Neural networks first treat sentences like puzzles solved by word order, but once they read enough, a tipping point sends them diving into word meaning instead—an abrupt “phase transition” reminiscent of water flashing into steam. By revealing this hidden switch, researchers open a window into how transformer models such as ChatGPT grow smarter and hint at new ways to make them leaner, safer, and more predictable.

by Dan Shipper in Chain of Thought Midjourney/Every illustration. Was this newsletter forwarded to you? Sign up to get it in your inbox. When an infant is born, they are completely dependent on their caregivers to survive. They can’t eat, move, or play on their own. As they grow, they learn to tolerate increasingly longer separations. Gradually, the caregiver occasionally and intentionally fails to meet their needs: The baby cries in their crib at night, but the parent waits to see if they’ll self-soothe. The toddler wants attention, but the parent is on the phone. These small, manageable disappointments—what the psychologist D.W. Winnicott called "good-enough parenting"—teach the child that they can survive brief periods of independence. Over months and years, these periods extend from seconds to minutes to hours, until eventually the child is able to function independently. AI is following the same pattern. Today we treat AI like a static tool we pick up when needed and set aside when done. We turn it on for specific tasks—writing an email, analyzing data, answering questions—then close the tab. But as these systems become more capable, we'll find ourselves returning to them more frequently, keeping sessions open longer, and trusting them with more continuous workflows. We already are. So here’s my definition of AGI: Become a paid subscriber to Every to unlock this piece and learn about: How AGI is defined by economic persistence The irreversible threshold of continuous operation Beyond the moving targets of Turing and OpenAI definitions The five essential capabilities of persistent agents A clear trajectory from seconds to perpetual runtime Upgrade to paid Click here to read the full post Want the full text of all articles in RSS? Become a subscriber, or learn more.

A topic-organized collection of 200+ LLM research papers from 2025

Learn how Genspark built a $36M ARR AI product in 45 days—with no-code agents powered by GPT-4.1 and OpenAI Realtime API.

Today, OpenAI, in partnership with Mandala Partners, is sharing the OpenAI AI Economic Blueprint for Australia. At a time when boosting productivity has emerged as a national priority for Australia, the Blueprint provides a clear, actionable plan for how Australia can unlock the full economic and social potential of artificial intelligence.

Retell AI is transforming the call center with AI voice automation powered by GPT-4o and GPT-4.1. Its no-code platform enables businesses to launch natural, real-time voice agents that cut call costs, boost CSAT, and automate customer conversations—without scripts or hold times.

Unify, an AI-powered GTM platform, uses OpenAI’s o3, GPT-4.1, and CUA to automate prospecting, research, and outreach. With hyper-personalized messaging and an always-on workflow, Unify helps teams generate pipeline at scale while focusing on high-impact customer interactions.

Research update on on applying local volume measurement to downstream tasks

Advanced AI can transform biology and medicine—but also raises biosecurity risks. We’re proactively assessing capabilities and implementing safeguards to prevent misuse.

We study how training on incorrect responses can cause broader misalignment in language models and identify an internal feature driving this behavior—one that can be reversed with minimal fine-tuning.

KV caches are one of the most critical techniques for efficient inference in LLMs in production.

We’re launching OpenAI for Government, a new initiative focused on bringing our most advanced AI tools to public servants across the United States. We're supporting the U.S. government's efforts in adopting best-in-class technology and deploying these tools in service of the public good.

OpenAI and Mattel are partnering to integrate AI into iconic brands such as Barbie and Hot Wheels, aiming to enhance creative development, streamline workflows, and create new ways for fans to engage.

