Meta Goes MAGA Mode +A Big Month in A.I. + HatGPT

Episode Summary

Executive Summary: Hard Fork returns from break by focusing on Meta’s dramatic rightward shift in speech policy, then surveys a packed AI news cycle featuring OpenAI’s O3 reasoning model, Sam Altman’s renewed AGI claims, Google Gemini 2.0, and China’s DeepSeek. The episode closes with HatchGPT on viral tech scandals, from AI personas to Honey, Siri privacy, Netflix live events, and Waymo weirdness.

Main Topics: Meta’s policy overhaul and rightward realignment (Priority: 5/5): The hosts analyze Meta’s staffing changes, board additions, and content moderation overhaul as a calculated bid to curry favor with the incoming Trump administration, emphasizing the end of fact-checking and reduced proactive moderation. The risks of reduced moderation on Meta platforms (Priority: 5/5): They argue Meta’s move from automated enforcement to user-reported violations will likely increase harassment, misinformation, and real-world harm, especially in groups and for vulnerable users like LGBTQ teens. OpenAI O3 and the accelerating AI capability race (Priority: 5/5): The discussion highlights O3’s strong benchmark results, especially on ArcPrize and coding benchmarks, as evidence that reasoning models may have broken through prior scaling limits. Sam Altman, AGI, and the 2025 AI narrative (Priority: 4/5): They unpack Altman’s claim that OpenAI knows how to build AGI, suggesting the industry is converging on the idea of a virtual AI coworker as the practical definition of AGI. Google Gemini 2.0 and DeepSeek’s efficiency breakthrough (Priority: 4/5): Google is framed as catching up in reasoning and research features, while DeepSeek signals that China may be closer to frontier models than U.S. chip-export policy assumed, using far less compute and cost. HatchGPT: platform scandals and the changing internet (Priority: 3/5): The segment covers Meta’s AI personas, Apple Siri’s privacy settlement, Netflix/WWE’s live-streaming ambitions, the Honey affiliate scandal, and a Waymo car circling a rider in a parking lot.

Key Arguments: Meta’s policy changes are not minor tuning but a wholesale acceptance of Republican/right-wing critiques of content moderation. Removing proactive moderation and relying on user reporting will leave abusive, hateful, and conspiratorial content unchecked in the spaces where it is most dangerous. Fact-checking mattered for harm reduction even if it touched only a small fraction of posts, because the content it did reach shaped beliefs at scale. OpenAI’s O3 suggests reasoning models may have found a new scaling path through test-time compute rather than pretraining alone. The industry’s near-term definition of AGI may become a useful AI employee that can do real work, especially software engineering. DeepSeek shows that frontier-like AI systems may be achievable with far less money and less advanced chips than U.S. policy makers assumed. Many of the most important tech changes are driven by commercial incentives, but the political and cultural consequences can be severe and lasting.

Data Points: Meta fact-checking end date: March - Meta told fact-checking partners their contracts will end in March, after which fact-checking on Meta properties ends. OpenAI O3 ArcPrize score (limited compute): 75.7% - O3 scored this on an evaluation where compute spending was capped at $10,000. OpenAI O3 ArcPrize score (unlimited compute): 87.5% - O3 reached this score when allowed to spend far more on computing, reportedly over $1 million. GPT-3 ArcPrize score: 0% - Referenced as the 2020-era baseline on the Arc AGI-1 public training set. GPT-4o ArcPrize score: 5% - Referenced as last year’s result before O3’s jump. O3 Code Forces rating: 2727 - Presented as roughly equivalent to the 179th-best human competitive coder. Number of states suing Meta over child safety: 41 states and D.C. - Cited in the discussion of how the new moderation changes may increase Meta’s liability. DeepSeek V3 training cost: $5.5 million - Reported estimated cost for training DeepSeek’s new model. DeepSeek V3 parameter count: 680+ billion - Used to compare the model’s size with Meta’s Llama series. Meta Llama model parameter count: 405 billion - Described as the previous open-model benchmark DeepSeek surpassed in size. Apple Siri privacy settlement: $95 million - Apple agreed to settle a long-running Siri privacy lawsuit. Siri settlement payout per device: Up to $20 per device - Claimants could submit for up to five devices. Netflix/WWE deal value: $5 billion - The decade-long agreement for WWE Raw streaming on Netflix. Number of homes Netflix can reach via WWE: ~280 million homes - Estimated reach from the WWE partnership.

Pivotal Quotes: "We are going to stop using them, and we’re going to go back to human beings who don’t even work for us or have any training or expertise." — Casey Newton: Describing Meta’s shift away from proactive automated moderation toward user-reported enforcement "If you could hire an AI as a remote employee to be a great software engineer, I think a lot of people would say, OK, that’s AGI-ish." — Sam Altman (quoted by Casey Newton): Explaining the practical direction OpenAI appears to be taking in defining AGI "People do not want to spend time on networks that are full of violence and harassment and abuse and gore and porn." — Casey Newton: Arguing that moderation is also commercially necessary, not just politically contested

Implications: Listeners should expect more polarized social platforms, weaker moderation on Meta, and a faster AI race centered on agentic workers and reasoning models. The episode suggests 2025 could bring major product shifts, new safety risks, and deeper scrutiny of tech companies’ incentives and political alignments.