Developing projects with AI

[WhyFhy]

Thanks to Elon's greed, the era of people using AI to vibe code projects will be ending soon. 

Now that SpaceX has an IPO they need to make money for shareholders.  They are going to help this by raising the price of their Grok service a good ten or twenty fold.   Businesses that have grown to rely on these subscriptions will continue to pay - the rest of us will be limited to trial/minimal usage.

And before you think you'll just move to ChatGPT or some cheaper service, they will all be doing this together.   OpenAI needs a bailout.   Grok needs a AED.   Just like Uber, the days of their subsidization is over. 

This was always the direction where AI was heading though. All these companies didn't spend hundreds of billions of dollars so that they could give us free/cheap service. They wanted to make a shitton of money from it and become powerful and control the global market.

It was like the drug dealers who give some poor bastard a sample. They don't do it because they like him They want to get him hooked and addicted so that they can have a consistent source of revenue and become filthy rich...

And as long as there is no proper competition going on (there is Chinese stuff like DeepSeek but it is not enough) they can charge whatever they want and prices won't break.

Closed lab moats have to stay on their toes as open-source agentic tooling, tokenization efficiencies, and WebGPU compute steadily bridge the capability gap. As open weights and local bare-metal setups mature, they offer a grounded counterweight that makes centralized API lock-in much harder to sustain in the long run.
​Did you see Manus before Meta acquired them? Aside from them having to sell it back and stuff, Manus actually counters your entire argument. Big lab saw a Chinese model threat and immediately whipped out the checkbook, then the Chinese govt clawed it back (Meta still dropped the ball with it imo).
​Based on how the attention mechanism works, simply observing features and providing shape can result in whatever you want nowadays. I can observe a frontier model's capabilities and have a skeleton of that feature built locally based on the shape I see. I can browse an arXiv whitepaper for a concept or philosophy on agentic frameworks and have it deployed before the guy who wrote the thesis.
​These features can't really be patented, copyrighted, trademarked, etc., because they are essentially mirrors of user intent based on prediction engines. My 2 sats.

[Vod]

Just like Uber, the days of their subsidization is over. 

This was always the direction where AI was heading though. All these companies didn't spend hundreds of billions of dollars so that they could give us free/cheap service. They wanted to make a shitton of money from it and become powerful and control the global market.

I'm surprised the prices haven't shot up yet.   The bubble keeps growing and it's not supported by usage.

I recently switched from Kimi back to Gemini.  The western world AI is much better for me.  Not only is the speed much faster, but it's better trained on the applications and services I use.   I'm certainly not loyal to any one...
a) I started with OpenAI
b) I moved to DeepSeek
c) I moved to Gemini
d) I moved to Kimi
e) I moved to Gemini

I know in the next couple years we'll see one model from each side start to fight for supremacy.  What I don't know is what happens to the internet itself (and bitcoin) during/after this fight.   If I don't post here again, is it by my choice or AI?

[asvsfs]

Just like Uber, the days of their subsidization is over. 

This was always the direction where AI was heading though. All these companies didn't spend hundreds of billions of dollars so that they could give us free/cheap service. They wanted to make a shitton of money from it and become powerful and control the global market.

I'm surprised the prices haven't shot up yet.   The bubble keeps growing and it's not supported by usage.

I recently switched from Kimi back to Gemini.  The western world AI is much better for me.  Not only is the speed much faster, but it's better trained on the applications and services I use.   I'm certainly not loyal to any one...
a) I started with OpenAI
b) I moved to DeepSeek
c) I moved to Gemini
d) I moved to Kimi
e) I moved to Gemini

I know in the next couple years we'll see one model from each side start to fight for supremacy.  What I don't know is what happens to the internet itself (and bitcoin) during/after this fight.   If I don't post here again, is it by my choice or AI?

Still the most reliable for coding is Claude, my team only using it.
GPT Astra is insane just came out. but I would still stick to Claude for reliable coding, for personal project Gemini Flash might be good enough.

[Vod]

Still the most reliable for coding is Claude, my team only using it.
GPT Astra is insane just came out. but I would still stick to Claude for reliable coding, for personal project Gemini Flash might be good enough.

The AI companies are offering a lot more for their subscriptions.   Gemini also gives me 5TB of Google Drive space, YouTube Premium and Google Health Pro among other services I already use.  It's just a bonus they are so much faster and cheaper than the Chinese ones (right now)

[pooya87]

Claude seems to be performing better than others for solving complex problems. So far in my experience OpenAI is sometimes garbage. For example I tried to write a music piece a while ago with the help of 3 AIs with a full orchestra and everything. Claude performed fantastically while OpenAI just produced garbage. DeepSeek was good too but the result was simpler (I think its training data is lacking).

When it comes to subsequent fixes, again OpenAI was garbage (it kept deleting instruments!) while DeepSeek and Claude both performed correctly.

[bitmover]

I recently switched from Kimi back to Gemini.  The western world AI is much better for me.  Not only is the speed much faster, but it's better trained on the applications and services I use.   I'm certainly not loyal to any one...
a) I started with OpenAI
b) I moved to DeepSeek
c) I moved to Gemini
d) I moved to Kimi
e) I moved to Gemini

I know in the next couple years we'll see one model from each side start to fight for supremacy.  What I don't know is what happens to the internet itself (and bitcoin) during/after this fight.   If I don't post here again, is it by my choice or AI?

I use the paid version ot Gemini (Pro)
For coding, they allow me to use Gemini Flash 3.8, Claude Sonnet and Opus and GPT OSS

For coding, Claude Opus is the best, and the second best os Claude Sonnet . They are better in everything, but a little bit slower. But they solve the problem when it is complex.
Unfortunately,  they are very expensive. My tokens ends pretty quickly. This is why i try to use flash for the basic stuff first.

AI allow me to code much faster, and to accomplish things I would have to research a lot to do.

[BogdanXI]

That matches my experience: model choice matters, but I think there is another problem that becomes more visible as a project gets larger. Even a strong model can make a locally reasonable change while losing track of decisions made earlier in the project.

If an AI-assisted project is developed across many sessions, how do you verify that a later change is actually based on the intended project state rather than incomplete or lost context?

Git preserves the files and history, but does it provide enough evidence about how the change was produced and verified?

[nobswebdev]

I am using AI for coding and it is awesome. I can do 10x more than before AI era. And 10x is when I manually review all changes to match my coding standards. If I would just wibe code everything then productiveness would be 100x. The problem with wibe coding is that AI produce slop code. It might work for small application, but if you want to build bigger system technical debt would grow over time and AI will lost itself at some point with their own slop. That make me more optimistic about future of my job, because this human touch is still needed.

[Vod]

The problem with wibe coding is that AI produce slop code.

That's by design.  "fully give in to the vibes, embrace exponentials, and forget that the code even exists"

Allows you to quickly write something as proof of concept but it doesn't bother you with details.  But if you want code written exactly your way, you must train your AI properly and then watch it each step of the way.  Each AI has methods to control how often each is needed.

[BogdanXI]

I agree. I also think human review is still necessary. But this raises another question for me: even if a human reviews every change, how do we preserve the state of the project itself, rather than just its files?\n\nImagine a project going through hundreds of AI sessions over several months, using different models and possibly different developers. Git shows what changed, but it does not necessarily show which decisions were considered authoritative, which assumptions were verified, or why a particular project state was accepted as correct.\n\nSo I am wondering: for large AI-assisted projects, is Git + human review actually enough, or do we need a separate mechanism that records and verifies project state independently of the current AI session?