The Arm-based superchip pairs up to 20 CPU cores with a Blackwell GPU carrying 6,144 CUDA cores and 128GB of LPDDR5X at up to 300 GB/s, aimed at agentic AI workloads on the desktop.
The RISC-V AI chip company's oversubscribed round brings total funding past $2.19 billion, with ARK Invest, Fidelity, Morgan Stanley and T. Rowe Price among the investors. Capital is earmarked for volume manufacturing of Galaxy systems and commercial expansion.
A Google DeepMind agent autonomously picks Turing machines, spots patterns in execution traces, and proposes compact acceleration rules verified with Lean proofs — new confirmed rules for 11 previously unresolved machines. An early example of agents choosing their own problems on uncomputable territory.
Early-fusion multimodal models pretrained from scratch on 100M image-text pairs show language pretraining is not required for strong T2I generation; Halcyon-0.52B beats LlamaGen-XL/8.
WildAWM is a scalable pipeline for synthesizing controllable environments to train generalist computer-use agents, avoiding the security risks of learning directly on real systems.
Post-training safety degrades when later training omits safety data and often costs capability; the paper proposes inserting safety data as a mid-training phase instead of a post-hoc layer.
Across 22 benchmarks, simply raising MLLM inference budget yields only modest gains on tasks needing grounded perception; coupling visual agents with tools and memory closes the gap.
131 protein-structure tasks and 4,249 rollouts show frontier agents retrieve structural metadata well but compute geometry from coordinates far less reliably.
A reinforcement-learning-trained model that automatically detects errors and failure points in AI agent traces, aimed at debugging production agent workflows.
The open-source project lets Android APK apps run on macOS and drew significant Hacker News attention within hours of release.
Norforge released Cairn, giving security teams visibility into agent-based attacks as they investigate AI-driven intrusions.
An AI legal assistant designed to only cite case law it can verify, targeting hallucination risk in legal research.
An open-source realtime talking-avatar system that runs locally on consumer GPUs with no cloud service required.
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:
