something_else

I agree. Direct experience, however, will give you an accurate perspective on AI. Anybody who has used this tech beyond consumer level has direct experience of what it is capable of.

It is only an open question for you because you haven't actually used the tech or really seen what it can do

This is completely outdated knowledge If you are dumping code snippets into a ChatGPT window that's probably the case. I suspect that's what this study looked at. If you are using a coding agent like Claude Code it is more like a 300%+ increase in efficiency when used correctly. Much more in some cases. ----- I'm sorry but I think your knowledge on this topic is both outdated and quite incomplete, and while I appreciate that you want to avoid falling into group-think I get the sense that your anti-mainstream bias is clouding your judgement a bit here. I remember that when everybody thought AI was a bit gimmicky you were a massive promoter of AI tools (I vaguely even remember you calling them smarter than most people) and now that they're mainstream you have gone 180 and started calling them overhyped and unintelligent despite the fact they are orders of magnitude more intelligent than they were at that point in the past. They are overhyped, but not nearly as much as you think. Certain industries like software engineering, most IT jobs, support agent work, any kind of data analysis, marketing, search engines, SEO and tons more... all unlikely to ever be the same again. Even as a result of less than perfect AI imagery, industries are changing. I went to an AWS conference recently and Skoda did a presentation where they said that almost all of their newest still images of cars were AI generated. They were talking about their methods for getting the AI to produce consistent images which did not have oddities or inaccurate depictions of the target car in them, and most people in the audience could not tell the difference between the real car and the AI generated one in the end results. Skoda are not going hire a team of people to go into the dessert and take photos of a new car anymore, they'll just get a few people to use their AI tools to do it.

Ok well even if we say a newborn is intelligent, that is coming from billions of years of instincts encoded in it, which is essentially just data. Intelligent behaviour is always coming from data and experience in some way or another. I don’t see how a newborn with a billion years of evolutionary data encoded in it is all that different from an AI model encoding a billion gigabytes of human knowledge It produces a different kind of intelligence but it is still intelligence

Are they? A newborn isn’t that intelligent and that’s basically a human without any data

I mean yes, I think where we disagree is why this matters. More data, more training and more experience makes you smarter, the same is true for humans.

They can. It will translate the core concepts from languages it was trained on into the syntax of another language. It's better when it has more examples of the unseen language's syntax, but the same can be said of human intelligence. If it had modelled the concepts of fish, sword, head and ocean then it could make a pretty decent effort at combining all of that together into a swordfish. Image models are still quite a bit behind text models in capability but they can still combine two known concepts into a novel one to a degree, just not as well

I disagree. They are able to strategise, plan, answer questions, figure out where to get the data they need from, and come up with new solutions to problems in completely novel scenarios, that is more than memorisation. If LLMs were relying purely on memorisation then they would not function very well as reasoning engines, yet they do. The core concept they are relying upon is actually pattern recognition, not memorisation. Wikipedia is more akin to memorisation than an LLM is. Here is why their pattern recognition is so good: The entire internet (which at this point is comparable to what Anthropic models are being trained on) is estimated to be around 200 zettabytes. So lets say all human knowledge from the past 2000 years comes out at 200 zettabytes as a very crude estimate. Claude Opus is around 1TB in size. 200 billion times smaller. Yet somehow it is still able to make a pretty good attempt at giving you information from that entire 200ZB corpus. The only way this is possible is with intelligent abstraction of all of the core concepts and patterns contained within that 200ZB until it can model most of them pretty well with only 0.0000000005% of the amount of data. It is this abstraction of concepts that I feel most people would call 'intelligence'. Most human measures consider the ability to identify abstract patterns to be a core component of intelligence as it is one of the fundamental building blocks of doing anything useful with intelligence.

It can spawn subagents with any model, Fable, Opus, Sonnet or Haiku. If you don't specify, I think it tries to pick a sensible value. But Fable may decide it wants to put Fable on the job and if you have a few Fable subagents running under a Fable session agent it will burn through your quota very very quickly. I like to play it safe and specify that Opus should be used for subagents. Though I also do what you do sometimes as well; ask Fable to create an MD file and then put that into Opus. This works better for really big tasks where having a papertrail of MD specs can be useful.

Calling it a chatbot at this point is disingenuous. It functions as a very capable reasoning engine in most real world applications beyond consumer level

This is why a lot of people are scared, and unconvinced that this is a bubble where the outcome is going to be a stock market crash. AI seems to be good enough where it's very possible that it will upturn the entire economy by wiping out 50%+ of our white collar jobs. Is that a stock crash? I guess in a sense.

This is very much the case for image generation, it is almost good enough to be amazing for specific applications but always falls just short. I promise you that this is not the experience with coding agents in most cases. They're flexible enough to do exactly what you want much quicker than you could without them. They make mistakes, but if you point them out and give them more direction, they'll adapt and do what you want. Right now I'm working on a quick prototype for a new editor interface we are building at work and it's the most creative I've felt during working hours in a while. I have ideas for features I want to add, I ask the AI, it implements them, they work. My mind is entirely focusing on ideas, and implementation is secondary. This code will never reach production, but in ONE day I have an extremely detailed, functional prototype that proves the tech is viable for a production-level implementation. This prototype alone would have taken a team of four people 2-4 weeks to complete three years ago.

Yea, exactly. It's going to be interesting to see how this pans out. We've completely restructured our entire team now as a result of AI. Instead of having two squads of 4 devs each working on one feature, now every dev is working on one feature in a 'pod' with a designer and a product manager. Even if the feature is massive, it gets one dev. Because of AI assisted coding this is manageable, and actually we're all loving it because we get a ton of autonomy and ownership over features that we didn't get before. For a while, when we were working in squads of 4, it felt like devs became meat interfaces to Claude. The product team came up with a feature, designer designed it, tech lead broke it down and wrote tickets for us and then we basically pasted the tickets into Claude. Now each dev has complete ownership over the lifecycle of their feature, manage their own tickets, breaking it down however they like, architecting a solution, working closely with product and designers etc. I think this may be how the future of software engineering is going to be, moving away from the complexities of writing code and more towards system design, consultation, and collaboration.

I think as time goes on the token cost will go down a bit, but yea they're very subsidised by VC and a desire to dominate the market right now. Also, right now devs are not particularly efficient with tokens, we're often just throwing the biggest models we can afford at every problem. Where I work we pay $200/m for each dev to have Claude, and with Opus 4.8 you have to really try to hit the limits. With Fable it's really easy to hit limits. But you can get around this by asking Fable to do the planning, and then ask it to spawn Opus 4.8 subagents (or even Sonnet or Haiku) to actually implement the code. This way you get the brains of Fable coordinating cheaper models to do the bulk of the token-heavy work. I suspect as time goes on and token costs begin to reflect reality, we'll see more devs learning efficient usage patterns like this.

Agree 100%. This is why I'm not scared for my job, but why I WOULD absolutely hate to be a new/junior software developer right now. It's going to become quite hard to build the deep experience required to guide AI tools when you're exposed to them right from the beginning.