The week in AI: The rise of 'AI-pimping' and 'PDF to brainrot' tools

In partnership with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence - we pass along the news, useful resources, tools and services; we highlight the top research in the field as well as exciting developments in open source. Even if you aren’t a machine learning engineer, we’ll keep you in touch with the most important developments in AI.

NEWS & OPINION

Concerning AI trends continue to emerge: ‘AI-pimping’ and ‘PDF-to-brainrot’

-------------------------

Now that the US presidential election is over, we can finally shift our AI doomsaying lens off of election disinformation, political social media bots and AI-generated videos of Trump and Biden sharing ice cream cones. Unfortunately, there’s not much rest for the AI doom-weary these days.

To start, AI-generated influencers are flooding Instagram and adult social media platforms, often leveraging stolen content from real-life models to craft and monetize virtual personas. According to WIRED, what started as a niche or smaller scale problem has now (perhaps unsurprisingly) exploded into full-scale industry. The accounts typically rely on advanced AI tools to replace faces in videos and/or generate entirely synthetic profiles.

While some of the accounts investigated disclose their AI origins, others deceptively present themselves as real individuals - often repurposing stolen videos and photos. Guides from those currently profiting (Professor EP claims to have made $1m in 6 months; FanVue backs his numbers) are now teaching aspiring “creators” how to exploit this technology for financial gain. Meanwhile, platforms like Instagram face criticism for being unable - or unwilling - to curb the trend. Major offenders can and sometimes do get removed, but content moderation on these platforms is not nearly well-equipped enough to deal with the issue.

Alexios Mantzarlis, former principal of trust and safety intelligence at Google, warns that this "blended unreality" where bots dominate platforms could be our social media future: “It felt like a possible sign of what social media is going to look like in five years. Because this may be coming to other parts of the internet, not just the attractive-people niche on Instagram. This is probably a sign that it's going to be pretty bad.”

Meanwhile, AI-powered PDF to Brainrot “study” tools are attempting to profit from TikTok’s viral trend of layering monotone narration over ASMR or footage from video games. These tools allow users to upload text, such as a PDF or study guide, which is then summarized or read aloud by an automated voice over footage of ASMR or video games in an attempt to keep students engaged with the study material. The article outlines the emerging popularity of the trend and tools, while questioning their legitimacy and effectiveness. PDF to brainrot tools might actually help some students study. Still, it’s hard not to wonder what it will take to capture attention spans in the future if this is where we’re at already.

Your go-to AI probably got an update of some sort this week - here’s what’s going on in LLMs

-------------------------

This week saw many incremental LLM improvements from almost all the major players, as well as France’s Mistral AI and China’s DeepSeek staking a claim to belong on that list. If you have found yourself using only one LLM over the past few months, now would be a great time to experiment/branch out: a) all the frontier models are currently quite good - gaps in capability, reasoning and task handling vary from model to model but even DeepSeek’s open models can outperform ChatGPT or Claude on certain tasks, b) many companies are currently not pay-walling their most powerful models/features in order to compete with OpenAI’s massive subscriber base, and c) performing more complex tasks and building with these models is becoming increasingly accessible for non-developers.

• ChatGPT: OpenAI has introduced Advanced Voice Mode for ChatGPT’s web platform for all paid users. The feature, initially available on iOS and Android apps only, allows users to interact with ChatGPT via voice input directly in their web browser. If you haven’t tried advanced voice mode, it’s definitely worth a try. It’s convenient and the interactions feel different than traditional text-based outputs.

  • Additionally, GPT-4o got an upgrade - the model’s creative writing ability has improved to be more natural and engaging, with tailored writing to improve relevance and readability. The model also works better with your uploaded files, providing deeper insights and more thorough responses.

• Mistral: French AI startup Mistral has transformed its Le Chat platform into a full-on comprehensive workspace rival to ChatGPT. They’ve added a ton of new features to Le Chat and also upgraded the base models powering the chatbot (which are state-of-the-art) - and everything is free to try/use, as Mistral currently has an incredibly generous free tier.

  • They also just released Pixtral Large, a multimodal model that outperforms top models on math reasoning and real-world tasks (also beating Gemini 1.5 Pro and GPT-4o on chart and document understanding). The model features a 128K context window capable of simultaneously processing 30 high-res images or a 300-page book.

• Gemini: Google already most likely knows everything about you, but now Gemini can finally save and remember information about you. It’s called “Saved Info” and is similar to Memory in ChatGPT. But unlike ChatGPT’s auto-saved memories, you explicitly tell Gemini what to remember about you. Memory in chatbots feels primarily like a way to “humanize” the AI, but there are practical uses here, too - like getting personalized recipe outputs if you’re a vegetarian, etc.

  • Gemini also got a new standalone iPhone app featuring Gemini Live voice conversations (same as Advanced Voice Mode above - Gemini and ChatGPT are the only models that currently offer this), image generation capabilities, and broader integration with Google services on iPhones.

