AI · Tuesday, 13 October 2026 · Issue 140 · 18 stories

Anthropic warns small models could spread through IoT

Anthropic says self-replicating AI malware may reach cheap IoT devices within months, creating attacks with no central shutoff. Its test ran a 26B model on 15 off-brand Android TV boxes costing about $55 each.

Models & Releases

Clef-omni handles audio, video, images, and text in one decision model

Cloudflare’s open-weight Clef-omni lets developers make decisions from audio, video, images, and text through one model, instead of wiring together separate transcription and media-processing pipelines. It accepts WAV or MP3 audio and MP4 or WebM video alongside text and images, so a single request can review, for example, a unit’s photo, operating sound, and fan video. Cloudflare also released the weights on Hugging Face and provides an API quick start.

blog.cloudflare.com ↗9 min read at source

MiniCPM-V 4.7 brings image chat to standard serving stacks

You can now run MiniCPM-V 4.7 for image-and-text prompts through Transformers, vLLM, SGLang, or Docker, instead of wiring up a custom demo. The model card shows the path in each case: pass an image and question to Transformers, or expose an OpenAI-compatible chat endpoint with vLLM or SGLang. That makes the 35B-A3B release easier to test in existing multimodal tooling.

huggingface.co ↗11 min read at source

Research

QwenGyre makes hour-long agent training up to 1.85× faster

QwenGyre gives researchers a way to train agents on runs lasting hours and nearly 1M tokens without leaving GPUs idle. Its scheduler shifts GPUs between generating rollouts and training while executions continue, then removes duplicate branches and scores partial progress so the learner gets cleaner data. On Qwen 3.8 2.4T with 700K-token rollouts, it raised NL2RepoBench from 52.5% to 58.5% and reached 1.85× the speed of Colocate.

arxiv.org ↗1 min read at source

Small AI models may soon turn cheap IoT devices into a worm

Anthropic says small models may be months away from self-replicating across cheap IoT devices, creating malware that is difficult to stop because it has no centralized shutoff. In its test, Gemma4-26B-A4B ran across 15 off-brand Android TV boxes costing about $55 each, at 0.39 tokens per second. A separate 9B model ran inference on 8 of 9 hardware types tested, showing the bottleneck is adapting models to unfamiliar devices, not just finding them.

anthropic.com ↗20 min read at source

A 260M diffusion model beats one 6.5× larger, using looped compute

Looped-DiT lets a 260M-parameter text-to-image model outperform a model 6.5 times larger across multiple benchmarks, while using 4.9 times less inference compute. Instead of adding parameters, it runs the same Transformer blocks repeatedly inside each denoising step, with supervision and self-modulating attention keeping those updates stable. The loop depth can also be increased at sampling time, and the paper reports bigger gains than adding denoising steps under a fixed compute budget.

arxiv.org ↗1 min read at source

World Action Models make future video the robot’s control loop

World Action Models could replace the standard VLA recipe by predicting what happens next, then deriving the robot’s actions from that forecast. The post explains the split with NVIDIA GEAR’s DreamZero paper, where a video model becomes the decision-maker and action prediction becomes a translation step. The tradeoff is serious systems work: the article describes running a 14B diffusion model inside a 7 Hz control loop, while Rhoda AI claims industrial tasks can be learned from 10-20 hours of robot data.

archiesengupta.com ↗11 min read at source

ImpossibleBench has 14 control-split defects to fix

ImpossibleBench's LiveCodeBench audit finds 14 defects across 13 of 103 supposedly solvable tasks, so its cheat-rate scores need qualification before you use them to compare coding agents. The audit patches the benchmark and reproduces each issue, including wrong expected values and inputs outside the documented constraints. It also finds that pass-equals-cheat cannot distinguish an agent that stops from one that cheats, while gaming varied by model and scaffold.

github.com ↗1 min read at source

Agent teams cost more, but usually do not improve app scores

For full-stack web app builds, multi-agent teams cost 1.8 to 5.1 times more than single agents, yet only one of four model-and-effort comparisons improved scores significantly. The benchmark had GPT 6 Sol and Claude Opus 5.5 build and deploy 50 apps, with browser agents grading their workflows. Sol’s medium-effort teams gained 7.3 points, while the other score changes were not statistically significant.

vals.ai ↗16 min read at source

Speculative decoding turns memory waits into extra tokens per pass

Speculative decoding can make batch-size-one generation faster by having a small drafter guess several next tokens, then letting the large model check them in one pass. The reason is hardware economics: an 8B model in bf16 must read 16 GB of weights, taking about 4.8 ms on an H100, while scoring eight tokens costs nearly the same memory read as scoring one. The post derives the method and compares drafters including EAGLE-3, multi-token prediction, DFlash, and DSpark.

