AI News

Federated fine-tuning of on-device large language models (LLMs) faces a significant computing burden. To overcome this limitation, split learning (SL) has emerged as a promising solution, which offloads the primary training workload to a powerful server. However, SL requires exchanging high-dimensional activations and gradients between clients and the server, resulting in prohibitive communication costs. To overcome this challenge, we propose SplitLite, a communication-efficient split federated LoRA fine-tuning method that exploits the low effective rank structure of consecutive-epoch activation and gradient residuals. Our key finding is that, when LoRA uses rank $r$ updates in parameter space, the activation and gradient residuals of the same data sample between adjacent epochs also exhibit effective rank-$2r$ and rank-$4r$ structures, respectively. By revealing this property, SplitLite transmits only quantized truncated singular value decomposition (SVD) residual factors, thereby significantly reducing both activation uplink and gradient downlink traffic. Extensive experiments on the GLUE benchmark across a series of advanced on-device LLMs demonstrate that our method reduces activation uplink communication costs by up to 93.5\% and total communication costs by up to 83.7\%, without performance degradation.

Space robots operate in extreme environments where hardware degradation can critically compromise traditional control strategies. While continual reinforcement learning offers a promising mechanism for online adaptation, it inherently requires access to a reward signal during deployment. However, precise reward computation in space is often infeasible due to the lack of external tracking systems and the overall complexity of the environment. To address the challenge of unobservable rewards, we introduce a reward-free continual learning framework that leverages latent-state world models. By pre-training a model-based agent across diverse simulations, the world model learns a robust predictor of the reward structure within its latent space. Upon deployment to an environment with severe hardware degradation, we freeze the observation encoder and reward predictor to update only the transition dynamics of the world model through unsupervised rollouts. By training the policy entirely on imagined trajectories generated by this updated world model, the agent adapts to altered dynamics without receiving new rewards. We demonstrate our approach across simulated planetary traversal, orbital navigation, and precision assembly tasks subjected to severe morphological failures.

Open source AI is booming, according to Hugging Face CEO Clem Delangue. The company has grown into something like a GitHub for AI in recent years, where AI builders can share and download open models and datasets, now used by roughly half the Fortune 500. Delangue has seen the same story play out again and again: companies start out on frontier APIs, but as they scale, the costs push them towards open source models. On this episode of TechCrunch's Equity podcast, Rebecca Bellan talked to Delangue about why the open vs closed source fight matters in the wake of Anthropic’s halted Fable release, and why he's worried about the possibility that a handful of big companies could end up controlling everything. Chapters: 00:00 Intro 00:33 Breaking down open source growth data 04:34 What's driving the open source resurgence 08:47 Who’s using Hugging Face, and how? 10:28 China overtakes the US in open model downloads 16:34 Safety, access, and the risk of AI power concentration 24:03 Hugging Face's approach to legal risk 28:00 Turning down Nvidia 31:47 Underinvested opportunities: local AI, bio, robotics

At Hot Chips 2026, Arm is showcasing their AGI data center CPU, the company's first complete commercial CPU design. Based around Arm's Neoverse V3 CPU cores, AGI is designed to power the next generation of agentic AI servers

Wizstar helps founders, creators, and makers build an expressive, accurate digital ambassador that looks and moves like you. It gestures naturally, interacts with objects, and maintains precise lip sync—even during head turns or things in front of your face. Create product updates, training, localized videos, and branded content without being on camera every time. Bring your digital presence to life and scale it across markets, languages, and channels.

HyNote is your all-in-one AI note-taking solution, supporting multiple platforms and document formats. It captures every key detail from your meetings and work files, instantly delivering summaries to turn scattered information into actionable insights.

Start using HyNote for free now!

Supernova connects your startup’s live data to Claude and Codex, so anyone can ask questions, investigate performance, and run complex analysis in the AI tools they already use. Connect Stripe, HubSpot, PostgreSQL, and 30+ other apps, then analyze revenue, pipeline, customers, usage, and operations without waiting on engineers or moving everything into a traditional BI stack.

Great paper on multi-agent systems for code review.

It's challenging to know how many coding agents to use to address a problem.

The default fix for weak agentic code review is more agents. In turns out that scaling agents to a large number gives diminishing returns on

onemcp unifies all your Model Context Protocol (MCP) servers behind one portal endpoint. Connect your AI client once — native OAuth and Code Mode, no credential sprawl or context bloat.

AI-assisted UI testing: natural-language visual checks plus deterministic WCAG accessibility audits via axe-core - gojiplus/layoutlens

Many people discovering the importance of open source harnesses, sovereignty, and controlling your own stack today.

Hermes Agent

The chipmaker's minority investment and collaboration with the infrastructure developer highlight how private capital and expertise are tackling the power and site bottlenecks holding back American AI growth.

