- AI Toast
- Posts
- Reason why ChatGPT hallucinates
In partnership with
Welcome to AI Toast!
Today’s Menu:
AI Hallucinations, explained
Anthropic’s $1.5 Billion Settlement
Apple plans AI Search Engine
Use NotebookLM for studying
Superhero Tools for building
Quick AI Toasts: Altman says social media now feels “fake”
Total read time: 5 minutes, enough for you to grab a Starbucks.
OpenAI cracks AI's hallucination code
Despite their growing abilities, language models like ChatGPT sometimes confidently generate false answers, these are called “hallucinations.” OpenAI’s latest research explores why this happens and suggests ways to fix it.
Hallucinations are plausible but incorrect statements models make, often choosing to “guess” rather than admit uncertainty.
Standard training rewards accuracy but penalizes abstaining models get more points for lucky guesses than for saying “I don’t know.”
This creates a system where guessing is incentivized, even when the right approach would be humility and expressing uncertainty.
Pretraining teaches models only to predict the next word, not to distinguish truth from fiction, leading to errors with low-frequency facts.
Accuracy-driven evaluations overlook that some questions are inherently unanswerable for any model, keeping hallucination rates alive.
Penalizing confident errors more (and rewarding honest uncertainty) in training and tests can reduce hallucinations.
Smaller models might better know their limits and say “I don’t know,” while larger models can get overconfident and guess wrong.
Hallucinations stick around because we reward guessing. OpenAI’s fixes aim to teach models to be honest about what they don’t know. Read full report here.
Personal Take: As AI tools get better, teaching them humility may be the key to making them genuinely reliable. The future of language models should be about but knowing when to pause and say, “I’m not sure.”
Marketing ideas for marketers who hate boring
The best marketing ideas come from marketers who live it. That’s what The Marketing Millennials delivers: real insights, fresh takes, and no fluff. Written by Daniel Murray, a marketer who knows what works, this newsletter cuts through the noise so you can stop guessing and start winning. Subscribe and level up your marketing game.
Image Source: Capacity Media
Anthropic, the AI startup, has agreed to pay authors a staggering $1.5 billion to settle a lawsuit over pirated copyrighted books. This deal marks a turning point for AI, and content ownership.
Anthropic admitted to downloading unauthorized copies of books to train its AI models.
The settlement means a payout of at least $3,000 to each author whose work was used.
Authors called this agreement a “critical victory” after risking a courtroom battle.
The case could have cost Anthropic over $1 trillion if it had gone to trial.
This move brings financial and ethical stakes of AI training sharply into focus.
The industry is now on alert, every major AI company faces mounting legal and moral scrutiny regarding training data.
The verdict sets a precedent: future AI innovation must factor in creator rights and fair compensation.
This is not just a legal headline, it’s a wake-up call for the entire tech world.
Personal take: Going forward, the AI boom will be built on trust & compensation. This settlement is just the first volley; the future of AI will be defined by how well innovation and artistic ownership find common ground.
Image Source: Apple
Apple is gearing up for a major overhaul of Siri, transforming it into a powerful AI-driven search assistant. This update is set to launch in Spring 2026 with iOS 26.4.
Siri will deliver clear responses instead of sending users to websites.
Siri will deliver clear responses instead of sending users to websites.
The new search experience will include an interface that makes use of text, photos, video and local points of interest.
Search results will be condensed into concise explanations.
Apple is partnering with Google to boost Siri’s intelligence, after weighing Anthropic’s Claude. Apple will process searches on-device where possible, using its own foundation models to protect user data.
While launching in Siri first, the feature may later roll out to Safari and Spotlight.
Personal take: This upgrade aims to make Siri a more conversational and capable tool, rivaling AI models like ChatGPT and Perplexity. With its focus on privacy and potential for personalized answers, Siri could become a more attractive option for users seeking information.
How 433 Investors Unlocked 400X Return Potential
Institutional investors back startups to unlock outsized returns. Regular investors have to wait. But not anymore. Thanks to regulatory updates, some companies are doing things differently.
Take Revolut. In 2016, 433 regular people invested an average of $2,730. Today? They got a 400X buyout offer from the company, as Revolut’s valuation increased 89,900% in the same timeframe.
Founded by a former Zillow exec, Pacaso’s co-ownership tech reshapes the $1.3T vacation home market. They’ve earned $110M+ in gross profit to date, including 41% YoY growth in 2024 alone. They even reserved the Nasdaq ticker PCSO.
The same institutional investors behind Uber, Venmo, and eBay backed Pacaso. And you can join them. But not for long. Pacaso’s investment opportunity ends September 18.
Paid advertisement for Pacaso’s Regulation A offering. Read the offering circular at invest.pacaso.com. Reserving a ticker symbol is not a guarantee that the company will go public. Listing on the NASDAQ is subject to approvals.
Superhero Tools (Save hours)
Zapier: Automates your daily workflows by connecting over 8000 apps, letting you move data and trigger actions across your favorite tools without any coding.
beehiiv: Helps you build, grow, and monetize newsletters and websites, with powerful tools for custom audience engagement and smooth subscriber management.
Replit: Lets you build, collaborate, and deploy apps or websites right from your browser, with AI features to help code, test, and launch faster.
Wan.video: Gives you free AI-powered tools to quickly create and enhance stunning videos from images or text, perfect for beginners and advanced users alike.
Veo 3: Generates high-quality AI videos with synchronized audio, letting you bring ideas to life with realistic sound, motion, and dialogue.
Quick AI News Bites
ASML is now the largest shareholder of Mistral AI after investing $1.5 billion in the French startup’s latest funding round, aiming to make Europe a stronger player in AI technology. The move will give Mistral a whopping $11.7 billion valuation and boost European tech independence.
Pray.com is making waves by posting AI-generated Bible videos online, attracting millions of viewers but also raising tough questions from religious scholars. While young audiences love the creative "Marvel Universe" take, some critics say it turns sacred stories into just flashy entertainment.
Snapchat just launched an “Imagine Lens” that lets subscribers create unique AI-generated images using text prompts, like “make me an alien” or “create a grumpy cat.”
Sam Altman says social media now feels “fake” because bots are everywhere and even real users sound like AI. He worries it’s almost impossible to tell if posts are from humans or automated accounts.
How to use NotebookLM
Image Source: Google
Visit NotebookLM website.
Create a new notebook notebook and upload your sources.
Add PDFs, URLs, and notes to create a knowledge base of content you want to study or synthesize.
Use natural language questions, generate summaries, and organize ideas, all within the NotebookLM interface.
With the latest update, you can now create customizable flashcards and quizzes directly from your materials. If a quiz question stumps you, just tap "Explain" — NotebookLM will generate an in-depth summary in the chat.
Tailor the platform to your workflow, whether you’re researching, writing, or brainstorming.
After uploading sources, access the new flashcards/quizzes option from your notebook summary screen.
To create quality reports, select “Create Report” and customize the structure and output as per your requirements, choose preferred language, format, and even the style or tone.
Use dynamic suggestions for faster workflows: Let NotebookLM recommend report types based on your source material, then edit or prompt as needed for your specific use-case.
Boost revenue and gain new customers by partnering with us
Reach over 35K AI enthusiasts with your product.
Join our newsletter to connect with tech professionals, investors, engineers, managers, and business owners worldwide. DM now!
Got feedback, a story worth toasting, or a wild tech question? I’d love to hear from you; just reply, or find me on X.
Cheers,
— Poonam Soni