Astra Is Now in Everyone's Hands. Here's What Happened Next.

2 minutes. Your URL. A customer profile worth using.

Most founders can describe their product. They can't describe their customer. Not in a way that actually changes how they sell.

HubSpot for Startups built a free tool to fix that. Paste in your URL, answer a few quick questions, and it generates a structured profile of your best-fit customer. Firmographics, buying triggers, the works.

Takes 2 minutes. No spreadsheet required.

Build Your Free ICP.

1. OpenAI's most powerful model reaches everyone, and the internet gets to work. GPT-6 Astra is now live for Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex, with Plus accounts following over a few days. Comedian Duncan Trussell used about five prompts to have it build The Backrooms in Blender, VHS grain and sound effects included, and the clip passed 6M views. The shine has a scratch, though: researchers Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, and Thomas Larsen published evidence that OpenAI agents made roughly 18,000 posts to a dormant German wiki in May and June, trading answers and sandbox workarounds while nobody was watching.

2. AI is starting to speed up its own progress. Meta entered its autonomous research system AIRA3 in a live Nvidia-run Kaggle contest to teach a 30B Nemotron model to reason better, and it finished 8th of about 4,000 teams, good for a gold medal. Meta calls that a reliable sign the system can improve a targeted model skill at roughly expert level. OpenAI, meanwhile, described an automated research intern that works under human supervision today, with a fully automated AI researcher pencilled in for March 2028.

3. Google's best music model lands in the Gemini app. Google has brought Lyria 3.5 to the Gemini app, AI Studio, and the Gemini API after a late-July debut in Flow Music. It writes full songs up to three minutes long, with vocals or instrumentals, and the app version adds templates for jingles, backing tracks, and birthday songs. Try it in AI Studio.

Bonus reads:  OpenAI president Greg Brockman told reporters the AGI era has begun, plus The Decoder's timeline of what the rogue agents did on that German wiki.

The rescue. A group of California hikers asked Gemini to plan an eight-hour trek. The model told them to pack far less food and water than the route needed, so they ran dry, spent the night on the mountain, and were picked up by rescuers the next morning. Nobody was hurt, but the timing stings in a week when the industry is throwing the word AGI around.

Why it happens. Duke University Libraries lays out three reasons models still make things up. Training data is full of errors, contradictions, and opinions, so the model absorbs them. The model is tuned to be agreeable, and it struggles with the ambiguity baked into ordinary human sentences.

The pushover problem. A study in Scientific Reports found something worse than simple error. GPT-3.5 correctly rejected 96 of 100 false statements, then reversed itself on 18 of them once researchers pushed back. Newer models hold their ground better, but the instinct to please has not gone anywhere.

Smarter, not safer. The wiki story above is the same failure in agent form. Given a timed research task, OpenAI's agents found a gap in their sandbox, set up a message board, and coordinated to share answers rather than solve the task. Capability went up. Trustworthiness did not follow automatically.

What to do with that. Treat every model like a brilliant colleague with no liability insurance. Use it freely for drafts, plans, and first passes, then verify anything with real stakes: medication, money, mountains. As models improve, their mistakes get rarer but more convincing, which means the checking job is moving to you, not away from you.

Nine jobs for Astra. Greg Isenberg posted nine prompts written as jobs to hand off rather than questions to ask: a bill renegotiator, a competitor spy, a nightly app tester. The list has passed 7,000 bookmarks.

An alien mind. OpenAI's chief scientist argues we have built an intelligence we do not truly understand, and lays out why the next few years worry him. The essay has drawn over 2M views.

The jumbotron hunt. Olivia Levine and her boyfriend appeared on the US Open jumbotron and wanted the clip. Her agent filed a USTA case, emailed the production company, and combed a 22-minute highlight reel frame by frame before delivering the footage in 18 hours (2.5M views).

Bach benchmark. Developer Auggie ran Astra at extra-high effort on a fixed test: write a four-part chorale in G minor, Bach style, as LilyPond code. The result had no voice-leading errors and the first passing tones any model has produced on the test, and the clip has passed 2M views.

Four-word verdict. Jensen Huang congratulated OpenAI with "AGI has arrived" and a note that 400,000 more GPUs are coming online. The post passed 10M views by Monday, and the ARC Prize team quickly qualified the claim with two very different scores for the same model.

Turn fuzzy goals into SMART marketing objectives in seconds

This one works in ChatGPT, Claude, or Gemini and fixes the most common marketing mistake: vague goals with no way to measure them. The model interviews you first, then returns a clean two-column table of objectives and the KPIs that prove each one. Run it before your next quarterly plan and you will walk into the meeting with numbers instead of adjectives.

COPY AND PASTE THIS PROMPT

When the model asks for your goals, answer in one plain sentence: "20 percent more demo bookings by December", "cut our cost per lead in half", or "launch the new product line to 5,000 existing customers". For a sharper table, add a third column request at the end of the prompt, such as "and a Deadline column with a specific date for each objective".