Anthropic and OpenAI Pick Their Lanes

Hello, and happy Sunday! It was another jam-packed week both in AI and at Every. Alex Duffy explains Anthropic's and OpenAI's briefs to the U.S. government on their AI plans, Google's major robotics advances, and the newest benchmarks. Here at Every, Gareth Edwards's chronicle of the untold story of Jack Tramiel, the man who made Commodore into a household name, evoked scores of memories from readers who grew up with the company's computers, and Vivian Meng explains how Every Studio made, messed up, and rectified a pricing update on our email product, Cora. Read on for more on all of that as well as everything we published this week.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

Release notes

The future of AI comes into focus—and benchmarks are key

With each round of announcements, we’re getting a clearer picture of the future—and realizing just how important it is to keep careful tabs on what these models can (and can’t) do. This week industry heavyweights Anthropic and OpenAI sent detailed AI action plans to the U.S. government, signaling where each company is placing its bets. Elsewhere, there were major advancements and spirited debates about how we benchmark and understand these powerful tools. Let’s get you caught up, then we’ll break it all down.

This week’s major developments

AI action plans: Anthropic and OpenAI set their lanes

• In its plan, Anthropic emphasized national security and model safety, calling AI a “critical national asset,” in part because, “Powerful AI technology will be built during this Administration.”
• OpenAI’s plan focused on broad economic growth, proposing a comprehensive framework that ensures AI-driven prosperity and “freedom of intelligence,” arguing: “AI could drive significant increased productivity over the next decade.”
• Both strongly advocate for investing in infrastructure, workforce development, and stricter export controls to maintain the U.S.’s AI leadership over competitors like China.

Google’s robotics upgrade

Google’s Gemini 2.0 Flash now natively handles images. For the first time, Gemini can edit photos via chat, and generally has a much better understanding of pictures and video than prior models. I made the below image with the prompt:

“Can you turn half of this photo of a headshot of a man in glasses into a hand-drawn style, like the image of the girl? The man should be split down the middle, with the left half being photorealistic with a black background, and the right half being a hand-drawn drawing of him in a minimal single-line style with black marker (his glasses, hair, shirt, and tie) on a white background. Keep everything on the left half the same.”

It took me a few tries to get exactly what I wanted.

Make email your superpower

Not all emails are created equal—so why does our inbox treat them all the same? Cora is the most human way to email, turning your inbox into a story so you can focus on what matters and getting stuff done instead of on managing your inbox. Cora drafts responses to emails you need to respond to and briefs the rest.

Try Cora today

Want to sponsor Every? Click here.

• Google Robotics is making use of Gemini's advanced multimodal understanding and massive context windows, creating robots capable of generalizing real-world tasks better than ever. The future in which embodied AI physically transforms our world is inching closer.

Optical character recognition (OCR) takes on new importance

• Mistral launched an OCR model claiming industry-leading accuracy for extracting information from PDFs, graphs, and handwritten documents—critical given an OCR market worth tens of billions, as well as troves of valuable data locked in PDFs.
• LlamaIndex quickly challenged Mistral’s claims with benchmarks suggesting its workflows perform better, highlighting an ongoing—and necessary—debate around accurate measurement that extends beyond just document understanding tasks.

Developer tools

• Anthropic released a text editor tool, simplifying iterative AI edits to files making it easier to build a “Cursor for X.”
• OpenAI introduced “new tools for building agents” in the form of updated APIs and SDKs. 

Why this matters (and what’s next):

OpenAI and Anthropic are continuing to pick their lanes, as I reported last week—Anthropic doubling down on safety and specialized applications like coding, while OpenAI aims for mass adoption and broad economic impact. There are three areas in which the companies agree, though:

• It is imperative the U.S. maintains global AI leadership against adversaries like China.
• Broad AI adoption will lead to economic prosperity, but investment in infrastructure and workforce development is a must.
• Very powerful AI is here, more is coming, and preparing for this is a priority for national security. These models are already powerful enough to change our world and are only getting better. 

Taken at face value, these statements are alarming. The only way to understand how true they are is by objectively measuring what these models can (or can’t!) do—and doing that often. Benchmarks!

This week's debate over Mistral’s OCR illustrates how valuable—and challenging—it is to measure AI performance. OCR isn't niche: It impacts critical sectors like finance, healthcare, and government, where reliably pulling information from documents directly adds value. Neither LlamaIndex nor Mistral began with OCR, though: Llamaindex started with tools to make coding with language models easier, while Mistral trained models. They both gravitated toward the space for a simple reason: It’s a huge revenue opportunity, and return on investment is key. Each team's ability to perform better on benchmarks now has very real stakes, and both are incentivized to publish their own benchmarks where they do best. 

