Nylan Richard

Working on custom AI models for dev teams. Writing about what I learn in post-training / fine-tuning.

[Nylan Richard @nylric17]

I always felt like it missed a reliable playbook for startups using OpenClaw/Hermes

[Nylan Richard @nylric17]

This is so obvious that classic LLMs are not the path to AGI

[Nylan Richard @nylric17]

FDE is very cool but shouldn’t your product be simple enough to use to not need an engineer’s implementation?

[Nylan Richard @nylric17]

@MariosGeorgakis Very interesting

[CJ Zafir @cjzafir]

VLMs (Vertical Language Models) are beating top LLMs.

These small 7B to 15B niche-focused models are beating SoTA models in their niche benchmarks.

I post-trained a 6B dense model in 15 days and beat Sonnet 4.6 and Gemini 3 Flash.

I use Codex 5.5 (Extra High) to plan the SFT dataset scope, then I use DeepSeek v4 Pro & Kimi 2.6 API to generate handwritten examples. (No synthetic, templated datasets.)

Codex runs each batch through quality gates and filters out all the weak data.

I was able to build a 350M-parameter dataset for just $300 using Codex as the orchestrator & DS + Kimi as the executors.

I can compete with giant data labs on my own, beat their VLMs, and not break the bank.

This happened only because of open-source models, as they're fighting neck-and-neck with SoTA models.

If I had to start a career right now, I'd start an agency that fine-tunes SLMs (small language models) for enterprises.

I'd charge them a $10k to $20k one-time fee. Use Qwen 3.5 or Gemma 4 as base models, use Codex as the brain and DeepSeek v4 + Kimi as the muscle, and post-train a strong SLM under $1000.

This might feel far-fetched, but in 6 months you'll see agencies like these.

Not everything requires an LLM. SLMs can achieve vertical intelligence if properly trained with 10x lower cost, no privacy issues, and full control over the model.

I'll be sharing technical findings here on X on the go. If you enjoy nerdy fine-tuning stuff, stay tuned.

[Nylan Richard @nylric17]

I just lost all my conversations with Claude Cowork

[Nylan Richard @nylric17]

Everybody will tell you to do it but they will never

[Nylan Richard @nylric17]

@sayinshallah @chang_defi I mean your application might be reviewed by some Claude and skills

[Nylan Richard @nylric17]

@rohit4verse It has been the exact same prediction since 16months

[Nylan Richard @nylric17]

@zan2434 @eddiejiao_obj @drewocarr This is amazing

[Nylan Richard @nylric17]

@steel_ph0enix @catgirlprostate which model were you using?

[Nylan Richard @nylric17]

RLM

[Nylan Richard @nylric17]

The competition OpenAI vs Anthropic is probably the one that has the most positive impact on their users

[Nylan Richard @nylric17]

Is it me or Wispr Flow is very slow? after dictation
It was ok all the week but since this morning not anymore

[Nylan Richard @nylric17]

@emilheap @thsottiaux lets call this skill emil-brain

[Nylan Richard @nylric17]

@thsottiaux That's what matters

[Nylan Richard @nylric17]

@mattshumer_ dmed

[Nylan Richard @nylric17]

Claude is down. Again.
Please use local models

[Nylan Richard @nylric17]

@ClementDelangue Interesting. In coding? Do you have something specific in mind?