How a former DeepMind researcher raised at a $300M pre-seed valuation before launching a product
A former DeepMind researcher raised $300 million at pre-seed valuation before launching a product, leveraging over a decade of AI development experience. Andrew Dai believes visual AI represents a major frontier in artificial intelligence based on his work contributing to systems that informed ChatGPT's development.
The Download: OpenAI unveils GPT-Red and heat pumps rise in the US
OpenAI developed an LLM super-hacker called GPT-Red to identify vulnerabilities in its models before deployment. The system acts as a sparring partner to test and improve the safety of other AI models.
NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI
NVIDIA introduced the T3000 and T2000 modules based on the Thor architecture to support general-purpose robots and autonomous machines in real-world deployment. These compact, power-efficient AI supercomputers are designed to run foundation models at the edge for mass-market robotics applications.
Agentic orchestration: Enterprise AI organizations have a deployment problem, not a platform problem — and most are calling chatbots agents
Anthropic's Claude is the primary platform for 40% of 101 enterprises surveyed for agent orchestration, more than double its nearest rivals. Most deployed "agents" function as chatbot wrappers rather than true orchestrated systems, with enterprises prioritizing multi-step execution reliability and model quality over advanced control mechanisms.
Amazon AGI director says AI agent reliability, not capability, is blocking enterprise deployment at VB Transform 2026
Eighty-five percent of enterprises pilot AI agents, but only five percent deploy them to production because reliability issues cause agents to fail in real-world conditions despite passing internal evaluations. An Amazon AGI director identified four reliability dimensions,consistency, robustness, predictability, and safety,needed for enterprise deployment, citing an example where a QA agent read serial numbers correctly for two months before a software change made its vision encoder sensitive to screen position.
Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer
OpenAI created GPT-Red, an LLM designed to identify vulnerabilities in its models by simulating cyberattacks. Training GPT-5.6 against this adversarial system produced what the company claims is its most robust model release.
Whatnot acquires Shaped to power real-time live shopping recommendations
Whatnot acquired AI startup Shaped to enhance its real-time recommendation and search capabilities for livestream shopping. The acquisition will strengthen Whatnot's personalization features as the platform expands into additional product categories.
PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B That Run on Laptops and Phones
Ternary Bonsai 27B uses weights of negative one, zero, and positive one at 1.71 bits per weight, achieving an ideal size of 5.9GB. A 1-bit variant with binary weights also ships under Apache 2.0 license for running on laptops and phones.
OpenAI’s first hardware device is reportedly a screenless speaker that can move
OpenAI is developing a screenless speaker with mechanical elements that can move independently. The device is designed to function as a physical companion that manifests ChatGPT capabilities.
Apple opens its new Siri AI to everyone with the iOS 27 public beta
Apple released iOS 27 public beta, making its revamped AI-powered Siri available to iPhone owners ahead of the official fall launch. The public beta gives users early access to the new assistant and other features before the software's official release.
Nemotron Labs: How Open Models Give Enterprises and Nations AI They Can Trust, Control and Customize
NVIDIA released Nemotron, an open model designed for enterprises to customize and deploy AI systems tailored to specific business needs. The model enables organizations to build domain-specific applications while maintaining control over accuracy, trust standards, and workflow improvements.
AWS and Bluesight build AI for hospital 340B compliance
Bluesight developed an AI layer called Prism that connects hospital pharmacy and compliance data across its product suite. Prism Assistant for ControlCheck is now generally available and operates across 20 health systems.
Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency
Performance per watt measures how many tokens an AI infrastructure system generates within a fixed power budget. This metric directly determines an AI factory's revenue and profitability since power represents the inescapable constraint limiting token generation.
Mistral AI Releases Robostral Navigate: An 8B Model Enabling Robots to Navigate Complex Environments Using a Single RGB Camera
An 8 billion parameter navigation model enables robots to follow plain-language instructions using only a single RGB camera without requiring LiDAR or depth sensors. The model achieves 76.6% success on R2R-CE validation through pointing methods and online reinforcement learning techniques.
Skyfall AI Releases MORPHEUS: A Persistent Enterprise Simulation Benchmark That Makes Continual Reinforcement Learning Necessary Under Structured Non-Stationarity
MORPHEUS is a persistent enterprise simulation platform where worlds never reset and agents face parameterizable regime shifts evaluated across six metrics. Standard reinforcement learning algorithms like PPO, HER, EWC, and LCM all perform significantly below the theoretical upper bound on the platform.
