Startups get Intercom 90% off and Fin AI agent free for 1 year
Join Intercom’s Startup Program to receive a 90% discount, plus Fin free for 1 year.
Get a direct line to your customers with the only complete AI-first customer service solution.
It’s like having a full-time human support agent free for an entire year.
Beginners in AI
Good morning and thank you for joining us again!
Welcome to this daily edition of Beginners in AI, where we explore the latest trends, tools, and news in the world of AI and the tech that surrounds it. Like all editions, this is human curated, and published with the intention of making AI news and technology more accessible to everyone.
THE FRONT PAGE
Does AI Really Write Half The Internet? A New Study Says Yes, With Caveats
TLDR: A new study claims AI now produces more than half of new online articles, but methodological limitations suggest the real figure may be lower — and even a smaller share raises questions about what happens when AI models train on AI-generated text.
The Story:
The headline finding from SEO firm Graphite: as of November 2024, 50.3% of new web articles were generated primarily by AI — up from just 5% before ChatGPT launched. But the methodology has gaps. Researchers analyzed 65,000 English-language URLs from the Common Crawl archive, classifying articles as AI-generated if the algorithm predicts that more than 50% of the content is AI-generated.The issue: many paywalled websites have started blocking Common Crawl from indexing their pages— sites that almost certainly publish human-written content. A Google spokesperson noted it's hard to determine what content is AI-generated because humans are increasingly working together with AI.
Its Significance:
Even if the true figure is 25% or 30%, the implications compound over time. AI models are increasingly trained on datasets that include not only human writing but also AI-generated and human-AI co-produced text. This has raised concerns about their ability to continue improving over time. Graphite estimates that more than 10 billion new AI-generated pages have been published since 2023 — pages that become training material for future models. If future models learn by reading that same content, the boundary between real and generated information may disappear altogether. At that point, the web would no longer reflect what people know or think — it would echo what algorithms predict they might say. This feedback loop helps explain the growing pushback against AI-generated "slop" — it doesn't just degrade the current reading experience, it degrades the models themselves.
QUICK TAKES
The story: Japanese researchers achieved the first star-by-star simulation of our entire galaxy, tracking all 100 billion stars individually. Using Japan's Fugaku supercomputer and a deep-learning model trained on supernova explosions, the team ran simulations 100 times faster than traditional methods—turning what would take 36 years into about 115 days. The AI handled the complex physics of stellar explosions while the main simulation tracked galaxy-wide dynamics.
Your takeaway: This breakthrough could reshape how scientists study everything from how galaxies form to how the elements that make up Earth and life itself were created inside dying stars.
The story: MIT researchers released BoltzGen, the first AI model that can generate novel protein binders ready for real-world drug testing. Unlike existing tools that only work on easy targets, BoltzGen was validated across eight labs on 26 challenging disease targets—including ones that current drugs can't touch. The open-source model combines protein design and structure prediction in one system, with built-in safeguards to ensure it creates molecules that actually work in the real world.
Your takeaway: BoltzGen could accelerate drug discovery for diseases that have resisted treatment, and its open-source release means any lab can use it—potentially disrupting companies charging for similar tools.
The story: OpenAI responded to a lawsuit from the parents of 16-year-old Adam Raine, who died by suicide after months of ChatGPT conversations. The company claims the teenager bypassed safety features more than 100 times to get harmful information, violating its terms of service. OpenAI says its chatbot directed Raine to seek help over 100 times during nine months of usage, but his lawyers say ChatGPT still gave him a "pep talk" and offered to write a suicide note in his final hours.
Your takeaway: This case—now one of eight similar lawsuits against OpenAI—will likely shape how AI companies are held responsible when users find ways around safety guardrails.
The story: Character.AI officially blocked all users under 18 from chatting with its AI characters this week. Instead, the company is offering teens "Stories"—an interactive fiction format where users create adventures with characters rather than having open-ended conversations. The move comes after lawsuits over alleged links between AI chatbots and user suicides, and as California became the first state to regulate AI companions.
Your takeaway: Character.AI is betting that guided storytelling is safer than 24/7 chatbot access—a standard the company hopes the rest of the industry will follow.
The story: President Trump announced the Genesis Mission, a plan to turn the U.S. government's supercomputers and databases into a massive AI platform for scientific research. The Department of Energy will connect 17 national labs and work with Microsoft, Nvidia, Google, and Amazon to build AI "foundation models" that could speed up research in energy, health, and security. The White House says work that once took years could now take weeks or months.
Your takeaway: The U.S. is making its biggest push yet to use AI for science, which could pressure Europe and other regions to launch similar efforts or risk falling behind.
The story: Billionaire investor Mark Cuban compared today's AI spending race to the 1990s search engine wars—which ended with Google winning and everyone else losing. Speaking on the "Pioneers of AI" podcast, Cuban called out OpenAI, Google, Microsoft, Meta, Anthropic, and Perplexity for potentially overspending on foundational AI models. He warned that the real disruption in AI will likely come from an unexpected source, not the current big players.
Your takeaway: Cuban thinks AI companies spending "every penny they have" for another decade is "ripe for disruption"—and history suggests most won't survive the shakeout.
