I do agree with most of the things you said, but I just want to clarify myself a bit.
I have found the 3.7 Thinking model to do exactly what you just said, even to the point I wanted to throw my MacBook out the window. Multiple times.
I 'fixed' most of its problems by using extremely detailed instructions that it takes into account for every single prompt I throw at it, and one of the rules is basically just that it will not absolutely do anything else I don't specifically tell it to do.
I write these extremely long and specific rules by hand for each project. Almost no projects share the same rule sets.
Though, I will have to also try the o1-pro and see if I need to make a switch.
I have used the 3.7 Thinking model to write major code for projects that have done/or are currently doing low-mid 5 figures a month in profit.
By technical skills, I'm nowadays around the Senior-engineer level, and my own skills definitely come into play when I use the 3.7 model for coding, but I could easily argue that my own contributions are actually so little that an actual Junior level engineer could more than easily recreate everything I have done with the help of 3.7.
Also, I do get your point about AI not being fully able to make maintainable code to some degree.
AI might not currently be able to write maintainable code very well, but it really depends on the project and if maintainable code is even needed either by design or by how financially wise it is to make the time investment to do that
Are you trying to code software for a million-dollar ARR company, or are you just creating small-scale 5-figure MRR apps?
This is really something that changes from person to person, but I personally find it much more financially wiser to create slightly spaghetti code with an AI tool quickly compared to hiring an actual engineer.
I can churn out MVPs in mere hours and throw them out to the public the day after. If the project does well and needs updating, I might consider hiring an actual engineer at that point or getting them to even make the same app from scratch but better.
Do I want to wait a few weeks to a month for an actual engineer to build an app well, or do I want to create it myself with AI with garbage code and get it public in a few days?
The latter one is the one I'm going with as I'm not a technical guy by my trade. I'm a marketer; the faster I get something to market the more money I make.
It does not matter to me one bit how well the software is written as long as every feature works like I want, and at that point, I can market the hell out of it.
This is just my stance on this and this is the wisest way for me to do it, but this of course changes per person. The story is completely different for someone who might actually want to build something that actually lasts more than the time it takes for the app to generate $100k or so in profits.
However, I could probably even debate that churning projects quickly with AI is much, much wiser for the bigger portion of people here as that's the way they can their foot-in-the-door to business that generates money.
At that point, can we call a project that is written with AI and generates let's say, $100k a year, a hobbyist project?
If AI is more than able to do the majority of the code for projects of this scale today, how can we be so sure that AI won't replace devs in the near future?
Either way, I do agree with the majority of your points - just throwing a few debates here.
Yes, definitely. I am pro-AI coding. I'm just showing the current limitations.
o1-pro is significantly better. It also doesn't go off the rails like sonnet 3.7.
3.7 is better for one-shotting quickies and o1-pro is better for on-going development.
I was testing with 3.5. o1-pro would have done better. I'll test it on that bug to see if it gets it.
But, none of this is important..
I understand LLMs really quite deeply. I can write neural networks with just numpy. I can explain the math behind support vector machines, kernel regression, linear regression and RNNs. I can tell you when you should use each and the advantages to them. I can explain the problems you'll face going into high dimensions with kernel regression and how to use PCA to reduce the dimensionality of your data.
99% of people right now who are jumping on the hype don't even know what a vector is. They just come out with social terms like "this is AI copium", or "cope harder bro". People write some prompts for chatgpt and start calling themselves "AI engineers".
I sit here, spending hours digging into math trying to understand everything. Watching videos, reading books and posing endless questions to LLMs. I've spent so many hours going deep on this stuff.
And my conclusion is that I don't believe for one second that the super smart experienced AI guys who understand this 10x deeper than me and could write full transformers without even numpy off the top of their heads, believe that AGI is around the corner.
These companies have agendas. They want investment. As long as the general market consensus is that AGI is close and whoever gets it first wins big, then investment will flow like it did in the dot com boom.
It's so unbelievably clear to me that AGI is impossible with LLMs based on how they fundamentally work.
