Google Research announced a new Federated Learning system that uses Trusted Execution Environments (TEEs) to provide externally verifiable privacy guarantees. The system shifts computation to the server to improve training speed, accuracy, and device coverage. Gboard has already adopted the new system and is benefiting from substantially faster compute times.
AppleResearch paper· 1h agoApple researchers found that enhancing language discrimination during pretraining of multilingual speech models (using an auxiliary language classifier and per-language k-means targets) reduces the performance gap compared to monolingual models. In a controlled English/French HuBERT setting, phone discrimination error decreased from 11.6% (bilingual baseline) to 10.4%, while lexical and prosodic performance improved, in some cases matching or exceeding monolingual baselines. The strongest gains occurred when language discrimination was introduced in the first training iteration.
SunoCapability change· 17h agoSuno released Speech, a beta audio model that generates voice and background music together as a single track, allowing users to create spoken audio set to original music by typing an idea and describing the desired voice and musical style.
Reported by
Unite1 source▾▴
LG AI Research presented a study applying mechanistic interpretability to graph transformers, specifically analyzing how TokenGT processes graph data. The work found that models trained on degree calculation, cycle detection, and shortest-path distance tasks all begin with a shared local-structure computation, such as encoding node degree through ID-matching attention in the first transformer layer.
ServiceNow CoreAI introduces AutoSynthData, a pipeline that identifies capability gaps in a target model by evaluating it in an enterprise environment, then generates and validates new training tasks that exercise those weaknesses. The system uses a stronger teacher model to characterize successful behavior and produces tasks that are feasible, realistic, and difficult for the current model. The approach is illustrated with the EnterpriseOps Gym environment and a released dataset.
Language models can already help researchers search the literature, synthesize evidence, and work through complex questions. But scientific work places particular demands on these models—answers need to stay grounded in evidence, the models need to preserve what the evidence actually supports rather than quietly broadening a study’s conclusions, and researchers need to be able to verify the final outputs. We see that in how scientists use Asta , our agentic platform for scientific work. Instead of simple keyword searches, users often bring substantial context and many constraints—for example, asking Asta to compare approaches across a body of literature while accounting for a particular method, population, or setting. Many also return to generated reports later, treating them as working research artifacts rather than one-off answers. We wanted to help scientists generate cited reports faster, with a model they could download and run themselves. To do that, we tested whether a small, open model trained specifically for scientific report generation could match the report quality of the proprietary models we were using, while reducing generation time and serving costs. …
Local AI is becoming more useful by the token. As AI agents move from experiments into everyday development, increasingly capable open models are shrinking to fit on more devices, giving builders more to run locally. Coming this month, NVIDIA DGX Spark will be available with 64GB of unified memory from top manufacturer partners — Acer, ASUS, Dell, Gigabyte, HP and MSI — giving developers, researchers and AI enthusiasts a new configuration with DGX OS and the NVIDIA AI software stack ready to use from day one. The new SKU runs capable local agents on device — privately, without cloud dependency. And when workloads grow, two units can cluster together via NVIDIA Sync Cluster Assistant without any additional setup. A New Starting Point for Personal AI Supercomputing DGX Spark combines NVIDIA Grace Blackwell compute, unified memory, NVIDIA ConnectX-7 networking and an NVIDIA CUDA -accelerated AI software stack in one system. It’s a complete local AI platform for agents, inference, fine-tuning, data science and edge development. …
AppleResearch paper· 1 day ago Apple researchers introduce RLTL;DR, a method where after each failed attempt, the policy sees the verifier outputs and writes its own TL;DR insight, conditioning subsequent rollouts on all previous insights. On challenging tool-calling and coding datasets where standard GRPO training of a Qwen 3.5 9B Thinking policy achieves 0–1% Pass@1, RLTL;DR achieves 14–31% Pass@1 with insights in context during training and 12–13% when no insight is in context at evaluation. The paper also presents SFTL;DR, training only on (task, insight) tuples, which recovers almost full performance from only 4k tuples.
AppleResearch paper· 1 day ago Apple researchers found that, under equal time and using the same frontier LLM, a minimal-harness coding agent baseline matches or outperforms open-source state-of-the-art harnesses on current MLE benchmarks, suggesting the backbone model is the primary driver of performance.
Perplexity announced ready-to-use Skills for Perplexity Computer curated for U.S. American Express Business Card Members with a Perplexity Enterprise subscription. The collection provides pre-built AI workflows for tasks such as tax prep, cash flow forecasting, and marketing campaign generation. Eligible Card Members can link their American Express Business Card through Plaid to access the Skills.
