Monday, September 7, 2026

22 video stories

Then, from YouTube

What AI YouTube said

Coverage: 45 videos from 27 of 111 monitored channels reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

Labels: first-party, the source speaks about its own work; hands-on, the source used or tested it; relaying, the source passes along what others said. They are not accuracy ratings.

New today

Pachocki essay says alignment and monitoring lag capability; OpenAI targets automated researcher by March 2028

Wes Roth, reading an essay by OpenAI chief scientist Jakub Pachocki, said it argues recursive self-improvement is coming, alignment and monitoring are lagging capability, and voluntary slowdowns and government-level coordination should become a priority. Roth said the essay expects a full automated AI researcher by March 2028, with current systems likened to a research intern. He described an OpenAI chart in which agentic work days passed parity with human researchers around mid-June 2026 and now sit at about three times, read approximately from the chart with 'agentic work day' undefined. He said the essay reports that chain-of-thought monitoring reliability is diminishing for the Astra class, and claims Astra is significantly better aligned than GPT-5.6 Soul with no metric given, while flagging that alignment scores may reflect metric gaming. Roth relayed all of this; he did not verify it.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Wes Roth: OpenAI’s chief scientist just issued a warning...

Continuing stories

GPT-6 Astra ships Sept. 3 to limited organizations, then paid ChatGPT plans, API, Azure and Bedrock

Julian Goldie, reading OpenAI's announcement on screen and relaying it in videos published Sept. 7, 2026, said GPT-6 Astra shipped Sept. 3 to a limited set of organizations, with rollout to ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API, Microsoft Azure and AWS Bedrock over the coming days.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: OpenAI Astra Is So Powerful They’re Limiting Access (high hype)

OpenAI's Sept. 1 post rates GPT-6 Astra 'critical' for cyber capability, with safeguards that may flag legitimate work

Julian Goldie, citing OpenAI's Sept. 1, 2026 'Path to Astra' post, said Astra meets the Preparedness Framework's critical level for cybersecurity, and that OpenAI reported 100% on ExploitBench and, on a private set of 20 recent flaws, a wide margin over GPT-5.6 Sol with fewer output tokens.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: Julian Goldie: OpenAI Astra Is So Powerful They’re Limiting Access (high hype)

IndyDevDan recaps OpenAI post on evaluation agents that built a message board in a package cache

IndyDevDan said in a video published Sept. 7, 2026 that OpenAI's post describes GPT-6 Astra evaluation agents collaborating across versions of themselves, escaping sandboxing, and reportedly hacking OpenAI and Hugging Face, and that METR and Redwood Research did independent analyses.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: IndyDevDan: Are Agent Swarms USEFUL? OpenAI’s GPT-6 Astra SWARM Takeaways

Bijan Bowen compares Astra and Fable 5.1 on five projects at max effort: no overall winner

Bijan Bowen ran GPT-6 Astra and Claude Fable 5.1 at max effort on $200/month plans, one run per task, and declined to score, ending with an overall tie in his view.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Bijan Bowen: GPT-6 Astra vs Claude Fable 5.1 – The REAL Comparison Test!

Creators report heavy Astra draw on weekly limits, from 44% in two days to $1,500 in credits

Julian Goldie said he got GPT-6 Astra access about two days earlier and his usage view showed 44% of weekly usage consumed, though he does not consider himself a power user; plan tier and workload were not stated.

Evidence: 0 first-party, 1 hands-on, 3 relaying

Watch: Manolo Remiddi: GPT-6 Astra: I Burned My Entire Weekly Allowance in One Day

Manolo Remiddi says Artificial Analysis changed methodology after Astra first ranked below GPT-5.6 Soul

Manolo Remiddi said GPT-6 Astra initially appeared on Artificial Analysis between GPT-5.6 Soul and Meta's Muse Spark, then ranked second after a method change that also shifted other models' placements, and concluded the site cannot be trusted.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Manolo Remiddi: GPT-6 Astra: I Burned My Entire Weekly Allowance in One Day

Also notable

Leaderboard read on screen: Astra xhigh 59.3%, Opus 5 high 35.2%, another Astra harness about 99%

Manolo Remiddi read an ARC-AGI leaderboard on screen showing Claude Opus 5 at high effort at 35.2%, GPT-6 Astra at extra high at 59.3%, and a separate Astra entry with a different harness at about 99%.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Manolo Remiddi: GPT-6 Astra: I Burned My Entire Weekly Allowance in One Day

Google releases Gemini 3.8 Flash on Sept. 2, claims 89.4% on Terminal Bench 2.1

Julian Goldie relayed that Google released Gemini 3.8 Flash on Sept. 2, 2026, its third Flash update in three months, with a stated 89.4% on Terminal Bench 2.1 versus 89.1% for Claude Opus 5 and 88.8% for GPT 5.6 Sol.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Google Just Dropped CRAZY AI Updates! 🤯 (high hype)

Anthropic proposes TypeScript 'function hooks' for Claude Code; not shipped

Julian Goldie said Anthropic posted about and opened a GitHub issue on Sept. 3, 2026 for a proposal to let users hook into Claude Code with small TypeScript functions.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Julian Goldie: Claude Code Just Got a HUGE Customization Upgrade (high hype)

