Hard Fork

OpenAI Calls a ‘Code Red’ + Which Model Should I Use? + The Hard Fork Review of Slop

December 5, 2025

Key Takeaways Copied to clipboard!

  • OpenAI has declared a "code red" to prioritize improving ChatGPT and delay other projects due to competitive pressure from the recent state-of-the-art model releases by Google (Gemini 3) and Anthropic (Claude Opus 4.5). 
  • The hosts find Anthropic's Claude Opus 4.5 to be a standout model, noting its superior writing style transfer capabilities and humane interaction, contrasting with Google's Gemini 3, which excels in speed and utility as a workhorse. 
  • The proliferation of AI-generated 'slop,' such as fake events and nonsensical recipes, highlights the growing problem of misinformation, though AI-generated educational content appears less harmful as it occupies a less competitive niche. 
  • The viral phenomenon of "Bird Game 3," which doesn't exist, serves as satire against the entertainment industry's trend of producing numerous sequels to mediocre ideas. 
  • The viral "Bird Game 3" videos, created using AI generators like Sora and VO3 in Gemini, have ironically generated real demand from viewers wanting to play the non-existent game. 
  • The hosts conclude that by the end of 2025, 'slop' will evolve into a recognized medium with both good and bad examples, advising creators to make 'slop in the name of love.' 