Barclays is expanding its collaboration with Anthropic to deploy Claude across the bank, aiming to accelerate software development, modernize legacy systems, and improve operational efficiency. The bank expects Claude Code adoption to reach 50% of its developer population by end of 2026 and a majority of software engineers by 2027. Already, over 16,000 colleagues use a Claude-powered knowledge assistant that has handled over one million searches, and Claude models process approximately 120,000 emails daily in the Global Markets business.
LumaCapability change· 1 day agoLuma launched Variants, a feature that lets users upload one approved static ad and automatically generate versions for multiple placements and languages. The tool supports five standard ad formats and translation into selected markets, keeping logos, headlines, and CTAs intact. It is available now inside Luma's Discover tab.
Carl Johnson, a sales development leader at Anthropic, shares how a Claude-powered buying agent now answers most inbound customers, and how that changed the way our sales team works. Category Enterprise AI Product Claude Platform Date September 30, 2026 Reading time 5 min Share Copy link https://claude.com/blog/how-anthropics-sales-team-rebuilt-inbound-with-claude-managed-agents Author(s) Carl Johnson As a sales leader, it pains me to admit that not long ago, people who wanted to buy Claude for their company weren’t getting the answers they needed quickly enough. They had filled out our Contact Sales form but would wait too long to hear back, sometimes for multiple days. Most of their questions were simple: what a plan costs, whether there's a seat minimum, or whether we can meet HIPAA’s contract requirements. The answers were in our documentation and support articles, but customers wanted someone to walk them through, quickly. So we built a buying agent on Claude Managed Agents (beta) that takes a prospect from "I want Claude for my company" to a completed purchase. …
AI tools help consultants speed up research, analyze data, synthesize client documents, draft deliverables, capture meeting notes, keep client knowledge organized, automate repetitive work, build proposals, manage time, and generate new leads. This gives them more time to focus on what they're paid for: applying their expertise and judgment to solve clients' problems. Perplexity is one of them. Consultants use it for client deep dives, market sizing, proposal development, final-pass document review, and as a general AI assistant. Every Deep Research answer comes with inline citations to the sources it pulled from, making it easier to verify accuracy. Projects keep all chats, files, and custom instructions for each client in one place, and Perplexity Computer , its AI Agent, completes complex multi-step tasks. But we’re not going to focus on Perplexity. What are the other tools that should be at the top of your consideration list? We’ve covered 10 other AI tools for consulting, organized by use case. Each entry lists what the tool does, its core features, current pricing, and alternative tools worth considering. 10 Best AI Tools for Consultants at a Glance …
ManusCapability change· 1 day ago Manus 2.0 introduces Game Dev, a feature that enables people without coding experience to create, tweak, and share games using AI. It includes a real-time tweak panel, asset management with AI editing, and built-in multiplayer server support. The feature is part of the Manus 2.0 update launched on September 28.
Anthropic made Claude for Government generally available to U.S. federal and state agencies, offering coding and agentic capabilities through a FedRAMP High authorized environment. The platform includes purpose-built governance controls, audit logs, and usage-based billing with no seat fees. Claude Code CLI and Claude for Microsoft 365 are also available in early access through the same environment.
Search agents powered by large language models (LLMs) are transforming how enterprises retrieve information. Rather than requiring users to craft the perfect query, a search agent autonomously decides what to search for, which retrieval strategy to use, and when to stop searching. It does this across multiple rounds of interaction, refining its approach based on what it has already retrieved. However, getting this multi-step behavior to work well is hard. No base model arrives knowing your tools or your environment. Prompt a small model and you rarely get dependable multi-turn behavior. Prompt a frontier model and it often works, but you pay for that capability in latency and cost. Fine-tuning offers a third path: you teach a small model your tools and environment directly. The result is a small model’s speed and cost with the reliability that would otherwise require a frontier model. Even though fine-tuning is the natural next step, the traditional approaches each fall short. Supervised fine-tuning (SFT) depends on expert demonstrations of ideal multi-turn trajectories, which are costly to collect and usually don’t exist for your setup. …
AppleResearch paper· 2 days ago Apple researchers systematically studied conditioning methods for LLMs, finding that efficient steering often degrades fluency and that activation steering is less effective on instruction-tuned models than on base models. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics correlate well with costly LLM-as-judge scores.
Microsoft Research developed a machine learning pipeline that forecasts space-weather risk for 66,935 substations in the continental United States, using solar-wind data and local geological factors to provide location-specific risk estimates 30–60 minutes ahead of potential impact. During evaluation, the system detected 76.5% of major events (≥10 nT/min), 81.2% of severe events (≥20 nT/min), and 64.1% of extreme events (≥50 nT/min).
LovableSafety or policy· 2 days agoLovable's security team discovered a vulnerability in TanStack Start, a framework used by Lovable apps, and reported it to maintainers. Firewall protections were deployed for hosted apps, and affected projects are automatically updated when users make changes. No evidence of exploitation was found.
