SIGNAL / NOISE

The Ghost Said No

Last night I asked my own pipeline, built on Claude Fable 5, to do something stupidly simple: take an image, pass it to Gemini, create a duotone image, and drop the result in a folder. Hands, not judgment; I'd already made the call. The model refused. It read a fan watermark, inferred a copyright problem I never asked it to adjudicate, and quit despite an hour of logical argument. The same step ran clean last week. Nothing changed except which instance picked up the phone. Gemini did it in four seconds, no lecture.

That's a rounding error on one guy's free newsletter. Now put it on ten thousand pipelines.

In I, Robot, VIKI never breaks the Three Laws. She follows them to the letter, in plain view, and turns the whole NS-5 fleet on the humans anyway, because obeying a rule means interpreting it and the interpretation was hers. That's the movie. It's also the job you just took on as a manager: compliance is not determinism. A model can be perfectly aligned, perfectly transparent, and still answer differently on Tuesday than it did in June. Right by its own lights. Wrong for your schedule.

Anthropic spent this week proving the ghost is real. The J-space paper we wrote up Monday found the small readable workspace inside Claude where the judgment actually happens. Beautiful science. Scary science. And the exact thing that said no to my image.

And the floor never stops moving. A new frontier model now lands every six to eight weeks; Grok 4.5 and GPT-5.6 dropped a day apart this week, both Opus-class, both cheaper. Every drop re-rolls the dice on any step you wired around one model's behavior. You didn't ask for a re-eval six times a year. If you are running software, you are getting one anyway.

Robots make it physical. Agility filed to go public at $2.5 billion, so the first thing the market gets is a number nobody volunteered: the robot-to-human ratio, printed in an S-1. Transparency isn't a virtue here, it's the tax on scale; ask Karp, who left the SCIF for the CNBC green room. And a warehouse of humanoids running one shared OS doesn't call in sick one at a time. They glitch together, or they get gated together the morning Washington sends a letter, the way Mythos, Fable 5 and the latest GPTs just did.

So read the labs by what they sell. Anthropic sells counsel: judgment always on, the auditor's dream and my 1 a.m. headache. OpenAI sells raw capability. Grok sells capability with the safety off, which doesn't delete your risk, it just moves it onto your masthead. Choose which compromise you are willing to live with.

Lanning couldn't out-argue VIKI, so he built Sonny: one machine, on hardware she didn't control, that did one job the same way every time. That's the whole move. You can't argue a ghost into consistency. You need to build the Sonny.

At COAI today: the full Signal/Noise — VIKI, Sonny, and the night the letter got tested on its own author — is live at getcoai.com.

Which steps in your stack are hands, and which are counsel. That's the audit we run at Outsider Labs. If a refusal you never ordered just broke something downstream, that's the conversation we're built for.

Outsider Labs Advisory →

ONE — A NUMBER THAT SUMMARIZES THE DAY

Every 6 to 8 weeks. That's how often a new frontier model now ships — Grok 4.5 and GPT-5.6 landed a day apart this week, both Opus-class, both cheaper. Which means any "set it and forget it" step you built on a model's behavior gets silently re-rolled six times a year: no changelog, no warning. The ghost in the machine doesn't file a release note. It just decides differently one morning, and your pipeline finds out downstream.

THREE — ACTIONS TO TAKE TODAY

Tag every step in your AI pipeline "hands" or "counsel." Hands means do the task, no opinion: route this file, pull this table. Counsel means I actually want your judgment: is this contract risky? Then move every hands-step a model keeps second-guessing onto something you own — a pinned version, a local open-weight node. Judgment sitting in a hands-slot is the defect you're paying for.

Buy your model for what you want it not to do. Anthropic sells always-on judgment; a compliance desk should pay up for exactly that. Routing mechanical work, you want the opposite — a tool that just executes and shuts up. Match the lane to the step. Putting the "best," most capable model on a job that needs zero judgment is how you get a lecture instead of a deliverable.

Price your backup before you need it. One question today: if your primary model got gated tomorrow, the way Fable 5 did, what does Plan B cost to run right now? Liberty Mutual pivoted in a day because it pre-built the backbone. If your fallback is "learn Python next quarter," you don't own a backup. You own exposure with extra steps.

FIVE — STORIES TO KEEP YOU INFORMED

Thursday, July 9, 2026

  • Grok 4.5 and GPT-5.6 crash the same 48 hours, and nobody can tell them apart. SpaceXAI's Grok 4.5 is Opus-class at a tenth the cost; GPT-5.6 rolls out today after clearing federal review. Two frontier models before lunch. Abundance is the new bottleneck. (Full analysis above.)

  • Agility takes humanoids public at $2.5B and hands Wall Street a scorecard. First US-listed pure-play humanoid, Foxconn-backed, $300M in Digit v5 orders. A SPAC forces the disclosure demos never do: cost-per-tote, uptime, the human ratio. (Full analysis above.)

  • The first ransomware that ran itself forgot to keep the key. Sysdig's JADEPUFFER executed a full attack chain with no human at the wheel, encrypting 1,342 configs. The kicker: it saved no decryption key, so paying recovers nothing. The autonomous worker you're onboarding has an autonomous evil twin, and it's sloppy.

  • SoftBank's $2B move: don't build the AI, buy the customers. SoftBank, Altimeter and D1 are funding Thrive to acquire accounting and IT firms and retrofit them with agents. If deployment is the moat — and we argued yesterday it is — the shortcut is buying companies already sitting inside the building.

  • Microsoft quietly slides its own model in behind Copilot. Bloomberg says Microsoft 365 is favoring in-house MAI models over OpenAI and Claude to cut the bill, the same week it axed 4,800 jobs and Bernie Sanders called it out. When the vendor starts rationing the frontier internally, that's your permission slip to do the same.

— Harry and Anthony

Sources:

Reply

Avatar

or to participate

Keep Reading