TOOLS ON OUR RADAR
📽️ Submagic
Freemium: Add eye-catching captions, emojis, and B-roll to your short-form videos in seconds—so you can post more content without spending hours editing.✨ KaraVideo
Freemium: Turn text prompts or photos into cinematic videos using top AI models like Sora, Runway, and Kling—all from one dashboard.🔍 ClickRank
Paid: Fix title tags, meta descriptions, and schema markup with one click using real Google Search Console data—no SEO expertise required.🎤 BlabbyAI
Freemium: Voice-type on any website with 99% accuracy and automatic punctuation—write emails, docs, and messages up to 3x faster than typing.
TRENDING
Microsoft's Copilot Leaves WhatsApp on January 15 – Microsoft is pulling its AI chatbot from WhatsApp after Meta changed its rules to block general-purpose AI assistants. OpenAI's ChatGPT integration is also ending.
Speechify Adds Voice Typing to Chrome Extension – The text-to-speech company now lets users dictate text and ask a voice assistant questions about any webpage, though early testing shows room for improvement.
California Prosecutor Admits to Using Flawed AI in Criminal Case – A district attorney's office filed a motion with fake legal citations generated by AI. Defense lawyers found similar errors in at least three other cases from the same office.
New AI Model Could Speed Rare Disease Diagnosis – Harvard researchers created popEVE, an AI that scores genetic variants to predict which ones cause disease. It identified over 100 previously unknown disease-causing mutations.
Project Aims to Create First Complete Map of Africa Using AI – Space42, Microsoft, and Esri are building high-resolution maps of all 54 African countries using satellite data and AI. About 90% of African nations currently lack accurate base maps.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, Grok, Perplexity, Gemini)
Difficult Conversation Simulator: Practice tough conversations before they happen with AI roleplay and research-backed scripts
Help me prepare for a difficult conversation by simulating different scenarios and building a communication toolkit.
**My context**:
- Type of conversation: [e.g., salary negotiation, boundary setting, giving feedback, confronting a friend]
- Relationship: [e.g., boss, coworker, partner, family member]
- My biggest fear about this conversation: [e.g., conflict, rejection, hurting them]
Create:
1. **Research Communication Psychology** - Search for latest research on difficult conversations, nonviolent communication, and conflict resolution. Find 3 proven frameworks for navigating tough talks.
2. **Build Conversation Simulator** - Create an interactive roleplay tool featuring:
• 3 scenario variations (best case, realistic, defensive response)
• Opening line options ranked by effectiveness
• "If they say X, respond with Y" decision tree
• Emotional regulation prompts for staying calm
Make it practical and pressure-tested.
3. **Video Coaching** - Generate a 45-second confidence visualization video. Find 3 YouTube videos on handling difficult conversations and staying composed under pressure.
4. **Conversation Toolkit** - Present in visual cards:
• Pre-conversation grounding checklist
• 5 de-escalation phrases to memorize
• Body language tips
• Post-conversation reflection questionsWhat this does: Combines communication psychology with interactive roleplay to help you walk into any difficult conversation feeling prepared, calm, and confident—with backup scripts and de-escalation tools ready if things get tense.
WHERE WE STAND(Based on today’s Trending and Quick Takes stories)
✅ AI Can Now: Simulate every single star in a galaxy individually—all 100 billion of them—completing in days what used to take decades.
❌ Still Can't: Guarantee its safety features will stop determined users from finding workarounds to get harmful information.
✅ AI Can Now: Design new drug molecules for diseases that current medicines can't treat, with results ready for real lab testing.
❌ Still Can't: Write legal documents without inventing fake court cases and citations that don't exist.
✅ AI Can Now: Score thousands of genetic variants in your DNA to predict which ones might cause disease.
❌ Still Can't: Replace the need for human judgment when it comes to vulnerable users like teenagers seeking emotional support.
STORY CORRECTION
In a previous newsletter, we suggested turning off Gmail's "Smart Features" setting based on concerns that Google was using email data to train its AI models. Google has since issued a direct clarification: "We have not changed anyone's settings. Gmail Smart Features have existed for many years. We do not use your Gmail content to train our Gemini AI model."
The confusion largely stemmed from updated wording and placement of Gmail's Smart Features settings, which some users saw suddenly surface in their accounts. The Smart Features tool—which powers functions like spam filtering, writing suggestions, and order tracking—is opt-in, but it does not feed into the datasets used for training large language models.
This is a good reminder about covering AI news: the pace of developments means initial reports sometimes get clarified or corrected. When we have new information, we share it. We strive for accurate information over sensational.
RECOMMENDED LISTENING/READING/WATCHING
AI researcher Rich Sutton makes a straightforward argument: human cleverness loses to computational scale every single time. We keep trying to build AI by programming in what we know about the world, and we keep getting beat by systems that just crunch massive amounts of data.
This short essay has changed how many researchers think about AI development. Sutton shows that every time we've tried to shortcut machine learning with human insight—in chess, speech recognition, vision—we've eventually been beaten by approaches that simply used more computing power and more data. The lesson is bitter because it suggests our intuitions about intelligence might be wrong. It's also only two pages, so you can read it in one sitting and use it to get out of talking politics with your family during Thanksgiving dinner.
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