Verified, self-healing knowledge layer for AI coding agents — ranked search with trust verdicts (STRONG/WEAK/STALE/REBUILT), session auto-sync, grep guard. Orchestration over Graft-style semantic memory backends. - ArihantDeva/heimdall

Large language models (LLMs) often produce fluent but incorrect answers with unwarranted confidence. A central limitation is that standard LLMs represent uncertainty through a single predictive distribution, conflating epistemic ignorance with genuine ambiguity. We introduce Credal Large Language Models (CLLMs): an ensemble of LoRA adapters induces a credal set whose lower and upper probabilities expose the spread of plausible predictive distributions rather than collapsing to a single softmax output. From this representation we derive two complementary commitment scores. Credal Token Commitment (CTC) is a token-space score that combines lower-bound support, credal width, and intersection entropy, computed without additional generation. Semantic Commitment Consistency (SCC) extends commitment to semantic space using sampled completions, with SCC-Gap measuring the mismatch between token-level and semantic-level support. We evaluate hallucination detection, calibration, selective prediction, and reasoning on Gemma-2-9B, Llama-3.1-8B, and Qwen2.5-7B across OpenBookQA, CoQA, TriviaQA, and ARC-Challenge. CLLM is the best method on QA accuracy at competitive expected calibration error, and CTC tracks the best hallucination AUROC within 1.5 pp on most settings without additional generation. On selective prediction at 80% coverage, CLLM with SCC reaches 99.0% accuracy on OpenBookQA, and on ARC-Challenge CLLM with Csem confidence achieves <= 0.6% ECE across the three backbones.

Group-based reinforcement learning methods such as GRPO for large language models avoid training a critic by sampling multiple responses for each prompt. A reliable critic could instead estimate token-level advantages from one response, but standard critic-based training recipes are often unstable. We study this instability and develop \textbf{Best-Practice Critic Optimization (BPCO)}, a recipe that combines DPPO, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized policy advantages, and length-adaptive generalized advantage estimation. Because the critic is used only during training, BPCO can also condition it on reward-defining information, such as a reference answer or grading rubric, that is hidden from the policy. Controlled experiments isolate the effect of each design choice. Across mathematical reasoning tasks with models ranging from 1.5B parameters to 30B-A3B mixtures of experts, BPCO improves a strong critic-based baseline consistently, and matches or exceeds a group-based baseline while sampling one response per prompt. The same recipe also improves learning with rubric-based rewards. These results show that a carefully designed critic provides a reliable alternative to group-relative advantage estimation. Code is available at https://github.com/QPHutu/golden_critic

XPeng raised over $900M for its robotics unit, advancing humanoid robot IRON's development, aiming for mass production and global commercialization by 2027.

Nvidia's dedicated inference accelerator Groq 3 LPX enters full production to supercharge AI agents - SiliconANGLE

Your agents forget everything the moment a session ends. You stuff more context into every prompt, burn tokens on redundant history, and still ship responses that feel like amnesia with extra steps. Actx0 is the memory layer your agents are missing — a drop-in infrastructure that stores what matters, retrieves it in milliseconds, and keeps working across sessions, agents, and apps. Built for production teams who care about latency, cost, and control.

Stop overpaying for LLM inference and juggling multiple API integrations. Router.com provides a single API endpoint that routes every request to the lowest-cost model that meets your performance threshold—backed by Ramp's financial visibility. Start routing smarter today at router.com!

Google Antigravity - Build the new way

Rust project. Local Free AI transcription based on whisper, parakeet, MoonShine, SenseVoice, Zipformer. - GitHub - AntoineChatry/Dictata: Rust project. Local Free AI transcription based on whisper, parakeet, MoonShine, SenseVoice, Zipformer.

Delegate Grok Bot coding work to your own local Claude Code / Codex CLIs via a custom MCP connector - HarjjotSinghh/locum

Travel just two hours west of Beijing by train, and you'll find yourself surrounded by the rolling grasslands and ancient cinder cones of Inner Mongolia. Over the last few years, Inner Mongolia has become the hottest place in China to build AI data centers.

In Ulanqab, nearly 100 data centers have been opened or begun construction since 2016, with Chinese companies committing to projects totaling 12.5 gigawatts of capacity. The city's high elevation and cold winters make it an energy-efficient location for data centers, while its proximity to Beijing and low electricity costs attract companies like DeepSeek, ByteDance, and Alibaba to invest in their own infrastructure rather than renting from cloud providers.

Chinese ⁠smartphone maker Honor’s self-developed humanoid ⁠robot Lightning ⁠ran ​the 100m in 9.32 seconds, ⁠beating the human world ⁠record of 9.58sec