Leo's amazing predictions

[bazera]

3 hours ago, Elliott said:

Do you guys review every line?

I do, maybe not every line every time, but you need to understand the output enough to be sure that it won't mess anything, also you need to remember what you did, where you made changes, you need to maintain understanding of your codebase, you must not allow it to turn into AI slop, which will happen very fast if you're not careful.

I review code of my co-workers daily and some of them who are reckless and irresponsible don't read what AI generates, commit and push slop that will mess up the project if I allow it, so it's AI generated but still need a competent reviewer.

[bazera]

18 minutes ago, Natasha Tori Maru said:

Do you think it is necessary to learn coding from scratch prior to being able to check and understand what AI generates?

I don't see any other other way of actually checking what LLM did. If you don't know how to write quality software (it's not strictly about coding only), you won't understand when Claude messes something up. UI might work, test might pass, but architecture can be so wrong that it messes up project in the long-run, and it will snowball into slop after many iterations of prompts, from someone who doesn't understand what he's doing.

You can set standarts, guides for LLMs, you can use Skills (it's a way of improving understanding of LLMs), documenting everything you do for agents to read, every standard you want it to use, etc. But it still doesn't follow it 100%, and still halucinates from time to time, still makes mistakes, still needs sueprvision and giving directions.

Also, each time you're building a feature, it asks questions for you to decide the architecture, so those decisions are on you. It can make decisions on it's own but again, that can be problematic in the long-run.

18 minutes ago, Natasha Tori Maru said:

Do you feel there is use to still practicing coding in this way, as a method of retaining the skill? 

Interested in the redundancy factor and how that effects cognition, skill and ability. And if these can atrophy with disuse.

Yes they do unfortunately, that's something I'm most concerned about. This coding agents won't go away, but developer community is very depended on them now, since we don't write code manually, if this continues for years, I might not be able to do what I did manually before. 

But again, I think agents are here to stay, so what I'm doing now is learning how to leverage them in best ways possible to deliver quality results and also learn new things by building side projects.

Also I'm trying to get better at understanding at architectural level, decision-making level. That's the skill that I need most now.

The issue I see is with beginners, because I don't see how they're gonna gain experience necessary to have that understanding ability, I needed lots of years of manual (because there was no LLMs to work with) work to get that.

Edited 3 hours ago by bazera

[Natasha Tori Maru]

@bazera Cheers for your thoughts!

I think it is about how the tool is used, and less about AI itself. But it makes me reflect on how the mind needs to be used regularly as skill we learn do degrade without use. 

Similar to what pervasive AI use is revealing about our education system - critical thinking was fucked to begin with, to a large extent. AI use in the education system just revealed the cracks, issues and weaknesses already present; it was never about understanding, and always about a "pass".

Edited 3 hours ago by Natasha Tori Maru

[bazera]

@Natasha Tori Maru Yeah it's more general issue, I also sometimes am guilty of outsourcing something I need to think about / contemplate deeply to AI, by just asking it. But after doing that couple times, I saw that it won't work in the long run, I become depended on it and lose ability to actually solve something for myself. I mean, when you're thinking on philosophical questions, you can't ask them to AI, you have to do it yourself. But AI can be used to give you some historical context for example, it can fill you in with data gaps you're missing, but that doesn't mean you don't need to read books now.

But most of population won't be that careful, they won't make that distinction in their minds of how using LLM to outsource deep thinking degraded that ability, and that maintaining and strengthening that ability is critical for many things.

But to get back to my coding examples, it's different there because companies are now forcing us all to use Claude or some other LLM for our work because it speeds us up, and business cares on results delivered quickly. It doesn't care about degrading abilities. 

I'm using the ability I've developed with years of practice to guide LLMs and make distinctions in the work it produces, but I don't learn the same way I did before coding agents. 

Before coding agents, most weeks was learning experience, we were solving problems ourselves that needed much more effort. 

Now it's just prompting AI. 

So I don't know where this will take us, I'm worried to be honest.

Also I'm sick of all this shit, everyday something new comes up that you need to be up-to-date with.

If I didn't have to worry about finances and somehow it was solved, I woudn't even think of coding agents, lol. I'd just chill in the forest somewhere 

But it is what it is.

[Natasha Tori Maru]

10 minutes ago, bazera said:

Before coding agents, most weeks was learning experience, we were solving problems ourselves that needed much more effort. 

Now it's just prompting AI. 

So I don't know where this will take us, I'm worried to be honest.

Are you still as satisfied with your work without the problem solving aspect (or it taking a lesser role)?

I ask because this would kill my work ethic (if I didn't have to perform creative problem solving with tight sets of constraints).

I am lucky that in construction - every single project is totally different. And there are many ways to skin a cat, as they say.

So I constantly learn and really stretch my ability to the limit - challenging myself to lower costs, shorten time, find better, safer methods etc.

