AI News 2026-09-25

$ cat top-story.md

Google Preps to Launch Its First AI Chips Into Space Under Project Suncatcher

On October 1, Google will send a prototype satellite carrying four Trillium TPUs into orbit aboard SpaceX's Transporter-18 rideshare — the first in-orbit test of Project Suncatcher, a moonshot exploring whether satellite clusters powered by near-constant solar energy could one day host scalable ML compute. Built with Planet, the mission will measure how the chips hold up under radiation, vibration and thermal extremes, ahead of a 2027 test linking two satellites via laser interconnects.

Why it matters: If off-planet TPU clusters prove viable, it reframes AI infrastructure scaling once terrestrial power and land become the real bottleneck — and Google's hardware/software choices here tend to trickle down into the commodity inference stack you actually run.

// source: Google · Hacker News · Reuters

## 📦 Models & Releases

▸ OpenAI Releases MentalHealthBench, Built With 80+ Licensed Psychologists

An open benchmark of 1,215 synthetic mental-health conversations — from everyday check-ins to acute crises — paired with 5,262 expert-authored rubric criteria, co-created with psychologists and psychiatrists across 22 countries. Why it matters: A free, expert-grounded way to test how any model (including a self-hosted one) handles sensitive conversations, not just another leaderboard number. // OpenAI

▸ OpenAI Shuts Down the Sora API

The Videos API and all Sora 2 model aliases/snapshots were pulled on September 24, following the Sora web/app shutdown back in April; existing content can still be exported from sora.chatgpt.com/sunset. Why it matters: If any script or workflow of yours calls Sora, it broke this week — check your dependencies now. // OpenAI Help Center

## 🛠️ Tools & Repos

▸ ai-engineering-from-scratch Tops GitHub Trending With a Free 523-Lesson Curriculum

A from-scratch AI engineering course spanning 20 phases, from math and classical ML up through agents, multimodal systems and autonomous AI, with runnable code in Python, TypeScript, Rust and Julia. Why it matters: A structured way to actually understand what's under the tools you use daily instead of just calling their APIs. // GitHub Trending (+347 today)

## 🤖 Agents & MCP

▸ Hindsight Rockets Up GitHub Trending as "Agent Memory That Learns"

Combines semantic, keyword, graph and temporal retrieval into retain/recall/reflect operations, consolidates facts into deduplicated observations, and claims state-of-the-art results on LongMemEval; ships with MCP support, 60+ framework integrations and a Postgres backend. Why it matters: A self-hostable alternative to plain RAG-as-memory for agents that need to remember and reason across sessions, not just retrieve chat history. // GitHub Trending (+1,668 today)

▸ Strands Agents Ships a Production Harness SDK With No Hosted Control Plane Required

Apache-2.0 Python/TypeScript SDK with turn limits, token budgets, structured output, multi-agent patterns, MCP support and model portability across Anthropic, OpenAI, Bedrock and Gemini — usable as a ready-made harness or as building blocks for your own agent loop. Why it matters: Worth comparing against Claude Code's own subagent primitives if you're building a custom agentic tool that has to run more than one model provider. // GitHub Trending (+455 today)

## 🏠 Self-Hosting & RAG

▸ vLLM Ships vllm-metal for Concurrent Serving on Apple Silicon

A new community plugin (contributed by Docker, MLX-backed) brings vLLM's paged, continuously-batched serving to Apple Silicon: GGUF/hybrid-model support, LoRA, speculative decoding, structured outputs, pipeline parallelism across multiple Macs and a faster M5 prefill path. Why it matters: The first real vLLM-grade serving story for Apple Silicon, instead of falling back to llama.cpp or raw MLX for anything on a Mac. // vLLM Blog

▸ NVIDIA Model Optimizer Unifies Quantization, Pruning and Distillation for vLLM

A single library to quantize (FP8/NVFP4), prune, distill or speculative-decode your own Hugging Face/PyTorch/ONNX models, exporting checkpoints straight into vLLM, TensorRT-LLM or SGLang — or just pull pre-optimized checkpoints from Hugging Face. Why it matters: A plausible way to shrink or speed up a self-hosted Gemma on vLLM without hand-rolling your own quantization pipeline. // GitHub Trending

## 📰 Worth Reading

▸ Google, OpenAI and Anthropic Plan a Joint "Frontier AI Standards" Body

The three labs are reportedly forming the Standards Authority for Frontier AI (SAFA), a FINRA-like self-regulator for testing and auditing frontier models, and have approached former White House AI advisor Sriram Krishnan to lead it, targeting a launch by early 2027. Why it matters: Comes right after a string of autonomous-agent security incidents — a sign the labs would rather self-regulate than wait for regulators to move first. // BankInfoSecurity

▸ Trail of Bits on What Six Months of AI-Assisted Security Audits Actually Taught Them

A senior auditor walks through using agents throughout a real zkVM audit — building custom tooling and formal models before code review even starts — and where agents genuinely caught real vulnerabilities versus where they didn't. Why it matters: A grounded, practitioner-level account of using agents for security work, useful if you're weighing Claude Code or similar for your own review pipeline. // Trail of Bits

AI News — September 25, 2026