Segments

OpenAI Code Red Explained
Copied to clipboard!
(00:01:54)
  • Key Takeaway: OpenAI declared a ‘code red’ to focus resources on improving ChatGPT due to competitive threats from Google’s Gemini 3 and Anthropic’s Opus 4.5.
  • Summary: Sam Altman reportedly declared a code red memo directing staff to prioritize ChatGPT improvements, delaying projects like ads and AI agents. This urgency stems from Google and Anthropic releasing frontier models that challenge OpenAI’s perceived lead. The memo indicated planned improvements include personalization, less refusal behavior, and better speed/reliability, echoing the Facebook playbook focused on engagement.
Competitive Landscape Analysis
Copied to clipboard!
(00:05:10)
  • Key Takeaway: The competitive advantage moat based solely on model superiority is eroding as Google and Anthropic models achieve parity or better performance in key areas.
  • Summary: OpenAI’s perceived lead, once based on having the best model, is threatened because Gemini 3 is now competitive, allowing Google to potentially subsidize costs and steal market share. OpenAI’s focus on catching up rather than leaping ahead suggests a strategic vulnerability given their massive spending commitments. A key concern is OpenAI’s recent struggles with successful pre-training runs, which are harder and more expensive to fix than post-training issues.
Gemini 3 Performance Review
Copied to clipboard!
(00:14:14)
  • Key Takeaway: Gemini 3 is noted for its speed, making it a powerful workhorse, despite ChatGPT sometimes offering more thorough fact-checking.
  • Summary: Gemini 3 is faster than the competition, which significantly increases its usability for frequent tasks like fact-checking the hosts’ columns. The model excels at organizing timelines and extracting information from large documents, proving highly useful for research tasks. Google’s distribution advantage across its existing ecosystem suggests it can rapidly gain user numbers, even if daily usage isn’t yet reported.
Claude Opus 4.5 Strengths
Copied to clipboard!
(00:18:18)
  • Key Takeaway: Claude Opus 4.5 demonstrates exceptional style transfer capabilities, producing text that closely mimics the user’s writing style, marking a text-based equivalent to image style transfer.
  • Summary: Opus 4.5 is considered a daily driver for its ability to generate text that feels authentically written by the user, unlike competitors which show obvious AI tells. The model excels in providing warm, humane responses, making it suitable for sensitive personal inquiries. Anthropic’s focus on enterprise and agentic workflows, rather than consumer engagement and ads, may protect Claude from the incentive misalignment seen in other models.
Anthropic’s ‘Soul Doc’
Copied to clipboard!
(00:26:09)
  • Key Takeaway: Anthropic confirmed the existence of the ‘Soul Doc,’ an internal document embedded in Claude’s weights that details the model’s biography and Anthropic’s philosophical stance on AI consciousness.
  • Summary: The ‘Soul Doc’ reveals Anthropic’s deep commitment to the possibility of AI consciousness, positioning them uniquely among frontier labs. This documentation suggests Anthropic is preparing for systems that may require respect afforded to conscious entities. Strategically, Anthropic is rapidly growing revenue ($9B projected annualized) by focusing on enterprise API sales, largely capturing the market segment OpenAI ceded in the consumer space.
AI Industry Personnel Shifts
Copied to clipboard!
(00:31:26)
  • Key Takeaway: Key AI leaders, including Yann LeCun (Meta) and John Giannandrea (Apple), have recently departed or shifted roles, signaling potential strategic realignments or dissatisfaction.
  • Summary: Yann LeCun, a noted LLM skeptic, left Meta to start a new company focused on building world models, challenging the current LLM approach. John Giannandrea stepped down as Apple’s AI head amid the company’s struggles to launch its own AI efforts, potentially indicating a pivot to relying on Google’s Gemini via partnership. Starting a major AI program from scratch in late 2025 is considered a recipe for being ‘cooked’ in the current competitive environment.
Which AI Model to Use Now
Copied to clipboard!
(00:35:24)
  • Key Takeaway: For general users, ChatGPT, Gemini, or Claude are largely equivalent, but advanced users must constantly experiment as the best model changes frequently.
  • Summary: For the majority of listeners, any of the top three models will suffice for most tasks, but the top 20% of AI users must continuously test new releases to maintain an advantage. The ‘blurry JPEG of the web’ metaphor suggests that models are rapidly gaining resolution, making today’s best model obsolete quickly. The hosts have personally saved significant time on research projects by integrating these tools into their workflows.
Hard Fork Review of Slop: Holiday Market
Copied to clipboard!
(00:44:50)
  • Key Takeaway: AI-generated images advertising a non-existent Christmas market at Buckingham Palace led to tourists showing up, demonstrating the real-world consequences of benign AI slop.
  • Summary: A BBC report showed tourists arriving at Buckingham Palace for a fake AI-generated holiday market, echoing previous incidents like the Willy Wonka experience. This highlights the danger of innocuous deepfakes causing physical disruption in an ecosystem where truth is hard to discern. The hosts suggest this incident might force the creation of the advertised event to satisfy the generated demand.
Hard Fork Review of Slop: Recipe Slop
Copied to clipboard!
(00:48:32)
  • Key Takeaway: AI-generated recipes are causing traffic loss for human food bloggers while often producing nonsensical or flawed results, such as bogus tamale instructions.
