Samsung has agreed to abandon its turnkey approach and produce the base die of Nvidia's custom NVHBM memory on TSMC's 3nm process, with Micron and SK hynix reportedly joining a standardized single-process flow. NVHBM targets roughly 30% more bandwidth than HBM4E at up to 15% lower power for Nvidia's Feynman-generation GPUs.
TSMC reported Q3 2026 revenue of $46.87 billion, up 41.1% year over year, with net profit up nearly 70% to roughly NT$763 billion — its sixth consecutive quarterly record. The result suggests the AI accelerator buildout still has room to run despite margin-outlook concerns.
An open-source MCP server lets AI agents read and send WhatsApp messages through a free Cloudflare Workers deployment, with no self-hosting required.
A new method encodes documents into reusable embeddings distilled into 16 memory tokens, beating full-context inference on FactKG (84.33% vs 80.43%) while using roughly 108x fewer tokens, and improving agent performance in TextWorld.
New work generates verifiable data-analysis tasks, files and rubrics from scratch using known generative mechanisms; RL training on these environments yields significant benchmark gains, with frontier models still weaker at interpretation than computation.
A benchmark of 95 end-to-end forward-deployed-engineer delivery tasks across 20 customer scenarios, scored on real business-metric uplift with scripted stakeholders. The best of 13 model configurations scores 0.56 and resolves only 14/95 tasks; discovering unstated constraints is the main bottleneck.
A System-1 spatial decision model scores candidate answers via native attention matching with no generated tokens. Built on Qwen3.5-4B it tops eight models with 65.64% overall, ranking first on 7 of 12 spatial benchmarks while running 2.5x faster than the fastest autoregressive baseline.
Supervises VLAs on explicit physical state transitions between current and future states instead of only observation-action mappings. It achieves SOTA-competitive results on LIBERO and RoboTwin and outperforms pi-0.5 in real-robot experiments.
A new cross-platform terminal emulator that ships with strong out-of-the-box opinions on configuration and rendering, aimed at replacing stock terminals.
An open-source vector database that keeps all durable state in S3, making it a lightweight single-node option for AI retrieval workloads.
An open-source engine that lets AI agents interoperate across protocols and services, positioning itself as a neutral translation layer between agent ecosystems.
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:
