by Dan Shipper in Chain of Thought DALL-E/Every illustration. It’s the last day of Every’s think week—our quarterly time to dream up new ideas and products that can help us improve how we do our work and, more importantly, your experience as a member of our community. In lieu of publishing new stories, we’ve been re-upping pieces by Dan Shipper (who’s been on hiatus from writing his regular Chain of Thought column to work on a longer piece) that cover basic, powerful questions about AI. Last up is his piece about how humans should think in a world with thinking machines. We'll be back with a new piece in your inbox on Monday.—Kate Lee Was this newsletter forwarded to you? Sign up to get it in your inbox. A world with thinking machines requires new thinking styles. Our default thinking style in the West is scientific and rationalist. When was the last time you heard someone talking about a hypothesis or theory in a meeting? When was the last time, when sitting down to solve a problem, you reminded yourself to think from first principles? When was the last time you tried an experiment in your work or personal life?  Even the frameworks we use to understand business are scientific: It’s unlikely that Harvard Business School professor Michael Porter would have looked for or found five “forces” governing business without physics as inspiration; Clay Christensen’s jobs-to-be-done framework is close to an atomic theory of startup ideas.  We romanticize science and rationalism because it's been so successful. Since the Enlightenment, when Galileo, Newton, Descartes, and Copernicus began to think in this way, we have used rationalism to generate modernity. It's where we get rockets and vaccines from, and how we get computers and smartphones. But new technologies demand new thinking styles. As the AI age unfolds, we are shifting away from what former Tesla and OpenAI engineer Andrej Karpathy calls Software 1.0—software that consists of instructions written by humans, and which benefits from a scientific, rationalist thinking style.  Instead, we're moving into Software 2.0 (a shift that Michael Taylor recently wrote about), where we describe a goal that we want to achieve and train a model to accomplish it. Rather than having a human write instructions for the computer to follow, training works by searching through a space of possible programs until we find one that works. In Software 2.0, problems of science—which is about formal theories and rules—become problems of engineering, which is about accomplishing an outcome. This shift—from science to engineering—will have a massive impact on how we think about solving problems, and how we understand the world. Here are some of my preliminary notes on how I think this shift will play out. 1. Essences vs. sequences In a pre-AI world, whether you were building software or teams, or writing books or marketing plans, you needed to strip the problems you were facing down to their bare elements—their essence—and work your way forward from there. In building software, you need to define your core user and the problem you want to solve; in writing books, you need a thesis and an outline. In a post-AI world, we are less concerned with essence and more concerned with sequence: the underlying chain of events that leads to a certain thing to happen. Language models do this when they predict what word comes next in a string of characters; self-driving cars also do this when they predict where to drive next from a sequence of video, depth, and GPS data.  To understand this better, consider the case of a churn prevention feature for a SaaS business in a pre-AI world. In order to automatically prevent a customer from churning, you needed to define what a customer who might churn looked like with explicit rules—for example, if they hadn’t logged into your app in a certain number of months, or if their credit card was expiring soon. This is a search for essences. In a post-AI world, by contrast, you don’t need to explicitly define what a customer who is about to churn looks like, or which interventions you might use in which circumstances.  All you have to do is identify sequences that lead to churn. For every customer who churns, you can feed their last 100 days of user data into a classifier model that categorizes inputs. Then you can do the same for customers who haven't churned. You'll create a model that can identify who is likely to churn, in all of their many thousands of permutations, without any rules. This is what it means to search for sequences.  2. Rules vs. patterns Become a paid subscriber to Every to unlock this piece and learn how: Pattern recognition replaces rule-based thinking Intuition-driven approaches eclipse process-centric methods Creative work evolves from sculpting to gardening Predictions become more valuable than explanations 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.

by Dan Shipper in Chain of Thought DALL-E/Every illustration. The world has changed considerably since our last ”think week” five months ago—and so has Every. We’ve added new business units, launched new products, and brought on new teammates. So we've been taking this week to come up with new ideas and products that can help us improve how we do our work and, more importantly, your experience as a member of our community. In the meantime, we’re re-upping four pieces by Dan Shipper that cover basic, powerful questions about AI. (Dan hasn’t been publishing at his regular cadence because he’s working on a longer piece. Look out for that in Q2.) Thus far we’ve re-published his jargon-free explainer of how language models work and his piece about how language models function as compressors—or summarizers—of text. Today we’re sharing how language models function as the opposite—as text expanders.—Kate Lee Was this newsletter forwarded to you? Sign up to get it in your inbox. You can’t get energy for free. It can be neither created nor destroyed, just moved around. That's more or less what computers were able to do with text on their own for a long time. Barring a disk failure, text was always conserved, often moved around, sometimes crudely transformed.  But they almost never created it. Other than doing a spell check, if you were seeing text on a computer, it was probably because some human, somewhere, had typed it. Language models changed this entirely.  Now, you and I can type a few sentences into ChatGPT and watch it expand, character by character, line by line, into something new—composed out of thin air, just for you. Language models take your text and stretch it into a different shape, like glass heated and blown through a tube.  What had previously been an inert collection of bits—a line of characters extending across a screen—is now something different, something potentially alive. When you feed a piece of text to a language model, the text is like an acorn turning into a tree. The acorn itself contains instructions for the tree it will become, and the language model becomes rich dirt, water, and warm summer sun. In short, language models are free energy for text. Let’s talk about how we can use that function for creative purposes. A world where every question contains an answer Become a paid subscriber to Every to unlock this piece and learn about: The transformation of questions into instant answers Three expansion types: comprehensive, contextual, and creative How expansion differs from compression in practical application The creative power of controlled unpredictability 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 conversation with Jai Malkani, Head of AI and Product, Customer Obsession at Uber.

A look into state-by-state adoption and how gaps might impact workforce readiness.

by Dan Shipper in Chain of Thought DALL-E/Every illustration. The world has changed considerably since our last ”think week” five months ago—and so has Every. We’ve added new business units, launched new products, and brought on new teammates. So we’re taking this week to come up with new ideas and products that can help us improve how we do our work and, more importantly, your experience as a member of our community. In the meantime, we’re re-upping four pieces by Dan Shipper that cover basic, powerful questions about AI. (Dan hasn’t been publishing at his regular cadence because he’s been on a longer piece. Look out for that in Q2.) Yesterday we re-published his jargon-free explainer of how language models work. Today we’re re-upping his piece about how language models function as compressors—or summarizers—of text.—Kate Lee Was this newsletter forwarded to you? Sign up to get it in your inbox. I want to help save our idea of human creativity. Artificial intelligence can write, illustrate, design, code, and much more. But rather than eliminating the need for human creativity, these new powers can help us redefine and expand it.  We need to do a technological dissection of language models, defining what they can do well—and what they can’t. By doing so, we can isolate our own role in the creative process.  If we can do that, we’ll be able to wield language models for creative work—and still call it creativity.  To start, let’s talk about what language models can do. The psychology and behavior of language models The current generation of language models is called transformers, and in order to understand what they do, we need to take that word seriously. What kind of transformations can transformers do? Mathematically, language models are recursive next-token predictors. They are given a sequence of text and predict the next bit of text in the sequence. This process runs over and over in a loop, building upon its previous outputs self-referentially until it reaches a stopping point. It’s sort of like a snowball rolling downhill and picking up more and more snow along the way. But this question is best asked at a higher level than simply mathematical possibility. Instead, what are the inputs and outputs we observe from today’s language models? And what can we infer about how they think?  In essence, we need to study LLMs’ behavior and psychology, rather than their biology and physics. This is a sketch based on experience. It’s a framework I’ve built for the purposes of doing great creative work with AI. A framework for what language models do Language models transform text in the following ways: Compression: They compress a big prompt into a short response. Expansion: They expand a short prompt into a long response. Translation: They convert a prompt in one form into a response in another form. These are manifestations of their outward behavior. From there, we can infer a property of their psychology—the underlying thinking process that creates their behavior: Remixing: They mix two or more texts (or learned representations of texts) together and interpolate between them. I’m going to break down these elements in successive parts of this series over the next few weeks. None of these answers are final, so consider this a public exploration that’s open for critique. Today, I want to talk to you about the first operation: compression. Become a paid subscriber to Every to unlock this piece and learn about: The multi-dimensional nature of AI compression How compression powers creative work AI as the next evolution of cultural technology 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.

Can frontier LLMs earn $1 million from real-world freelance software engineering?

OpenAI and Guardian Media Group announce content partnership to bring Guardian news content to ChatGPT.

A conversation with Andrea Ellis, Chief Financial Officer of Fanatics Betting and Gaming.

A conversation with Fiona Tan, Chief Technology Officer of Wayfair.

Rogo scales AI-driven financial research with OpenAI o1

NASA is expanding testing of distributed autonomy for its first satellite swarm, Starling. The post NASA boosts science objectives for Starling swarm appeared first on SpaceNews.

OpenAI and Schibsted Media Group announce content partnership to bring Guardian news and archive content to ChatGPT.

We aired our first-ever television ad during the Super Bowl to pique people’s curiosity and help us all realize how AI can open up new possibilities for us, create more fulfillment in our lives, and make us more productive, just as all the tools that came before AI did for those who came before us.

OpenAI looks forward to engaging with global leaders on AI’s role in shaping innovation and economic prosperity.

Data residency builds on OpenAI’s enterprise-grade data privacy, security, and compliance programs supporting customers worldwide.

Methods and Strategies for Building and Refining Reasoning Models

The largest deployment of ChatGPT to date will expand the use of AI in education and help the United States build an AI-ready workforce.

ChatGPT and personal tutoring