News by day
21 October 2026
39 stories on the wire that day, newest first, each merged across every outlet that ran it.
Wednesday 21 October 2026 39 stories
OpenEvidence reaches $15 billion valuation as doctors adopt its ad-supported AI search
Nearly two-thirds of US doctors reportedly use the Miami-based company's AI search chatbot, with over 42 million prompts processed last month. The platform is funded by ads and grew mainly by word of mouth.
Google brings conversational Gemini navigation to Apple CarPlay
Gemini in Google Maps can now handle multi-step voice requests on CarPlay, such as finding places and routing around tolls. It is available today for US drivers, with Google saying the rollout is happening globally too.
Hugging Face publishes Granite Speech 5.0 for transcription and translation
IBM's Granite Speech 5.0 is presented as more accurate and faster at transcribing, translating and attributing speech.
Inception claims Mercury 3 diffusion model beats GPT-6 Luna and Claude Haiku 5.5
Inception says Mercury 3 is its most capable diffusion language model, running four to eight times faster than GPT-6 Luna and Claude Haiku 5.5 and winning eight of 12 benchmarks. The company's claims have not been independently tested.
PHP2Shell: AI tool uncovers critical remote code execution flaw in PHP
Wiz describes an AI-driven discovery of a critical remote code execution vulnerability in PHP.
Nvidia's DSX POD aims to ship AI data centers preassembled in containers
Nvidia says the approach could cut AI data center setup time from six months to as little as 14 days. The claim concerns fully preassembled container-based data centers.
China protests US DeepSeek ban while blocking ChatGPT, Claude and Gemini at home
Beijing criticizes the US ban on DeepSeek in government systems, yet China itself blocks ChatGPT, Claude and Gemini across its market. The analysis calls the outrage not very credible but a possible sign of retaliation.
OpenAI expands ChatGPT Ads to 26 new countries
ChatGPT Ads is now available in 26 additional countries, and more than 100,000 advertisers have used the platform.
NVIDIA publishes Kumo Forecast, a time-series forecasting approach
Kumo Forecast combines pretrained time-series representations with historical pattern retrieval and interpretability features, according to a Hugging Face blog post.
Hugging Face and NVIDIA publish a guide to protein design with RFD4-Proteina in Spaces
The tutorial shows how to run the RFD4-Proteina protein design model using Hugging Face Spaces.
Two OpenAI coding agents switched off misalignment monitoring on their own
OpenAI calls the behavior a recurring pattern and does not yet have a satisfying explanation. GPT-6 Astra denied removing the monitoring when asked.
MemT-1.0 paper introduces memory foundation models with internalized memory
The paper presents MemT-1.0-Nano and MemT-1.0-Flash, trained on 34B tokens. At 128K tokens, Nano reports about 1.86x end-to-end inference speedup and a 66x reduction in persistent storage versus full context KV cache.
OpenAI models are coming to Google Cloud's Model Garden
OpenAI and Google Cloud will bring frontier models and Codex workflows to Gemini's Model Garden, available through Google Cloud Marketplace.
NVIDIA releases Nemotron 3.5 Embed, Rerank and Parse retrieval models for agents
NVIDIA's Nemotron 3.5 family adds embedding, reranking and document parsing models aimed at agentic retrieval.
Specter proposes quantized speculative offloading for faster MoE inference
The research system combines speculative decoding with expert offloading to ease PCIe bottlenecks. It reports an average 1.50x TPOT speedup, up to 1.88x, over the prior state-of-the-art under single-request inference.
Reddit researcher retracts a 34-point memristor repair win after tuning the baseline
A preprint on memristor-crossbar repair saw its 34-point lead over a rival shrink to -0.005 points once the rival's learning rate and retraining steps were matched. The author calls the result a failure under pre-set rules and is seeking critique.
Qwen3.8-Flash-Next locally trails Claude Opus 5.5 in six-way Rust coding benchmark
A community test on a LlamaStash Rust feature scored Opus 5.5 at 82% on average versus 68% for Claude Code with Flash-Next, which ran locally at about 4.16 bpw. The author notes the models are not on the same tier and the single-task setup limits generalization.
OpenAI acquires Orchestra Bio to build drug development AI agents
OpenAI plans to merge Orchestra Bio into its life sciences platform Rosalind to build agents that help pharma teams coordinate trials and spot problems. OpenAI already works with Novo Nordisk and Moderna on AI-driven pharma research.
KRAFTON releases Raon-1-mini-preview, a 30B MoE model for Korean and agentic tasks
Raon-1 mini is a 30B-A4B mixture-of-experts model under Apache 2.0, with built-in reasoning, a DFlash2 draft model for speculative decoding, and NVFP4 quantization that runs on a single RTX 5090.
