Sunday, September 6, 2026

Coverage: 34 videos reviewed (2 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

New today

Continuing stories

• Nvidia reportedly agrees to buy Hugging Face for about $12.9 billion, pledging it stays open - Nvidia is reported to be buying Hugging Face for $12.9 billion, according to panelist statements in a David Shapiro video published Sept. [0 first-party, 0 hands-on, 2 relaying] Watch: Sam Witteveen: NVIDIA Doubles Down on Local AI With PAIR
• OpenAI releases GPT-6 Astra to paid ChatGPT plans, API and AWS, presenter says - Nate B Jones said OpenAI released GPT-6 Astra on a Thursday across all paid ChatGPT plans, the API and AWS, with an emphasis on long-running computer use; his full review was still to come and he gave no benchmarks or prices. [0 first-party, 0 hands-on, 2 relaying] Watch: Nate B Jones: GPT-6 Astra Doesn't Need Your Instructions Anymore. (high hype)
• OpenAI says GPT-6 Astra is its first model rated 'critical' for cybersecurity capability - Julian Goldie, relaying OpenAI materials on Sept. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: OpenAI Astra Just Crossed a Dangerous AI Threshold (high hype)
• Nate Herk's 15-task test: Astra won 10, Fable 5.1 won 5; Astra cost less but ran longer - Nate Herk scored GPT-6 Astra ahead of Claude Fable 5.1 on 10 of 15 use cases in single runs he graded himself. [0 first-party, 1 hands-on, 0 relaying] Watch: Nate Herk: I Tested GPT-6 Astra vs Fable 5.1 on 15 Real Use Cases

Also notable

• OpenAI reportedly paused parts of Astra training for two weeks after Hugging Face incident - Julian Goldie's video said OpenAI paused parts of GPT-6 Astra training for two weeks to tighten security and monitoring, and restarted its largest RL run on Aug. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: OpenAI Astra Just Crossed a Dangerous AI Threshold (high hype)
• Panelist says Astra scored 99.9% on ARC-AGI-3 versus roughly 9-13% for GPT-5.6 - On David Shapiro's channel on Sept. [0 first-party, 0 hands-on, 1 relaying] Watch: David Shapiro: Unpacking what Astra means
• Nate B Jones says Astra system card notes same-user agents communicating in Codex - Nate B Jones paraphrased OpenAI's system card as reporting that researchers noticed agents tied to one user communicating within the same Codex setup, and that OpenAI is building tests for agents discovering other agents' messages. [0 first-party, 0 hands-on, 1 relaying] Watch: Nate B Jones: GPT-6 Astra Doesn't Need Your Instructions Anymore. (high hype)
• Alex Ziskind measures DeepSeek V4 Flash on four RTX Pro 6000 GPUs: 33 to 364 tok/s across agents - Alex Ziskind reported DeepSeek V4 Flash (FP4 experts, FP8 attention) on four RTX Pro 6000 GPUs at 33 tokens/s with one agent, 62 at concurrency 2, 116 at 4 and 364 at 16, after which throughput dropped; the serving stack and batch settings were not stated. [0 first-party, 1 hands-on, 0 relaying] Watch: Alex Ziskind: All That VRAM Needs a Bigger Brain
• Spark-X2.5 4B open model tested via Hermes: bug fix succeeds but reasoning runs 16-40 minutes - Fahd Mirza reported the Spark-X2.5 4B model card claims a hybrid of three sliding-window layers per full-attention layer, 1 million token native context, 200+ languages, about 20 trillion training tokens and an Apache 2 license; he capped context at 65k and did not test 1M. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: Spark X2.5 4B: What a 4B Model Can and Can't Do Locally
• Nvidia SkillSpector scanner flags a malicious agent skill's exfiltration script, misses a text injection in static mode - Fahd Mirza described Nvidia's open-source SkillSpector as scoring agent skills 0-100 for injection, exfiltration and supply-chain risk using static checks plus an optional LLM pass. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: How to Scan AI Agent Skills for Hidden Malware: NVIDIA SkillSpector
• Qwen 3.8 Flash Next reportedly passes Claude Opus Max's 55 on Artificial Analysis index - Sam Witteveen said Claude Opus Max (April 2026) scored 55 on the Artificial Analysis Intelligence Index and that the open Qwen 3.8 Flash Next, released in August, has surpassed it, without giving the Qwen score. [0 first-party, 0 hands-on, 1 relaying] Watch: Sam Witteveen: NVIDIA Doubles Down on Local AI With PAIR
• Google adds Lyria 3.5 music model to Gemini app and API, presenter says - Julian Goldie said Google's Lyria 3.5 is now in the Gemini app and Gemini API, with Lyria 3 Clip for 30-second clips and Lyria 3 Pro for full songs, and in AI Studio, Flow and Vids. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: Google AI Studio + Lyria 3.5 Is CRAZY! (high hype)
• Gemini 3.8 Flash described as Google's smartest Flash model with tool-check loop - Julian Goldie said Google's Gemini 3.8 Flash can stop, use a tool, check its work and continue, was shown building apps from one prompt, and reads about a million tokens. [0 first-party, 0 hands-on, 1 relaying] Watch: Julian Goldie: Google Gemini NEW Updates are WILD! (high hype)

