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This is completely outdated knowledge If you are dumping code snippets into a ChatGPT window that's probably the case.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 ----- I'm sorry but I think your knowledge on this topic is both outdated and quite incomplete, and while I appreciate that you want to avoid falling into group-think I get the sense that your anti-mainstream bias is clouding your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.

Not really, we've relaxed our PR review guidelines because we're producing a lot of high quality code. Code review for us now is something like: Manually test Manually review big picture architecture choices Manually review critical code paths Do deep-reviews using AI tools — an adversarial approach works well here, where you have an AI agent reviewing and another AI agent trying to refute it's review findings, then it produces a final list of findings for the reviewer to analyse Design review This seems to be working quite well for us. Our bug/error rate in new features has gone down.

I have definitely noticed a change in my personality. It's very hard to judge whether it's for the better or worse. My friends described me as a bit more 'subdued' which makes sense as I have a much clearer thought process and can relax more during conversations now instead of being a cortisol-fuelled word-vomiter. There are elements of increased anxiety for me, in particular during the first 2 hours. If something stressful happens during the come-up the anxiety can be locked in for the rest of the day which sucks. Oddly I relate to the worse intuition a lot. I think people with ADHD learn to rely on intuition quite heavily because it is 'quick' and requires little mental effort. This is great sometimes, but also leads to a lot of careless errors because intuition can lack precision. On meds I'm more inclined to choose to think something through than rely on my intuition.

I've recently been diagnosed with ADHD. When I look back on my life, especially childhood and early adulthood, it really should have been much more obvious. However it never really occurred to me that a lot of my troubles could be caused by ADHD until the last year or two. I thought a lot of the stuff I struggled with was just 'normal' and that everybody was like that. Anyway, I've been prescribed Elvanse (Vyvanse for people in the US) and it will be arriving shortly. I have tried Ritalin and Vyvanse a couple of times before (just as a one off) and based on that I feel like consistently taking these meds has the potential to be life changing. It was truly insane to experience what it is like to have a mind that just does what you want it to do instead of fucking around all of the time and self-sabotaging you at every turn. I also notice that they significantly improved my social anxiety and made me realise that most of my social anxiety came not from shame, but from a lack of ability to trust my mind to pay attention and behave during conversations. I'm curious if anybody here both has ADHD, and has experience with ADHD meds and the effects they had? Thanks

@Leo Gura The company I work for is generating about 6 million per year in revenue with about 15 staff members, we are within the top 10 competitors in our space. Arguably not world class per se but certainly well above a home project. We're in the top 1% of revenue for a UK company. Every single feature we have implemented in the past year was done with heavy AI assistance. Most of our developers (some with 15-20yrs experience and who are world class in the frameworks we use) haven't manually written a line of code in 6 months.

@Leo Gura I respect the point this dude is trying to make, but he is presenting this as a duality of “to vibe code or not to vibe code” which is an (arguably intentional) misrepresentation of the issue. The fact is that the majority of dev teams are incorporating powerful AI coding tools into their systems and workflows now. If done right most of the risks he warns about simply vanish. As a concrete example, the company where I work has quadrupled our rate of shipping features in the past year. This is also despite 2 out of our 8 developers leaving. Those features are some of the most complex we’ve ever built, yet our error/bug rate has not increased substantially as a result of using AI coding agents to build them, in fact it has decreased. Where I agree with him is that coding standards are extremely important. But as long as you define your coding/architecture standards to the AI model once it will largely follow them for the rest of time. When it doesn’t you can nudge it in the right direction. The fact is that these coding agents are incredibly powerful and they are never going to go away even if they only ever stay at their current level of capability. They do not replace developers per se, but they allow companies to operate with a much smaller and more dynamic team of developers than they would have needed before, hence people’s deep concern for the software job market

I agree. There is usually some element of communication, ethics, or a human factor to many jobs that I suspect AI won’t be able to replace anytime soon. But I think you would be surprised how capable these agentic models are. It is leaps and bounds above the AI image generation. Like not even in the same ballpark of impact. Read my description of what these agents can do from earlier in the thread.

@Leo Gura Have you tried out any agentic models like Claude Code? I feel like outside of tech nobody has really seen these models much and what they can do. They are not just chatbots, they’re a tier above. They are extremely autonomous and can perform actions, gather new data, plan and orchestrate stuff etc.

I think we pay $100-200/m per developer, something like that. We're on a subscription, not metered by tokens. Devs here are on average using about $500-$1000/m in metered tokens, although we don't really do much to conserve them. It's relatively reasonable for the value it provides tbh. We have to de-vibe it sometimes, but it's not that much effort. We cycle through with Claude and help point it in the right direction to produce the best output. Sometimes it does stupid stuff, but most of the time it's really good. It helps that we have solid CLAUDE.md files and that we establish strong architectural patterns already, so Claude has lots of good examples. The work has shifted more towards the development process, systems and big picture stuff rather than writing code.

This is a good question. In short, yes, they can. They need oversight but not nearly as much oversight effort as it would have taken to do the task yourself. I can give a very concrete example of how this works for my own field/job which is software engineering. This is what has happened at the company where I work: Before AI adoption Previously we were a team of 8 developers, which was split into two squads of 4. Each squad works on 1 project at a time, so the entire company is working on 2 projects at a time. Those projects would typically aim to take anywhere from 2 weeks to 2 months to complete and would be for small to medium sized features. Occasionally we would have 'large' features which we would set aside 3-6 months for a team to implement. After AI adoption 2 devs left and the company opted not to replace them, so we are a team of 6 now. We have restructured into 'pods' where each of our 6 devs works on a single feature at a time with the help of AI, a human designer + human product manager. The AI agent writes the vast majority of the code and the developer guides it with some oversight and prompting. Usually one other dev will review the work before it is released, but this is probably an average of 30 minutes worth of effort of review per dev per day. We no longer distinguish much between feature sizes. Almost any feature we could want to implement can be implemented within 1, maybe 2 months. We are shipping 6 medium/large features in the same amount of time we would previously have shipped 2 medium features, all while having 2 less devs employed.

I don't know if you've ever seen proper agentic AI working. Most of the chatbots you see embedded in Google, Gemini, ChatGPT etc. are just that: fancy chatbots that are good at writing text. Once you put those models in an agent harness, which allows the LLM to loop, talk to itself, think, use tools, operate autonomously for long periods of time etc. they are incredibly competent. A lot of people who aren't in tech haven't seen these agents work. It's extraordinary. You can give them a task (a single prompt of a few lines of text for what you need them to do) and the process it will follow after this (fully autonomously) could look something like: talking to itself to come up with a very rough overall approach to take brainstorm ideas connect to tens of external services to pull in the information it needs research what the best practices / standards are for this task fire of 5 subagents primed with different skillsets to go off and do some more research in different areas spawn 5 more subagents to adversarially critique or summarise the output of those first subagents compile all of those findings + brainstorming down into a plan execute that plan via connecting to external services or modifying text/code/data in whatever mediums it needs to to achieve the end result Once they have an end result, they'll test it extensively and check that what they implemented works Perhaps it will spawn some more subagents to adversarially critique and verify the end result until it fixes every problem it has found with its own implementation Publish/push/commit/save the end result of that task for you with a tidy summary And all of this takes like 10-30 minutes. To finish a task that may have taken a human a week. It isn't perfect and it will make mistakes, which is why it typically works best when done in combination with someone who is competent in the field. But in the past you would have one senior working with 3 juniors and a mid-level, now you may just have a senior working with one AI agent.