Funding, IPOs & Market Signals
OpenAI's $1.25T round. OpenAI raised $11 billion so far in a new funding round at a $1.25 trillion valuation, below earlier hopes of $30 billion at $1.4 trillion after tech stocks dropped. The round may still reach $20 billion.
Anthropic IPO wobbles. Investors are balking at Anthropic's $1.5 trillion IPO valuation, with some suggesting $1 trillion; missing October revenue and ongoing losses are spooking them, and the roadshow could slip.
AI spend plateaus. Ramp data shows token spend grew only 1.5% in August–September versus 42% in July, suggesting price competition and savvier customers are capping OpenAI's and Anthropic's revenue growth.
Midterm backlash hits AI. Voters approved 40 of 41 data center ballot measures restricting development; newly elected Democrats are expected to push more AI and data center regulation, triggering a sharp tech selloff.
Meta pauses AI terms. After backlash, Meta indefinitely delayed commercial terms that would have required business page owners to allow AI training on linked external sites like news publishers.
Open Model Releases
Qwen Flash Next tames. New Unsloth Dynamic v3.1 and ISTA-DASLab quants plus Strata SSD offloading let Qwen Flash-Next run on 8GB VRAM/32GB RAM with strong quality, 50 tokens/s decode across a 131K context. ISTA-DASLab's quants face reproducibility criticism from Unsloth, and Basalt switched to Unsloth equivalents.
Ripple-65B Apache 2.0. Ricursive Intelligence previewed Ripple-65B, a dense 65B open US model scoring 80.52 on TerminalBench 2.1; it was trained on 12T tokens for $32M.
Inkling 2 Small debuts. Thinking Machines released Inkling 2 Small, a 594B MoE with 17B active parameters, roughly matching Gemini 3.8 Flash on the Artificial Analysis Index but costing six times more per task.
Arete-27B values post-training. NousResearch post-trained Qwen3.8-27B into Arete-27B, cutting CCP-framed answers from 22% to 0.9% while preserving general and agentic capability.
Edge models shrink further. Scylla's Band v4 offers a 131M multilingual singing TTS under Apache 2.0; Mirage is a 22MB vision-language model for phones; Underdog's Woof is a 32MB speech model running in-browser.
- Scylla's Band v4, open source TTS with code to run everywhere(opens in a new tab)
- Mirage: A 22 MB visual-language model that runs scene descriptions + object detection + OCR + segmentation + spatial relations analysis + extractions + decisions, in a single pass.(opens in a new tab)
- Underdog shows off a 32MB speech model in new demo(opens in a new tab)
Agents, Tools & Platforms
Harness leaderboard updates. Kimi Code, Hermes, Pi, Claude Code, Codex and others were benchmarked against Terminal Bench 2.1 and Deep SWE 1.1, giving a snapshot of current agentic coding performance.
Open WebUI computer control. Open WebUI 0.13.0 and Open Terminal 0.17.2 let models drive a sandboxed Linux/Windows VM; a QFN 3.8 model used mouse coordinates to draw a dog in MS Paint.
DevSecOps agent controls. GitLab 19.5 makes its MCP server generally available, letting Claude Code, Cursor and other agents access GitLab tools with least-privilege controls. AWS made Continuum generally available to find, validate, and patch code vulnerabilities.
Meta Muse Spaces. Meta launched Spaces for its Muse agent: customizable app-like experiences that pull live data from Gmail, Google Calendar, Notion, etc. and support tracking, planning, and editing.
Always-on agents arrive. Perplexity's Teammates run on their own VMs for ongoing projects for $200/month Max users; Amazon lets select US customers text Alexa to shop, and both show agent interfaces moving to messaging.
Research & Safety
Anthropic acceleration warning. METR's 11-week review found Anthropic's model improvement rate roughly doubled, with about 80% from AI used in its own research, crossing Anthropic's own Responsible Scaling Policy threshold.
Weight editing patches flaws. Experiments show Opus 5.5 can edit weights to fix prompt injection and hallucination flaws in trained models, useful for patching but not recursive self-improvement.
SPLADE sparse retrieval. A 0.1B SPLADE model outperformed Qwen3-Embedding-0.6B in a Korean/English RAG evaluation, making local retrieval feasible on 8GB PCs.
Robotics simulation push. Hugging Face introduced FoundationWorlds, starting with ToyBlocks, for controlled robot learning and evaluation. Physicl raised $10M seed to help robot builders train with realistic 3D simulations.
Open security skills stripped. Kimi K3.1 and Tencent Hy4 are removing cybersecurity, reverse-engineering, and vulnerability-patching capabilities from open weights, worrying local defenders.
Data Centers & Hardware
AMD data center first. AMD CEO says it is a data center first company, so cheap consumer GPUs are low priority; however, some RTX 6000 prices have begun to drop from $18K peaks to $16.5K.
Crusoe sells to Google. Crusoe sold its unfinished Wyoming data center project to Google, which plans a $17B campus; highlights how hard it is for independents to deliver huge projects on time and budget.
NVIDIA hardware stolen. Six pallets of NVIDIA hardware worth $11.4 million were stolen en route to Texas.
Amsterdam DC protest. Protesters broke into a Microsoft data center construction site in Amsterdam, scaling a crane and placing anti-AI banners, amid Europe-wide data center opposition.
China spy plot. China's MSS allegedly recruited Megaspeed executive Alice Huang to spy on Nvidia CEO Jensen Huang and push for advanced chip access.
That's everything for this day – about a 5-minute read.