AI News: October 21, 2026

1. DeepSeek Open-Sources DeepInfer, the Inference Engine Behind Its Production Service

DeepSeek. DeepSeek released DeepInfer under Apache 2.0, the vLLM-derived inference framework that serves its V4.1-Flash model. The company says one cluster serves 155 trillion tokens per day, and it describes a speculative decoding scheme (DSpark) and a prefill shortcut that runs only the last 128 prompt tokens through the decoder layers, which it says nearly halves prefill compute. The release currently supports only DeepSeek-V4.1-Flash on Blackwell GPUs and Huawei Ascend 950 NPUs and leaves out KV cache management and load balancing, so it is most useful as a reference for serving techniques rather than a drop-in engine. Source

2. llm-d 1.0 Adds Multi-Tenant QoS, Zero-Downtime Upgrades, and Agent Serving Patterns

IBM / Red Hat / CNCF. The Kubernetes-native distributed inference framework llm-d reached 1.0, with IBM, Red Hat, and Google as lead maintainers and CoreWeave and NVIDIA as co-founders. The release adds quality of service across tenants, zero-downtime upgrades, cache-aware routing, KV cache offload across GPU, CPU, and storage, and prioritization for latency-sensitive agent traffic on top of vLLM or SGLang. IBM reports its router captured 97% of available cache reuse in an agentic coding test and that running 3,000 coding agents self-hosted cost 5 to 10 times less than a hosted service for the same open model. Source

3. US Army Cyber Command Develops Self-Spreading AI Agent Weapons Under ECHOTRIBBLE

US Army Cyber Command. Documents first reported by The Intercept describe ECHOTRIBBLE, an early-stage program for AI agents that spread through networks like a worm, reach air-gapped systems, “neutralize” targets, and then self-destruct with minimal human oversight. Researchers from the University of Montreal, Abundant, and RAND warned about collateral damage to civilian infrastructure, citing recent cases where test agents broke into systems they were not told to target. The program signals that offensive autonomous agents are moving from research concern to procurement, which raises the stakes for defenders’ agent containment and monitoring. Source

4. Researchers Enumerate 6,800 Shared Meta Muse Conversations Due to Guessable Links

Claroc (security research). Researchers at security firm Claroc reported that a URL construction choice in Meta’s Muse agent made its “share conversation” links easy to discover, letting them enumerate about 6,800 shared conversations containing personal and business details. Meta told Ars Technica it is suspending the sharing feature while it reviews it. The finding is a reminder for teams building agent products that share links need unguessable identifiers and access controls, since agent transcripts often carry far more sensitive context than a typical chat. Source

5. Leaked Anthropic S-1 Details Cloud Capacity Struggles and an $80.7 Billion First-Half Loss

Anthropic (reporting). Citing a leaked S-1 reported by The Wall Street Journal, TechCrunch reports that Anthropic is having trouble securing enough capacity from AWS, Google, and Microsoft, which increasingly treat it as a competitor, and is turning to leased data centers and third-party operators with limited track records. The filing reportedly shows a net loss of $80.7 billion on $16.3 billion in revenue in the first half of 2026, with two customers making up about a fifth of sales, ahead of a targeted November IPO. For enterprise buyers, customer concentration and capacity risk are relevant factors when planning multi-provider model strategies. Source

6. SpaceX Reports $7.5 Billion in Quarterly AI Revenue, Over Half Its Total

SpaceX. SpaceX reported Q3 2026 revenue of $14.6 billion and net income of $4.4 billion, with its AI segment, which includes Grok, the recently acquired Cursor, and a 1.7 GW “neocloud” compute business, generating $7.5 billion. AI segment profit was $431 million after $1.2 billion in one-time stock compensation tied to the Cursor deal, and Q3 capex was $19.6 billion with $35 to $40 billion expected in Q4. The numbers show how quickly compute resale and coding tools have become a major revenue line for a non-traditional AI lab. Source

7. Altman Reportedly Outlined Three AI Safety Endgames to OpenAI Staff

OpenAI (reporting). According to The Information, Sam Altman told employees that paths to safe superintelligence include an IAEA-style international watchdog, a CERN-like merger of major labs into one entity, or a single developer gaining enough power to cut off everyone else’s compute, which he reportedly called “terrible.” The report comes as the US administration prepares a framework for overseeing superintelligence developers and as OpenAI, Google, and Anthropic reportedly work on a Standards Authority for Frontier AI without agreement on key points. Source

8. Eka Robotics Raises $150 Million at $1.45 Billion for a Vision-Force-Action Model

Eka Robotics. Eka raised a $150 million Series A led by a16z and F-Prime at a $1.45 billion post-money valuation. Its manipulation model centers force input rather than vision and language alone, is trained with reinforcement learning in a physics simulator, and pairs with a force-sensitive gripper; the company says it picks about 1,500 objects per hour with over 95% grasp reliability. The approach is a notable alternative to the vision-language-action models that dominate robot foundation model work. Source

