NVIDIA takes the retrieval crown, Zyphra open-sources a brain-signal model, Sakana trains a network without backpropagation, and Google Cloud makes a case for killing RAG. Meet Voxtral: Mistral's full audio stack, built for voice agents
A 380M-parameter masked diffusion autoencoder for scalp-EEG that reconstructs, denoises, and upsamples brain signals. The upgrade: variable-length inputs from 0.5 to 30 seconds (0.125s token grid) instead of fixed 5s windows, plus reconstruction across arbitrary channel/time subsets. Fully Apache 2.0.
So what: One model serving both a snippet and a 30s stretch removes real friction from BCI research. Open-weighting a brain-signal model pushes noninvasive thought-to-text out of closed labs.
Voxtral Transcribe delivers the lowest word error rate of any transcription API. Speaker diarization, word-level timestamps, and context biasing across 13 languages, with recordings up to 3 hours per request. Need live audio? Voxtral Realtime streams sub-200ms, open-weights under Apache 2.0 and deployable on edge.
sponsored
A hands-on build: a stateful agent that runs a venue-operations workflow end to end. MongoDB Atlas for data + vector search, Voyage embeddings for retrieval, and LangGraph to orchestrate the planning / tool-calling loop into a durable graph.
So what: A clean reference for the emerging "database + embeddings + graph orchestration" agent stack — the pattern most production agents are quietly converging on.
sponsored
An open-sourced agent that ditches vector databases and embeddings entirely. Instead it runs continuous LLM consolidation on Gemini 3.1 Flash-Lite — ingesting, connecting, and compacting context on the fly (1M input / 64K output) so the model just remembers rather than retrieving chunks.
So what: If a cheap enough model can consolidate memory in-context, the vector-DB + embedding pipeline becomes optional plumbing. A direct bet that falling inference cost eats RAG — from Google itself.
A biologically-plausible learning rule: Error Diffusion trains Dale-compliant dual-stream (excitatory / inhibitory) networks with no backpropagation. It reaches 96.7% on MNIST and 61.7% on CIFAR-10 using task-dependent credit assignment instead of a global backward pass.
So what: Backprop's global weight transport is the "brains don't do this" objection to deep learning. A local, Dale-respecting rule hitting competitive accuracy is a real step toward neuromorphic-friendly learning.
Hey — new lesson. Last week we covered how AI remembers things across a chat (agentic memory). But what happens when the model was never trained on your data at all — your docs, last week's news, your customer database?
That's the gap RAG fills. Here's the whole idea in one sentence, before any jargon.
|
Lesson 09 · The one idea RAG never teaches the model anything. It finds the most relevant text and pastes it into the prompt — right before the model answers. |
Everything below falls out of that: no retraining, instant updates when a doc changes, and answers that can cite a source.
1 · The closed-book problem
|
Every model has a knowledge cutoff — a date after which it knows nothing. It also never saw your private files. So on its own it's taking a closed-book exam: it can only answer from what it memorized during training. Ask it "What's our refund policy?" and the honest answer is: it has no idea. Worse, a model that isn't grounded will sometimes invent a plausible-sounding one. |
|
✗ Without RAG — closed book Answers from training memory only. Can't see your data. Guesses when it doesn't know. |
|
✓ With RAG — open book Looks up the relevant text first, then answers from it — and can cite where it came from. |
2 · What RAG actually is
|
RAG = Retrieval-Augmented Generation. Three moves: Retrieve the relevant text → Augment the prompt with it → Generate the answer. No retraining. It's a really good search, followed by a normal answer. Everything in the pipeline below exists to serve one job: find the right text to paste. |
3 · The pipeline, step by step
Example: you want AI to answer questions about a 200-page employee handbook.
|
↓
|
↓
|
↓
|
↓
|
4 · RAG vs its neighbors
Three tools, three different jobs. The one-line test is at the bottom.
|
Prompt engineering · cheapest Changes how you ask. Best for pulling better answers out of knowledge the model already has. Like asking a sharper question on the exam. |
|
RAG · updatable Changes what the model knows by handing it external text. Best for private docs, live data, domain Q&A. Swap the knowledge base anytime — no retraining. The open book. |
|
Fine-tuning · most expensive Changes how the model behaves — tone, format, learned patterns. Best for a consistent voice or a specialized output style. Like sending it back to school. |
|
The test: RAG changes what the model knows. Fine-tuning changes how it behaves. They're teammates, not rivals — production systems often use both. |
5 · "Is RAG dead?"
Context windows keep growing — 1M tokens is now the baseline for frontier models, with some reaching 2M–10M. So the recurring claim is: just dump the whole knowledge base into the prompt and skip retrieval. Two reasons that doesn't hold.
|
1. Big windows lose the plot A model's usable context is smaller than its advertised one — retrieval quality tends to fall off well before the limit, and facts buried in the middle get missed ("lost in the middle"). Sending only the relevant chunks keeps the signal high. |
|
2. You pay for every token you send Long context — send the whole doc set (~248K tokens/question): RAG — send only the matches (~8.5K tokens/question): Roughly 29× fewer tokens for the same question, measured on a document-QA benchmark (arXiv, 2606.20898). One vendor puts the end-to-end range at 8×–82× cheaper (LightOn, Nov 2025). The bar is drawn to scale, so the RAG sliver is genuinely that thin. |
|
Verdict: RAG didn't die, it specialized. Long context handles small doc sets in one shot. RAG handles enterprise-scale knowledge cheaply and precisely. The strong systems use both. |
6 · How RAG evolved
|
Classic RAG · coined 2020, mainstream 2023–24 Chunk, embed, search by similarity, generate. Simple and effective, but weak on multi-step questions that need to connect facts. |
↓
|
GraphRAG · Microsoft, 2024 Organizes knowledge as a graph of connected entities instead of flat chunks. Better at "how do these things relate?" questions that span many documents. |
↓
|
Agentic RAG · 2024–25 An agent decides what to retrieve, when, and whether the results are good enough. It can rephrase the query, search multiple sources, and iterate — RAG meets agentic memory. |
|
Key takeaways → RAG doesn't retrain the model — it pastes the right text into the prompt. → Pipeline: chunk → embed → store → retrieve → generate. → RAG adds knowledge; fine-tuning adds behavior. Use both. → It grounds answers in real sources, so it hallucinates less and can cite. → Big context windows didn't kill it — cost and retrieval quality keep it alive. |
|
When to actually reach for RAG Any time the knowledge changes — company policies, product docs, news, live data. You update the knowledge base and the answers change instantly, with no retraining. If you instead need a consistent style or format, that's a fine-tuning job, not a RAG one. |
|
Marktechpost · AI Dev Signals — a short course for AI engineers. Numbers verified against primary sources, July 2026: GraphRAG (Microsoft, 2024), RAG (Lewis et al., 2020), token/cost figures (arXiv 2606.20898; LightOn). Context-window and cost figures move fast — the "is RAG dead" section will date first; re-check before reuse. |