A final note on benchmarks

Two great examples of creative approaches to testing model performance were released this week: 

• Jina AI shows that models get worse when you give them too long of a prompt. This is why no one has solved the problem of “chatting with your data”—if you put too much information into a prompt, performance starts to degrade.
• Anthropic adopted another game as a benchmark with Factorio.

But what will truly identify the best model for you is the benchmark you make yourself. Compare model responses on the same prompt in OpenAI’s brand new prompt playground and see which model you like best. Why? 

All benchmarks are undergirded with the belief that AI is transforming our digital world, but Google’s robotics advances hint at the next leap: bringing AI’s disruptive impact into the physical world. By using models that understand visual contexts directly (like Gemini 2.0 Flash), robots become not just smarter but significantly more adaptable to real-world environments. Check out some videos of these robots in action:

• “Dunking” a basketball
• Spelling with Scrabble tiles
• Folding origami

Google is seeing substantial improvements: Gemini Robotics more than doubled performance on its open-source benchmark of 400 multiple choice questions that rely on vision, spatial reasoning, and task reasoning, among other things. If digital AI has made knowledge and creativity dramatically cheaper, physical AI—robotics—might do something even more remarkable: lower the barriers to transforming our built environment.—Alex Duffy 

Knowledge base

"How Commodore Invented the Mass Market Computer" by Gareth Edwards/The Crazy Ones: Before Steve Jobs became a tech demigod, there was a man named Jack Tramiel. He was a Holocaust survivor who built Commodore into a computing empire through sheer force of will and ruthless business tactics. His “religion” of vertical integration and relentless cost-cutting created the Commodore 64 computer, which sold an astonishing 12 million units in 12 years and launched countless tech careers by making computers affordable for working-class families. Read this for a deep dive into how one man's obsession with control and affordability shaped our digital world. 🎧 Listen to an audio version of this essay on Spotify or Apple Podcasts.

"AI Is Breaking Media. Here's How Publishers Can Survive." by Evan Armstrong/Napkin Math: A new era of media is arriving faster than you can say "Hey ChatGPT." AI is simultaneously making content production dirt cheap while stealing attention and traffic from publishers. The future belongs to prolific creators who can build genuine relationships with audiences while using AI to create content at volumes that once required entire teams. Read this if you want to understand how media companies will survive in a world where both writing and coding approach zero cost.

🎧 "Prompt Your Way to Personal Growth" by Dan Shipper/AI & I: Executive coach Steve Schlafman uses AI as a personal therapist, dream interpreter, and self-awareness too. The result: he’s able to understand his mind better. Watch this episode of AI & I if you're curious about how AI can help you access parts of yourself you didn't know existed—even if you're a skeptic. The episode includes a live dream analysis and visualization that might just change how you think about introspection. 🎧 🖥 Watch on X or YouTube, or listen on Spotify or Apple Podcasts. 

"What Parenthood and AI Taught This Productivity Expert About Reclaiming His Time” by Scott Nover/Superorganizers: In 2020, Gaurav Vohra was revolutionizing email by ignoring it—well, only checking it twice a day. Fast forward to today: Superhuman’s Vohra says things are more complicated. He has a toddler and a startup advising business, but uses AI to help him reclaim his time. Read this if you want to understand why Vohra believes most people are drastically underestimating AI's ability to speed up their work.

“When an AI Tool Finally Gets You” by Lucas Crespo/Source Code: Most AI design tools treat creativity like a slot machine—input prompt, receive image, try again. But real design is messy, non-linear, and iterative. Flora, a new intelligent canvas, combines multiple generative capabilities in one workspace, letting designers explore ideas simultaneously rather than bouncing between different tools. Read this if you're interested in how thoughtful product design can make AI genuinely useful rather than just technically impressive.

Sentiment analysis

Gareth Edwards’s chronicle of Jack Tramiel and the history of Commodore triggered the memories of a flood of people who grew up with the company’s computers. Here are just a few:

“To a working-class kid trying to work my way through college, that $120 computer and an old $5 TV as a monitor opened a whole new world for me. Learned ML, then assembly, then eventually went on to get a CS degree and a tech career.” 

"My mother was a secretary and my father worked in a factory. There's no way we could afford an Apple II or an IBM PC. That Commodore 64, though—that was the best $100 they ever invested in my future."

“So much has been said about Steve Jobs and yet I almost never hear anything about Tramiel. Both were apparently absolutely ruthless businessmen, but I will always respect Tramiel’s ‘for the masses, not the classes’ approach to computers.” 