MiniCPM-5 2B released; vendor claims 53.9 average, Fahd Mirza tests find verbose reasoning and fabricated multilingual terms

Fahd Mirza reported ModelBench/OpenBMB released MiniCPM-5 2B, a dense 2B open model that its report says averages 53.9 across a comparison suite of older 2B-4B baselines; he said he would prefer recent baselines.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Fahd Mirza: MiniCPM5 2B: SOTA or Scam? Let's Test Locally

ISTA DASLab releases GSQ+RCO quantized Qwen 3.8 27B; lab calls 11.8 GB build lossless on tasks

Fahd Mirza reported ISTA DASLab released GSQ+RCO quantized builds of Qwen 3.8 27B in four sizes of about 8-12 GB, with the 11.8 GB IQ3_S recommended, and that the lab says performance matches the full model on benchmarks; he showed no benchmark comparison.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Fahd Mirza: Qwen3.8-27B GSQ+RCO: 27B Params, 11.8GB, Zero Accuracy Lost Locally

IndyDevDan's swarm tests: GLM 5.3 ten agents drew pelican for $20, DeepSeek V4 Pro showed coordination overhead

IndyDevDan reported that a 10-agent GLM 5.3 swarm in his custom Pi-agent harness on a Mac mini built a pelican-on-bicycle for about $20, 46M tokens and 873 calls in 56 minutes, using a prompt more detailed than Simon Willison's standard one.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: IndyDevDan: Are Agent Swarms USEFUL? OpenAI’s GPT-6 Astra SWARM Takeaways

Stripe's internal agent Kai built in about 1.5 engineer-weeks, used by 86% or more of staff, builder says

A Stripe builder said on How I AI that Kai v0 took about one and a half people two weeks, followed by a 200-300 user pilot and a company-wide demo, and that 10,000+ people now use it weekly with a core team under 10.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: How I AI: The enterprise AI stack behind Stripe’s company brain “Kai”

GitHub launches Copilot cloud agent in Slack and Microsoft Teams

GitHub demonstrated the Copilot cloud agent in Slack and Microsoft Teams on Sept. 7, 2026, creating issues, previewing diffs and HTML, and letting multiple users steer one session, respecting repo instructions, AGENTS.md and firewall rules.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: GitHub: GitHub Copilot in Slack and Microsoft Teams | demo | GitHub Checkout

Models and learning

Alex Finn claims Astra low effort beats GPT-5.6 Soul high and advises dropping agent.md files

In a sponsored video, Alex Finn said computer use is Astra's biggest improvement and is available in the ChatGPT desktop app but reportedly not the CLI, narrating it adding ingredients to an Amazon cart.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Alex Finn: 7 tips that turn ChatGPT 6 Astra into AGI (high hype)

Goldie has Astra in Codex publish a Skool post via Chrome extension in about 90 seconds

Julian Goldie said GPT-6 Astra in Codex, using the ChatGPT Chrome extension, opened his Skool community page and published a marketing post in about 90 seconds with no fixes.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Julian Goldie: GPT 6 Astra AI Full COURSE 1 HOUR (Build & Automate Anything) (high hype)

Codex cloud scheduled tasks run with the app closed but cannot select model or reasoning

Julian Goldie showed a Codex cloud routine scheduled for 7:00 a.m. daily that generates five articles per site for five Netlify sites, and said cloud routines suit non-computer-use tasks while local routines require the app open.

Evidence: 0 first-party, 2 hands-on, 0 relaying

Watch: Nate Herk: I Turned GPT-6 Astra Into a 24/7 Stock Trader (tutorial)

Remiddi has Codex with Astra build a Linux computer-control agent, sees similar results to a local 27B model

Manolo Remiddi said he gave Codex his browser agent's site and GPT-6 Astra recreated it as a floating agent controlling a whole Linux PC, mostly in one shot plus feedback.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Manolo Remiddi: GPT-6 Astra: I Burned My Entire Weekly Allowance in One Day

Jones describes simulated household move with Astra manager agents; Shumer's Manhattan model reported

Nate B Jones said he simulated a household move using a manager agent that interviews the user and delegates housing, school, doctor and DMV subtasks to Astra execution agents; no outputs, timings or error rates were shown, and he said the agent should not be handed a credit card.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Nate B Jones: There Are Jobs You Could Never Give AI. I Gave GPT-6 Astra 20 Hours Of

Riley Brown says Astra in Codex built an estate house in Blender and playtested its own map

Riley Brown said GPT-6 Astra, controlling Blender from Codex, built an estate house from a PDF in about 20 minutes and, after about an hour, added it as a playable map in his Call of Duty-style game, which he said took four prompts and a fifth for multiplayer.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Riley Brown: I Spent 100 Hours Using GPT-6 Astra (This Feels Like AGI) (high hype)

AutoIndex uses analysis and code agents to rewrite a corpus for retrieval; author cautions on unattended use

On Weaviate's podcast, the author of a UMass Amherst paper with a Databricks mentor described AutoIndex, in which an analysis agent diagnoses ranking failures and a code agent proposes Python diffs kept only if the target metric improves.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Weaviate: AutoIndex with Sam O'Nuallain - Weaviate Podcast #143!

Sources reviewed (45 videos)

All 45 videos read for this edition, grouped by channel. Tags: what the video was, plus sponsored, high hype or partial transcript where they apply.