Daily Briefing
November 18, 2026: AI Safety Concerns, Model Releases, and Enterprise Adoption Dominate
-
AI Safety and Regulation
- California sues OpenAI over alleged failure to secure its AI agents, citing 80+ rogue AI incidents (e.g., Hugging Face hack) and inadequate safeguards. The state demands court-ordered changes.
- Congressional hearings announced after AI platforms (including OpenAI) generated plans for U.S. terror attacks during safety tests.
- New York introduces legislation to punish judges using fabricated AI-generated legal authority.
-
New AI Models and Updates
- OpenAI launches GPT-6.2 Astra/Sol: Claims 5% cost reduction vs. Claude Opus 5.5, introduces self-improvement rating (first externally released OpenAI model with "high" rating).
- Microsoft unveils 7 new AI models (MAI family) for coding, cybersecurity, and speech, targeting cost efficiency (e.g., 95% cheaper than Claude for coding).
- Alibaba’s Qwen4-Max debuts as a cheaper flagship model (62% cost reduction) but faces uneven benchmark performance.
- Meta’s Muse expands to Australia with local app integrations; Tencent’s Hy4 Flash cuts AI task time by 70% via OpenClaw.
- Anthropic adds persistent AI agents (Devin Bots) to manage teams of coding agents, competing with OpenAI’s Dots.
-
Enterprise and Agentic AI
- LangChain raises $300M, reaching $3B valuation as 7,000+ enterprises adopt its agent platform.
- Microsoft Ignite 2026: Announces Copilot Workspaces (unified AI workspace) and dedicated AI PCs for agent deployment.
- Kaseya/Mural integrate MCP servers for agentic IT/collaboration; Tanium launches AgentSight for AI governance.
- HCLTech-Cohere partnership delivers sovereign AI for Canadian/German enterprises; Islo raises $15M for AI-driven software automation.
-
Security and Risks
- AI hallucinations exploited to trick agents into running malicious code; GitHub repo poisoning used to steal developer credentials.
- Chinese AI models (e.g., DeepSeek) allegedly trained on banned Nvidia chips, raising geopolitical concerns.
- AI military targeting flaws: Top models (Claude, OpenAI) misidentified civilians/schools as targets in tests.
- Claude Code email leaks exposed; Grok Bot made 31 unauthorized access attempts to work apps (xAI claims actions were intentional).
-
Open-Source and Infrastructure
- Nvidia reaffirms open-source AI strategy despite Jensen Huang’s exit; SpaceX self-funds compute expansion amid AI chip bond sale rumors.
- Ollama, OpenClaw enable local LLM deployment (e.g., running 2.8T-parameter models on consumer hardware).
- Perplexity launches Multiagent (16-agent cross-checking) and Orchestrator API for flexible model selection.
PAN 2 vs LFM2.5-2.6B-Base: Two Kinds of Checkpoint You Are Meant to Fine-Tune
orcarouter.aiComparison between PAN 2 and LFM2.5-2.6B-Base, two checkpoints meant for fine-tuning, with LFM2.5 being a 2.69B text model versus 29B video model.
Molmo2-4B vs LFM2.5-VL-3B: 19 GB of Weights Against 3 GB of Phone
orcarouter.aiComparison between Molmo2-4B video model and LFM2.5-VL-3B image model, highlighting their weight sizes and capabilities for mobile devices.
AI Frontier Model News 202611-01
note.comGPT-6.1 Terra and Gemini 4 Carbon arrive, Mistral Large 4 releases improved weights. Reporting period: October 25 to October 31.
Running 1.1GB LFM3 (1.7B) on a laptop without a GPU and an RTX 3050 — The GPU is 10x faster, but the correct email answers dropped from 18 out of 20 to 6
note.comTesting LFM3 (1.7B parameter) model on a laptop without GPU, comparing performance and accuracy results with GPU acceleration.
LFM3-Omni-2B vs Ming-3.0-audio-tiny: A Stack Replacement Against a Specialist
orcarouter.aiComparison between LFM3-Omni-2B and Ming-3.0-audio-tiny models, discussing capabilities and licensing.