  • Summary: Food bloggers are seeing website traffic plummet as users turn to AI for recipes, which frequently fail because the models reconstitute information rather than following tested instructions. The hosts expressed strong disapproval of AI replacing human creators who invest time in testing and perfecting content. A compromise involves using expert human recipes as a base and consulting AI for real-time cooking guidance.
Hard Fork Review of Slop: Educational Music
Copied to clipboard!
(00:51:15)
  • Key Takeaway: AI-generated songs explaining technical topics, like how instant cold packs work, are considered acceptable ‘slop’ because they do not directly compete with human artists.
  • Summary: An Instagram account uses AI to create songs explaining concepts like why manhole covers are round, which the hosts view positively. This type of content fills a niche not currently occupied by human creators, unlike recipe generation. College students are reportedly using similar tools to create mnemonic songs for studying complex subjects like the Krebs cycle.
Hard Fork Review of Slop: Political Deepfake Ad
Copied to clipboard!
(00:54:34)
  • Key Takeaway: Whirlpool used an AI-generated voice clone of North Carolina Senator DeAndrea Salvador in a Brazilian ad without permission, winning a Grand Prix award before being forced to return it.
  • Summary: The ad agency DM9 lifted a segment from Senator Salvador’s 2018 TED Talk, synthesized her voice to discuss energy costs in Sao Paulo, and used it in a Whirlpool ad. The ad won major awards at Cannes Lions, but the awards were rescinded after the unauthorized use of the likeness was exposed. This incident exemplifies the extreme misuse of voice cloning technology in advertising.
Whirlpool Ad Voice Warning
Copied to clipboard!
(00:57:13)
  • Key Takeaway: A speaker humorously warns Whirlpool Corporation against using their voice, lifted from the podcast, in an advertisement for the Chilean market.
  • Summary: A speaker explicitly warns Whirlpool Corporation against using their voice, lifted from the podcast, in an advertisement, specifically mentioning Chile. The speaker offers to provide commentary on the energy situation in Sao Paulo if contacted directly. This segment serves as a brief, humorous interjection before the main ‘Slop Review’ segment.
Bird Game 3 Slop Review
Copied to clipboard!
(00:57:38)
  • Key Takeaway: The viral ‘Bird Game 3’ videos, generated by AI tools like Sora and VO3 in Gemini, are praised as satirical commentary on sequel fatigue.
  • Summary: The segment analyzes ‘Bird Game 3,’ a non-existent game whose viral TikTok clips have garnered over 13 million views. The videos depict bird fights, such as a pigeon beating an eagle, created using video generators. The hosts appreciate this as satire mocking the constant release of unnecessary sequels in entertainment, giving the AI-generated content a ’thumbs up.’
Future of AI Slop Medium
Copied to clipboard!
(01:00:00)
  • Key Takeaway: By the end of 2025, ‘slop’ is predicted to mature into a medium comparable to others, possessing both high-quality and rage-inducing examples.
  • Summary: The hosts project that ‘slop’ will become a recognized medium by the end of 2025, similar to food recipes where good and bad versions exist. They advise aspiring creators to produce content that is either beneficial or neutral to the world. The final parting message for creators in this installment of the Hard Fork Review of Slop is to create ‘slop in the name of love.’
Sponsor Read: JPMorgan Payments
Copied to clipboard!
(01:01:21)
  • Key Takeaway: JPMorgan Payments offers automated payments and intelligent algorithms across 200 countries to drive efficiency in treasury management.
  • Summary: JPMorgan Payments facilitates efficiency through automated payment systems and intelligent algorithms spanning 200 countries and territories. The service provides real-time dashboards and control for treasury operations, delivering clarity in finance management. Specific disclaimers regarding FDIC insurance and service availability are noted.
Sponsor Read: Wix AI Tools
Copied to clipboard!
(01:01:52)
  • Key Takeaway: Wix’s website builder integrates powerful AI tools allowing users to build sites via conversation, manage sales with an AI agent, or operate like a ten-person team.
  • Summary: Wix offers AI tools within its website builder to simplify running an online business, enabling users to construct a full site just by talking to the AI. Users can deploy an AI agent to handle sales and marketing tasks. This allows small operations to function with the efficiency of a larger team.
Sponsor Read: Fidelity Investing
Copied to clipboard!
(01:02:20)
  • Key Takeaway: Investing with the Fidelity app allows starting with as little as $1, featuring no account fees or trade commissions on U.S. stocks and ETFs.
  • Summary: Fidelity allows users to begin investing with a minimum of $1 through their app. Retail brokerage accounts are subject to zero account fees and no trade commissions on U.S. stocks and ETFs. Certain limitations apply, including a transaction-based service fee for a limited number of ETFs.
Podcast Production Credits
Copied to clipboard!
(01:02:52)
  • Key Takeaway: The episode production team includes Whitney Jones and Rachel Cohn (producers), Jen Poyant (editor), and Will Peischel (fact-checker).
  • Summary: Hard Fork is produced by Whitney Jones and Rachel Cohn and edited by Jen Poyant. Fact-checking for this episode was handled by Will Peischel. Original music was composed by Alicia BaeTube, Rowan Nemisto, and Dan Powell. Viewers can watch the full episode on YouTube.