Bare with me on the boring mathy bits. I'll get to a point that might help you see what I'm saying(and anyone else taking the time to read this)
What they are doing is adjusting weights across all layers, query weights, key weights, value weights, output projection weights, FF1 weights, FF2 weights, layer norm 1 and layer norm 2. 8 weights per layer, with all of them being transformation matrices of shape (d,d) except the layer norms which are vectors (d,) which are used to normalize each input vector(token)
Those weights are just matrices. (d,d). In deepseek v3 that's 7168x7168
Now, in training, you'll input batches of training data, it'll go through the network and you get multiple outputs. The loss is calculated on each, then averaged.
Now, your loss, l, is what you differentiate with respect to each parameter.
So take just query weights, Qw, that's (7168,7168). You have a matrix (7168,7168) with dl/d_theta, where theta is the individual parameter, which is like a point in 51380224 dimensional space.
The matrix btw, is a transformation. It's not a vector. You can think of it like 51380224 hands in 51380224 dimensional space, that will each slightly adjust the shape of 7168 dimensional vectors passing through them
Now, each of these 6(ignore the layer norms) weights, Qw, Qk etc, need the right 51380224 hands.
Or another way to think of this is those 51380224 hands represent 1 single point in 51380224 dim space.
We need to find the point in each, that will result in the input vector(token) being transformed into the "right answer".
How that's done is, going back to the loss, think of that loss as height.
Every single point in 51380224 dimensional space has a height, so it's like a massive landscape with hills and dips.
We want to find the global minima. That's where we want to place our hands.
That global minimum point on the 51380224 dimensional landscape to place our hand that'll tweak the input tokens comes from training on huge amounts of data and making small adjustments.
Essentially what we're doing is having a super high dimensional representation of humanity's questions and answers. It's knowledge. It's style of writing. It's code.
Except it's done on ALL of these things, so it's the most likely representation of text, code, math, everything.
That's why it produces ok code at the function level, but if you try to have it craft something bespoke, it loses the plot with the structure.
Are you getting an intuition now as to why we won't ever get AGI from this?
But, you might also get an intuition as to how we can fine-tune to do very specific tasks, and if we create enough of these specalized agents we can have something very useful, resembling what an AGI might be able to do.
This is why I say it's simulated intelligence. It's simulating what intelligence looks like, based on seeing how intelligence represents in words and code.
It also generalizes fairly well, but that's not the same as real awareness/reasoning like humans have.
I also agree with using LLMs to create fast MVPs and making money. In business it's better to just get something out the door and make money.
Sorry for such a long boring post. I hope someone found it at least a little interesting though
I think there's a threshold that once past AI will blast past us. Because we don't fully understand intelligence we don't know where that threshold is and it makes it less likely to design something to break through it. There's also the motivation not to because people/companies want to maintain control. AI right now is like a giant library that's motivated to please humans, even if it's wrong and it inherits all the flaws of humanity in general and especially its creators.
See above
AI cannot write software, at least in one go, and cannot replace software engineers, for now.
But I am very positive about what AI can do, especially in coding field, or even making tremendous progress in software development.
I am not talking about AI replacing human stuff, I am just positive of what AI can do in the coding field.
Coding training data is clean, structured, rigid, easy to validate the results, which are all built for GPT.
As for software development, coding and software development are two things.
Coding is more like using AI to generate code, while software development is more like using AI to generate text or reasoning, and then generate code for each.
I can widely guess in near future, there will be dedicated workflow, or agent, or models trained for both the coding and the software development part.
With them combined, I am very positive about it.
See above
Its not AGI yet, when the AGI really comes, then coders are in real risks, I guess that would be another 3-5 for quickest, or 10+ years, who knows.
See above. AGI won't come with LLMs and I went into great detail above explaining why.
Those in the industry that say it will are mostly saying it to get more funding. They may have believed it in the past but they were just hoping LLMs would get so big and have so much data, that they would start to exhibit emergent properties of intelligence.
That seems unlikely with this current technology. Even if we could theoretically have a near infinite sized model with all the training data in the universe, it would still not have emergent properties, but just be producing based on existing patterns in the training data that aren't obvious to humans as being patterns there, and hence we see them as emergent.
