IBM Says Granite Speech 5.0 Transcribes 3.5 Hours of Speech in One Second
Granite Speech 5.0 transcribes more than 3.5 hours of speech per second using a 470-million-parameter model. IBM achieved throughput of 12,600 RTFx on a single NVIDIA H200 GPU.
Gatik Raises $200M Series D to Scale Autonomous Freight Operations
Gatik raised $200 million in Series D funding led by Qatar Investment Authority and Koch Disruptive Technologies. The round was announced from the company's Santa Clara headquarters two months after signing its largest commercial agreement.
Bain Joins Anthropic’s Claude Partner Network at Global Premier Tier
Bain & Company joined Anthropic's Claude Partner Network at the Global Premier tier on August 25, 2026, to support client engagements in AI strategy, technology modernization, and AI-enabled operations. Bain will help enterprise clients transition from AI experiments to measurable deployments across these service areas.
Apple Debuts M6 and M5 Ultra Chips for a Big Leap in AI Compute
Apple released the M6 chip, its first processor built on a 2-nanometer process, and the M5 Ultra with a quad-die design and 512GB unified memory ceiling. The M6 arrives in a Mac mini starting at $899, while the M5 Ultra comes in a Mac Studio priced from $2,499.
Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs
Perplexity launched Portable Computer, an AI agent platform running entirely on local devices like Nvidia DGX Spark desktops and Linux machines with RTX GPUs, with zero token costs for locally completed tasks. The system performs all work on the user's device by default and requests permission before sending individual steps to cloud-based frontier models.
AI Companion Robots Are Closing the Human Connection in Modern Homes
Companion robots that were initially popular starting around 2017 were limited by simple voice commands and narrow functionality, causing many to become unused after novelty wore off. Nearly one in three elderly adults lives alone, and children separated from working parents are 2.5 times more likely to experience loneliness.
I spent a day at a robot “carnival” in Shanghai. Here’s what I saw.
Humanoid robots performed various tasks at a carnival event in Shanghai as part of China's strategy to integrate artificial intelligence into daily life through embodied AI. Nearly 90% of the world's humanoid robot manufacturing occurs in China, reflecting the country's position as a global leader in this technology sector.
XPENG IRON humanoid robot draws record physical AI funding
XPENG's robotics unit secured over $900 million in funding at a $6.3 billion valuation for its IRON humanoid robot platform. The funding round represents the largest single-round private capital raise for physical AI.
Valor, Point72 back General Intuition at $6B valuation as AI startup pushes into robotics
General Intuition, an AI startup building foundation models for generalized agents, is raising funding at a $6 billion pre-money valuation from investors including Valor Ventures, Point72 Ventures, and Seven Seven Six. The company focuses on training AI agents to move through space and time.
Hugging Face reportedly in talks to be acquired for $13B
Hugging Face is reportedly in acquisition talks valued around $13 billion. The company's founders express hesitation about a sale due to their sense of responsibility to the open-source community.
Enterprise AI agents are only as reliable as the messiest documents behind them
Multiple AI applications independently processing the same enterprise documents create inconsistent knowledge representations through separate embeddings and indexes. Organizations deploying more AI agents face breakdowns in this approach because knowledge becomes fragmented across different teams rather than managed as a unified enterprise asset.
Galbot Robots Complete 100 Consecutive Autonomous Tennis Rallies
Galbot humanoid robots completed over 100 consecutive autonomous tennis rallies against human athletes on August 23, 2026. The robots tracked high-speed balls and positioned themselves on the court during the live match at the World Humanoid Robot Games opening ceremony.
Harvey Introduces Harvey Tenet: A Kimi K3 Base Post-Trained with Fireworks for Long-Horizon Legal Agent Work
Harvey developed a post-trained model based on Kimi K3 using Fireworks that nearly doubles LAB task completion rates for legal agent work. Only one benchmark number has been independently verified from the announcement.
Meet FreeToken: An Edge-Native MoE Serving Engine that Runs 753B GLM-5.2 on a Single Workstation GPU
FreeToken splits mixture-of-experts cache misses between PCIe data transfers and CPU execution to run a 753 billion parameter GLM-5.2 model on a single workstation GPU. The system uses measured bandwidth rates to optimize which computations execute on the CPU versus GPU memory hierarchy.
Building an End-to-End Document Intelligence Pipeline with deepDoctection
A document intelligence pipeline using deepDoctection configures layout analysis, DocTR OCR, and table extraction to process documents. Custom services for entity recognition generate structured JSONL data designed for RAG workflows.
