Network Automation Nerds

The episode contains a handful of genuinely useful technical details - GPU memory math for fine-tuning, the LoRA/QLoRA distinction, and the RLVR vs RLHF philosophical split - but they are diluted by lengthy biographical storytelling, basic concept re-explanation, and host filler that occupies the majority of the runtime.

“for 8 uh, billion parameters we need 16 gigabytes GPU only for the weights. For the weights of the model we need to factor in also the KV cache”

“They developed a method uh, called reinforcement learning with verifiable results. And they applied this method for domains where the results can be verified, which is math and code”

The networking-technology evolution parallel to AI evolution (STP→VXLAN EVPN mirroring prompting→RAG→agents) is a mildly interesting framing, and the DeepSeek RLVR contrast is substantive, but everything else - attention mechanism, RAG, fine-tuning rationale - is standard circulating content with no contrarian or first-principles angles.

“we had SP protocol uh, which solved a very clear problem within networking, redundancy and broadcast etc. And then uh, um. So it solved a problem in the first place and then it came with its own limitations”

“the Chinese researchers they uh, used a different approach. They said we don't need to seal the model capabilities and we need to um, let him find the best answer”

Eduard is a credible practitioner - 25 years in networking, NATO engineer background, actually fine-tuned multiple models, and participated in a group that built an LLM from scratch using Raschka's curriculum - but he is an early-stage founder with limited public track record and no verifiable production-scale deployments discussed in the episode.

“I um, was part of uh, a discord channel where a group of data scientists and practitioners, um, we studied together and we build this model”

“I build a small um, it's a contradictory terms from scratch. So I know how to build it”

The GPU memory calculation (8B params × 2 bytes = 16GB weights, ~24GB total) and the explicit naming of Google Colab, AWS, LoRA/QLoRA, Code Llama, and Qwen provide concrete anchors, but there are no benchmark numbers, no before/after fine-tuning metrics, no dollar figures, and no named customer deployments to substantiate claims of model improvement.

“for 8 uh, billion parameters model open source which is uh, it's a good one. I mean it has a lot of data. Ah uh, within this 8 billion parameters uh, we need let's say to be safe 24 gigabytes GPU which is available for free for example in Google Colab”

“code llama from meta was used. They used for training uh programming languages. So which means that natively this pre trained model, this baseline model uh knows um programming”

The host occasionally elicits useful clarification (e.g., pushing on reasoning vs. general models, fine-tuning vs. retraining, open-weight vs. open-source) but repeatedly opts for self-deprecating agreement instead of probing, never challenges any unverified claims, and allows the conversation to drift into lengthy biography with minimal redirection.

“I'm just nodding and agreeing without checking your math but I trust you”

“You talk very slow, but, uh, I think it's very thoughtful responses and very solid, uh, knowledge”