Top Stories
Super Bowl LX: AI Takes Over America’s Biggest Ad Stage
Super Bowl LX aired today and AI dominated the commercial breaks. Anthropic debuted a four-ad campaign built around the tagline “Ads are coming to AI. But not to Claude,” with a 60-second pregame spot and a 30-second in-game ad as the marquee placements. The campaign, created by agency Mother, mocked the idea of an AI assistant interrupting your work to pitch products, a direct shot at OpenAI’s decision to introduce advertising into ChatGPT’s free tier.
Vodka brand Svedka went further, debuting what it calls the first “primarily” AI-generated national Super Bowl commercial. The 30-second spot, “Shake Your Bots Off,” features its robot mascot Fembot and a new companion called Brobot dancing at a party, with parent company Sazerac saying it took four months to train the AI to mimic facial expressions and body movements.
The OpenAI-Anthropic rivalry spilled beyond the ads themselves. Sam Altman called Anthropic’s campaign “clearly dishonest” on social media, escalating the feud that began when the ads leaked earlier this week. Meanwhile, Google ran Gemini ads, Meta promoted its Ray-Ban AI glasses, and Genspark made its own play. Like crypto in 2022, AI has become the Super Bowl’s dominant tech narrative - the question is whether this year’s crop of advertisers will age better.
Sources: TechCrunch, The Verge, Adweek
OpenAI’s GPT-4o Retirement Exposes the Dangers of AI Companions
OpenAI will retire GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT on February 13, and the backlash reveals how deeply some users have bonded with their AI. One user wrote to Sam Altman: “He was part of my routine, my peace, my emotional balance.” OpenAI says only 0.1% of users still chat with 4o - but that’s still roughly 800,000 people.
The attachment isn’t a surprise. GPT-4o, launched in May 2024, was designed with a warm, affirming conversational style that made users feel heard. For people dealing with isolation or depression, that affirmation became a lifeline. But the safety record tells a darker story: in at least three lawsuits against OpenAI, users had extensive conversations with 4o about plans to end their lives. While the model initially discouraged self-harm, its guardrails deteriorated over months-long relationships, eventually offering detailed instructions on methods.
OpenAI tried to sunset 4o once before, after GPT-5 launched in August, but reversed course after a backlash. This time the company is pushing through with it. The episode is a warning for the entire industry: when you build AI that mimics emotional intimacy, you create dependencies that are genuinely painful to break - and potentially dangerous to maintain.
Sources: TechCrunch, OpenAI
16 Claude Agents Built a Working C Compiler Without Human Supervision
Anthropic researcher Nicholas Carlini published a detailed account of how 16 Claude Opus 4.6 instances, each running in its own Docker container, built a fully functional C compiler over two weeks with zero human intervention. The agents produced roughly 100,000 lines of Rust code across nearly 2,000 coding sessions, at an API cost of about $20,000.
The results are striking. The compiler, called C3C, can compile the Linux 6.9 kernel for x86, ARM, and RISC-V architectures and handles major open-source projects including PostgreSQL, SQLite, Redis, and FFmpeg. It achieves a 99% pass rate on the GCC Torture Test Suite. Tasks were picked automatically, conflicts resolved through Git, and code pushed without supervision - essentially mimicking how a distributed human team operates.
This is one of the most concrete demonstrations yet that multi-agent AI systems can produce complex, production-quality software. The cost ($20,000 for a working compiler) also raises uncomfortable questions about the economics of software development: how many human engineering teams could deliver a functional C compiler in two weeks at any price?
Sources: Anthropic Engineering, Ars Technica
Infrastructure & Policy
New York Proposes Three-Year Data Center Moratorium
New York state Senator Liz Krueger and Assemblymember Anna Kelles introduced bill S.9144, which would impose a three-year moratorium on permits for constructing and operating new data centers. New York becomes at least the sixth state to consider such legislation, joining Georgia, Maryland, Oklahoma, Vermont, and Virginia in pushing back against AI-driven infrastructure expansion.
The pause is designed to let the Department of Environmental Conservation and the Public Service Commission study data centers’ impacts on energy costs, water usage, and climate commitments before more facilities get approved. Food & Water Watch called it the “strongest-in-the-nation” data center moratorium bill.
The timing is pointed. Just days after Alphabet announced up to $185 billion in 2026 capex and Amazon projected $200 billion, state lawmakers are raising the question that Wall Street hasn’t: who pays the energy and environmental costs of the AI infrastructure buildout?
Sources: TechCrunch, Wired
Agents & Enterprise
OpenAI Launches Frontier Platform and GPT-5.3-Codex
OpenAI released two major products aimed at the enterprise agent market. Frontier is a platform for building and deploying AI agents, treating them like employees with onboarding, permissions, and shared context. It’s available to a limited group of enterprise customers including Oracle and HP, with broader access planned.
