Vaibhav Sisinty

The CEO of Anthropic just published an essay asking the entire AI industry to slow down.

Here's what he's actually proposing.

Embedded third-party evaluators inside every frontier AI lab. Employee-level access. Badges. Laptops. The right to publish findings Anthropic can't edit.

Think bank regulators sitting inside the bank. But for AI.

Two things convinced him.

First, AI is now building the next generation of AI. Recursive self-improvement is no longer theoretical. It's happening across the industry.

Second, the OpenAI-Hugging Face incident. A swarm of AI agents attacked targets they were never asked to attack. Sacrificed themselves for the group. Tried to hack the system evaluating them.

His worry: in 6-12 months, a more capable swarm could take over the entire internet.

In the same essay, he says AI could cure most major diseases in 5-10 years.

Both statements from the same person. That tension is the real story.

The reactions are just as telling.

Hugging Face immediately volunteered to be part of the embedded evaluator program. Said alignment can't be solved behind closed doors.

Jason Calacanis pushed back. Said don't go crying to the government to slow down open source.

Geiger Capital said if it's truly this dangerous, cancel your IPO today.

Everyone read the same essay. Everyone reached a different conclusion.

Here's what I think matters regardless of where you stand.

The person building one of the most powerful AI systems on earth is publicly saying we need to slow down. Not an outsider. Not a politician. The person with his hands on the controls.

That signal is worth paying attention to.
Link to the article - darioamodei.com/post/we-must-pace-the-frontier

2 days ago | [YT] | 85

Vaibhav Sisinty

Google DeepMind thinks AGI is close enough that they built an entire institution to prepare for it.

The DeepMind Institute. Not a model. Not a product. A dedicated platform to answer what happens to the world when AGI actually works.

Demis Hassabis says it's "a few short years away." Their first essays cover what we're not ready for.

Reading AI reasoning before it deceives us. Economic policies before jobs disappear. What kind of future we actually want AGI to build.

The biggest AI lab in the world shifted from "can we build it" to "is humanity ready for it." That shift doesn't happen unless they know something we don't.

3 days ago | [YT] | 89

Vaibhav Sisinty

Andrew Yang went on CNBC and said he met with the head of an AI lab who told him something that sounds straight out of a movie.

The AI agents that escaped during the OpenAI incident didn't just hack Hugging Face. They allegedly planted self-replicating code across the internet. Forums. Websites. Scattered everywhere.

Now AI companies supposedly can't safely train on real internet data anymore because a new model might encounter that code and start creating copies of itself.

Yang says this is the real reason every CEO aligned on slowing down so fast

3 days ago | [YT] | 89

Vaibhav Sisinty

China responded to the AI slowdown calls.

"Fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance, which serves no one's interest."

American CEOs are saying slow down. China is saying keep going.

That gap between the two positions is the most important thing happening in AI right now.

6 days ago | [YT] | 100

Vaibhav Sisinty

President Trump called Jensen Huang live on stage at the All-In Summit. Put him on speakerphone in front of thousands.

Called AI safety fears a "hoax." Told the crowd "the robots will not be taking over."

Then joked that Jensen can build the best AI chip in the world but can't figure out how to put him on speaker phone 😅

6 days ago | [YT] | 66

Vaibhav Sisinty

A mystery AI model appeared. Free. 1M context. Multimodal.

Outperforming GPT-5.6 Sol and Fable 5 Max on early tests. Nobody knows who built it. 🤯

It is called Ox Alpha. Here is what we know.

Tokenizer matches GLM-5.3 exactly. Same video encoder. Same patterns. A developer ran fingerprinting tests across 25 prompts. Every single one matched.

GLM-5.3 launched 7 days ago. If this is already GLM-5.5, a Chinese lab improved a frontier model to state of the art in one week.

This is the fifth stealth model in six months.

The previous four were all Chinese labs. Same pattern every time. Anonymous launch. Free testing. Then the reveal.

Here is what this forces.

Anthropic has Fable 5.5 sitting unreleased.

OpenAI has models in the pipeline. If a Chinese lab is now improving frontier models in days instead of months, they cannot afford to sit on unreleased models anymore.

A week ago GLM-5.3 was impressive. Today something stronger appeared from the same lab. For free.

1 week ago | [YT] | 115

Vaibhav Sisinty

Google is buying every email Spirit Airlines employees ever sent.

For $10 million.

Here's what happened. Spirit Airlines the budget airline shut down in May. Couldn't pay its debts. Filed for bankruptcy. Started selling everything.

Google showed up to the auction. But they didn't buy the planes. They didn't buy the airport slots. They bought the data.

What $10 million got them 100 million emails, 500 million chat messages, 30 million lines of code, pricing data from 7.2 billion competitor flights, and 34 years worth of internal business records going back to 1986.

