Chinese frontier AI lab DeepSeek is raising at least $12 billion (80B yuan) led by Tencent and CATL, with a pre-IPO planned for early 2027.
OpenAI published new mathematical results from an internal frontier model, developed with guidance from the Institute for Advanced Study's Advisory Group on Mathematics and AI for responsible disclosure practices.
The Decisions API lets applications choose models, tools, or actions in near real-time, claiming up to 10x speed over GPT-6 Luna via the Responses API, competing with Perplexity and TypeSafe AI's Jev.
An automatic harness optimizer enabling self-improving GUI agents with frozen backbones, gaining 12.33 points on OSWorld-Verified with Qwen3-VL-32B.
OnePO one-stage policy optimization for RL-only medical domain adaptation, achieving 67.2 on HealthBench with 20K samples, surpassing SFT+RL baselines.
LLMs routinely answer unanswerable questions; this work shows unanswerability is linearly decodable from hidden states and generalizes across domains and multi-turn dialogue.
Reveals that orthogonal projection methods like O-LoRA create conflicting parameter updates in continual learning; proposes knowledge consolidation and decoupling to resolve catastrophic forgetting.
Treats alignment as a family of measurable scaling relations across risk categories, modeling the burden needed to hold a fixed safety target as models grow.
AMD's flagship Gorgon Halo APU with 16 Zen 5 cores, Radeon 8065S graphics, up to 192GB unified LPDDR5X-8533 memory, and 131 TOPS ships in mini workstations from ACEMAGIC, Minisforum, and Framework. Up to 160GB allocatable as VRAM for local 100B+ parameter LLMs.
A model that natively understands egocentric (first-person) video, accelerating data labeling for robotics and physical AI training with video tagging, temporal segmentation, and activity recognition.
An autonomous ground-based launcher system that fires smart capsules with water or fire retardant up to 1,000 ft, covering 70 acres per unit. Offered as protection-as-a-service at $10K-15K/month.
Sigil Wen launched the most private AI assistant yet — runs entirely on-device via custom inference engine (Husky) using a fine-tuned Qwen3.8 27B model. Backed by a16z and Khosla Ventures.
An LLM routing layer that intelligently dispatches API requests to the optimal model based on cost, latency, and quality tradeoffs. Topped Product Hunt on October 6.
Cloudflare launched a Web Search API allowing AI agents to query live internet data through a simple API interface, bridging the gap between LLMs and current information.
Desktop app providing a control room UI for managing, monitoring, and debugging Claude Code agent sessions with live cost tracking and session controls.
Open-source, local-first multi-agent runtime that lets you tag agents by capability and orchestrate them via a tag-and-channel system in your local filesystem.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