Satya Nadella has issued a shocking warning to companies using AI
Microsoft's Satya Nadella warned companies that relying on proprietary AI models from major labs creates dependency risks. He cautioned that organizations should consider alternative approaches to avoid becoming trapped by single-vendor AI systems.
What Anthropic’s latest AI discovery does—and doesn’t—show
Anthropic published research exploring whether AI models can experience pain. The company, valued at nearly one trillion dollars, is known for publishing unconventional research on AI capabilities and consciousness-related topics.
How self-improving harnesses are rewriting the agent engineering playbook
AI agents can now write and optimize their own execution logic through new self-improving harness frameworks. These tools are becoming a primary focus in agent engineering, enabling systems to refine their performance automatically.
ACRouter picks the smartest AI model per task, beating Opus-only setups by 2.6x on cost
ACRouter dynamically routes prompts to different AI models based on task requirements, achieving 2.6 times lower costs than setups using only Opus. The framework uses a Context-Action-Feedback loop to track model performance and optimize routing decisions without requiring training of large models.
AI agent crawlers now need permission. Here’s how to get it
Starting September 15, AI agent crawlers will be blocked by default across parts of the web unless they receive explicit permission. Cloudflare announced this change on July 1, affecting bots that fetch pages in real time for users awaiting answers.
Building a Foundation Stack for General-Purpose Robots
X Square Robot proposes an integrated stack combining data, world models, and action models as the foundational recipe for general-purpose robots. The approach treats successful robot interactions as those that change the physical world as intended, not merely joint movements.
Stanford Researchers Introduce TRACE: A Capability-Targeted Agentic Training System That Turns Recurrent Agent Failures Into Synthetic RL Environment
TRACE diagnoses capability gaps from agent failure trajectories and synthesizes targeted training environments with LoRA adapters for each gap, improving performance by 15.3 points on τ²-Bench. The system achieved 73.2% Pass@1 on SWE-bench Verified through token routing across expert adapters.
Meet NeuroVFM: A New Neuroimaging Foundation Model Trained With Vol-JEPA on Uncurated Clinical MRI and CT Volumes
A neuroimaging foundation model called NeuroVFM was trained on 5.24 million clinical MRI and CT volumes using Vol-JEPA without requiring radiology-report labels. The model learns brain anatomy and pathology from uncurated volumetric medical imaging data.
Guide to Loop Engineering: How ‘autoresearch’ and ‘Bilevel Autoresearch’ Turn AI Agents Into Autonomous Machine Learning ML Research Loops
AI agents can execute autonomous research loops that eliminate manual back-and-forth interaction by automatically iterating through investigation cycles. The approach uses frameworks called autoresearch and Bilevel Autoresearch, which are demonstrated in Andrej Karpathy's repository and an associated research paper.
DeepSeek cut prices 75%. The 100x problem remains
DeepSeek reduced V4-Pro pricing by 75 percent, but agent systems consume tokens at rates faster than prices decline, creating substantial cost pressures despite cheaper models. A single user request in agentic workflows generates multiple billable operations through planning, retrieval, tool use, verification, and summarization loops, multiplying costs by approximately 100 times compared to simple chatbot responses.
Mira Murati’s Thinking Machines Lab Makes The Technical Case For Human-Centered AI Built On Customizable Model Weights
Thinking Machines Lab published an essay proposing human-centered AI through decentralized model ownership and alignment as technical challenges. The framework uses LoRA fine-tuning methods that allow teams to train and retain their own customizable model weights.
A Coding Guide to NVIDIA’s Tile-Based GPU Programming: From cuTile and Triton Kernels to Flash Attention
NVIDIA's tile-based GPU programming uses TileGym to operate on whole data tiles rather than single threads, with implementations falling back from cuTile to Triton kernels depending on available hardware. The tutorial covers vector addition, fused GELU, row-wise softmax, tiled matrix multiplication, and flash attention operations validated against PyTorch.
57% of enterprises have watched AI agents be confidently wrong. The fix is an agentic context layer, but who has one?
Fifty-seven percent of enterprises reported AI agents providing confidently incorrect answers due to missing or inconsistent business context in the past six months. Thirty-eight percent of enterprises rely on document retrieval as their primary method for giving agents business context, while seventy-five percent lack a dedicated agentic context layer to govern this information.
OpenAI introduces ChatGPT Work, a cloud-based AI agent that manages tasks across email, Slack and calendars
OpenAI launched ChatGPT Work, an AI agent powered by GPT-5.6 that executes complex multi-step tasks across email, Slack, calendars, and code repositories. The agent can gather context from connected apps to produce finished documents, spreadsheets, presentations, reports, and websites while working independently on projects for hours.