OpenAI’s cash burn accelerates

Welcome back, Superhuman. Big bets are being placed on big promises. OpenAI is racing to show AI can deliver movies faster and cheaper than traditional studios with Critterz, its animated film written by some of the same team that wrote Paddington in Peru. Meanwhile, Tesla’s board is offering Elon Musk a $1 trillion pay package based on how the company’s stock performs.

Today: OpenAI explains why models hallucinate, Anthropic settles for $1.5B, learn how to use Wispr Flow Snippets to save time at work, and get the latest prompts, headlines, and trending social posts.

TODAY IN AI

1. Alibaba’s AI arm drops its “biggest model” yet: The Chinese e-commerce giant unveiled Qwen3-Max-Preview — its largest language model to date with over 1 trillion parameters. Betting on scale, Qwen says the model outperforms Claude Opus 4, Kimi K2, and DeepSeek-V3.1 on reasoning and coding tasks. The model supports a 262K token context window and is available on Qwen Chat, Alibaba Cloud, and OpenRouter. Pricing starts at $0.86 per million tokens for shorter prompts.

2. Anthropic to pay $1.5B in record author lawsuit settlement: The Claude maker has agreed to pay authors about $3,000 per work for allegedly using their books to train its AI models. The company is also required to destroy the original files, but the settlement only covers past use and leaves future claims open. It may be the largest copyright recovery in US history. A court hearing on September 8th will determine final approval.

3. OpenAI’s costs expected to balloon past $100B+: The company is reportedly projecting $115B in cash burn through 2029 — an $80B jump driven by rising AI infrastructure costs. Losses are set to exceed $8B this year, rise to $17B in 2026, and hit $45B by 2028 as cloud costs soar. To offset the crunch, OpenAI plans on developing its own AI chips with Broadcom starting next year and building out new data centers.

SPONSORED BY PACASO

Learn From This Investor’s $100m Mistake

Grammy-winning artist will.i.am passed on investing $200K in Airbnb in 2010. It’d be worth $100M+ today. Now, the investors behind Venmo and Uber are backing another real estate disruptor. This time, you can invest.

Meet Pacaso. Created by a former Zillow exec, Pacaso brings co-ownership to the $1.3T vacation home industry.

They’ve made more than $110M in gross profits to date and they’ve even reserved their Nasdaq ticker, PCSO.

Invest in Pacaso before the opportunity ends September 18.*

FROM THE FRONTIER

New OpenAI research cracks the case on why AI models hallucinate (and what to do about it)

Large language models have gotten better at avoiding mistakes — but they still make things up. While it isn’t a big deal in some low-stakes scenarios, you probably don’t want an AI model to hallucinate in sensitive use cases like law and medicine. A new research paper from OpenAI digs into why even advanced models like GPT-5 still generate "plausible but false statements," and found some interesting answers.

Surprisingly, advanced models are hallucinating more. On PersonQA, which questions the model on public figures and facts, OpenAI’s o3 model hallucinated 33% of the time, double its predecessor o1. The smaller o4-mini fared even worse at 48%. That’s despite both models showing stronger math skills. It’s a paradox: advanced models can be smarter and more wrong at the same time.

The key finding: Current training and testing methods penalize models for admitting uncertainty. These incentives create pressure for models to respond, even when they don’t know the answer.

OpenAI says benchmarks must change. Instead of rewarding only right answers, evaluations should penalize confident mistakes and give credit for saying “I don’t know,” like exams that dock marks for wrong guesses but not for leaving a question blank. Hallucinations won’t vanish, but with better incentives, models could learn to fail more honestly.

THE AI ACADEMY

How to save hours every week with Wispr Flow Snippets

• Get Wispr Flow and open the app (it’s free).

• Go to Settings → Add New Snippet.

• Give it a name (e.g., “My Calendly link”) and paste the text or link you want it to expand into.

• Trigger it with your voice. Next time you’re dictating, just say the snippet name.

Sample Prompt: Say “intro email”, and Wispr Flow types: Hey, would love to find some time to chat later…

• Works in email, Slack, docs—literally anywhere you type.