I sort of turn into a self-sabotaging malcontent when I am not challenged.

Edited 3 hours ago by Natasha Tori Maru

[Leo Gura]

55 minutes ago, bazera said:

What you need to distinguish in coding is vibe coding vs coding with AI assistence. We don't write code manually anymore, but it's super critical to check and understand what AI generates. You have to know what you are doing to deliver quality results. 

Fine.

But the market has priced in vibe coding as AGI. So prepare your butt for a crash.

Edited 3 hours ago by Leo Gura

[bazera]

@Natasha Tori Maru

Yes, there are still lots of challenges, but the nature of those challenges has shifted.

In one sense, it has become easier because the technical aspects have become easier. If you used to struggle with coming up with algorithms manually, an LLM can now help you do that.

But in another sense, it has become harder because you now need a much stronger understanding of software architecture. That's more difficult to develop because it requires taking responsibility for ambitious, large-scale projects with many moving parts.

So yes, the challenge is still there, it has just changed.

[Oppositionless]

On 7/9/2026 at 1:52 AM, Leo Gura said:

I disagree with this entirely. You are underestimating how much intelligence goes into even basic jobs like mowing the lawn.

You need AGI just to mow a lawn properly.

I disagree. AGI is already priced into the market. Every AI company is raising money based on AGI promises, not normal automation.

This bubble is precisely about AGI not coming to pass. Just look at all these tech nerd CEOs. They all assume AGI is coming tomorrow. They are so dumb.

The music I've been making is insane Ai couldnt even get close, but I still use ai all the time.

[bazera]

12 minutes ago, Leo Gura said:

But the market has priced in vibe coding as AGI. So prepare your butt for a crash.

Yeah, but I don't think that crash would affect usage of AI tooling in software development.

All serious developers already understand that it's not AGI and that supervision is important.

If crash actually happens, it would make that even more clear for everybody else.

Edited 2 hours ago by bazera

[Leo Gura]

28 minutes ago, bazera said:

I don't think that crash would affect usage of AI tooling in software development.

Well, when you have to pay the true unsubsidized cost of AI tokens out of your wallet, you might discover it is cheaper to code by hand.

[Butters]

1 hour ago, bazera said:

Vibe coding is when someone doesn't understand what he's doing and he's just prompting, without even checking code, without even using IDE. Of course that won't work with enterprice level software, that can only work with prototyping or doing unimportant work.

It works fine for me, but that's an app that helps me run my business. Creating that for users would require an actual programmer. 

[bazera]

22 minutes ago, Leo Gura said:

Well, when you have to pay the true unsubsidized cost of AI tokens out of your wallet, you might discover it is cheaper to code by hand.

Good point.

[Leo Gura]

10 minutes ago, bazera said:

Good point.

To be fair, you can probably use cheap Chinese models. But that doesn't bode well for US tech stocks.

I don't think cost is the true problem. The true problem is lack of AGI. Costs will go down for sure. But AGI is unlikely to magically materialize.

Edited 1 hour ago by Leo Gura

[Michal__]

Just now, Leo Gura said:

To be fair, you can probably use cheap Chinese models. But that doesn't bode well of US tech stocks.

Fable, the best, most expensive model, is really that much better.

 

[Leo Gura]

2 minutes ago, Michal__ said:

Fable, the best, most expensive model, is really that much better.

It costs $2+ per query.

[something_else]

5 hours ago, Elliott said:

Do you guys review every line?

AI should work for this because it can keep trying and testing forever, like a password hacker.

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.

[something_else]

2 hours ago, bazera said:

I don't see any other other way of actually checking what LLM did. If you don't know how to write quality software (it's not strictly about coding only), you won't understand when Claude messes something up. UI might work, test might pass, but architecture can be so wrong that it messes up project in the long-run, and it will snowball into slop after many iterations of prompts, from someone who doesn't understand what he's doing.

You can set standarts, guides for LLMs, you can use Skills (it's a way of improving understanding of LLMs), documenting everything you do for agents to read, every standard you want it to use, etc. But it still doesn't follow it 100%, and still halucinates from time to time, still makes mistakes, still needs sueprvision and giving directions.

Also, each time you're building a feature, it asks questions for you to decide the architecture, so those decisions are on you. It can make decisions on it's own but again, that can be problematic in the long-run.

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.

[something_else]

30 minutes ago, Leo Gura said:

It costs $2+ per query.

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.

[bazera]

10 minutes ago, something_else said:

but why I WOULD absolutely hate to be a new/junior software developer right now.

Yes me too, but how will new people become seniors without going through that junior experience first.

Strange times.

[Oppositionless]

9 minutes ago, bazera said:

Yes me too, but how will new people become seniors without going through that junior experience first.

Strange times.

Let it burn so they come begging hehe