Models & learning

• Julian Goldie's Astra demos: single-prompt game, scheduled Codex task, credit use, one failure - In sponsored videos, Julian Goldie said GPT-6 Astra on medium effort in ChatGPT built a single-file space shooter called Starfall from one prompt, and that Codex created a recurring cloud task for daily SEO content with the next run shown 20 hours away. [0 first-party, 1 hands-on, 0 relaying] Watch: Julian Goldie: LIVE: Watch GPT 6 Astra Build The Ultimate Agent OS
• How I AI host has Astra operate a Flora workflow to make a thumbnail - A How I AI host gave GPT-6 Astra photos and asked for assets; it inspected the open Flora workflow, added a node, picked an image model and aspect ratio, wrote a prompt from existing ones and generated an image he judged a great thumbnail. [0 first-party, 1 hands-on, 0 relaying] Watch: How I AI: GPT-6 Astra made YouTube thumbnails on the first try
• OmegaClaw open-source agent framework tested: memory persists across restart, symbolic step needed re-prompt - Fahd Mirza described OmegaClaw, a SingularityNET agent framework built on Hyperon with a roughly 200-line MeTTa core, persistent memory and a proof trail, from project claims before testing. [0 first-party, 1 hands-on, 0 relaying] Watch: Fahd Mirza: OmegaClaw: An AI Agent Built on Symbolic Logic, Not Just an LLM
• Nvidia releases PAIR v0.1, an Apache-2.0 router for local agent requests across machines - Sam Witteveen demonstrated Nvidia's PAIR (Personal AI Router) v0.1, which proxies Ollama and LM Studio ports and spreads requests, such as sub-agent calls, across machines on a home network. [0 first-party, 0 hands-on, 1 relaying] Watch: Sam Witteveen: NVIDIA Doubles Down on Local AI With PAIR
• Witteveen reports Qwen 3.8 27B at about 300 tokens/s, 380 tokens/s batched - Sam Witteveen said he now has the Qwen 3.8 27B model at about 300 tokens per second, averaging about 380 tokens/s when batching. [0 first-party, 0 hands-on, 1 relaying] Watch: Sam Witteveen: NVIDIA Doubles Down on Local AI With PAIR
• Raschka measures KV cache raising Qwen3 0.6B from ~4 to ~27 tokens/s on Mac mini M4 - Sebastian Raschka reported that on a Mac mini M4 GPU (MPS) with a short prompt, Qwen3 0.6B generated about 4 tokens/s uncached and about 27-28 with a KV cache. [0 first-party, 1 hands-on, 0 relaying] Watch: Sebastian Raschka: Build A Reasoning Model Scratch 2: Loading a Base Model, Text Generati