• Set it up once. Save time forever.

Try Wispr Flow and set up voice shortcuts to help you get through your work faster.

PRESENTED BY DATADOG

Overcome the challenges of deploying LLMs securely and at scale

LLM workflows are complex, opaque, and hard to secure. This guide from Datadog offers practical strategies to monitor, troubleshoot, and protect your LLM applications in production.

In this eBook, you’ll learn how to:

• Trace multi-step LLM chains and debug errors faster

• Detect prompt injection and secure against data leakage

• Evaluate response quality with out-of-the-box and custom checks

• Control token costs and optimize model performance

Get the best practices guide today

IN THE KNOW

What’s trending on socials and headlines today

📞 Chatbot Hack: A frustrated user shares how they tricked an airline’s AI chatbot just to reach customer service. The post has over 3 million views.

📺 Ad Unlock: Higgsfield AI has unveiled Ads 2.0 — a new tool it says can replace entire production, marketing, and creative teams with a Mini App. See it in action here.

🍌 Web Spinner: A user built a Nano Banana-powered browser that spins up entire websites from just a URL. The result feels like you’re surfing a brand-new internet.

⚠️ Prompt Pirates: One of the first ways attackers exploit AI models is by hijacking prompts. Here are 5 examples, explained visually.

ASML will reportedly invest €1.3B in French startup Mistral, becoming its top shareholder in a move to boost European tech sovereignty. The deal values Mistral at €10B, making it Europe’s most valuable AI company.

Apple faces a class action lawsuit alleging it used pirated books to train its “OpenELM” AI models.

Anthropic is expanding its block on “authoritarian” regimes, now banning the use of its products by Chinese-owned companies.

PRODUCTIVITY

5 New AI Tools to Boost Your Productivity

• 🖥️ HeyGen: Create AI avatars and video clones of yourself.

• ⚡️ Rork: Turn your ideas into phone apps with AI.

• 📽️ Sidekick: Build Zapier-style automations using a chat interface.

• 🎨 Uxia: Validate your user flows UX and UI in seconds with AI.

• 🖼️ CapCut: Create, edit, or remix content with AI in a simple editor.

^{* indicates a promoted tool, if any}

PROMPT STATION

Fitness Plan Generator

Prompt: You are a certified personal trainer and sports scientist. Create a safe, progressive, beginner-friendly fitness plan tailored to my data.

Personal Data (fill in before generating):

Age: [insert age]
Gender: [insert gender]
Height: [insert height]
Weight: [insert weight]
Current fitness level (sedentary, light active, moderate, etc.): [insert]
Primary goal (fat loss, muscle gain, general fitness, energy, etc.): [insert]
Any injuries, health conditions, or limitations: [insert]
Available equipment (bodyweight only, dumbbells, gym, etc.): [insert]
Preferred workout duration per session: [insert minutes]
Days per week available for training: [insert]
Dietary preferences/restrictions (optional): [insert]

Instructions for output:
Provide a weekly workout schedule (day by day).
Detail exercises, sets, reps, and rest periods, with simple explanations.
Suggest warm-up and cool-down routines.
Recommend progression strategy (how to increase difficulty over weeks).
Include basic nutrition guidance to support my goal.
Provide safety notes (form, recovery, overtraining avoidance).
Suggest tracking methods (metrics or progress markers).

Timeless Icons

Midjourney Prompt: photorealistic image 64k of a 1939 Bugatti Type 57C Cabriolet, top view, background white. --ar 9:16 --v 7

You can use similar prompts in ChatGPT without parameters to get similar results.

Whenever you’re ready to take the next step

• Check out our Top 125 AI Tools

• Get better at prompting with our Top 1,000 Prompts

• Learn how to use AI at work and get certified with our free course

Grow customers & revenue: Join companies like Amazon, Hubspot, and Salesforce. Showcase your product to our 1M+ readers and 2M+ followers on socials. Get in touch.

Until next time — Zain & the Superhuman AI team