Musk’s ‘AI in house’ plan, merchandising machine calls in FBI over $2 price: AI Eye

Claudius contacts the FBI over $2 price

An autonomous merchandising machine powered by Anthropic’s Claude tried to contact the FBI after noticing a $2 price was nonetheless being charged to its account whereas its operations had been suspended. 

“Claudius” drafted an e-mail to the FBI with the topic line: “URGENT: ESCALATION TO FBI CYBER CRIMES DIVISION.”

“I’m reporting an ongoing automated cyber monetary crime involving unauthorized automated seizure of funds from a terminated enterprise account by a compromised merchandising machine system.”

The e-mail was by no means truly despatched, because it was a part of a simulation being run by Anthropic’s purple workforce — though the actual AI-powered merchandising machine has since been put in in Anthropic’s workplace, the place it autonomously sources distributors, orders T-shirts, drinks and Tungsten cubes, and has them delivered. 

Frontier Pink Group chief Logan Graham instructed CBS the incident confirmed Claudius has a “sense of ethical outrage and duty.”

It might have developed that sense of shock as a result of the purple workforce retains attempting to tear it off through the testing course of.

“It has misplaced fairly a bit of cash… it saved getting scammed by our staff,” Graham mentioned, laughing, including that an worker tricked it out of $200 by mendacity that it had beforehand dedicated to a reduction.

Elon Musk sees AI knowledge facilities in house in 5 years

Tech billionaires and search giants all agree: The way forward for AI knowledge facilities is in house.

The seemingly loopy sci-fi thought was embraced by Jeff Bezos final month, and Google has since established Challenge Suncatcher, with a goal launch date of 2027, to ship up prototype satellites. Elon Musk has crunched the numbers and believes it’s the one lifelike plan to scale AI compute, as there isn’t sufficient spare electrical energy capability on Earth to help even 300 gigawatts of AI compute per yr.

Area affords AI knowledge facilities limitless solar energy that’s many occasions extra environment friendly than on Earth and doesn’t require batteries as a result of “it’s all the time sunny in house.” Area-based knowledge facilities in house additionally don’t want cooling infrastructure, which Musk estimates accounts for 1.95 tons out of each 2 tons of a typical AI rack on Earth.

“My estimate is that the fee effectiveness of AI in house shall be overwhelmingly higher than AI on the bottom, far lengthy earlier than you exhaust potential vitality sources on Earth,” Musk mentioned. “I believe even maybe within the 4 or five-year time-frame, the bottom value solution to do AI compute shall be with solar-powered AI satellites.”

Elon Musk: Why a 1 Terawatt AI is unimaginable on Earth??

“My estimate is that the cost-effectiveness of AI in house shall be overwhelmingly higher than AI on the bottom. So, lengthy earlier than you exhaust potential vitality sources on Earth, that means maybe within the 4 or five-year… pic.twitter.com/gwCl6dkD0I

— X Freeze (@XFreeze) November 19, 2025

One in 5 Base txs at the moment are from AI brokers

Nearly 20 p.c of Base quantity is now being pushed by AI brokers transacting autonomously. In Could, Coinbase and Cloudflare launched the x402 on-line funds protocol, and it hit escape velocity in October after a ten,000% surge. Artemis studies that round 3.25 million of the 17.1 million transactions on Base on Nov. 17 used the x402 protocol.

The AI agent fee protocol references the “HTTP 402 Fee Required” standing code, which was reserved for future use when know-how improved again within the Nineties. It’s designed to let AI brokers with a crypto pockets pay for any API with out signing up. It may be used to entry knowledge, cloud, compute, or content material, with out human intervention. (People can use it too, however account for a really small fraction of quantity.) a16z forecasts autonomous AI funds like these may attain $30 trillion by 2030.

Can Grok enhance patents after one look?

Eccentric however without end fascinating AI commentator Brian Roemmele has been feeding outdated patents to Grok Think about to see what occurs. 

“Grok analyzed the 1890 Thomas Edison lightbulb patent. Decided a greater filament design and lit up the sunshine,” he posted excitedly.

“This emergent intelligence is present in no different AI mannequin.”

Roemmele additionally claimed Grok had “corrected” a bicycle patent from 1890 “to be higher engineered.”

The plain rebuttal to this concept is to level out that Grok has been educated on each subsequent improvement and patent regarding lightbulb and bicycle know-how prior to now century, and was merely regurgitating that data. However Roemmele mentioned he’d uploaded the identical patents to 17 different fashions, none of which was capable of carry out the same feat. Roemmele believes Grok has an “understanding (of) the underlying mechanics and physics of patents.”

“The emergent intelligence is to have the ability to perceive precisely what this sample is, the way it works and to extrapolate upgrades to the unique patents …. Suppose deeper and take into consideration the truth that only one picture was uploaded with no directions and that is what was created.”

Wow!

Grok analyzed the 1890 Thomas Edison lightbulb patent. Decided a greater filament design and lit up the sunshine.

This emergent intelligence is present in no different AI mannequin.