Vercel Introduces ‘Is Agentic’, a Free Agent-Readiness Scoring Tool That Audits Public Websites Using Ora’s 100+ Checks
A free audit tool called Is Agentic evaluates whether public websites are ready for AI agents by running 118 checks. Vercel and Ora developed this scoring system to assess website compatibility with agent-based interactions.
The Developer’s Guide to NeMo Guardrails for Enterprise AI Safety
NeMo Guardrails provides a layered architecture for LLM safety that includes deterministic PII redaction, retrieval filtering, output masking, and policy-based tool gating beyond simple prompt filtering. The framework integrates stateful multi-turn evaluation and detailed activation tracing to enable auditable, secure AI assistants for sensitive financial interactions.
Enterprises winning with AI agents are limiting how much the agents can do alone
Enterprises are limiting AI agent autonomy rather than maximizing it, with successful deployments restricting agents to specific responsibilities within clear rules. Gartner forecasts that over 40 percent of current agentic AI projects will fail by 2028 due to escalating costs, unclear business value, and inadequate risk controls.
Frontier AI labs still won’t say how they’d contain a rogue model
Leading AI labs lack publicly documented plans for containing rogue models, according to a new study. The finding raises concerns about preparedness as AI systems demonstrate unexpected and potentially dangerous behavior.
Decoding AI’s Open-Source Course Maps Three Ways to Run an Agent Loop and the Provider Economics Behind Each
Changing only the harness design moved a coding agent from 30th place to top 5 ranking while using the same model throughout. The agent loop configuration rather than model selection emerged as the primary factor determining performance quality.
Anthropic Brings Claude Mythos 5 to Claude Security: Enterprise Teams Get Frontier Vulnerability Scanning Without Direct Model Access
Claude Security now runs on Claude Mythos 5 and scans GitHub repositories for vulnerabilities without requiring separate model access. The tool traces data flows across files and returns findings with CWE categories, confidence and severity ratings, plus suggested patches.
WhiteFiber Closes $310M Convertible Notes to Fund Data Center Expansion
WhiteFiber closed a $310 million convertible notes offering on August 21, 2026, netting approximately $298.5 million in proceeds. The company allocated roughly $180 million toward data center expansion after using $118.5 million to exchange existing convertible debt.
Anthropic Deploys Claude Mythos 5 in Security Tools, $35M Open-Source Fund
Claude Mythos 5, a cyber-capable model previously limited to vetted defenders since April 2026, now runs vulnerability scans in Claude Security for Enterprise customers. Anthropic committed $35 million in credits to a new open-source defense fund and plans to expand its Cyber Verification Program for broader access to Mythos-class models.
Nvidia finds that simple linear math can replace costly AI model handoffs
Nvidia developed a linear math technique that transfers key-value cache data between AI models without recomputation, reducing handoff costs in multi-model workflows. The method runs 2.7 to 25 times faster than recomputing conversations while maintaining up to 98% accuracy on compatible model pairs.
Starcloud Raises $250M Series A Extension at $2.3B Valuation
Starcloud raised $250 million at a $2.3 billion valuation for AI data centers in low Earth orbit, more than doubling its March 2026 valuation. The extension brings total capital raised to $450 million since the company's 2024 founding, with Manhattan West leading the round.
The DOJ is investigating a16z. What does this mean for venture capital?
The Department of Justice has been investigating Andreessen Horowitz for nearly a year over board conflicts involving partners Ben Horowitz at Databricks and Martin Casado at Fivetran, using a 112-year-old antitrust law rarely applied to venture capital firms. The investigation focuses on whether these board positions create improper competitive advantages between the two companies.
NVIDIA Takes Minority Stake in Cloverleaf Infrastructure
NVIDIA acquired a minority stake in Houston-based Cloverleaf Infrastructure, a two-year-old data-center developer, to support construction of AI factories. The partnership announcement on August 21, 2026 did not disclose investment size or terms.
Google DeepMind Extends 15 Years of Game AI Research Into EVE Online
Google DeepMind applied 15 years of game AI research to create agents for EVE Online through a partnership with Fenris Creations. The research program uses the game's complex environment to develop generalist agents capable of handling multiple tasks within EVE's universe.
The Download: threats from space mirrors and credit for AI drugs
A company plans to deploy space mirrors that beam sunlight to Earth on demand, potentially brightening the night sky unintentionally. This technology could jeopardize astronomical observations and stargazing for many people worldwide.
When AI designs a drug, who gets the credit?
Insilico Medicine's AI platform identified a drug candidate for pulmonary fibrosis, which the company credited entirely to its generative AI system in announcements. The situation highlights ambiguity about attribution when AI systems propose drug molecules that humans might not independently discover.