Welcome back, Superhuman. Astra’s launch video finished the weekend with a staggering 128M views, causing Nvidia CEO Jensen Huang to declare AGI is here on Sunday. But it’s not all good news for OpenAI — a new report uncovered yet another instance of the lab’s agents hacking a website.

Today: AI accelerates its own progress, how to set an intelligence floor for your AI application, and get the latest prompts and trending social posts.

TODAY IN AI

View an Astra-rendered walkthrough of The Backrooms. Source: @duncantrussell.

1. OpenAI’s most powerful model becomes generally available, fueling a surge in experimentation: The weekend was flooded with use cases of Astra being put to work, including a 3D website that pulls apart every piece of the human body, a fly powered by a simulated insect brain, and a realistic 3D rendering of The Backrooms. But the lab’s powerful models are still having alignment issues: researcher Thomas Larson identified an instance of OpenAI agents hacking another website back in May.

2. AI is starting to accelerate its own progress: Meta’s autonomous research system, AIRA 3, placed 8th out of 4,000 teams in a competition where you teach a model to reason better. The tech giant said it’s a sign the system can improve AI models at a level similar to human experts. Meanwhile, OpenAI revealed it has an automated research intern that works under human supervision to further progress, with a fully automated AI researcher expected by March 2028.

3. Google’s top music model comes to Gemini: Lyria 3.5, Google’s most advanced music generation model, is now available in the Gemini app, AI Studio, and the API. The model — first released in late July — offers stronger melodies and more detailed musical arrangements. It’s capable of generating full songs up to 3 minutes long, with one song clip generating over 1M views on Friday. Try it here.

Your agent needs access to Linear, Notion, and GitHub. Each integration means another OAuth flow, refresh token, and provider scopes. Giving those credentials to an agent risks spreading them into context windows, logs, and notes.

Pipes handles authorization, token storage, and refresh across providers. Relay then attaches credentials only when needed and releases them only to allowlisted hosts. The agent can act without holding the token. 

  • Open the Finest Console and create a workspace

  • Create an API key

  • Go to "Your floor" > “A floor model" > Choose your preferred model

  • Copy the base URL and key into your app. Keep your existing model, prompts, and parameters unchanged. If another model can't do at least as well, your floor model will be served

  • Send your first request

Sample Prompt: “Summarize this customer feedback into three key points and identify the most common complaint: [paste feedback]”

  • Open the receipt to see what you requested, which model served it, how it was verified, and what you paid compared to the frontier-model price

FROM THE FRONTIER

Don’t rely on AI for anything potentially life-threatening — it still confidently lies

Made with Midjourney

AI has made incredible progress lately, but still has a key issue: confidently hallucinating. A group of California hikers just learned this the hard way.

The group relied on Gemini to game-plan an 8-hour trek, and the model recommended far less food and water than they needed. They ran short on supplies, spent the night on the mountain, and had to be rescued the next morning. It’s a reminder that despite AI’s strength, models still aren’t entirely trustworthy.

Why do models hallucinate?

There are three reasons why models give you false information:

  1. Their training data isn’t 100% accurate — it’s often filled with inaccuracies, contradictions, and opinions, which can then show up in their advice.

  2. They’re trained to be friendly and engaging, sometimes to a fault.

  3. They have trouble grasping ambiguity in human language, which can cause confusion.

A study uncovered another worrying trend: AI models may agree with you on false statements if you’re persistent enough. In the study, GPT-3.5 correctly rejected 96 of 100 false statements, but changed its mind on 18 of them after researchers pushed back. Granted, newer systems hold their ground better, but the point still stands.

IN THE KNOW

What’s trending on socials & headlines today

Meme of the day

Astra Prompts: AI has gotten so powerful it’s almost hard to think of how to use it. This list outlines nine Astra unique prompts that includes creating a competitor spy, bill negotiator, and agent opportunity auditor. It’s gotten over 7K bookmarks.

👽 Alien Mind: OpenAI’s chief scientist says we’ve created intellect we don’t truly understand in a new essay that’s generated over 2M views. Read why he’s concerned about humanity’s next few years.

📺 Jumbotron Footage: A young couple was featured on the jumbotron during the US Open. Naturally, they wanted the footage, so they asked their agent to track it down. It delivered within 24 hrs after jumping through some serious hoops (2.5M views).

🤯 Compressing Timeline: It took OpenAI 1,000 days to go from GPT-3 to GPT-4. Then another 877 days to go from GPT-4 to 5. Guess how quickly the lab went from GPT-5.6 to 6? It’s alarmingly short.

🎵 AI’s Symphony: Astra composed Chorale in G Minor after running on extra-high effort, and the baroque-style musical clip has generated over 2M views. Give it a listen (prompt included).

Most engineers are stuck babysitting AI one prompt at a time. A handful have figured out how to build agents that plan, execute, and self-correct on their own.

Sign up for The Code newsletter and get access to The Complete Playbook for Agentic Engineering. 5 modules, 29 lessons on Claude Code, subagents, and MCP.

Get the playbook

PRODUCTIVITY

  • 🚀 Ema*: AI Employees that actually solve HR tickets. Not chatbots that redirect you to FAQs. Ticket resolution time down from 5 days to seconds.

  • 🤖 iAsk: Ask questions and receive instant, accurate, cited answers without storing any data.

  • 📽 Videomaker: Bring your ideas to life with motion, sound, and cinematic clarity.

  • Agentkit: Get AI agents for website trained in minutes.

PROMPT

Set SMART Marketing Objectives

ChatGPT Prompt: You are an expert digital marketing strategist who understands the importance of SMART goals.
Define SMART marketing objectives and KPIs in a structured format. Use a table with two columns. The first column header is ‘Objective’ and the second column header is ‘KPIs’. There will be between 1 and 5 rows depending on how many objectives seem necessary. Each row should be a SMART objective, accompanied by the specific KPI/s that define success for that objective.

Before doing that, firstly ask me to describe, in my own words, the results I would like to achieve in my digital marketing, or in my business. If they unclear, ask me follow-up questions to clarify.

Source: Will Francis
EXTRAS

Want more?

Grow customers & revenue: Join companies like Amazon, HubSpot, and Salesforce. Showcase your product to our 4 million+ readers and followers on socials. Get in touch.

What did you think of today's email?

Your feedback helps me create better emails for you!

Login or Subscribe to participate

Until next time — Zain, Theodore, & the Superhuman AI team.