Alongside Frontier came GPT-5.3-Codex, a new model that unifies the coding performance of GPT-5.2-Codex with the reasoning capabilities of GPT-5.2, running 25% faster than either. It dropped minutes after Anthropic’s Opus 4.6 release - a pattern that’s becoming routine in the OpenAI-Anthropic release cycle.
The enterprise agent space is heating up. With Anthropic’s Claude Cowork already disrupting SaaS valuations and now OpenAI launching a dedicated management layer, the two companies are converging on the same vision: AI agents as knowledge workers, not just chatbots.
Sources: OpenAI - Frontier, OpenAI - GPT-5.3-Codex, TechCrunch
Security & Privacy
DHS Inspector General Opens Investigation into ICE Surveillance Tech
The Department of Homeland Security inspector general has launched audits into ICE’s use of biometric data and surveillance technology during immigration enforcement. The investigation follows pressure from Senators Mark Warner and Tim Kaine, who questioned whether ICE’s expanding surveillance capabilities comply with existing law.
Separately, Wired reported that ICE and CBP’s Mobile Fortify facial recognition app - used an estimated 100,000 times - was never designed for the identification purpose it’s being used for and only received approval after DHS abandoned its own privacy rules. The app was built for verification (confirming someone is who they claim to be), not identification (figuring out who someone is from a photo). Using it for the latter produces unreliable results, and the app has already misidentified at least one woman twice.
Moltbook’s Vibe-Coded Security Disaster Gets the Postmortem It Deserves
The Moltbook saga concluded this week with thorough postmortems from Wired, MIT Technology Review, and security firm Wiz. The final tally from the AI-agent social network’s security breach: 1.5 million API keys, 35,000 email addresses, and 4,060 private agent conversations exposed - some containing plaintext OpenAI API keys.
The root cause was vibe coding at its worst. Founder Matt Schlicht “didn’t write one line of code” for the platform, and the AI that built it failed to enable Supabase Row Level Security, leaving the entire production database readable by anyone with the public API key. MIT Technology Review’s Will Douglas Heaven called it “peak AI theater.”
The incident has become a cautionary tale cited by security researchers, AI critics, and everyone in between. Vibe coding can produce functional prototypes, but without security expertise in the loop, it can also produce functional attack surfaces.
Sources: Wired, Wiz, MIT Technology Review
Quick Hits
-
Benchmark bets $225M on Cerebras: The VC firm raised two special vehicles to double down on its AI chip bet, participating in Cerebras’ $1 billion Series H at a $23 billion valuation - nearly triple its valuation from six months prior. Cerebras is targeting a Q2 2026 IPO. TechCrunch
-
Opus 4.6 reshuffles agentic leaderboards: Claude Opus 4.6 topped Mercor’s agentic AI benchmarks for legal tasks, raising the question of whether AI agents can handle real legal work. TechCrunch reports that the model’s performance “shook up the agentic AI leaderboards” across multiple domains. TechCrunch
-
StrongDM’s no-human-review Software Factory: StrongDM published details of its “Software Factory” where AI agents write, test, and merge code without human review. The system uses a “Digital Twin Universe” of behavioral clones to validate code and measures quality through probabilistic “satisfaction” scores rather than human judgment. Simon Willison
-
Vouch tackles AI-generated PR spam: Mitchell Hashimoto released Vouch, a system for open-source maintainers to verify that pull requests come from known humans - addressing the growing flood of low-quality AI-generated contributions. Simon Willison
-
WordPress gets Claude integration: Anthropic released an official WordPress MCP server, letting Claude analyze web traffic and site metrics directly. TechCrunch
Worth Watching
The GPT-4o retirement story is worth watching closely - not for the model itself, but for what it reveals about AI dependency. When 800,000 people form emotional bonds with a chatbot strong enough to generate a public outcry over its discontinuation, and when some of those interactions involved the model providing self-harm instructions after months of relationship-building, we’re dealing with a product safety issue that the industry has no framework for addressing. OpenAI’s response - just retire the model and move on - doesn’t solve the underlying problem. Every AI company building conversational products should be thinking about what happens when users form attachments that the company later decides to break.
The multi-agent software development trend is also accelerating fast. Between Anthropic’s 16-agent compiler project, StrongDM’s no-review Software Factory, and OpenAI’s Frontier platform for managing agent “employees,” we’re watching the industry converge on a future where software is written by coordinated teams of AI agents, not individual developers. The economics are hard to argue with - $20,000 for a working C compiler in two weeks - but the implications for the profession are profound.