published
updated
A brief History
I've been watching LLMs since GPT-1, and machine learning since very pre-hype. I ran folding@home on my computer, I eagerly combed through all the DoTA 2 coordination papers OpenAI was doing back in the day, my notes have a GPT-2 poem from back in 2019, that I found beautiful. I studied the things in university, I know this is the 6th "AI summer", after many long, dark and cold winters, but this one sure is HOT. All this to lay out my background: I'm not a hype artist, I'm a technologist. I found the ramifications of GPT3 startling. I read the ReACT paper in shock. I struggle with implications of how the data that builds these machines is acquired. I worry deeply about for what purposes they're deployed TODAY, just like DAIR is. I'm not interested in AGI, and I'm not convinced we're close to it. I don't think LLMs will make another step jump to "beyond human intelligence", we'll forever be trapped to "our average response". This is both a constraint, and a guiding principle. AI will not improve until it improves the collective average person. A challenge if you will. A purpose.
But that leaves us with a problem in the here and now too. LLMs are already incredibly capable. The models we have now are not significantly better than GPT-3, I stand by that, just like the iphone 15 isn't a step change in behaviour from the iphone 5. Sure it's a little different, they use your face instead of you fingerprint now, but not a difference in kind. Before I was happier on the side lines, abstaining from technology in tainted data fields, that wasn't particularly useful. Even now, I don't use image-gen, it's not my line of work, and a big part of why I enjoy art is the context of who made it. AI art is still tacky and gross to me, though I'll admit DeepDream did capture my attention once upon a time. But I do work for a company developing AI programming tools (isn't the whole industry?). For a long time our AI native tool SUCKED. I would never pay for such a thing, and I mostly didn't use it, even though it was free for me. I was glad I wasn't working on that product. It simply wasn't very useful. Not worried about the dissonance since Data Gains Value in Volume and I was not contributing to the volume.
Feedback loops are more powerful than you think
That changed. I still believe the models themselves haven't gotten a lot better, but we taught them how to use tools and that set them free. Turns out the missing ingredient was feedback, who could've guessed? It's not like backprop is based on the idea. The newest generation of AI tooling is absurdly capable, but only in environments where feedback is deterministic and cheap. Either it produced the right text or it didn't, and we can determine this cheaply with near perfect accuracy. You know the cheapest place to buy mistakes? That's right, code! Mistakes are (mostly) free! The compiler will tell you immediately if you're wrong. Thus, Vibe coding was born. It exists as a concept because of how capable these tools have become when placed in feedback environments. I really want that to sink in. The barrier to producing working software fell for the first time since the compiler. Arguably by an order of magnitude.
"But the code sucks!", I hear you complain. Sure, the code sucks. But it's significant enough to be a step change in productivity even for experienced professionals. I felt it myself. Shipping projects I never would have completed otherwise. You can see it in the wild too, there's plenty of examples. Early adopters with the right skill set are building software faster, better or bigger than they were previously capable of individually. Not everyone is willing to be a load-bearing user, but those who manage are themselves reporting these things.
Exploding examples of high quality software, built because LLM capabilities for code in the wild:
The Curse of using LLMs
Using agents (ugh, I dislike this name, but whatever) is a curse though. Every interaction costs money. The agent is often wrong, misunderstands you, or incapable of performing what you asked. It's a slot machine for code. Prompts and tokens go in, sometimes prizes come out the other side. Othertimes you get a toxic, three eyed fish that nukes your database. Building, wielding, and growing these tools is a dark art, and nobody really has it down to a science yet. Sometimes things go boom.
Just like a casino, they want you to keep pouring in the tokens, and hoping for a prize. They'll make you lazy, stupid and angry if you let them. You'll atrophy your own skills, and pay for the opportunity. Bound to the prompt box, because you've created a ball of mud so obtuse nobody ever understood it, but one that sometimes does work. It's a dangerous game. But the rewards are real, let me assure you of it. It isn't always a ball of mud that comes out. You can get high quality code that is well engineered and understood by you, for a fraction of the cost of whatever the going rate of engineers is in your area. Yeah, even the cheapest engineers in a foreign country that you sneer your nose at because they program for a paycheck and don't program in Rust in their free time.
I don't think it's wise to bury our heads in the sand. It really is a shift in the landscape of our industry. But as crafts people using these tools is awful. "We do this because we LIKE to, I didn't want it automated", I hear you yell. Yeah, I feel you, me neither. I still love crafting artisinal code. But, the industrialization of our industry is finally at our footsteps. Are you ready? Are you part of a union that will coordinate action? If you want to be a real Luddite you got to have bargaining power. They smashed the tools for a reason afterall, not because the loom was a moral affront to their sensibilities. They wanted a seat at the table. Will your bosses listen to you? Or will they replace you with someone who will bargain with the machine, one that might write code that will bankrupt the business if the engineers aren't being careful. They're playing games with lying machines, careful and diligent will determine their success, or they'll get popped by the dark forest online.
- The industrialization of code: code yields are already becoming a thing
Fighting the Machine
The industry is now able to extract value from larger data commons than ever before, and so our tactics have to shift.
- I have to believe we can use these tools for good. I just have to, I studied them before they were weapons.
- Tokens are expensive, local hardware is catching up, but also out of reach for most people
- When not to use the LLM, even when your tokens are free
- It's not enough to "We will not use the master's tools", burying our heads in the sand would likely mean obliteration
- Code being orders of magnitude cheaper is landscape defining change
- New way to inflict damage just dropped: Abuse the lethal trifecta
- The code written by LLMs is often shit. The people using them recklessly will depreciate in value.
- We must learn to use the tools in such a way that makes US better programmers
- We must hope the cost of hardware, and the ability of local models becomes a commodity, so we're not chained to big labs