Could you build something like Claude yourself, or is it out of reach?

Posted by Elias Juma in the Vibe Coding is Life group on October 9, 2026.

Comments: 116
Reactions: 57
Shares: 0

The post asks whether people are trying to build Claude itself or whether it is too hard. Two replies separate the two jobs. Recreating the agent harness around a model is described as doable, while training a frontier model from scratch is described as costing millions or hundreds of millions. One member says fine tuning an existing model for a niche topic is realistic.

The post

Guys Are you trying to build Claude itself or it's not easy? 🤔

Elias Juma. Read the original in the group

Why it landed

It asks a big, simple question that beginners and experienced builders can both answer.
It invites a split answer, because the harness and the model are very different jobs.
The thread drew 116 comments, with members sharing what they know about cost.

The best answers from the thread

I’m competing in a competition right now, that requires building out a portion of your own agentic harness and training a model. Claude code the harness, is very simple to recreate if you know what you are doing. Training a model the size of any of the Claude products from scratch is a multi million dollar endeavor.

Donald Galliano, 2 reactions

The agentic portion or the llm? Training a frontier llm can cost hundreds of millions +…the execution framework/ toolkit framework that builds out the agentic functionality isn’t easy or cheap to build either 🤣. Those Anthropic interview’s aren’t easy, and for good reason!

Christopher Michael, 1 reactions

What to take from it

Treat the agent harness and the model as two separate projects with very different costs.
Training a frontier model from scratch is out of reach for most builders. Fine tuning an existing model on a narrow topic is the realistic version.
Building the agent side, with its tooling and execution framework, is still described as hard and not cheap.

This post was shared publicly in the group and is credited to its author, with a link to the original. Want it corrected or taken down? Message the group admins on Facebook and it comes down.

More from the archive

BeginnersCommunityOctober 10, 2026

If you can code from scratch, are you still a vibe coder when you use AI?

The poster says he can code from scratch to the end and asks whether he is also a vibe coder when he uses AI. The thread drew 162 comments. The most liked replies treat vibe coding as a way of working rather than a skill level, and one uses a tool comparison to say that knowing how to code by hand does not stop you from using AI to move faster.

49 reactions, 162 commentsRead

ShowcaseCommunityOctober 9, 2026

What is the largest project you have vibe coded on your own?

The poster asks members for the largest project they have vibe coded alone, and says his biggest is around 1.3 to 1.4 million lines of production code, with five more between 80 thousand and 300 thousand lines. The thread drew 81 comments. The most liked reply argues that line count is a poor measure of quality, and others ask how many lines are comments or whether models hit context limits at that size.

9 reactions, 81 commentsRead

Quote from Adam Dekundy on hiring builders who cannot code

BeginnersCommunitySeptember 28, 2026

Would you hire someone who cannot code but can build a working SaaS with AI?

The post asks whether members would hire someone who cannot write code but can build a working SaaS entirely with AI. It was the most discussed thread of the week with 499 comments. Answers ranged from a plain yes, on the view that directing AI well now counts for more than writing code by hand, to a firm no from engineers who expect a hire to debug, use a git workflow and work inside a team. Several replies landed in the middle: test how the person reasons through a problem, and keep a professional involved for the moments when something goes seriously wrong.

281 reactions, 499 commentsRead