“So many Mountain Dew-soaked weekend nights dialing into local BBS's to chat with a stranger or download a game. We felt like true futuristic hackers in the late eighties on our Commodore 64.” 

“The first computer I ever encountered (aged 10) was a PET. I was hooked, and I absolutely loved my VIC-20. Amazing what could be done with just 3.5KB of RAM! Jack Tramiel changed my world.”

---

From Every Studio

Non-native English speaker? Unwrite, an app made by Every entrepreneur in residence Naveen Naidu, helps you write English confidently, clearly, and naturally—exactly like a native speaker. No awkward phrasing, no copy-pasting. Just real-time, personalized improvements wherever you type. We're looking for non-native English speakers to test it out. Sound like you? Sign up.

Unwrite. Source: Naveen Naidu.

Swipe through your newsletters. Every creative lead Lucas Crespo is experimenting with a new way to read newsletters without distractions. Swipe through, read, and click into an Every essay if you find it interesting. Feeling lucky? Draw a random Every essay from the bag. Swipe away and happy reading.

Cora pricing update and what we learned. This week, we announced a pricing change for our email management tool Cora. Running AI at scale to process over 3 million emails so far isn’t cheap—our AI works hard and it eats a lot (of compute, and money).

So we’d crafted a pricing update to support Cora’s growth—free for your first email account, and $20 per account after. We even tried to make you smile with our email, too (because nothing softens a price increase like a well-placed joke, right?). 

Then, we sent the email. And, well… yeah. We did not make you smile, that’s for sure. 

The reaction was clear from you all: This wasn’t the right move. And honestly? We hear you. If we were using a free tool we loved and suddenly got hit with a price tag, we’d be frustrated, too.

We went from, “Sweet, the price update is live!” to… “Ohhh. Ohh man,” and quiet, contemplative staring into the void. (Hello darkness, our old friend.)

But here’s what we realized. That so many of you cared to tell us why you were upset? That meant something. It meant Cora matters to you. And that’s something we take to heart, so, thank you.

Just like in jazz, there are no bad notes—it’s about what we do next. And we’re so grateful you’re here for our tunes. 

To make it right, we rolled back the pricing change. We’re:

• Giving all Cora users 3 months free while we work on making it even better
• Improving how Cora categorizes your emails and making sure it’s even more secure
• Promising to bring you into the discussion more in big decisions like these going forward

We’ll keep building, listening, and making Cora better—without any more crazy surprise solos. Thanks for jamming with us so far.—Vivian Meng

---

Alignment 

Likes don't matter. Performing for attention is soul-crushing work, which is why I wish I'd found Arthur Brooks's framework for creativity sooner. In the Atlantic, Brooks describes his three-question checklist before publishing: Does my work glorify the divine, does it uplift others, and is it personally interesting? I think this simple process prevents most people's worst creative impulses: publishing rage-bait for easy clicks, creating shallow content just to maintain visibility, sharing half-formed thoughts to seem relevant. I've since adapted it for my own checklist: Is it truthful? Does it serve beyond myself? Is it meaningful to me? The hierarchy matters—truth before service, service before self-interest. Rather than limiting creativity, I’ve found these boundaries focus it. I measure success on my terms, not on likes and shares. This way I become more authentically myself, creating from a place of integrity rather than performing like a circus animal for digital bananas.—Ashwin Sharma

Collaborative filtering

A year in AI. We know it can be hard to stay on top of the latest developments in AI given the pace at which it’s developing—especially for founders who want to build with the technology. So we’ve partnered with New York-based venture capital firm Company Ventures on AI Review, a year-long conversation series on the current and future state of AI designed to provide ongoing, actionable knowledge. On March 26, Dan Shipper will kick off the series with Company Ventures partner Matthew Harrigan for a talk by Google’s Steven Johnson, the editorial director of NotebookLM (who was a guest on AI & I last year). In addition to talks with people like Databricks chief AI scientist Jonathan Frankle, future events will include product demos days and hackathons. Learn more and register to attend the first event.

The past and future of Every. Dan sat down with Lightspeed Venture Partners’s Michael Mignano for an in-depth discussion of Every’s history and strategy on the Generative Now podcast. 🎧 🖥 Listen to or watch the episode.

---

Hallucination

What if you could choose your own algorithm?

Alex Duffy is the consulting lead and a staff writer at Every, where he writes about empowering people with AI tools and technology. You can follow him on X at @theheroshep and on LinkedIn, and Every on X at @every and on LinkedIn.

We build AI tools for readers like you. Automate repeat writing with Spiral. Organize files automatically with Sparkle. Write something great with Lex. Deliver yourself from email with Cora.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.