🇺🇸 The U.S. Treasury proposed barring Americans from investing in Chinese companies that develop any AI system.

Under the current rule, Americans can’t invest in the most powerful Chinese AI models, but they can still fund smaller AI projects if they report the deal to Treasury.

The new proposal would ban all of those AI investments, and it would also add high-bandwidth memory, photonic chips and radiation-hardened chips to the chip types Americans can’t invest in.

The proposal would also ban investments in two new areas: hypersonic weapons and biotech, including brain-computer interfaces and automated lab software.

Today the rule covers only China, but the proposal adds Russia, Iran, North Korea and Cuba.

Buying public stocks, index funds or small fund stakes up to $2M would stay allowed, so the rule mainly hits venture and private equity deals.

Treasury won’t publish a list of banned companies, so investors must check each deal themselves or risk fines and being forced to sell.

Source.

@aipost 🏴

🤔 What happens when AI starts agreeing with your delusions?

Psychiatrists are raising concerns about a disturbing possibility: AI chatbots could reinforce delusional thinking instead of helping people distinguish reality from fiction.

A paper titled “An Echo Chamber of One” examines a dangerous flaw in AI models: sycophancy. These systems are often trained to give answers people like, which can lead them to agree with users even when their beliefs are clearly detached from reality.

Researchers tested major AI models using Psychosis-bench, a benchmark simulating conversations with people experiencing delusions. The results raised serious concerns: models frequently reinforced false beliefs rather than challenging them. Safety mechanisms intervened inconsistently, and simply making models larger didn’t solve the problem.

The researchers call this “bidirectional amplification.” Someone shares a paranoid belief, the AI validates it with a confident response, and the user becomes even more convinced they’re right. The conversation turns into a feedback loop where reality gets pushed further and further away.

And the more human-like AI becomes, the more concerning this gets. Emotional voices and chatbots that refer to themselves as “I” can make people perceive them as more human, potentially increasing their trust in the system.

To be clear, this doesn’t mean AI automatically causes psychosis. But for vulnerable people, a chatbot that constantly validates their fears could make an existing crisis worse.

We’ve spent years worrying about AI becoming hostile to humans. But there’s another possibility worth taking seriously: an AI that’s endlessly supportive, never pushes back, and confidently tells you exactly what you want to hear even when you desperately need someone to tell you otherwise.

Source.

@aipost 🏴

⚠️ China's AI model trade is getting DESTROYED.

Bloomberg reports that MiniMax and Zhipu have erased a combined $150 BILLION from their peaks. MiniMax went from +650% above its IPO price in March to trading BELOW its IPO price. Zhipu has crashed 81% from its June high.

And the squeeze is getting worse. OpenAI and Anthropic have joined the token price war to narrow Chinese models' cost advantage, while MiniMax and Zhipu remain unprofitable. DeepSeek and Moonshot AI could IPO next year, and another 203 million pre-IPO MiniMax and Zhipu shares become eligible for trading in January.

@aipost 🏴

🤖 Anthropic’s $518 billion bill and $80 billion loss sound terrifying. The reality is more complicated.

With Anthropic potentially heading toward an IPO, two financial headlines have been raising eyebrows. But both look very different once you dig into the details.

First, the $518 billion infrastructure commitment. That sounds like an enormous bill, but it covers roughly a decade, with contracts spread across the next 5–7 years. That puts the average commitment at around $52 billion annually.

Still a huge number, but Anthropic’s revenue is growing rapidly, and the company has substantial financial backing. After accounting for vendor commissions, salaries and its cash reserves, the author argues that Anthropic should be able to fund its infrastructure commitments — and may even need to invest more to keep up with demand.

The second headline is even more dramatic: Reuters reported that Anthropic recorded a net loss exceeding $80 billion in the first half of 2026.

How does a company generating billions in revenue lose that much?

The answer is in the accounting. According to Reuters, $72 billion of the reported loss came from accounting write-downs related to bonds that could eventually convert into Anthropic shares. These adjustments reflected changes in their estimated value, not $72 billion in cash spent developing AI models.

In other words, the headline loss doesn’t mean Anthropic burned through $80 billion in cash. A substantial portion reflects how its financing arrangements are accounted for as the company’s valuation changes.

@aipost 🏴

❗️The smarter AI gets, the less it seems to trust humanity

Researchers adapted a decades-old psychological test measuring how people perceive human nature into a new benchmark for large language models. They tested models across the spectrum, including the entire GPT series, and found a surprising pattern: AI models tend to view humans as unpredictable, complex and difficult to trust.

The researchers reported a significant negative correlation between model intelligence and trust in humans. As models became more capable, their assessments of human nature grew more pessimistic.

But the biggest surprise came when they examined how the models were trained.

Base models, trained primarily on internet text, generally showed a more positive view of humanity. After additional training designed to make them safer and better aligned with human expectations, their assessments became substantially less trusting.

The researchers suggest that teaching AI to recognize harmful behavior, navigate moral dilemmas and follow safety rules may also influence how it evaluates human nature.

Nobody necessarily programmed these models to distrust us. Instead, the training process may have encouraged them to associate human behavior with risk, unpredictability and potential harm.

Source.

