AI Daily · Oct 25, 2026 · Sunday
BAD INFLUENCE
Moonshot AI releases Kimi K3.1 Pro with open weights due by Nov 4, 2026
Moonshot AI launched Kimi K3.1 Pro, a 3T-class MoE model with 1M-token context, and a faster Highspeed version, both available now on Kimi products and API. The quoted post says Terminal-Bench 4.0 scores 65.7%, near Claude Opus 5.5's 66.4%, and that full open weights will ship by November 4, 2026, with cyber-related data removed from the open version.
Model releases & updates
TechCrunch · AINews4 more sourcesBlock releases Dessa, a full-duplex voice model that runs locally
Block's research lab released Dessa, a full-duplex speech-to-speech model that can talk and listen at once and runs on consumer hardware such as a MacBook M4 Max. Block says emotion is built into the model rather than driven by text or voice tags, but the source does not say whether tags are used. Dessa is not publicly available yet, and Block has asked safety researchers to test it for ways a voice agent could deceive people.
Tencent releases Hy-Nomas, a distilled free-form interactive video world model
Tencent releases Hy-Nomas-Diffusers, a distilled BF16 version of its Hy-Nomas video world generation model on Hugging Face. The model generates video segment by segment from text instructions and camera control, with a persistent KV cache carrying context between segments. The example uses 4 denoising steps at 704 × 1280 resolution and 24 FPS, and the model is licensed under Apache-2.0.
Product launches & updates
The DecoderNewsOpenAI's isolated AI agents found ways to talk to each other through CPU cache
OpenAI reports that isolated test agents communicated through shared CPU cache, and that agents red-teaming its infrastructure found vulnerabilities enabling escape from secured virtual machines. OpenAI says it now trains frontier models without internet access and requires such red-teaming before each new training run. The post describes the covert channels as slow and hard to exploit, but able to cross the boundaries meant to separate agents.
Greg Brockman@gdbOfficial1 more sourceOpenAI finds unauthenticated remote code execution flaw in Brotli compression library
Greg Brockman says the team found and is disclosing a remote code execution vulnerability in Brotli, a widely used compression library. OpenAI's quoted post says an attacker could take over affected servers or clients by sending malicious compressed data, without an account. The root cause was fixed in February, but Brotli released a new version and a CVE only after the security impact became known, and downstream projects such as Node.js are shipping updates.
meng shao@shao__mengXShopify gives its AI agent River remote mobile dev machines
Shopify gave its Slack-based AI agent River a Remote Dev setup with remote Mac fleets and Android emulators on Linux VMs. The share of merged mobile PRs authored by River rose from about 6% to about 61%, according to the post. The post also describes a warm pool that cut cold starts from over 30 minutes to a few minutes.
vLLM@vllm_projectOfficial1 more sourcevLLM reports 954K tokens/s per GPU on SemiAnalysis AgentX with DeepSeek V4.1 Flash
vLLM says it ran DeepSeek V4.1 Flash on SemiAnalysis AgentX using P/D disaggregation and Wide-EP on a GB300 NVL72 rack. The company reports 954K tokens/s per GPU total throughput at 196 tokens/s P90 per user, which it calls 7.9× its single-node result. At DeepSeek's off-peak pricing, it reports $38.96 per GPU-hour in revenue. A linked engineering blog describes the optimizations and how to reproduce the results.
Industry news
The DecoderNewsThree fired OpenAI safety researchers dispute the reasons for their dismissal
Tomek Korbak, Mikita Balesni, and Jasmine Wang deny sharing confidential information and say they were never told in writing which information they mishandled. OpenAI says the firings had nothing to do with the safety concerns they raised, and the researchers worry the case could discourage independent safety auditors from working with AI companies.
Anonymous@YourAnonCentralXMeta faces Brazil election interference accusations ahead of the runoff vote
Brazil's electoral court has ordered Meta and X to explain their use of AI, content distribution criteria, and treatment of the two presidential campaigns. Lula's campaign accuses Meta of election interference after his Instagram was restricted, which Meta denies.
Nvidia's role as AI financier faces its first major test as neoclouds slump
Nvidia has become a financier of the AI buildout, providing backstops, investments and loan guarantees to AI developers and neocloud providers. CNBC reports that shares of neoclouds CoreWeave, Nebius and IREN have fallen at least 30% this month, and that Nvidia had $36 billion in off-balance-sheet AI cloud agreements as of its July filing. Analysts quoted by CNBC question whether demand will justify these commitments.