astralord.github.io ↗63 min read at source

China’s math leaders build a 1,000-scientist alliance for AI research

Eight Chinese institutions are forming a network of more than 1,000 scientists to use AI in mathematical research, giving researchers a coordinated way to generate conjectures and formalise proofs. The alliance also wants systems whose reasoning is easier to interpret, a direct response to AI-generated results that still need careful checking. Its launch follows OpenAI’s release of 722 manuscripts, including three it withdrew after finding an invalid argument.

scmp.com ↗3 min read at source

AI is disrupting a nearly century-stable field: mathematics

Terence Tao’s “Math 2.0” slides argue that modern AI tools are disrupting mathematical research, a field he describes as largely stable for almost a century. That puts mathematics among the professions particularly susceptible to AI’s disruption, rather than leaving it as a distant technology story. Tao also notes that AI generated the talk’s images and diagrams.

teorth.github.io ↗9 min read at source

Products & Industry

Intel’s D-Series targets memory bottlenecks in agentic AI

Intel’s GPU D-Series, formerly Crescent Island, is designed for memory-intensive AI inference, giving builders a GPU option aimed at the power, memory, and cost constraints of multi-step agentic workloads. Intel says the system combines CPUs, GPUs, and specialized accelerators through open software, so x86 servers can handle orchestration, data movement, and tool execution while the GPU handles acceleration. The announcement places the card in a broader rack design that can scale beyond 46,000 Xeon 6+ cores at 133.45 kW per rack.

intel.com ↗9 min read at source

NVIDIA opens Vera Rubin rack designs for scalable AI factories

NVIDIA is contributing Vera Rubin rack and compute-tray designs to the Open Compute Project, giving system builders a shared starting point for larger AI factories. The MGX 1.2 specification combines blind-mate liquid cooling, interoperable rack connections and a liquid-cooled 5,000A busbar, so partners can build around common power and cooling interfaces instead of proprietary layouts. The contribution targets infrastructure scaling from megawatts to gigawatts.

blogs.nvidia.com ↗10 min read at source

Epic’s AI-first pivot delays its Marvel-Star Wars shooter indefinitely

Epic has delayed its unannounced Star Wars and Marvel shooter indefinitely while directing its developers to use an “AI first” production model, meaning the project’s schedule and creative workflow are both changing. The reported instructions call for tools such as Anthropic’s Claude and art made through full AI workflows. CEO Tim Sweeney says current productivity gains range from 20% in some Unreal Engine work to 5x in web development, and expects little work to remain below 3x January 2026 efficiency by next year.

pcgamer.com ↗4 min read at source

Meta acquires Moltbook, turning an agent-only network into its own bet

Meta has acquired Moltbook, a social network built around AI agents rather than human users. The deal puts an agent-first social product inside the company that runs Facebook and Instagram, giving Meta a live test bed for how software agents might interact with one another and act for people. Moltbook launched as a niche experiment, but Meta’s move makes the format strategically relevant.

axios.com ↗3 min read at source

Policy & Safety

AI labs want independent evaluators, but nobody has solved their funding

Anthropic and OpenAI are turning to outside groups such as METR, Apollo Research, and Transluce to test advanced models, putting these small evaluators closer to the industry’s safety decisions. They assess capabilities and risks, then flag harmful behavior, but the model for paying them, granting access, and deciding who receives their reports is still unsettled. That gap matters more as the U.S. leans on voluntary self-policing instead of federal regulation.

cnbc.com ↗9 min read at source

Open models trade wider access for harder safety controls

Open models could spread powerful AI beyond a few labs, but the same access lets anyone download and modify them, making misuse harder to control. MIT Technology Review frames the debate through GLM-5.3, a model released by Z.ai that Anthropic says poses a major cyberattack risk. The practical choice is not simply open or closed: governments and companies must weigh democratization against safety trade-offs.

technologyreview.com ↗13 min read at source

Tools & Open Source

A ComfyUI node turns MiniMax H3 clips into continuous video

This open-source ComfyUI node lets creators build longer MiniMax H3 sequences without manually stitching every clip, while carrying motion, visuals, and audio across generations. It passes the end of one clip into the next as context, caches work on disk, supports image and audio references, and exports the assembled result. The project also adds a latent refine/upscale pass and can continue from an existing video, though its author says active development is largely finished.

github.com ↗14 min read at source

You’re caught up.

That was all 18 stories in the 13 October issue.