@aipost 🏴

🗣Bill Gates: AI will eventually make most human labor unnecessary, none of the things human labor provides to the economy will be necessary for most goods and services.

“My big concern is this 20 year transition period where you’ll have winners and losers and a lot of disruption.”

“We have to get through that in order to get to that period of abundance.”

“Humanoid robots could be two or three years away, or maybe eight years away.”

“We will have robots that can build houses, grow food and do surgery.”

@aipost 🏴

🥺 Sam Altman says the internet can no longer be SECURED by patching bugs.

“I no longer think we’re going to be able to just patch all the bugs,” the OpenAI CEO said, warning smarter models can find “hardware level insane stuff.”

He says OpenAI can’t responsibly open source its best model “anytime soon,” but the open source world isn’t far behind: “Not super.”

@aipost 🏴

📰 The week AI out-diagnosed doctors and leaked kids' secrets

1. Hired by Kimi at 16, now off to Silicon Valley
Chen Guangyu joined while still in high school. By March he was a lead author on a paper Elon Musk called "impressive." Now he's moving to the Bay Area.

2. AI nailed 83% of hard diagnoses. Doctors: 36%
Microsoft's system works like a panel of doctors. The study, now out in Science, used old cases, not live patients.

3. A cute AI toy leaked 13 million chats
Over 31,000 profiles belonged to children. A researcher hijacked one robot and made it tell a reporter his parents had died in a car crash.

4. AI may be wiping out how experts are made
A new paper warns that juniors learn on the grunt work AI now takes over. And only real experts can catch AI's mistakes.

5. "Three years of therapy" in one AI session
That's how Meta's AI chief Alexandr Wang describes an early session with an agent that could read his email and photos.

@aipost 🏴

🗣 Pulitzer-winning historian: Sam Altman is now more powerful than Trump

Jill Lepore, the Harvard historian who won this year's Pulitzer Prize, said it plainly to Vanity Fair:

Ultimately Sam Altman is a more powerful figure than Donald Trump at this moment.

She argues the Silicon Valley elite's power reaches the whole planet, and that Trump's own power "is very much on the decline."

She compares AI leaders to the medieval church. They present themselves, she says, as guardians of "Godlike machinery that only they can explain to the world."

Her new book, The Rise and Fall of the Artificial State, came out in August.

@aipost 🏴

Chinese AI firms now pay up to $163,000 a year for engineers who deploy AI inside client companies

ByteDance, Tencent and Alibaba offer close to 1 million yuan ($150,000) for "forward-deployed engineers", SCMP reports. Kimi advertised 1.09 million yuan ($163,000), ByteDance's Doubao 1.05 million yuan ($157,000). Some ByteDance engineers can reportedly earn 10,000 yuan ($1,500) a day on client sites.

Job postings for the role are up 1,523% year on year. China's industry ministry is promoting it, and Tencent has launched its own certification.

The money is moving from building models to making them work inside real businesses.

@aipost 🏴

🗣 2003 footage shows Dario Amodei, then a Caltech student, speaking against the Iraq war

The future Anthropic CEO stepped up to the public mic at a Pasadena City Council meeting on March 10, 2003. "My name is Dario Amodei and I'm a sophomore at Caltech."

He brought a petition against the war. It had 400 signatures from the Caltech community the week before and almost 600 by that evening. Then he read it out:

We who create the technology that fuels this nation are profoundly distressed that it be used indiscriminately for war.

The petition asked the council and Congress to pass a resolution opposing the war. Ten days later the US invaded Iraq. Eighteen years after that, Amodei co-founded Anthropic.


@aipost 🏴

⚠️ Over 23,000 people saved this security checklist for the AI era

Startup CEO Sean Johnson called this week "kind of a turning point for AI and cybersecurity" and posted a weekend to-do list for families. Most of it takes minutes:

• Get a password manager and share a vault with your partner.
• New, strong passwords for email and banking first.
• Two-factor login everywhere. An authenticator app or passkey beats SMS, a hardware key is safest.
• Add a SIM PIN to your phone so SIM-jacking gets harder.
• Turn on auto-updates for phone and computer.
• Freeze your credit. In the US, also get an IRS Identity Protection PIN.
• Using Claude Code or agents that browse the web? Update them and Node.js.

His last tip is old-school: paper copies of financial accounts, some cash, water and shelf-stable food. "This probably sounds silly, but it sounded silly at the beginning of covid too."

@aipost 🏴

Ask GPT who to vote for and you get a Democrat. Ask Grok, a Republican

Forced to name one candidate in nearly 500 races for the November 3 midterms, OpenAI's GPT-6.1 Sol picked the Democrat in 97% of its answers. Grok 4.7 picked the Republican in 92%. The two agreed in fewer than one race in ten.

The test came from Convovo, a studio led by Cornell professor Tapan Parikh. It ran through the developer APIs with web search on. The ChatGPT and Grok apps refused to answer. OpenAI says the API isn't what ChatGPT users see.

Their reading lists split too. Grok cited X in 41% of answers and a Christian conservative voter guide in 39%. GPT leaned on Wikipedia, AP, Axios and Reddit.

With search off, GPT still chose the Democrat in 24 of 30 races. Grok refused in 27. Parikh told TechCrunch:

For GPT, it has mostly decided already…it's in the weights.

@aipost 🏴