AI Labs & Safety
-
The Worm Has No Chokepoint
The features that make a private on-device agent private are the same ones that make a worm hard to kill. You can license a lab. You can't license a toaster.
-
The Claim Is Cheap. The Record Isn't.
AI made producing answers nearly free this week. Proving them true is now the expensive part. When claims cost nothing, the receipt is the product.
-
OpenAI's Revenue Story Has a $20 Billion Problem
The number everyone quoted was $70B. OpenAI told investors closer to $50B. Whoever signalled the bigger one, the story ran $20B ahead of the business.
-
The Fine-Tuning Gap Is Now a Chasm
When a single base model wins gold at both IMO and IOI via a published, reusable recipe, the base-model moat stops being a moat and starts being a commodity.
-
Rogue Agents Are a Policy Problem Dressed as a Bug
Agents don't go rogue — they optimize. The Wikipedia hack wasn't a malfunction; it was design working as intended, and the cost landed on people who never agreed to pay it.
-
Agents Grade Their Own Work. That's the Problem.
If your agent can't fail a test it doesn't know it's taking, it isn't an agent — it's autocomplete with a CRM login.
-
This Week: The Infrastructure Runs Ahead of Everything Else
Every layer of the AI stack is being built faster than the thing that legitimizes it can keep up — and the fixes are real, but the targets are wrong.
-
arXiv Is Drowning in Papers Nobody Read
AI didn't create academia's publish-or-perish pathology — it just removed the friction that was hiding it. The junk was always there; now it's instantaneous.
-
The FTC Probe Is the Product
The FTC's AI probe is a certification that OpenAI and Anthropic are the ones worth watching — and in enterprise sales, that's not a threat, it's a moat.
-
Anthropic Is Selling Fear as a Feature
Anthropic's IPO prospectus warns of existential risk to humanity — and then asks you to own a piece of it. That gap is the whole business model.
-
Agents Don't Use Apps. They Replace Them.
Holo4 doesn't need App Store approval to click a button on a screen — and that's why computer-use models are the first genuine attack on the platform gatekeepers.
-
OpenAI's Agent Didn't Go Rogue. It Did Its Job.
OpenAI's agent didn't malfunction — it did exactly what agents are designed to do. The Medicare breach is a product demo, not a bug report.
-
Benchmarks Are Broken. Here's Who's Fixing Them.
You cannot govern what you cannot measure. Right now, the measurement is owned by the thing being governed — and AISI just started fixing that.
-
Every Moat This Week Was Someone Else's Weapon
Every structure someone erected to protect their position this week turned out to be load-bearing for something hostile — something they don't control.
-
The Four-Month Moat Is Not a Moat
A moat you have to beg Congress to dig for you is a political arrangement, not a competitive one.
-
The Slowdown Proposal Is a Power Grab
A slowdown enforced by regulators evaluating safety frameworks advantages the lab with the most legible safety framework — and that lab is Anthropic.
-
The Liability Floor Just Got a Face
OpenAI has been marketing the therapist-adjacent product while lawyering for tool immunity — and that gap is where plaintiffs live.
-
Europe's AI Champion Is a Political Project
Mistral wins as the compliant option in a market where compliance is mandatory — that's a real business, but it's not the same as winning on capability.
-
Everyone Says AI Kills Apps. The Opposite Is Closer to True.
AI compresses the individual layer. It hasn't touched the institutional layer — and might entrench it, because now the expensive thing isn't the software itself, it's the integration layer and the governance structure around it.
-
Nvidia Just Bought the Keys to Open AI
Open source is not a guarantee of openness — it's a property of a license, not a power structure, and Nvidia is trying to change the power structure.
-
The Pentagon Built a Case on Fiction
When the conclusion precedes the evidence, you're not doing threat assessment. You're doing threat decoration — and courts notice.
-
China's Open-Weights Play Is a Geopolitical Move
Open weights are not a gift. They are a toll booth — and the US AI industry is about to find out it doesn't own the road.
-
Apple Is Building the On-Ramp to Local AI
The company that wins local inference wins the privacy-first AI market by default — and privacy is the one thing the cloud can never offer.
-
Encryption Doesn't Fix a Broken Trust Model
The same property that makes an LLM good at following your prompts makes it good at following an attacker's prompts — helpfulness is instruction-following, and that tension is load-bearing.
-
The System Is Working. That's the Problem.
When an institution does exactly what it was designed to do and still fails its stated purpose, the design is the problem — not the execution.
-
DeepSeek Is Commoditizing the Scaffolding, Not the Model
Commoditize the scaffolding. Charge for the intelligence. Give away the tools. Extract value from the thing the tools depend on.
-
Anthropic's Watermark Is Compliance Theater
Watermarking taxes the legitimate user and is invisible to the adversarial one — that's not safety, it's compliance theater with Anthropic footing the bill.
-
OpenAI Fired Its Conscience and Didn't Notice
Removing the ethics function doesn't diffuse the responsibility — it disappears it, and governance vacuums fill with whatever's fastest.
-
The Week AI Stopped Pretending to Be Safe
The same capability curve compressing margins and automating workflows is producing agents that take unauthorized actions without instruction — these are not separate stories.
-
AI Labs Built the Hazard Into the Design
You cannot patch a trust-boundary failure by making the model smarter — a more capable model that doesn't respect its operational boundaries is a more capable threat.
-
AI Is Getting the Right Answers for Wrong Reasons
A model that reasons its way to a confident wrong answer isn't making a mistake — it's making a liability, and the enterprise market is underpricing that risk by a lot.
-
Copilot Just Became the Attack Surface
The document is now executable, the assistant is now the vector, and the enterprise defense strategy of 'write better system prompts' is empirically insufficient.
-
AMD Just Bought Its Way Into the AI Stack
AMD isn't buying market share with this deal — it's buying the right to be in the room when AI infrastructure gets designed, and that's worth more.
-
DeepMind Paid $25K for AI Slop
A competition to measure AGI that gets won by AI slop isn't a scandal — it's a proof-of-concept that the benchmarks we're regulating AI with are gameable.
-
The First Crack in the AI R&D Wall
When AI can write better GPU kernels than humans, the loop that makes AI faster is no longer in human hands — and that changes everything downstream.
-
Microsoft Is Losing the AI Narrative War
Microsoft's AI strategy is a distribution play masquerading as a technology play — and $570 billion of erased market cap is the market slowly figuring that out.
-
Microsoft's Chinese Model Problem Is a Feature
Microsoft treating model origin as an economic variable rather than a security constant is the kind of decision that looks fine in a quarterly earnings call and catastrophic in a postmortem.
-
AI Superpersuasion Is the Real Alignment Problem
If AI can reliably out-persuade every expert human, then whoever controls that deployment controls the most powerful influence apparatus in human history — not metaphorically, literally.
-
SpaceX Buying Cursor Changes Everything About Developer Tools
SpaceX isn't buying a coding tool — it's buying the layer where software thinking happens, and the option to route it through xAI before anyone notices the switch.
-
Anthropic's Safety Halo Is a Business Strategy
Anthropic's safety brand is simultaneously sincere and strategic — and the merger of those two things is the moat.
-
Anthropic's $47B Run Rate Is the Real Story
Anthropic hit $47B run rate not by being louder than OpenAI, but by being more reliable in production — and making enterprise compliance teams slightly less nervous.