In this post, we will study inductive biases of the parameter-function map of random neural networks using star domain volume estimates. This builds on the ideas introduced in Estimating the Probability of Sampling a Trained Neural Network at Random and Neural Redshift: Random Networks are not Random Functions (henceforth NRS). Inductive biases To understand generalization in deep neural networks, we must understand inductive biases. Given a fixed architecture, some tasks will be easily learnable, while others can take an exponentially long time to learn (see here and here).

by Dan Shipper in Chain of Thought DALL-E/Every illustration. Inspired by recent AI & I guest Nadia Asparouhova, Dan Shipper is away on a Jhana meditation retreat. In lieu of him saying, well, anything this week, we’re republishing his piece from last fall about the importance of context in building products. It’s a topic on which Thesis writer Tina He recently co-wrote a paper in regards to LLMs—we’ll have more to say on it soon.—Kate Lee Was this newsletter forwarded to you? Sign up to get it in your inbox. A common piece of startup advice from the last 10-15 years is: "Make something people want." It resonates because it effectively distills the core job of a founder into a phrase. Unfortunately, it’s not very actionable. It’s sort of like telling an investor: “Beat the market.” In both of these cases, the obvious next question is: “How?” A dominant framework for thinking about this question is to consider customer needs. Under this worldview, people simply buy the products they need. Or to put it another way, they “hire” products to do specific jobs for them. If you use this framework, you often implicitly adopt the view that people have an objective, discoverable set of problems that they need solved. And that if you look closely enough, you can map out those problems, connect them to needs, and thereby identify products to build. This is a scientific approach to the building of products. Something previously mysterious becomes an empirical activity: You go out and discover needs, and then create ways to fulfill them. This means talking to customers—on the street, through surveys, on the internet, in bars—however and wherever you can find them. A naive way to do this is to grab a potential customer by the collar and straight-up ask them, “What do you want?” Any experienced founder can tell you this doesn’t normally work. You get shrugs or, if you’re in Manhattan, the occasional, “A fuckin’ raise!” People don’t have a list of needs at hand. Instead, an experienced founder acts more like an ethnographer. Their goal is to build up a detailed understanding of their target customer: what they do when they wake up in the morning, how they go through their inbox, and how they respond when they see that their boss replied to their long, detailed, late-night email with the single word “No.” The idea is to be objective. You ask questions to gather neutral data. That data tells you the truth about the world. You don’t want to fool yourself with leading questions and trying to confirm your priors. But as any experienced founder knows, this process is fractal. The deeper you get into asking questions about a person’s day, the more detail there is to observe. And true needs or jobs—the ones that are burning or unfulfilled—get lost in the process of mapping the infinite contours of an individual life. What actually works? Drop the objectivity. Instead, lead with a particular idea or bias. Once you present to a customer your own view of the world or a perspective on what you intend to build, everything changes. Suddenly, the squishy mess inside their head hardens into visibility. It’s like a non-Newtonian fluid subjected to force: a liquid until you press on it with the strength of your own perspective. You’ve probably had an analogous experience in other areas of your life. Consider the question: “What do you want for dinner?” Or, “Which movie do you want to watch?” Or, “What kind of article do you want to read?” For me, I draw a blank. But if you ask me about the food at Sukh or what I thought about Inside Out 2, I can and will unfurl an incredibly long string of sentences with minute detail about why Sukh is the best Thai in the city—the hor mok in particular—and how Inside Out 2 is catnip for therapy nerds with, by the way, a totally unsatisfying story arc. Why the difference in my ability to respond? Needs arise in context—they are not context-free facts Become a paid subscriber to Every to unlock this piece and learn about: How to uncover latent customer needs Product development as creative act Cultivating founder intuition Upgrade to paid Click here to read the full post Want the full text of all articles in RSS? Become a subscriber, or learn more.

OpenAI introduces its Outbound Coordinated Disclosure Policy to guide how it responsibly reports vulnerabilities in third-party software—emphasizing integrity, collaboration, and proactive security at scale.

Announcing the Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text

OpenAI is fighting a court order at the demands of The New York Times and plaintiffs, which involves retention of consumer ChatGPT and API user data indefinitely. Learn how we’re working to uphold user privacy, address legal requirements, and stay true to our data protection commitments.

Our latest report featuring case studies of how we’re detecting and preventing malicious uses of AI.

"In projecting language back as the model for thought, we lose sight of the tacit embodied understanding that undergirds our intelligence." –Terry Winograd The recent successes of generative AI models have convinced some that AGI is imminent. While these models appear to capture the essence of human