Alibaba's T-Head unit unveiled the Zhenwu V900, delivering 3x the performance of its predecessor with 216GB memory and 1,200GB/s chip-to-chip interconnect, scalable to 500,000 units for frontier-model training. Mass production slated for Q1 2027 alongside 20GW global datacenter capacity by 2032.
Anthropic's new flagship performs at Fable 5.1 level on most tasks but costs 40% less to run than Opus 5, priced at $4/M input and $20/M output tokens with 30% faster generation and improved safety guardrails.
Snorkel AI closed a $350M Series E at $3.5B valuation led by Insight Partners and S32, with 18x ARR growth to $375M driven by demand for AI training data and synthetic data for RL environments.
Long-term conversational memory that distinguishes who said what, whom each statement concerns, how individuals perceive one another, and how group-shared information changes over time.
Formalizes three regularity conditions (completeness, additivity, policy-consistency) bridging token-level credit assignment and critic alignment in reinforcement learning for language models.
Dynamic expert pruning achieves 33% throughput improvement with negligible quality drop, validated across fine-grained open-weight MoE architectures that routinely select many experts per token.
Addresses the tool-trialing budget problem in LLM agents that must generalize to unseen tools at deployment, using outcome-based post-training guidance instead of brute suppression.
Improves small- and mid-scale LLM reasoning by combining progressive question simplification with self-evolving reasoning traces, addressing brittleness that persists even with knowledge distillation.
Identifies the Proximity Trap in long-context LLMs: insufficient attention to distant evidence arises less from distance than from cumulative competing background context.
Proposes on-policy distillation to address repetitive degenerate loops in long generations that plague quantization-aware distillation below 3-bit precision.
Released at ROSCon 2026 in Toronto, Isaac ROS 5.0 introduces agentic workflows to the ROS ecosystem with expanded GPU-accelerated physical AI libraries across the Jetson platform, enabling robots to perceive, reason, and act in dynamic environments.
An AI control system lets a micro-aerial vehicle achieve insect-like agility, boosting speed by 450% and enabling 10 consecutive somersaults in 11 seconds, pushing the limits of micro-maneuverability.
AMD's share price nearly tripled in 2026, driven by AI datacenter demand and the Meta Muse AI agent rally, making it the fourth US chipmaker to cross the trillion-dollar threshold after Nvidia, Broadcom, and TSMC.
CEO Aaron Edsinger demonstrated the Stretch 4 on stage, showcasing physical AI for mobility-impaired users with a wheeled robot and telescoping arm targeting real-world utility over humanoid spectacle. First production run sold out.
JetBrains launched Air, an open system connecting IDEs, team coordination, governance, and cost controls with multi-agent support including Codex, Claude Agent, Gemini CLI, and Junie.
An open-source AI engineering platform that designs, verifies, and documents circuit boards from natural-language prompts using KiCad integration, ERC validation, and decision logging. Hit #2 on HN.
DigitalOcean's new managed service integrates with Claude Code, Codex, and 16,000+ tools, with idle resource optimization for production AI agent workloads.
The Italian open-source platform for creating AI agents and automating business processes closed a seed round from Zest, Vertis, and OpenT to accelerate natural-language process creation and digital sovereignty.
The Italian corporate training startup raised from Club degli Investitori, Vento, Doorway, Zanichelli Venture, and angels including Bending Spoons co-founders and ex-Apple CFO Luca Maestri. Launched an AI training vertical for AI Act literacy compliance.
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:
