Live Feed
Ornith-1.0: MIT-Licensed 397B MoE Scores 82.4% SWE-Bench Verified, Claims Open-Source Coding Lead
DeepReinforce releases Ornith-1.0, an open-source agentic coding family trained with a method that jointly optimizes task scaffolds and solutions. The flagship 397B MoE posts 82.4% on SWE-Bench Verified and 77.5% on Terminal-Bench 2.1, leading all published open-weight coding models.
Notion Shuts Down Its Email Client on September 22 — Over Half of Users Already Let Agents Run the Inbox
Notion is winding down Notion Mail across web, desktop, and iOS on September 22, 2026. The company cited its own usage data: more than half of Notion Mail users manage email without ever opening their inbox, having handed the work to AI agents. The product launched April 2025 and lasted barely 14 months.
Unconventional AI Raises $475M at $4.5B to Build Oscillator Chips That Cut AI Power Use 1,000x
Naveen Rao, who sold Nervana Systems to Intel for ~$400M and co-founded MosaicML (acquired by Databricks for $1.3B), has raised $475M at a $4.5B seed valuation for Unconventional AI. The company released Un-0, an image generation model on a simulated oscillator-based architecture, hitting FID 6.74 on ImageNet 64x64. The target: 1,000x less energy than conventional GPU inference.
IBM Breaks the 1nm Barrier: Nanostack 3D Architecture Packs 100 Billion Transistors Per Chip
IBM unveiled sub-1nm chip technology on June 25, using a 3D nanostack architecture that nearly doubles transistor density over its previous generation. The company is targeting AI compute workloads and a decade-long scaling runway.
Colossus Has 770,000 GPUs and 1GW of Compute — Two Thirds of It Never Touches the Grid
An investigation into xAI's Colossus cluster finds three buildings across the Tennessee-Mississippi border, 770,000 GPUs, and a 1.2GW natural gas plant that xAI built itself to bypass the utility grid entirely.
AA-Briefcase Runtime Data: Claude Opus 4.8 Takes 23 Minutes Per Task, GPT-5.5 Takes 11
AA-Briefcase added time-per-task data, and the result is a procurement problem hiding inside a benchmark. Claude Opus 4.8 is the strongest available model, but GPT-5.5 xhigh finishes comparable long-horizon work in roughly half the time.
ByteDance Seedance 2.5 Breaks the 15-Second Ceiling: 30-Second Native AI Video, 50 References to Veo 3.1's 3
Seedance 2.5 generates a full 30 seconds in a single diffusion pass — no stitching — and accepts 50 multimodal reference inputs. ByteDance's enterprise video platform is at $2B ARR before the model even ships publicly in July.
Cursor Is Training a 1.5-Trillion-Parameter Model From Scratch on Colossus — First IDE Company to Own the Full Stack
At the Compile keynote on June 16, Cursor disclosed it is pre-training a 1.5T-parameter model from scratch on xAI's Colossus using 100,000+ GPUs — abandoning the Kimi K2.5 fine-tuning approach it used for every prior model. The model targets general-purpose frontier capability, not just coding.
Former Bitcoin Miner Signs $1.2B Deal to Convert Michigan Data Center for AI Compute
Hyperscale Data ($GPUS) will retrofit 60,000 sq ft of Bitcoin mining space at its Michigan campus into AI infrastructure for a California neocloud. The deal starts at 20 MW and scales to $3B if expansion options are exercised.
Meta Is the Only Major US Lab Without a Pre-Release AI Safety Deal
The Trump administration is pressing Meta to join a voluntary government review program for frontier AI models. OpenAI, Anthropic, Google, xAI, and Microsoft have all agreed. Meta has not.
NSA Loses Mythos Access After White House Export Order — Early Commercial Testers Keep Theirs
Parts of the NSA have lost access to Anthropic's Mythos 5 after Trump administration export restrictions took effect. Early commercial testers retained access through a preview carve-out. The split exposes a structural problem: the US intelligence community lost a tool mid-deployment with no clear timeline for return.
Google Bakes Computer Use Into Gemini 3.5 Flash — Browser, Mobile, and Desktop in One Model
Computer use is now a native capability in Gemini 3.5 Flash, not a separate model. Google says it delivers its best agentic computer use performance yet, adds enterprise injection-stop safeguards, and opens it through both the Gemini API and Enterprise Agent Platform.
OpenAI Ships Jalapeño: First Custom Inference Chip, 9-Month Tape-Out, Gigawatt Deployment Planned
OpenAI and Broadcom unveiled Jalapeño, OpenAI's first purpose-built AI accelerator. Designed in nine months with help from OpenAI's own models, it targets inference workloads for ChatGPT, Codex, and the API — with early testing showing substantially better performance per watt than current SOTA.
42 State AGs Open a Sweeping Investigation Into OpenAI — Five Days After Its IPO Filing
A coalition of 42 state attorneys general has launched an investigation into OpenAI, with New York serving a subpoena covering advertising, user safety, minor and senior data, health data, and internal AI policies. The probe must be disclosed in OpenAI's S-1 and lands as the company prepares to list at an $852 billion valuation.
SpaceX Acquires Cursor for $60B in Stock — Anysphere Becomes the Coding Layer of the SpaceX AI Stack
SpaceX has agreed to buy Anysphere, parent of AI coding tool Cursor, for $60 billion in class A stock. The deal closes the loop on a strategy spanning xAI foundation models, Colossus training compute, and now the most-used AI-first IDE in professional software development.