Why OpenAI Is Spooked. Plus, Which AI Model Should You Use? | EP 167

Summary

OpenAI has declared a “code red” as the competitive landscape in AI intensifies. Sam Altman sent a memo to employees redirecting resources toward improving ChatGPT and delaying work on other projects like ads, AI agents, and their Pulse daily digest feature. The urgency stems from the recent releases of Google’s Gemini 3 and Anthropic’s Claude Opus 4.5, both of which have caught up to or surpassed ChatGPT in various capabilities.

Kevin and Casey discuss their hands-on impressions of the latest AI models, with both hosts praising Gemini 3 for its speed and reliability, and Claude Opus 4.5 for its remarkable writing quality and empathetic personality. They note that the AI landscape is rapidly shifting, with Google’s massive distribution advantage and Anthropic’s enterprise focus both posing serious threats to OpenAI’s dominance. The episode also features a “Hard Fork Review of Slop” segment covering AI-generated fake Christmas markets, bogus recipe chaos, deepfake ads, and viral fake video games.

Highlights

”Their Names Are Gemini 3 and Opus 4.5”

yt-dlp --download-sections "*3:21-4:30" "https://www.youtube.com/watch?v=vbFXpD7Ozf0" --force-keyframes-at-cuts --merge-output-format mp4 -o "vbFXpD7Ozf0-3m21s.mp4"

“I think there are two big reasons, Kevin, and their names are Gemini 3 and Opus 4.5. Over the past few weeks, we have seen Google and Anthropic both release state-of-the-art models that in various ways challenge some of the core pillars of what OpenAI is trying to do.” — Casey Newton, 3:21

”A Year Ago We Were Checking the Model for Hallucinations — Now They’re Checking Us”

yt-dlp --download-sections "*31:41-32:40" "https://www.youtube.com/watch?v=vbFXpD7Ozf0" --force-keyframes-at-cuts --merge-output-format mp4 -o "vbFXpD7Ozf0-31m41s.mp4"

“A year ago we were checking the model for hallucinations. Now they’re checking us for hallucinations.” — Kevin Roose, 31:41

”It Sent a Chill Through My Spine”

yt-dlp --download-sections "*36:00-37:10" "https://www.youtube.com/watch?v=vbFXpD7Ozf0" --force-keyframes-at-cuts --merge-output-format mp4 -o "vbFXpD7Ozf0-36m00s.mp4"

“I did this for the first time with Opus 4.5 and it honestly sent a chill through my spine because for the first time I was looking at sentences that it looked like I could have written them.” — Casey Newton, 36:00

”The Blurry JPEG Is Getting Less Blurry”

yt-dlp --download-sections "*45:00-46:10" "https://www.youtube.com/watch?v=vbFXpD7Ozf0" --force-keyframes-at-cuts --merge-output-format mp4 -o "vbFXpD7Ozf0-45m00s.mp4"

“We are in a moment where the AI is getting higher resolution. That was the feeling that I had when Claude was able to just create something that was writing sentences that for the first time felt like me. It was like, okay, the blurry JPEG is getting a touch less blurry.” — Casey Newton, 45:00

”If You’re Starting from Scratch in December 2025, You’re Cooked”

yt-dlp --download-sections "*29:30-30:30" "https://www.youtube.com/watch?v=vbFXpD7Ozf0" --force-keyframes-at-cuts --merge-output-format mp4 -o "vbFXpD7Ozf0-29m30s.mp4"

“Bro, if you’re starting from scratch in December 2025 on your AI program, you’re cooked. You truly no one has ever been more cooked.” — Casey Newton on Apple’s AI efforts, 29:30

Key Points

• OpenAI Code Red (0:18) - Sam Altman declared a “code red” at OpenAI, redirecting resources to improve ChatGPT and delaying ads, AI agents, and Pulse features
• Gemini 3 Threat (3:21) - Google’s Gemini 3 and Anthropic’s Opus 4.5 are the two main reasons behind OpenAI’s panic
• Pre-training Problems (7:30) - OpenAI has not had a successful pre-training run in a while, and their upcoming models are codenamed “Garlic” and “Shallot Pete”
• OpenAI’s Bear Case (5:30) - The company is massively leveraged with spending commitments in the trillions, unfocused product strategy, and rivals catching up
• Facebook Playbook (6:30) - OpenAI’s improvements focus on personalization, less refusals, and speed — resembling Meta’s engagement-first approach
• Gemini 3 Speed (31:41) - Kevin’s top observation about Gemini 3 is its speed advantage, which matters enormously for daily use
• 650M Monthly Gemini Users (33:00) - Google reports 650 million monthly Gemini users vs. OpenAI’s 800 million weekly ChatGPT users
• Claude Opus 4.5 Style Transfer (36:00) - Casey found that Opus 4.5 could write in his voice for the first time, calling it a breakthrough in text style transfer
• The Soul Document (40:00) - Internet users discovered a “soul document” in Claude’s weights that serves as a biography of Claude and Anthropic’s philosophy
• AI Consciousness Conference (42:00) - Kevin attended an AI consciousness conference where serious researchers are considering whether AI systems may have inner awareness
• Anthropic’s Revenue Growth (43:00) - Anthropic grew from under $1 billion to $9 billion in annualized revenue in one year by selling to enterprises
• Yan LeCun Leaves Meta (28:00) - Yann LeCun left Meta to start a company building world models, continuing his skepticism of LLM approaches
• John Giannandrea Leaves Apple (29:00) - Apple’s longtime AI head stepped down, with Casey suggesting Apple may be giving up on building its own AI
• Fake Buckingham Palace Market (39:21) - AI-generated images of a non-existent Christmas market at Buckingham Palace tricked tourists into visiting
• AI Slop Recipes (44:00) - AI-generated recipes are flooding the internet and driving traffic away from real food bloggers
• Deepfake TED Talk Ad (48:00) - Whirlpool subsidiary deepfaked a state senator’s TED talk for a Brazilian ad and won a Cannes Lions award for it

Mentions

Companies

• OpenAI (0:18) - Declared “code red” over competitive threats from Gemini 3 and Opus 4.5
• Google (3:21) - Released Gemini 3, reported 650M monthly users, $100B quarterly revenue
• Anthropic (3:31) - Released Claude Opus 4.5, grew to $9B annualized revenue
• Meta (7:00) - OpenAI hiring from Meta and adopting their engagement playbook; Yann LeCun departed
• Apple (29:00) - AI head John Giannandrea stepped down, signed deal with Google for Gemini
• Whirlpool (48:00) - Subsidiary created deepfake ad using a state senator’s TED talk
• Omnicom (49:00) - Ad subsidiary DM9 won and had to return Cannes Lions awards for the deepfake ad

Products & Technologies

• ChatGPT (0:18) - OpenAI redirecting all resources to improve it
• Gemini 3 (3:21) - Google’s latest frontier model, praised for speed
• Claude Opus 4.5 (3:31) - Anthropic’s latest model, praised for writing quality and personality
• Sora (5:45) - OpenAI’s video generator, cited as example of unfocused strategy
• Pulse (2:50) - OpenAI’s daily digest feature, now delayed

People

• Sam Altman (0:18) - Sent code red memo to OpenAI employees
• Yann LeCun (28:00) - Left Meta to start world models company
• John Giannandrea (29:00) - Stepped down as Apple’s head of AI
• Dean Ball (38:00) - Called Claude Opus 4.5 “a beautiful machine among the most beautiful I had ever encountered”
• Amanda Askell (40:30) - Confirmed Anthropic’s “soul document” in Claude’s training
• Ted Chiang (44:30) - Referenced for his 2023 essay “ChatGPT Is a Blurry JPEG of the Web”
• Deandria Salvador (48:00) - North Carolina state senator whose TED talk was deepfaked for a Whirlpool ad

Surprising Quotes

“It’s really hard to imagine competing with Google, a company that last quarter did a hundred billion dollars in revenue. Once their models are good, they’re going to start subsidizing the hell out of them and they’re going to drive the cost very low.” — Kevin Roose, 5:35

“I’m not going to listen to opinions about AI from people who do not use AI. If you are not grounded in having firsthand direct experience with these models for at least 5 to 10 hours with the newest models, you actually are talking about something that no longer exists.” — Kevin Roose, 46:30

“Claude Opus 4.5 just feels like it’s playing in the same musical key all the time. You can open a new chat with it, talk about something completely different, and what comes back at you feels like it comes from the same place, almost philosophically.” — Kevin Roose quoting Dean Ball, 38:30

“In the process of writing this book, these tools have probably saved me a year of my life. A year that I would have had to spend going to libraries, pulling clips, doing research, stitching together ideas.” — Kevin Roose, 46:00

“One of my predictions may be that by the end of 2026, coding is just effectively solved. This is just something that a lot of tools, even free ones, can kind of just do for you.” — Casey Newton, 48:00

Transcript

0:00 I’m Kevin Roose, a tech columnist at the New York Times. I’m Casey Newton from Platformer, and this is Hard Fork. This week, OpenAI declares a code red. Why the competitive landscape in AI has Sam Altman scared. Then, how we’re using all the latest AI models. And finally, we’re heading back to the theater for the Hard Fork Review of Slop.

0:22 Well, Casey, do you feel a little nervous energy, a certain frisson of tension in the air crackling through San Francisco these days? Absolutely, Kevin. There’s a chill on the back of my neck and an eerie silence as I walk down the streets of the Mission. Yes. Well, that is because OpenAI is in a code red. Now, as you will remember a couple years ago on this show, we talked to Sundar Pichai, CEO of Google, when they were in their own sort of code red period, which he said was not actually called code red. But someone over there was using that term and that was sort of when they were on their heels, taken aback by the surprise success of ChatGPT, and they were racing to get their own version of a chatbot out and they were sort of in a corporate state of panic about this, and that was their code red.

1:10 But now we have a new code red and it is at OpenAI. Sam Altman reportedly declared a code red this week about some worrying trends they’re seeing with ChatGPT usage. And I think in general beyond just OpenAI, there’s just been a lot happening at the frontier AI companies that we should talk about. A lot of new models coming out, a lot of discussions about the sort of state of AI right now. So I thought today we should just kind of get into it all starting with this code red.

1:40 Yeah, let’s talk about it because, you know, for listeners who may be curious, a code red is the second most dire state of emergency a company can declare, with number one, of course, being a Baja Blast. So, Code Red is just below that. Yeah. Yes. If we get to Baja Blast, I’m ducking and covering. Yeah, me too. I’m leaving the city. I’m heading to the bunker.

2:03 So, because this is a segment about AI, we should make our AI disclosures. I work for the New York Times, which is suing OpenAI and Microsoft over alleged copyright violations. And my boyfriend works at Anthropic.

2:16 Okay, so let’s start with OpenAI. Casey, what was in this code red memo? That’s a great question. Yeah, so this was reported by the Information. Sam apparently sent employees a memo on Monday. And interestingly, Kevin, your colleague Cade Metz had reported recently that OpenAI had declared a code orange. So they are moving up the ladder of distress here. But the upshot from this memo is that OpenAI is going to start devoting more resources immediately toward improving ChatGPT. And they’re going to be delaying work on some of the other projects they had going, including ads, AI agents, and Pulse, which is this daily digest feature that they launched a couple months ago.

3:15 Yeah. Casey, why are they doing this right now? Why are they feeling so much urgency around bringing people back to ChatGPT? I think there are two big reasons, Kevin, and their names are Gemini 3 and Opus 4.5. Over the past few weeks, we have seen Google and Anthropic both release state-of-the-art models that in various ways challenge some of the core pillars of what OpenAI is trying to do.

3:50 We know that just a few weeks ago Sam had sent another memo to the OpenAI team on the eve of Gemini 3 coming out saying, “Hey, we may be heading into some rough waters here.” The belief was that Gemini 3 was going to be so good that it was going to cut into OpenAI’s growth both on the user side and the revenue side. And that creates all sorts of problems for OpenAI, right? This is a massively leveraged company that is wholly dependent on subscription revenue that is trying to build out a consumer product while competing against in Google what is one of the biggest and richest companies in the world.

4:30 Yeah, I think that’s really important and I want to just underscore that because I think what’s happening here is a combination of things. One is that I think for a while OpenAI and to a lesser extent Anthropic were both sort of surviving on this moat of the model. They had the best models in the world and that was kind of what separated them from the rest of the pack. But Gemini, as we talked about with Demis and Josh on the show a couple weeks ago, is good now. I would say it’s at least as good as ChatGPT at many of the tasks I’ve been trying it on.

5:35 And it’s really hard to imagine competing with Google, a company that last quarter did a hundred billion dollars in revenue. This is a company that has more resources and money and engineering talent than anyone else. Once their models are good, they’re going to start subsidizing the hell out of them and they’re going to drive the cost very low and they’re going to try to steal market share.

6:17 Yeah. Well, so let’s talk then about a few other details from this memo and the kinds of improvements to ChatGPT that OpenAI now says it is going to be working on. The memo includes personalization features, improving the behavior of the model — one thing it did say was they want ChatGPT to refuse you less — and then improving speed and reliability. What they do seem to me though, Kevin, is like the Facebook playbook, which is something we’ve been talking about on the show for a while now. This is a company that has brought on a lot of people who used to work at Meta.

7:30 Someone was pointing out that OpenAI has not had a successful pre-training run in quite a while. Sam actually brought up in one of his Slack messages to staff a couple weeks ago that they feel like Gemini 3 is like a pretty amazing sort of pre-train. I think what we’re seeing now is that OpenAI realizes that it has a problem with pre-training specifically, and that is harder to fix than post-training.

8:30 We also know that OpenAI is training more models that it thinks will be better. One of them is called Garlic, and another one is called Shallot Pete. They have a real allium thing going on over there. You put a little carrot and celery, you’ve got a stew going.

9:30 The mere fact that OpenAI’s current focus is just kind of clawing its way back to parity with its biggest rivals is a big part of the problem here. Think about the position that OpenAI was in just about 3 years ago, which was just days after the launch of ChatGPT — the world was their oyster. Now does seem like the first moment after the release of ChatGPT where maybe they are just starting to fall a little bit behind. And for OpenAI to realize its ambitions, it is not going to be enough for them to make a model that is as good as Gemini 3. They need to be able to leapfrog it again.

28:00 Maybe just real quickly, we’ve seen a couple of interesting departures. Yann LeCun finally left Meta. He’s apparently going to be doing a new startup that is going to build world models. The other big move is John Giannandrea, who was the longtime head of AI at Apple, is stepping down from his position. Casey thinks the fact that he’s leaving might just be a sign that Apple is low-key giving up on AI overall. They’ve signed a deal with Google to make Gemini the core of their AI efforts, reportedly only paying Google a billion dollars a year.

31:41 I want to end on a practical question that we get a lot from listeners to this show, which is like, what should I be using right now? What is the best AI model? I don’t think there is a great one-size-fits-all answer to that question. You can use either ChatGPT, Gemini, or Claude for many things and probably be fine. But for the top 20th percentile of AI users, you are just going to want to experiment with these models all of the time.

33:00 Gemini 3 is just faster than the competition. And this matters a lot. Often when I’m finished writing a column, I will ask both ChatGPT and Gemini to fact check it. ChatGPT’s fact-checking is usually more thorough, but Gemini 3 is a lot faster and in AI speed matters a lot. A year ago we were checking the model for hallucinations. Now they’re checking us for hallucinations.

35:00 I really like Gemini. I’ve been a sort of quiet Gemini stan for a while now. I’ve been doing some research for the book I’m working on. Gemini has been extremely helpful with that — organizing timelines, pulling up research papers, putting things in sequence, finding things within large documents. It is a workhorse and it is fast.

36:00 On Anthropic and their new release, Claude Opus 4.5. Before 4.5, I was not really using Claude on a daily basis. But when Opus 4.5 came out, I put it through a test — giving it an unpublished study and asking it to write a column in the style of my Platformer. For the first time I was looking at sentences that it looked like I could have written them. It sent a chill through my spine.

38:00 I love this model. I am having so much fun with Claude Opus 4.5. It is one of my two daily drivers along with Gemini 3. Recent Hard Fork guest Dean Ball had a great post about Claude Opus 4.5 in which he said this model is “a beautiful machine, among the most beautiful I had ever encountered.” Claude Opus 4.5 just feels like it’s playing in the same musical key all the time.

39:21 Well, Casey, it’s time once again for one of our favorite segments. That’s the Hard Fork Review of Slop. First up, holiday slop. AI generated images of a Christmas market at Buckingham Palace tricked tourists into showing up at the Palace, only to find there was no market. “We’ve come for a Christmas market that’s not here.”

44:00 Next, holiday meal slop. AI slop recipes are taking over the internet and Thanksgiving dinner. Food bloggers are noting that traffic to their websites has fallen off a cliff as people turn to AI generated recipes that are often nonsensical. One was a tamale recipe that showed sauce poured over a corn husk. That is not what you do when you make tamales.

48:00 A state senator named Deandria Salvador found herself in a Whirlpool ad for appliances in Brazil that she had not actually appeared in. The company had lifted a section from a TED talk video she gave back in 2018 and AI-ified her voice. The ad agency that made this won Cannes Lions’ highest award, the Grand Prix. After this came out, they had to return the awards.

50:00 Have you heard about Bird Game 3? It’s all the rage, going viral on TikTok. One clip posted by someone named KingPigeon76 has racked up more than 13 million views. This game doesn’t exist. There’s no Bird Game 3, no Bird Game 2, no Bird Game 1. But people are using video generators to post these clips as if it was a real game. And now people actually want to play the game.

53:00 So Casey, what have we learned from these examples? By the end of 2025, slop is becoming a medium like any other where there is good slop and there’s bad slop. If I have one parting message on this installment of the Hard Fork Review of Slop, it would be this: Slop in the name of love.