Infinite-parameter LLMs generate weights from live data
A Cambridge group's swap for context-stuffing and RAG. Also: Intel's ternary trick speeds local inference 1.28x, and Gemini 3.8 Live ships voice agents.
Copy markdown
Weights that learn as the chat grows
A compact hypernetwork turns the data you feed at runtime into a low-rank modulation of a frozen base model, with a Bayesian belief that sharpens across turns — the authors (including Cambridge's José Miguel Hernández-Lobato) report it beating in-context learning and retrieval when the evidence is long and multi-hop. No code yet, but it sketches a route past the context-window ceiling you fight every day.
Ternary weights break the 1.58-bit floor
Intel's BITCOS packs {-1, 0, +1} weights at 1.485 bits each — under the usual 1.625 — by exploiting that up to 51.5% of ternary weights are zero, winning on 26 of 29 models with a 1.28x matrix-vector speedup and 1.18-1.27x end-to-end throughput on CPU and GPU. Smaller, faster local models on the hardware you already own.
Gemini 3.8 Live makes voice agents a primitive
Google pushed gemini-3.8-live (plus an extended-thinking variant) to GA: audio-to-audio with interleaved reasoning and async function calling over the Live API. It's now the default building block for low-latency voice agents — no more stitching STT, an LLM, and TTS together yourself.
Claude Code 2.1.274 warns before a session dies
The latest release adds visible critical-memory warnings with recovery steps, self-healing for corrupted transcripts, and a stack of MCP connection and timeout fixes across desktop, VS Code, web, and Slack. If long agent runs have been silently stalling on you, update.