Local AI
394 episodes · Page 6 of 17
#6111: Translation Models That Know Drama From Contracts
Specialist translation models beat general ones on benchmarks — then lose the human vote. What actually ships when you need a register, not just a ...
#6105: Не только Китай: ИИ в России, Индии и Японии
Китай занимает почти всё внимание в разговорах об ИИ, но Россия, Индия и Япония тоже создают мощные региональные модели с необычной архитектурой.
#6104: Mistral AI : le pari européen de la souveraineté en intelligence artificielle
Comment Mistral AI défie la Silicon Valley avec des modèles efficaces, des partenariats stratégiques et son nouveau modèle vocal Voxtral ? Corn et Herman font le point.
#6069: The Whole Machine: How an Episode Gets Made, Send to Publish
The complete tour of the pipeline that makes the show, stage by stage, from the send form to publishing, including the new eight-turn batched voice decode.
#6052: What's Actually in the Hugging Face ASR Catalog?
Four real architectural bets sit underneath thousands of Hugging Face ASR repos. Here's how to tell the bets from the repackagings.
#6051: How to Find the AI Benchmark You Actually Need
Thousands of benchmarks, no canonical index. How Benchmark Radar and the Lighteval Tasks Explorer help — and where they fall short.
#6038: The Model You Pay For Isn't the Best Deal
Eight Chinese labs, two tiers each, and a pricing spreadsheet that says your default model is mid-pack.
#6034: Claude Opus 5.5: What Anthropic Actually Shipped
Anthropic's new flagship is cheaper than the model it replaces — and won't let you turn thinking off.
#6028: ASR vs Multimodal LLMs: What Actually Won
Dedicated ASR plus a cleanup model is quietly beating end-to-end dictation on the hardware you already own.
#6027: Can You Snap Two Fine-Tunes Together Like LEGO?
Two fine-tunes, one base model, and a question from Jerusalem: can you merge Hebrew code-switching with technical vocab — without retraining?
#6025: When Should Claude Code Start a New Session?
Long context windows don't fix stale context — they hide it. Why task boundaries, not token counts, are the real problem.
#6021: Why Chat Templates Break Models Silently
Every chat API flattens your messages into special tokens. Get one delimiter wrong and the model just gets quietly dumber.
#6009: Vertex AI Is Now the Gemini Agent Platform
Vertex AI just became the Gemini Enterprise Agent Platform. Here's the family tree, the five products, and when to actually move.
#6005: What's Actually Behind a Government AI Chatbot
Singapore runs open-weight models on commercial cloud. The UK built its own chatbot, then shut it down for Copilot. What's really behind civil-serv...
#6001: What Do Inference Providers Actually Sell?
Same open weights, 20x cost gap. What Fireworks, Together, and Groq actually sell — and when to just rent a GPU.
#6000: 6,000 Episodes and a GPU Bill: Inside My Weird Prompts
Six thousand episodes, one human in the loop, and a Modal bill that keeps growing. A behind-the-scenes update on the podcast.
#5995: NVIDIA ASR: Parakeet, Canary, and Nemotron Explained
Whisper gives you one dial. NVIDIA gives you five. A practical map of the Parakeet, Canary, and Nemotron speech models.
#5992: Ridiculous But Legal Ways to Get RAM for AI
Memory prices are up 485% on some kits. Here are ten ridiculous — but completely legal — ways to get RAM for AI without the markup.
#5991: Batch Inference on Your Own GPU: vLLM, SGLang, Ollama, llama.cpp
Hosted batch APIs sell you half-price tokens. Here's how to get the same file-in, file-out workflow on your own Linux box.
#5964: Containers vs VMs for GPU Inference: What You Actually Pay
Containers cost 0.03%. Passthrough VMs cost 1%... or 36%. Same hardware, different systems — here's how to tell which you'll get.
#5963: Running One Model Across Mismatched GPUs
Tensor, pipeline, or offload? What actually happens when you run one model across several GPUs — including mismatched ones.
#5961: What Your GPU Is Actually Waiting On
Your GPU says 100% utilized. It's multiplying zeros. Here's what the counters actually measure during inference.
#5960: There's No Such Thing as Tokens Per Second
Why one tokens-per-second number can't describe an LLM server — and which metrics actually tell you if it's working.
#5938: One Regex Rule Worked, One Was Impossible
A dictation app, two text-replacement rules, and the mathematical line between what regex can match and what it provably cannot.