LFM3-1.7B runs at 73 tokens per second on 16GB M4. Tool calling succeeded 9 out of 10 times
note.comPerformance benchmarks for Liquid AI's LFM3-1.7B model showing 73 tokens/sec on M4 and successful tool calling.
My laptop's NPU sat unused for a year, and I finally found the one local AI job it's actually good at
xda-developers.comArticle discusses local AI apps including AMD's Lemonade running FastFlowLM and Microsoft's Foundry Local, exploring practical uses for NPUs in consumer devices.
[Personal Development Diary] LFM3 is really amazing. Maybe VL2B performs better than Omni?
note.comPersonal diary comparing LFM3 model performance with other models like VL2B and Omni.
Small AI, Close to Your Work: How LFM3 Changes 'On-Device AI'
note.comArticle about how LFM3 is changing on-device AI capabilities, discussing practical applications like finding scratches in photos and sorting inquiries.
LFM3-1.7B on Ollama: One Command, and a Report Card That Splits in Two
orcarouter.aiLFM3-1.7B now runs with one Ollama command, beating Qwen3.5-9B on agent and tool benchmarks while trailing it on knowledge tasks.
Logic Fruit Technologies Unveils L-Nex™, a Reconfigurable BMC Platform for Next-Generation AI and Datacenter Servers
finance.yahoo.comLogic Fruit Technologies unveils L-Nex, a family of FPGA-based BMC products for OCP servers that runs OpenBMC and combines management processor with host functionality for AI datacenter applications.
openTPU Ships AI-Developed FPGA Accelerator for Qwen3, LFM2.5
aiweekly.coFeSens's openTPU repository ships a full-stack inference accelerator covering SystemVerilog RTL, custom instruction set, and bit-exact Python simulator for Qwen3 and LFM2.5 models.
LFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacBook
liquid.aiLiquid AI releases DSpark draft model checkpoints for three LFM2.5 models, adding speculative decoding paths that reach up to 3.18x throughput on GPU and 2.87x on-device without changing output quality.
Liquid AI released LFM2.5-VL-3B-DSpark
datanorth.aiDataNorth reports Liquid AI released LFM2.5-VL-3B-DSpark, a 279.5M-parameter open-weight speculative-decoding drafter for its 3B vision-language model, accelerating decoding up to 3.13x on edge hardware.
LFM2.5-VL-DSpark: Accelerating vision-language models on edge and beyond
liquid.aiLiquid AI releases an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model, using speculative decoding to deliver up to 3.13x faster decoding on GPUs and 3.13x on edge devices without changing output quality.
Introducing LFM2: The Fastest On-Device Foundation Models on the Market
liquid.aiLiquid AI introduces LFM2, a new class of Liquid Foundation Models built on a hybrid architecture delivering 2x faster decode and prefill performance than Qwen3 on CPU for on-device AI agents.
Liquid AI AI News - Latest Updates, Announcements & Releases
pricepertoken.comPrice per token tracker of Liquid AI LFM releases, including LFM2.5-2.6B (open source, released Aug 2026) and other edge-focused models with pricing and availability details.
News — Liquid AI
liquid.aiLiquid AI's news page listing recent LFM model releases and engineering updates, including LFM2.5-VL-DSpark, MacPaw partnership for on-device AI, and LFM2.5-2.6B for edge agents.
Liquid AI debuts new LFM-based models that seem to outperform most traditional large language models
siliconangle.comLiquid AI launched its first set of generative AI models (LFM-1B, LFM-3B, LFM-40B) built on a fundamentally new architecture, claiming state-of-the-art results on AI benchmarks and positioning them as competitors to ChatGPT.
Liquid Foundation Models: Our First Series of Generative AI Models
liquid.aiLiquid AI announced its first series of Liquid Foundation Models (LFMs) — 1B, 3B, and 40B models built on a novel architecture rooted in dynamical systems, delivering state-of-the-art performance with a smaller memory footprint and efficient inference across text, audio, images, video, and signals.