• Perplexity: Pro users can now use a new one-click shopping feature that allows you to both research and buy products straight from Perplexity’s UI instead of just providing answers. You can tell Perplexity, “I am hosting a [enter event details]. Generate a shopping list with [additional details],” and you’ll get free shipping on all your Buy with Pro orders as a thank-you for shopping with Perplexity. They claim their product cards aren’t sponsored, but rather unbiased recommendations tailored to your search … but for what it’s worth, not even CEO Aravind Srivinas could tell you how Perplexity arrived at those recommendations.

• Claude: Anthropic has introduced a prompt improver and evaluator directly in the Anthropic Console, making it much easier to leverage prompt engineering best practices and build more reliable AI applications. Prompting best practices take a lot of time to develop/implement and often vary across model providers - the prompt improver leverages Claude to automatically optimize your prompts with chain-of-thought reasoning, structured rewrites, and example enrichment techniques. The prompt evaluator helps users benchmark prompts with ideal outputs and refine them iteratively.

• DeepSeek: The Chinese AI research firm just launched R1-Lite-Preview, a powerful new reasoning-focused model that matches OpenAI's o1 reasoning capabilities while showing a live, transparent chain-of-thought process. You can access the model through DeepSeek Chat for free, with premium reasoning features limited to 50 daily messages (basic chat remains unlimited). DeepSeek plans to open-source the complete R1 model in the future - and while lesser known in the West, open-sourcing this powerful Chinese model would send a warning shot to closed US AI firms.

A revolution in interactive storytelling: when books become games

-------------------------

We’ve covered Google’s NotebookLM quite a bit in recent weeks; it’s an outstanding AI-first notebook tool designed to summarize, organize, and create new interactive content out of your uploaded documents (including the ability to create dual-host podcasts from them, which went viral last month).

This week, NotebookLM’s advisor Steven Johnson published a blog post showcasing a fun new fusion of history, technology, gaming and storytelling. The post starts off with an embedded interactive ‘game’ that covers the rise of the modern detective during the early stages of fingerprinting technology (based on Johnson’s recent book, The Infernal Machine, which he uploaded) - powered by Google Gemini and some clever prompting.

The game places you in the role of Detective Joseph Faurot in 1911, tasked with solving a burglary using pioneering fingerprint analysis. You’re immediately immersed in the world of historical forensics from the book (and whatever relevant training data the LLM might have/contribute); but what’s really impressive here is how AI models can now instantly transform a static narrative into a dynamic, personalized adventure.

We played twice, and both playthroughs were quite a bit different (and interesting) based on our prompting. It’s not hard to imagine a somewhat more sophisticated setup where you can turn any setting, event, book, series or just about any IP into a full-blown, dynamic choose your own adventure.

Johnson created the game by uploading his book and using a single 400 word prompt - Gemini took care of the rest. One could ostensibly use any narrative text (fiction or nonfiction) and create an equally sophisticated game in a matter of minutes, just by slightly altering the wording of the prompt. To validate that idea, Johnson made a similar game out of the Cuban Missile Crisis Wikipedia entry, where the user plays as JFK trying to avoid nuclear war. The blog goes on to detail what makes these interactive adventures suddenly possible; suffice to say, it’s in large part due to Gemini’s massive context window - at 2 million tokens, you could actually upload the entire Harry Potter series with room to spare.

MORE IN AI THIS WEEK

• Elon Musk’s xAI now valued at $50b following fresh capital injection

• Coca-Cola causes controversy with AI-made ad

• ChatGPT outperforms doctors in diagnostic challenge - even when the doctors were allowed to use the chatbot

• US congressional commission proposes “Manhattan Project-like” initiative for government funding of AI

• Microsoft Ignite 2024: The biggest news in AI, Copilots, agents

• San Francisco Mayor-elect Daniel Lurie names OpenAI CEO Sam Altman to transition team

• Niantic is building a ‘geospatial’ AI model based on Pokémon Go player data

• Nvidia’s new AI chips are reportedly overheating in server farms

• Dozens of pages of correspondence between Elon Musk, Sam Altman and other OpenAI founders released in Musk v. OpenAI lawsuit

This isn’t traditional business news

Welcome to Morning Brew—the free newsletter designed to keep you in the know on the business news impacting your career, company, and life—in a way you didn’t know you needed.

Note: this isn’t traditional business news. Morning Brew’s approach cuts through the noise and bore of classic business media, opting for short writeups, witty jokes, and above all—presenting the facts.

Save time, actually enjoy business news, and join over 4 million professionals reading daily.

Check it out

TECHNICAL NEWS, DEVELOPMENT, RESEARCH & OPEN SOURCE

• Google DeepMind introduces AlphaQubit: an AI system that dramatically improves the ability to fix errors in quantum computers

• Cerebras video shows AI writing code 75x faster than world's fastest AI GPU cloud

• GenAI startup Writer introduces self-evolving models: the future of scalable AI

• AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably

• Stripe introduced a new agent toolkit that enables developers to integrate payments, financial services, and usage-based billing into agent workflows

• Anthropic publishes research proposing a statistical framework for AI model evaluations to measure language model capabilities better than current benchmarks

• LLaVA-CoT: the first vision language model capable of spontaneous, systematic reasoning, similar to GPT-o1

That’s all for this week! We’ll see you next Thursday.