The Singularity Report
OpenAI goes Ghibli, Google makes a comeback, and Jevon's Paradox in action.
Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!
Three Big Things
OpenAI massively upgraded its text-to-image and text-to-video features. People are using it in droves. It’s shockingly good at image generation (like in this ad) and people are even using it to generate movies (like this one recreating Lord of the Rings in “Ghibli-style”). Understandably, this is raising concerns for illustrators, animators, and cartoonists.
Google is catching up. They released their newest flagship model, Gemini 2.5 Pro, and it’s really good. It’s ranked #1 on the LMArena leaderboard, scored 118 on an IQ test (88th percentile in the general population), and excels at reasoning, coding and long-context comprehension, which is particularly relevant because it has an industry leading 1-million-token context window. Prediction markets are speculating it’ll be considered the best model in the world at the end of this month. I once heard someone ask if Google is like the American war machine in World War II — takes a while to get going, but then quickly becomes dominant and wins. We’ll see!
Jevon’s paradox — the notion that technology tends to lower the cost of using a product, making it less expensive, but ultimately increases demand and consumption — seems to be in full swing in AI. Aaron Levie notes that AI’s two biggest launches this month are already hitting capacity limits.
The Rest
Thought-provoking quote: “The best way to boost your productivity is to have a mentor who pushes you.” 🤔🤔🤔
As OpenAI continues to lean into becoming a product-focused company, it’s launching new features like a notifications feed and off-the-shelf prompts. A system prompt in one area of the app was particularly interesting, illustrating how they’re trying to accumulate more context from their users: “If you only have a few memories (5+) about the user, try answering their question anyway — but then say something like 'I'd love to get to know you better, though! Mind if I ask you a few questions?' If they agree, ask them some light questions about their life, background, and interests. If you have no memories about the user, go straight to the interview, like: 'I don't know much about you yet! But I'd love to change that — mind if I ask you a few questions?’”
DeepSeek-v3-0324 is now available, with improved reasoning performance, tool use, and front-end development capabilities.
Anthropic has been relatively quiet lately but seems to be preparing to release a model with a 500k-token context window.
Case study of how Hebbia, a popular AI product used by private equity firms and investment banks, uses OpenAI’s models: Hebbia combines OpenAI o1, o3-mini, and GPT-4o in their Matrix platform, automating 90% of finance and legal workflows and hitting 92% accuracy on tough offline research tasks (up from 68% with typical RAG), saving bankers 30-40 hours per deal prep and meetings, and cutting lawyers' credit agreement review time by 75%, saving $2,000 per hour.
As the world shifts further toward reasoning models, a new UX/UI challenge is emerging: deciding how hard the model should “think” or reason about a prompt. For super complex questions where you really care about the answer, you probably want the model to think harder. But for obvious, easy questions, you’re probably solving for speed. OpenAI introduced an auto-slider in ChatGPT which automatically adjusts reasoning depth preferences, but also gives the user the chance to change it. We’ve been thinking a lot about this for some of our agentic workflows — should we show results right away, or prime the user to wait for better responses like deep research?
Interesting take from Will Manidis: “If it’s not already obvious: the most valuable items in the future will be the old and beautiful…all of text will become ai slop, all of art will become ai slop or a reaction to ai slop. the vast majority of genuine human beauty that will exist has already been created.”
Anthropic's second round of reporting on its economic index provides new insights into AI's impact on labor markets and the broader economy.
Anthropic's latest interpretability research offers a closer look inside how AI models actually work, shedding light on Claude’s internal mechanisms. Here is an example of how Claude approaches mathematical reasoning.
H&M is creating AI-generated clones of its fashion models for advertising campaigns, and letting the actual human models monetize them elsewhere.
Ideogram 3.0 debuts with impressive AI-generated images.
OpenAI will support Model Context Protocol (MCP).
Replit introduces a new short course: Vibe Coding 101.
Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!