Top developments
OpenAI's models are now listed on Google Cloud, in Gemini's Model Garden via the Cloud Marketplace. Frontier models plus Codex programming workflows are available with standard Cloud billing, and purchases can draw down existing Google Cloud spending commitments (Reuters). This is less a capability change than a distribution one, and a sign of the hyperscaler truce era where rival clouds sell each other's models. It is genuinely useful if you deploy on Google Cloud, but it does not change what the models can do.
Kimi K3.1 Pro's promised weights are still unpublished, and the license is still unknown. As of this morning there is no K3.1 Pro repository on Moonshot's Hugging Face org (the latest repo remains Kimi-K3, updated Sept 2), and the Nov 4 date is a company statement, not a shipped artifact (Geotoolbox). The weights, not the date, are the operational reality for anyone planning to self-host or verify the benchmarks, so treat Nov 4 as a deadline to watch, not a done deal. This is a follow-up to yesterday's top story.
Frontier AI labs
No new frontier model releases, and the Oct 28-31 tracker pegs for Anthropic, OpenAI, and Mistral remain unconfirmed projections, not announcements. Anthropic's newsroom shows nothing pending beyond Oct 22 posts, so there is no announcement to cite yet. Worth watching next week, but nothing to act on yet.
MiniMax and DeepSeek have registered AI exports under China's new compulsory registration regime for 23 categories of AI technology, a week after Beijing made registration mandatory (SiliconNoon). This is a regulatory signal for open-weight distribution, not a technical development.
AV / embodied-AI research
Zoox is now running robotaxi service to Las Vegas airport, the second company after Waymo to operate at a US airport. Early user reports put fares roughly at Uber and Lyft levels ($23 to the MGM Grand, $26 to the Bellagio), and the service covers both airport terminals (Aviation Pros). This is another deployment milestone in what is becoming the most concentrated robotaxi market in the US, with Tesla, Uber, and Waymo all holding Nevada permits. It is still a deployment step, not a research result.
Quiet on the paper front: the arXiv driving feed shows nothing new since the Oct 22 batch already covered yesterday, and neither Bluesky account posted anything new in the window. No padding: the Oct 22 batch remains the current reading list.
Research/papers worth reading
Nothing new in this window. If you have not read them yet, yesterday's batch stands: TerraTransfer++ (demo-free end-to-end planner, 59.4 EPDMS on NAVSIM navhard), DriveMLP (lightweight MLP planner, 86.51 on nuPlan Val14 Reactive), NVIDIA's Foundation4D (unified perception and reconstruction), and the counterfactual study showing driving VLAs' stated reasoning is weakly coupled to their trajectories.
Agents/tools/models
Block's Dessa is a full-duplex voice model that can listen while it speaks, trained entirely on speech, and reportedly runs locally on a MacBook M4 Max (Unite.AI). On-device, full-duplex voice changes the privacy and latency math for voice agents, but the caveat is that these are company claims with no independent evaluation yet. Worth watching, not yet worth building on.
Cerebras published a writeup of serving a 3-trillion-parameter Kimi across 59 wafers at 1,000-2,000+ tokens per second, using pipeline parallelism, 4-bit weights, and an 8-bit KV cache (Cerebras). This is a concrete datapoint on what wafer-scale inference buys for trillion-parameter mixture-of-experts models, but the gap between the demo rig and publicly accessible Cerebras models (still capped around GPT OSS 120B and Qwen 27B class) keeps this in the impressive engineering, not yet a product bucket.
Sebastian Raschka released "Build Your Own Reasoning Model" as a full video course: eight videos, 11+ hours, implementing evaluation, inference-time scaling, RL, and distillation in Python and PyTorch on a small Qwen3 (Raschka). Genuinely useful if you want the reasoning-model recipe from first principles rather than from API docs.
A scrappy harness called GVS5H orchestrates five Qwen3.8-27B instances through a shared filesystem and hits 92.4% pass@1 on LiveCodeBench hard, roughly tied with Claude Fable 5's 90.4% (GitHub). The catch: giving models an agentic loop on a benchmark designed for single-pass evaluation is not a fair comparison, and LiveCodeBench is a saturated 2025-era benchmark. The real story is cost: frontier-class coding scores on hardware you already own.