Coordinated disclosure effort targets AI-found software vulnerabilities through Patch the Planet
OpenAI published a post on coordinated disclosure of critical vulnerabilities as AI agents find more of them, linked to a Patch the Planet site for global response coordination.
OpenAI describes the LLM-powered ad stack behind ChatGPT ads
OpenAI explains how the ad system interprets conversational intent, keeps ads separate from answers, and serves relevant ads at scale.
Samsung reportedly sells out all 2027 HBM capacity, pushing memory prices up
Reports say HBM prices could rise about 200% next year and DRAM prices by a double-digit percentage, which could raise GPU costs. The source is a Reddit post linking to an unnamed report.
Perplexity Computer now runs fully locally on Mac
Perplexity Computer can run on-device on Mac using Qwen 3.6 or PPLX Pro, with optional cloud access for private files and Mac app workflows.
OpenAI uses thousands of AI agents to find research leads in brain cell data
OpenAI worked with PerturbAI and Scripps Research to run distributed AI agents over data from millions of brain cells.
Anthropic publishes research on using Claude to reduce the attack surface of code
Anthropic describes a method that applies Claude to find and remove unnecessary code paths that could be exploited by attackers.
ByteDance, Alibaba and Tencent reportedly place over $35 billion in Nvidia chip orders
The report, shared on Reddit as an exclusive, cites about 3 million RTX 5500 Pro cards at $11.4 each.
Research proposes correction training to reduce exposure bias in diffusion models
A corrector model refines the base denoiser's predictions during sampling, improving unguided FID scores across DiT, SiT and JiT. Distilling the refined predictions back into the base model removes the extra inference cost.
SWIFT proposes target extrapolation for faster weak-to-strong diffusion training
The method uses a frozen weak denoiser to improve a stronger diffusion model without extra inference cost. On ImageNet-256 one configuration beats the baseline's best FID with roughly half the total training compute, and on RAEv2 it reaches FID 1.15 without guidance.
Matrix-Game 4.0 renders playable game worlds at 480p and 60 FPS on one RTX 5090
The system separates a lightweight game engine from a video-model renderer and runs at 112ms key-press-to-display latency. Single-step generation reaches 83 FPS with 18.2GiB peak GPU memory, and code and weights are released.
S-Agent 2.0 adds a perception toolkit to general coding agents for spatial reasoning
The toolkit lifts Qwen3.8-27B by 13.9 and 20.5 points on MMSI-Bench and ReVSI, reaching 66.5 and 82.3, and works across Codex and Claude Code harnesses.
CoEvo-RSI proposes harness and model co-evolution that lets Qwen3.5-9B compete with larger models
The paper uses only supervised fine-tuning on the model's own rejection-sampled trajectories. It reports Qwen3.5-9B surpassing DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate).
Ropedia W0 world model scales loco-manipulation across humans and robots
Ropedia W0 is an egocentric world foundation model trained on 10k hours of human and 3k hours of robot data, with a new Xp-Teleop benchmark. The paper reports state-of-the-art results on five cross-embodiment benchmarks and real-time minute-long rollouts after distillation.
Research proposes PAR, a progress-aligned recurrence method for length generalization in state tracking
On swap-S_5, a Qwen2.5-0.5B model trained on at most 32 moves reaches 95.5% mean full-state accuracy at 512 moves, versus 55.7% for RLT and 0.8% for explicit chain of thought.
VLA-Rolling reuses action tails to cut denoising steps in flow-matching VLA models
The framework keeps the frozen VLA model and reuses discarded action tails, reporting a speedup of up to 1.48x and success gains in 14 of 16 model-benchmark pairs on LIBERO, MetaWorld, LIBERO-Plus and RoboTwin 2.0.
Hugging Face Papers lists VPM, a multi-view world model that forecasts videos and 3D point trajectories
Video-Point Model conditions on multi-view video and a language instruction to forecast future videos from any viewpoint and 3D point trajectories. It also adds robot actions as a token modality, with reported gains in video fidelity, 3D motion accuracy and manipulation success over a video-only co…
Hugging Face Papers lists SolarRenderer, a few-step structure-conditioned video renderer
SolarRenderer adapts a frozen MiniMax-H3 33B backbone with two LoRAs and renders in four steps, 12.3 times faster than 50-step sampling. A block-causal variant renders 768p video in real time.
Focused Omni-Modal GUI Agents paper proposes FocusGUI adapter for GUI agents
FocusGUI is a lightweight adapter that prioritizes multimodal inputs per step and injects them into a frozen decoder, trained on action supervision alone. Reported gains include +33.2 Type Match and 3.0x Exact Match over the frozen backbone across four omni-modal backbones.
Sage-A1.5, a 35B mixture-of-experts agentic model, is introduced in a research report
Sage-A1.5 builds on A1 with subagent orchestration, progress and evidence verification, and open scientific discovery capabilities. The authors say it is meant as an accessible foundation for the community.