeah, thatâs right.
Anybody deploying AI into a production environment should have a human in the loop approval-like system. Now we have someone to blame ![]()