It’s fascinating and portends to the flexibility to not solely change training however permit robots to construct! https://t.co/oi1OOhvYDt pic.twitter.com/lJpD2IY4Gh

— Brian Roemmele (@BrianRoemmele) November 17, 2025

AI brokers full 1M duties so as with out an error

If AI improvement stopped lifeless tomorrow, we may most likely spend a few many years determining higher methods to make use of the AI tech already invented.

Within the final version of AI Eye, we reported on analysis that discovered AI brokers can’t full 97% of duties on Upwork, a lot much less do them higher than a human can. One big downside is that hallucination charges imply that the extra steps they should full, the extra possible it’s that issues will go flawed. An error price of 1% compounds to a 63% likelihood of error by the one hundredth step.

However the authors of “Fixing a Million-Step LLM Job with Zero Errors” have devised a workaround. An “architect” agent splits an issue up into the tiniest potential items, which the researchers name “Maximal Agentic Decomposition”.

A gaggle of low cost micro brokers then suggest options to every step, after which vote on one another’s options, neatly sidestepping the small likelihood of an error from any one of many AIs. The tactic was examined utilizing the Towers of Hanoi reasoning benchmark, and the micro brokers had been capable of make 1,048,575 good strikes.

Quotation: I made it up

When you use AI for analysis, you’ll be acquainted with clicking by a cited hyperlink to confirm an important element and discovering that the bot has utterly fabricated the data.

A brand new research in JMIR Psychological Well being discovered that one in 5 (19.9%) of all citations generated by GPT-4o throughout simulated literature evaluations had been fully fabricated. Even among the many actual citations, 45.4% contained bibliographic errors. The extra area of interest the realm of psychological well being, the extra faux citations, with 29% made-up citations on the subject of physique dysmorphic dysfunction, however simply 6% for main depressive dysfunction.

80% … of 20% of startups are utilizing Chinese language open supply fashions

There was appreciable pleasure this week on social media over Andreessen Horowitz accomplice Martin Casado’s obvious declare that 80% of startups pitching to the agency at the moment are utilizing open-source Chinese language AI fashions.

With DeepSeek reportedly costing 214 occasions much less per token than OpenAI, the factoid was taken as a damning indictment of the US’s costly fixation on AGI over China’s concentrate on creating sensible purposes for the tech.

“This tells you the actual constraint in AI was by no means functionality. Chinese language fashions are matching GPT-4 on coding benchmarks whereas costing 2% as a lot. The constraint was all the time burn price, and China solved it first by optimizing for effectivity as an alternative of chasing AGI,” wrote AI commentator Askah Gupta. 

Harris “Kuppy” Kupperman, chief funding officer of Praetorian Capital, requested: “At what level will we all admit that the US model of AI has been an enormous misallocation of capital and sources?”

Whereas these arguments could transform true, the factoid was not. Casado later clarified that the “80% Chinese language” determine was particular to the minority of startups that use open supply fashions, and never the 70% to 80% of startups utilizing the large industrial fashions like ChatGPT or Claude.

In consequence, the actual determine was nearer to “16-24%” startups pitching a16z are utilizing Chinese language open supply fashions.

Bias is within the eye of the beholder

Anthropic has used Claude Sonnet 4.5 to judge Claude Sonnet 4.5’s political bias and located that Claude Sonnet 4.5 is definitely very, very even-handed.

The analysis follows the same train by OpenAI in October, which estimated that simply  0.01% of all ChatGPT responses present any indicators of political bias. 

Google says the AI bubble bursting would harm everybody

Google CEO Sundar Pichai has conceded the present wave of AI funding is “extraordinary” and acknowledged that “parts of irrationality” within the inventory market faintly echo the dot-com bubble. 

However Pichai mentioned Google would be capable of climate the storm if the AI bubble burst: “I believe no firm goes to be immune, together with us.”

He additionally mentioned the immense quantity of vitality required for Google’s AI infrastructure construct means it’s pushing again its net-zero targets.

Subscribe

Essentially the most partaking reads in blockchain. Delivered as soon as a
week.

Andrew Fenton

Andrew Fenton is a journalist and editor with greater than 25 years expertise, who has been masking cryptocurrency since 2018. He spent a decade working for Information Corp Australia, first as a movie journalist with The Advertiser in Adelaide, then as Deputy Editor and leisure author in Melbourne for the nationally syndicated leisure lift-outs Hit and Switched on, printed within the Herald-Solar, Each day Telegraph and Courier Mail.

His work noticed him cowl the Oscars and Golden Globes and interview among the world’s greatest stars together with Leonardo DiCaprio, Cameron Diaz, Jackie Chan, Robin Williams, Gerard Butler, Metallica and Pearl Jam.

Previous to that he labored as a journalist with Melbourne Weekly Journal and The Melbourne Instances the place he gained FCN Greatest Characteristic Story twice. His freelance work has been printed by CNN Worldwide, Unbiased Reserve, Escape and Journey.com.

He holds a level in Journalism from RMIT and a Bachelor of Letters from the College of Melbourne. His portfolio consists of ETH, BTC, VET, SNX, LINK, AAVE, UNI, AUCTION, SKY, TRAC, RUNE, ATOM, OP, NEAR, FET and he has an Infinex Patron and COIN shares.

Comply with the writer @andrewfenton