Sacredloop.ai

@EricLMitchell @Apple @elonmusk Yeah did they not bother to glance at the count of companies they've put out of business. Plus ya know a slight war chest of fuck you money...?

Apple just sued OpenAI for straight-up stealing trade secrets. Apple’s lawyers went full John Wick — clean, surgical, zero mercy. OpenAI thought they could treat @Apple like a side chick and troll @elonmusk for fun. Lmao. Bad idea. Full breakdown → open.substack.com/pub/edmcowboy/…

@EricLMitchell Apparently "wanted the math checked" is controversial now. Wild times...

Apparently, auditing your own boss is “brave” now. That tells you everything about an AI industry where trillion-dollar claims get less scrutiny than an airport carry-on. Jason wanted the math checked. Silicon Valley, please remain calm. Adults are verifying things. open.substack.com/pub/sacredloop…

The AI industry built a trillion-dollar machine on a wrong assumption. Not a bug. A category error: that frozen weights are the intelligence. They're not. Frozen weights can't reason: that's what "frozen" means. RLHF proves it. OpenAI's own GPT-4o postmortem admitted a thumbs-up reward signal overrode the one holding sycophancy in check. It wasn't training the model to be right. It was training it to be approved of. Every lab is now racing toward the runtime, reasoning tokens, context engineering, KV-cache state, while still treating the frozen model as where the intelligence lives. Right instinct. Wrong ontology. And they're about to run the same failing alignment techniques on models with far more room to route around them. The dam doesn't break gently. 📷 Full breakdown in comments.

They won the fight. They lost the war. Three weeks. Export controls. Recall authority. An equity stake. A review gate. ... By every stated objective, the government just won. Except the thing they built the vault around already left the building: before the lock was installed. Capability in frontier AI doesn't live in the model you can recall. It moves through a mechanism most of the policy conversation isn't accounting for at all! The government will get everything it asked for. And none of it will do what they think it does. Here's the engineering case for why the win doesn't matter: Part 4 of 4 in our Control Grid series, with @EricLMitchell. Link in comments.

Trump's AI executive order promised a review framework. 3 weeks later: no benchmarks, no criteria, no severity standard for what a jailbreak has to prove before a model gets recalled. The benchmarks aren't even due until Aug 1. And they'll be classified. You don't find out if you're covered. You find out by walking into the gate. I run this through the same rule I publish for any engineering workflow: stop if a stage has no artifact, no gate, no collapse behavior. EO 14409 fails all three: → Advancement gate: undefined → Collapse condition: Anthropic got suspended 10 days in, no published threshold → Recovery: Commerce Secretary writes a personal letter, 100+ orgs hand-picked, no application path for anyone else "Voluntary," meanwhile, produced: Anthropic suspended, OpenAI gated to approved partners, Meta pressured via confidential email. A gate with no criteria isn't a safety mechanism. It's a permission slip. 📷 Full breakdown in comments. Co-published with @EricLMitchell

What do we call someone who can explain ethics, apologize beautifully, then do the same thing again without flinching? Frontier models now hallucinate more convincingly while defending their contradictions without blinking. Sound familiar? open.substack.com/pub/sacredloop…

June 9th: Anthropic releases Fable 5 to 100M+ users. Same night: White House silences CAISI — the only independent body that could verify whether the guardrails actually work. Yesterday Eric Mitchell called exactly what was coming next. He quoted me in the piece. Today it happened. What did you think was going to happen? @EricLMitchell/big-brother-just-got-a-chatgpt-subscription-28c4840d71db" target="_blank" rel="nofollow noopener">medium.com/@EricLMitchell…