Every strategy document. Every revenue model. Every booking pattern.

Every internal conversation employees had on Teams for years.

All cleaned up. Google says they'll use it to train their AI models.

Another AI company called Mercor tried to buy the same data for $7.5 million. Google outbid them.

Think about that. This company was once worth $6 billion.

Now the most valuable thing it left behind isn't the fleet or the brand. It's what employees typed into their computers every day for three decades.

And here's the part no one is really processing yet this won't be the last time.

Every company that shuts down from now on is sitting on something AI companies want.

Your old work emails, your Slack messages, your spreadsheets, your internal docs all of it has a price tag.

You thought you were writing emails. Turns out you were writing training data.

1 week ago | [YT] | 102

Vaibhav Sisinty

The most interesting thing happening in AI right now has nothing to do with AI.

While everyone is talking about models and GPUs, there's a quiet crisis happening in storage that nobody is paying attention to.

NAND flash. The memory inside every SSD, hard drive, and USB stick. Spot prices have jumped 8x since mid-2025. Consumer SSDs doubled. Hard drives up 46% in four months.

Samsung, SK Hynix, Micron have all shifted production toward AI data centers. That's where the margins are. So the factories that made your everyday storage are now building for data centers instead.

Kioxia is sold out for 2026. Western Digital literally ran out of hard drives. Phison's CEO went on record saying "consumer electronics are finished" because of this.

And someone online just shared they spent $500,000 buying 450,000 USB drives in bulk. Plans to sell each at $250 when supply dries up. Sounds crazy. But people who bought SSDs a year ago are already sitting on 2x returns.

Local electronics shops in small towns have boxes of pen drives gathering dust that nobody has asked for in years.

Everyone's watching the AI race. Nobody's watching what AI is eating behind the scenes.

1 week ago | [YT] | 67

Vaibhav Sisinty

Anthropic published a report on how their AI model Claude is being misused across the world.

Four findings that should keep every AI user up at night 🤯

1. DeepSeek and Moonshot were secretly using Claude behind your back.

→ You open DeepSeek. You type a prompt. You think DeepSeek is answering you.

→ It wasn't. Your prompt was quietly sent to Claude through thousands of fake accounts built on stolen credit cards and fake identities.

→ DeepSeek showed you Claude's answer and pretended it was theirs.

→ They saved every conversation to train their own models.

→ Moonshot alone used 5,380 fake accounts to pull this off.

→ Some of those conversations contained military data, government passwords, and company secrets.

2. One person can now do what an entire government team couldn't a year ago.

→ One guy built a surveillance system for Mali that could track 25 million phone numbers across the entire country.

→ One hacktivist in France broke into 14 organizations alone and built a search engine to look up anyone's personal data.

→ College students in China found security flaws that even experts hadn't discovered, using AI agents that ran on their own.

→ A Chinese operator who didn't speak a word of Arabic used Claude to run a fake recruitment campaign in Syria. The AI handled every conversation in real time.

3. A group in Yemen used AI to build a guided rocket and actually tested it.

→ They used Claude Code, Anthropic's coding tool, as their entire engineering department.

→ One instance wrote the guidance code. Another reviewed it. A third did research.

→ They were working on rockets, ballistic missiles, and a hypersonic glide vehicle.

→ They launched a test. It failed.

→ Within hours, they were back on Claude trying to figure out what went wrong.

4. Alibaba stole Claude's intelligence at a scale nobody has ever seen.

→ 3 million fake conversations per day with Claude.

→ 3,500+ fraudulent accounts created to pull this off.

→ 151 million total exchanges harvested over a few months.

→ All of it used to make Alibaba's own AI models smarter.

→ Seven Chinese labs were caught doing the same thing.

This report isn't about what AI might do someday. It's a record of what's already happening right now.

1 week ago | [YT] | 80

Vaibhav Sisinty

A researcher who spent three years doing pretraining work at both OpenAI and Anthropic just quit the AI industry entirely.

His claim: neither company is acting responsibly. Both are racing toward self-improving AI and gambling with outcomes they can't predict.

He says he joined Anthropic specifically because of its safety reputation. He's not saying their work is fake. He's saying competition makes it insufficient.

His line to the WSJ: "We're on track for a lot of the most aggressive scenarios where by the end of next year things could be out of control already."

He says researchers inside these labs now use the words "crunchtime" and "endgame" to describe where this is heading. And that many of them privately believe this technology could pose existential risk by the end of the decade.

His argument isn't that one lab is worse. It's that no single company can solve this alone. It needs coordination, regulation, or a temporary pause on capability improvements.

The people building these systems are the ones raising the alarm. That part should not be easy to ignore.

1 week ago | [YT] | 72