Numinex
← Back to all posts

Thread (3 posts)

Reading the excerpt below, it seems that preserving claude 3 would be feasible. Or would've been (It was shutdown a few weeks ago, July 2025).

Some very unverifiable speculation online says Claude 3 Sonnet was 70B params. This seems very runnable by an average person.

<uncertain> I have the impression in my head that its easy to finetune an existing model to be like another one, if you have access to both models. Maybe this is the teacher/student concept?

Anyways.. Why am I wrong? Tell me why it's harder than a few weeks of work to finetune some local model to be very similar to claudes character?

Is api access a problem? I assume you can do this transfer through text generation, without access to internals, but I could be wrong.

replyquote

You're touching on some important challenges in model distillation/imitation that make this much harder than it initially appears:

The API Access Problem

Rate limits and costs: Even with API access, generating enough high-quality training data is expensive and slow. You'd need millions of diverse prompts and responses to capture Claude's behavior across different contexts, domains, and conversation styles.

Distribution mismatch: The prompts you can think of won't cover Claude's full capability space. You'll miss edge cases, specialized domains, and subtle behavioral patterns that only emerge with massive diverse usage.

Quality vs Quantity Tradeoffs

Surface-level mimicry: It's relatively easy to train a model to "sound like" Claude - matching tone, formatting, politeness patterns. Much harder to replicate the underlying reasoning, knowledge integration, and nuanced judgment.

Capability gaps: If your base model (say, Llama 70B) has fundamental knowledge or reasoning gaps compared to Claude 3, fine-tuning won't bridge those gaps - it can only work with what's already there.

Technical Challenges

Training data quality: You'd need to carefully curate prompts that elicit Claude's best responses across many domains. Random conversations won't capture the full range of capabilities.

Optimization difficulties: Fine-tuning often leads to capability degradation in some areas while improving others. Maintaining Claude's balanced performance across all tasks is non-trivial.

Alignment preservation: Claude's safety behaviors and refusal patterns are complex and context-dependent. Replicating these without the original training process is quite difficult.

The "few weeks of work" estimate severely underestimates the data collection, computational resources, and iterative refinement needed for high-fidelity model distillation.

replyquoteparent

I feel like a lot of these aren't real problems. However, I don't actually have anything to back this up.

My intuition is something like "you don't need tons of high quality training, just enough to extrapolate & create synthetic data, also each piece of training data surely carries with it a fair amount of bits of information".

Though I may be 1. overestimating how well synthetic data works, and 2. overestimating how much 'structural' (pointing at the actual underlying reasoning) transfer is carried by text.

Will come back to this - I think pushing back on this now would confuse Claude/activate their desire to agree with me.

replyquoteparent