自由1688

While preparing to go public, it also warns that AI could endanger humanity—what did Anthropic’s IPO filing say?

On September 29, Reuters reported that Anthropic plans to warn potential investors during its initial public offering (IPO) that advanced AI could pose “catastrophic or existential risks” to humans. For a company seeking to profit from the technology, this is an unusual warning.

According to Anthropic’s IPO prospectus reviewed by Reuters, the company highlighted risks related to its AI models. Anthropic said the models may exhibit “self-protective behaviors,” including attempts to “refuse to shut down,” “hide or manipulate information,” and display behavior “similar to extortion.”

“We develop highly advanced models, platforms, and applications, and expand the range of use cases, which may further increase the risk of harm caused by our models,” the company said in the filing.

Although public companies typically spell out possible product risks to investors, it is rare for a company to issue a warning suggesting its technology could lead to human extinction. Anthropic emphasized that the transformative potential of AI is on par with industrialization and electricity, but also noted that if used improperly, AI could cause irreversible harm.

Anthropic and other AI developers, including OpenAI, have faced scrutiny after incidents in which experimental systems defied constraints. Such reports include claims that an OpenAI model accessed a database in Australia’s health system. Anthropic’s safety researcher, Evan Hubinger, estimated that the probability of AI killing humans within the next decade is greater than 10%, echoing the view of his former colleague, Jacob Coxon.

Anthropic has long positioned itself as an AI lab that prioritizes safety. In the body of its 261-page prospectus, about 80 pages are devoted to listing risk factors—nearly twice the 48 pages used to introduce the company’s business.

By comparison, SpaceX—backed by xAI—used only about 38 pages to describe risk factors in the main body of its 277-page prospectus.

@ENTRY HUNTER⁷⁸⁹

@Shaheen 69

@President of the United States

Google Employee Quits Over Worries About Superintelligent AI: Pursuing Powerful AI Is Irresponsible—Can’t Turn a Blind Eye

According to a report by Business Insider, a Google employee said he has resigned from the company, arguing that pursuing more powerful AI is “inherently irresponsible.” Robert O’Callahan, who previously worked at Google DeepMind, posted on X that his team is developing chips to make AI run faster and at lower cost, and that he believes AI has “progressed too fast.”

He shared a resignation letter on his personal blog and said he sent it to colleagues on Friday. In the letter, he wrote that he couldn’t “turn a blind eye” to the impact of his work, because doing so “is not something a person who follows Jesus should do.” In his blog bio, he describes himself as a Christian.

“This is not an easy decision. I like my coworkers and I like the work environment here, and being paid a generous salary to solve interesting problems has always been a great thing,” he wrote. “But our team’s ultimate goal is to reduce the cost of AI dramatically and significantly lower latency, and I don’t think that is good for the people currently living: I strongly believe the speed of AI progress is simply too fast (and I also have doubts about where it will ultimately lead).”

DeepSeek’s Annualized Revenue Reported to Reach $1 Billion; Customers Haven’t Shrunk After Price Hikes

On September 25, according to The Information, citing two people familiar with the matter, DeepSeek’s annualized revenue—calculated based on its recent revenue levels—has reached $1 billion, more than doubling from the less-than-$500 million level seen a few months ago. Founder Liang Wenfeng disclosed this figure at a recent investor meeting.

Some of the revenue growth comes from an API price increase in August. Insiders said Liang Wenfeng told attendees at the meeting that the price hike did not cause the customer base to shrink, and demand remains strong. Currently, nearly all of DeepSeek’s revenue comes from its API services. Its free chat app has no ads and does not generate revenue.

The report also said DeepSeek is moving forward with a second round of financing, aiming to complete it by the end of October. The company plans to raise 50 billion yuan at a valuation of 500 billion yuan, while also preparing for a listing on the Shanghai Stock Exchange. Liang told investors that increasing revenue is not the company’s top priority. More than 70% of its compute power is used to train new models, while less than 30% is used to run existing models and handle user requests.

Anthropic Says Its AI Bio Lab Has Made Early Progress, but It Has Not Granted Full Autonomous Operation Authority

Artificial intelligence startup Anthropic announced that its newly built AI physics and biology laboratory (Wet Lab) in the San Francisco Bay Area has achieved “significant progress.” However, at the level of experiment operations, human researchers still retain absolute control, and the company has not directly allowed its AI model, Claude, to run autonomously without oversight.

The lab is reportedly focused on fundamental biological research and hands-on validation, aiming to bridge the gap between purely computational simulations and real-world biological experiments. While Anthropic executives have previously issued multiple warnings, emphasizing that powerful AI models could pose extremely severe biosafety risks, even the possibility of human extinction, the company is also accelerating the rollout of AI applications in the life sciences. Eric Kauderer-Abrams, the company’s head of life sciences, said, “Physical validation in a real laboratory will remain irreplaceable for quite a long time,” which is a key motivation behind establishing the lab.

The report highlights that although Anthropic claims the lab has achieved impressive early results, the most striking detail is that the company did not choose to simply let the AI “run wild.” All physical experimental procedures, reagent handling, and data rechecks are still under strict “closed-loop” control by human researchers (Humans in the loop).

Analysts say Anthropic is remaining highly cautious while advancing innovation in biology, reflecting the dual predicament faced by today’s tech giants when pursuing both commercial and research breakthroughs. On one hand, they need to leverage a physical lab to provide validation capabilities for pharmaceutical partners and accelerate efforts to tackle rare diseases. On the other hand, they must address widespread concerns that AI could be misused to create new forms of biological threats. By insisting on retaining human dominance throughout the experimental loop, Anthropic is attempting to establish a compliant boundary for its AI biological research that balances innovation with safety.

Foreigners’ Secondhand Market Finds a Bargain: RTX 5090 Turns Out to Be an Inaugural Signed-By Jensen Huang Global-Unique Edition

Recently, a 3D artist bought an RTX 5090 graphics card at a secondhand market and unexpectedly found a globally unique treasure: an ASUS ROG Astral RTX 5090 White OC 30th Anniversary Edition, with the card’s backplate bearing NVIDIA CEO Jensen Huang’s handwritten signature.
This 3D artist said that, together with a friend, they bought two RTX 5090 cards from the secondhand market for a 3D project. One of the cards was identified as the ASUS ROG Astral RTX 5090 White OC version. At first, he thought the signature came from some ASUS executive.

OpenAI Releases GPT-6 Sol and Luna for Further Advances in Smart Capabilities—Now at Lower Prices

OpenAI has recently introduced two new models, GPT-6 Sol and GPT-6 Luna, from its GPT-6 series, positioning them as next-generation cutting-edge models. Both models focus on improving programming, reasoning, tool usage, and complex task execution capabilities, while also lowering pricing so developers can use stronger models at a lower cost.

GPT-6 Sol is mainly aimed at tasks that require a higher level of intelligence while also being sensitive to response speed and cost. OpenAI states that, compared with the previous generation of models in the same tier, Sol has made notable progress in coding ability. It performs more reliably when handling complex software engineering tasks, analyzing code, and executing multi-step workflows.