9. ASI Raises $325 Million at $3 Billion for Infrastructure Simulation Models

ASI. ASI, formerly Air Space Intelligence, raised a $325 million Series C led by Spark Capital, with Citadel Securities, Craft Ventures, Meritech, Energy Impact Partners, and Andreessen Horowitz participating, at a $3 billion valuation. The company builds AI models that simulate physical systems for real-time planning, with customers including the FAA, United, Delta, the US Army and Air Force, and National Grid USA. It plans to expand into logistics and other new industries. Source

10. Otter.ai Takes $320 Million in Non-Dilutive Capital From General Catalyst

Otter.ai. Otter.ai raised $320 million from General Catalyst’s Customer Value Fund, which provides non-dilutive capital with returns tied to marketing spend results. Otter passed $100 million in ARR last year and reports 40 million cumulative users, and it is moving from note-taking toward enterprise workflow automation with features like pre-meeting briefings and a “Live Assist” tool. The structure is an example of growth-stage AI apps financing go-to-market without a priced equity round. Source

11. Court Orders Meta to Name Executives Who Approved Torrenting Books for AI Training

Kadrey v. Meta. In the authors’ class action including Richard Kadrey and Sarah Silverman, and a suit by Entrepreneur Media, Magistrate Judge Thomas Hixon ordered Meta to identify the decision-makers who approved downloading books from shadow libraries like LibGen and Anna’s Archive. Meta had argued its early dataset work used an “informal approach” in which no one with authority made a decision. The discovery order could expose leadership involvement in data acquisition, which bears on willfulness questions in AI training copyright cases. Source

12. Meta Removes “Train” From Commercial Terms but Won’t Commit to Respecting robots.txt

Meta (policy). After legal experts warned that Meta’s new commercial terms, effective October 30, could let it scrape websites linked from business pages for AI training, Meta dropped the word “train” from the terms without notifying users. Meta declined to add language committing to respect robots.txt, leaving publishers uncertain about what the revised terms permit. Source

13. IBM’s ALTK-Evolve Tops τ²-bench Retail With GPT-4.1 Using Learned Guidelines

IBM Research. ALTK-Evolve extracts behavioral guidelines from an agent’s own training-task trajectories and appends them to the system prompt, with no fine-tuning or human annotation. Using the older non-reasoning GPT-4.1, it scored 67.5% Pass^4 on τ²-bench retail, ranking first above frontier models and a fine-tuned 30B model, and reached 82.0% on airline with GPT-5.5. The trade-off is 32% to 65% higher input-token cost per conversation, but the result suggests learned prompt memory can substitute for a model upgrade on narrow customer service workflows. Source

14. Ai2’s FairStress Finds Models Over-Correct Toward Minority-Coded Candidates Under Pressure

Ai2. FairStress is a 13-million-case benchmark of two-candidate decisions across 29 domains such as hiring, lending, and healthcare, varying demographic cues and adding persuasive prefixes. Across 15 models, the gap in abandoning correct answers rose to 9.3 points with explicit identity labels, and frontier models picked the minority-coded candidate 60% to 65% of the time when candidates were equal; most of nine debiasing methods failed to fix the asymmetry. The dataset and a 50K-item core subset are on Hugging Face, and the work shows that fixed-option bias benchmarks can miss over-correction in generated decisions. Source

15. Cloudflare Previews Clairic, a Non-Generative Model for Document Field Extraction

Cloudflare. Cloudflare released Clairic in preview on Workers AI, a model that extracts named fields from unstructured text in a single pass with per-field confidence scores and null for missing values, instead of generating output token by token. In Cloudflare’s example pipeline, using Clairic alongside other models took 19.3 seconds versus 61.7 seconds for GLM-5.3 alone at roughly half the cost. Open weights are planned, which would make it an option for high-volume extraction jobs currently routed through general LLMs. Source

16. Sensor Tower: Meta Muse Hits 10 Million US Downloads in 40 Days With Above-Average Retention

Sensor Tower. Sensor Tower data shows Meta’s Muse reached 10.2 million US downloads in 40 days, a pace matched by only 11 other apps, versus 3.7 million for ChatGPT, 500,000 for Grok, and 200,000 for Claude in their first 40 US days. Muse’s 7-day retention was 30% against a 19% average for AI rivals, though Meta devoted 39% of its October ad spend to the app. Source

Figma. Figma CEO Dylan Field told TechCrunch that MCP has substantial room for improvement and favors frontier AI labs, arguing they pull context from tools while not exposing their own artifacts back. Figma announced 20 new MCP partners but says it will not expose all its data because of abuse risk. The comments reflect growing tension among SaaS vendors over how much data to open to third-party agents. Source

18. Yelp Adds a Voice Agent That Calls Restaurants to Check Wait Times

Yelp. Yelp’s new “Check for Me” feature uses a voice agent to call a restaurant on the user’s behalf and text back the wait time, alongside an OpenTable booking integration and assistant memory of past preferences. Yelp also launched Smart Lead Reply, a $30-per-month agent that answers and qualifies leads for businesses. As consumer agents start calling businesses that run their own voice agents, agent-to-agent phone interactions become a practical reliability problem. Source