AMD detailed the Instinct MI455X silicon with 40.26 peak petaflops (MXFP4), up to 4x the MI355X. Helios racks offer 2.9 exaflops, 31 TB of HBM4 memory, and 1.7 PB/s bandwidth across 72 GPUs paired with EPYC Venice CPUs.
The AI robotics foundation model startup raised a $200M Series B extension from 8VC, Radical Ventures, Nvidia, USV, and Bezos Expeditions. Total round now stands at $600M at a $3B valuation.
Autonomous middle-mile freight company Gatik closed a $200M Series D, backed by Koch Disruptive Technologies, Millennium Management, ARK Invest, and Intact Private Capital. The startup has $600M in contracted revenue and 85,000 driverless orders on the books.
DiffusionOPSD converts image-level rewards into fine-grained denoising supervision, bridging RL alignment with on-policy distillation for text-to-image models.
Self-improving agents typically cap out at two meta-levels. This paper proposes a mechanism that lets systems recursively edit their own editing machinery, unlocking deeper emergent reasoning.
A family of universal multimodal embeddings from Tencent, enabling heterogeneous content to share a representation space for retrieval, recommendation, classification, and agentic systems.
Both crafted jailbreak prompts and post-deployment neuron pruning exploit a common weakness: safety-relevant activations are concentrated in easily removed components. NeuronGuard redistributes them robustly.
Branch sampling introduces local comparisons among intermediate decisions, addressing the sparse-outcome problem in multi-turn tool-calling RL for language models.
Alibaba's latest 27B MoE open-weights model is live on Together AI with a unified stack for both fine-tuning and dedicated model inference.
Kimi K3 leads in chat, reasoning, and coding agents; Qwen3.8-2.4T-A95B dominates vision; DeepSeek V4 Flash leads mid-size general purpose and function calling.
One-click setup for tool integrations across common applications, expanding Hermes Agent's skill ecosystem.
The fourth-generation data processing unit ties CPU, GPU, and network together, handling networking, security, and storage in every data center node as the trusted entity isolating host from fabric.
Built on Xe3P architecture with larger caches and deeper XMX engines, Crescent Island uses liquid-cooled chips and HBM4 memory targeting maximum AI flops-per-watt for agentic inference at scale.
Ars Technica reports from Beijing where humanoid robots broke sprint records but also crashed into barriers and caught fire, underscoring that robots remain engineered for single tasks rather than general-purpose autonomy.
Engineering prototype built around three internally developed XRING processors (O3, O100, D100) capable of running local models up to 120B parameters with a 16-core G2 Ultra NX GPU at 150W sustained.
Launched on Product Hunt (score 347). Delivers 4x depth and 2x coverage of PitchBook with 100+ event signals for agents to source deals and trigger outbound.
Shows real-time Claude usage meter and resets, optimizes tokens via context trimming, and carries conversations to another LLM when limits hit. 320 upvotes.
Analyzes user-agent conversations to find behavior drift, hallucinations, user frustration, and churn signals. Groups patterns and surfaces exact conversations behind each insight.
Provides a debit card that AI agents (OpenClaw, Claude) can use to make online purchases safely with one API call. Solves the payment-infrastructure gap for autonomous agents.
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:
