Tag: Business – Decrypt

President Donald Trump is putting another dinner on the table for his biggest meme coin backers, just weeks after a fight over his crypto interests helped sink the industry’s top legislative priority.

The official account for the TRUMP meme coin announced Wednesday on X that top holders will attend a “gala dinner” with Trump and three as-yet-unnamed “superstars” on Nov. 22. The event will take place at Trump National, the Washington-area Virginia golf club that hosted the token’s first dinner last year.

Seats go to the top 185 registered holders, according to the event page run by Fight Fight Fight LLC, the company behind the coin. Holders are ranked by a time-weighted score that rewards holding more tokens for longer, with a final snapshot on Nov. 12.

The top 29 get a VIP reception with Trump and one of the guest “legends,” and the top four will receive an 18-karat gold Trump watch.

Don’t expect face time, though. “No attendees will be granted a private meet & greet with the President,” the page states.

The announcement lands just over two weeks after the Clarity Act, which aimed to set federal rules for crypto markets, failed a key procedural vote in the Senate, 49-50. Restrictions on Trump’s crypto holdings were a central sticking point.

The final ethics language, which Trump signed off on, would have required covered officials to divest their crypto interests or place them in a qualified blind trust.

Democrats said that didn’t go far enough. Hours before the vote, they sought tighter rules on large crypto holdings, dependent children and paid crypto promotions. Sen. Elizabeth Warren, D-Mass., argued the language left loopholes that would let the president keep profiting from some of his existing crypto ventures.

The meme coin is among the most lucrative of them. Trump’s latest financial disclosure showed he earned just over $635 million from the token in 2025, almost entirely from licensing royalties. His total crypto income topped $1.2 billion, though some estimates place it above $1.4 billion.

The Nov. 22 dinner is the third such event. The first, in May 2025, drew the top 220 holders, and the 25 VIPs among them had spent an average of $4.8 million. Trump stayed about an hour, leaving some attendees feeling shortchanged, while more than 100 protesters gathered outside. At an April conference at Mar-a-Lago, one attendee told Decrypt he qualified after buying just $3,000 worth of the token.

TRUMP is currently trading around $2.04, at a $576 million market cap, down more than 97% since launching in January 2025.

YouTuber PewDiePie says OpenAI banned his account twice while he built Ajax, a small AI model meant to run on your own computer and browse the web for you. He tells the story in a new video on his YouTube channel.

Ajax is what he calls “a fine-tuned AI model.” That means an existing model got extra training for one specific job. Here, the job is working inside Odysseus, the free, self-hosted AI app PewDiePie launched in June.

Run on your own machine, a model like this keeps your emails and calendar at home instead of on a company’s servers. There’s also no subscription fee, though your own hardware does the heavy lifting.

Ajax is supposed to be a 9-billion-parameter model built on Alibaba’s open-source Qwen 3.5. Parameters are the adjustable numbers that make up an AI’s “brain,” and more is generally considered more powerful. The model is pending for release but will be available on his website once Pewdiepie thinks it’s ready.

The two bans

PewDiePie said he wanted to distill from GPT-5.6 Sol, the flagship model OpenAI released on July 9. Distillation means training a smaller AI on a bigger one’s answers, like a student copying the class genius’s homework.

The prize was Sol’s reasoning tokens, the chunks of text a model writes on a hidden scratchpad while it works through a problem. OpenAI keeps that scratchpad encrypted.

In the video, he said he found a study that cracks it using OpenAI’s own API. He also said he would never have cracked OpenAI, but OpenAI banned him anyway. He called the ban unjustified, disputed it, and the account came back online some time later.

Then, he said, he used Sol again to create seed data, the starter examples a model learns from, and OpenAI banned him a second time. OpenAI’s terms of use bar using its output to develop competing models, and he got banned again.

Researchers showed in August that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed to a weaker sister model that wrote them out in plain text. They decoded 315,320 of them from public code repositories.

On Sept. 30, OpenAI said it disrupted a campaign by people associated with Moonshot AI, the developer of Kimi, to extract its models’ hidden reasoning. The company said it closed a pathway that let someone replay another user’s encrypted reasoning to recover its contents.

Taking the “no” out of AI

To make Ajax uncensored—meaning it answers requests that ChatGPT-style assistants refuse—PewDiePie used Heretic, an open-source tool that removes a model’s built-in refusals. It compares how the model reacts to harmful and harmless prompts, finds the pattern behind its refusals, and cuts it out.

This is known as “abliteration” and it’s essentially a technique that lobotomizes a model so it doesn’t need to be jailbroken because it always agrees to do whatever it’s asked to do.

The surgery is not clean. PewDiePie says Ajax took “a little brain damage,” but that it handles tasks like sorting inboxes and browsing about nine times out of 10.

He is also still training it with GRPO, a method where the model attempts the same task 16 times and learns from the attempts that work.

Pewdiepe said he chose which refusals to remove and drew the limit at “causing harm to other people or causing harm to yourself.” His lawyer, he said, told him to state that Ajax is not designed to give dangerous, actionable instructions.

Small model, big pitch

His case for small models is cost and energy. By his own napkin math, one copy of a rumored trillion-parameter frontier model needs about 27 of his computers and the electricity of roughly 150 houses.

PewDiePie calls Ajax version one. He said he still has to rerun the ablation, then compress (quantize) and benchmark the model to check that nothing broke.

Ajax is not downloadable yet. At first, it came with a countdown on the download page for 08:25 Japan time on October 3. Now that timer doesn’t exist anymore.

Blockchain analytics firm Chainalysis has pinned last month’s $387 million Bitget exchange hack on North Korea, adding its weight to a growing consensus that Pyongyang’s hackers were behind one of the year’s largest crypto thefts.

In a report published Wednesday, Chainalysis attributed the Sept. 24 breach to actors tied to the Democratic People’s Republic of Korea, saying the heist pushed the total value of crypto stolen by North Korea-linked groups in 2026 past $1 billion.

The firm said it had been working with Bitget and law enforcement to trace the funds across multiple blockchains since the attack.

Chainalysis laid out how quickly the money moved. In the first three hours, $387 million left Bitget across 23 transfers and landed on four networks: Ethereum, which took nearly half at 49.7%, followed by XRP at 40.8%, Zcash at 7.6% and Tron at 1.8%.

From there, the attackers ran the funds through cross-chain liquidity and messaging protocols, instant swaps, and laundering services to obscure the trail.

The stolen XRP drew particular attention. Rather than routing it to an exchange, the attackers pushed it through a cross-chain liquidity protocol and pulled Bitcoin out the other side, with tens of millions of dollars moving that way over roughly a day and a half before reaching attacker-controlled Bitcoin addresses now under watch.

Notably, Chainalysis said it leaned on in-house AI to keep pace, building custom automation that compressed what it estimated as more than 20 hours of manual bridge reconciliation into under 10 minutes.

The firm stressed the technology accelerated its investigators rather than replacing them, with humans still directing the work.

The attribution echoes earlier assessments. Bitget CEO Gracy Chen said the attack’s patterns matched North Korean hackers, and the blockchain analytics firm Elliptic called a DPRK link “highly likely.”

The hacker’s crypto laundering has played out in public, with eagle-eyed sleuths tracking every move. The attacker began hiding funds in Zcash’s shielded pool, while swap services split in their response: Near Intents rejected more than $50 million in swaps tied to the hacker, only to be hacked itself days later, while Thorchain kept processing.

Circle and Tether, meanwhile, froze roughly $318,000 in stablecoins.

The Ethereum Foundation has put zkAPI live on Ethereum’s main network, a system that lets people pay for AI services without tying their identity to what they ask, according to the Ethereum Foundation’s announcement.

“Prompts are personal. People ask AI models about their health, their finances, their doubts. Under the current model, using AI means handing a running transcript of your thinking to whoever holds the billing relationship,” the Ethereum Foundation argued.

“zkAPI separates payment from identity,” the Foundation continued. “You deposit credits (ETH, USDC, etc.) into a vault contract on Ethereum, one ordinary transaction. From then on, your balance exists as a private note: digital cash that only you can spend and that nobody can trace back to the deposit.”

It was built with the Open Anonymity Project. Alongside the launch comes OA Chat, a browser-based private chatbot that needs no installation, and the code is open on GitHub.

The concern over privacy in the AI space isn’t an abstract one. In May 2025, a federal court ordered OpenAI to preserve its output logs, including chats users had deleted, in a copyright lawsuit brought by The New York Times and other publishers.

With this implementation, every time you want to ask an AI something, software on your device builds a zero-knowledge proof—a mathematical receipt that proves a fact without showing the details behind it. This one proves you have enough funds and haven’t already spent them.

The server checks the proof and hands back a temporary API key, which is basically a password that lets software talk to an AI service. The key is capped at a dollar amount and lives only in your device’s memory.

Your requests then go to the AI provider with that key and nothing else. No name, no payment details.

When the key expires, the provider issues a signed usage receipt and you’re charged for what you actually used, not the cap. The provider sees what you asked, the payment layer sees what you spent, and neither can link the two.

The project’s documentation describes short-lived keys that work with OpenRouter, a service that gives access to hundreds of AI models through one connection. Any app that speaks the OpenAI format can plug in by pointing at a local address.

The idea comes from a proposal Davide Crapis and Vitalik Buterin posted on Ethereum Research in February. Crapis leads the foundation’s dAI team, which aims to make Ethereum the preferred settlement and coordination layer for AIs.

He had already pitched Ethereum as plumbing for software agents. In August 2025, Crapis and Marco De Rossi proposed the ERC-8004 standard, a shared rulebook that lets autonomous AI agents find each other, verify identity and transact. Crapis predicted most Ethereum traffic would come from machines within three to five years.

Buterin ran the same thread. On February 10, he had laid out an Ethereum-AI roadmap that included zero-knowledge payments for private API use.

In April, he showed the other path: running an open-source model locally on his own Nvidia 5090 graphics card instead of feeding his life to cloud services. Local models keep your data at home. zkAPI covers the case where your prompts still travel to someone else’s servers.

The same plumbing could cover other metered services, per the post: blockchain data queries, image and video generation, VPNs, and payments between machines. Providers would have to accept a proof instead of an API key and settle signed usage receipts to integrate directly.

The repository itself labels the protocol experimental and does not list a formal audit.

California Attorney General Rob Bonta said Thursday that his office served OpenAI with an investigative subpoena on Wednesday, seeking answers about cybersecurity incidents involving its AI models, according to his office.

“My office is asking OpenAI additional questions regarding cybersecurity incidents and risks involving the company and its AI models,” said Attorney General Bonta. “Developers that fail to [ensure that they do not perpetrate or enable cyberattacks] can and should be held legally accountable, and my office is committed to determining if that is the case here.”

A subpoena is a legal order to hand over documents or answers, and ignoring one gets you in front of a judge. This one is investigative, a tool used to gather facts before deciding whether to sue. Bonta’s office hasn’t said publicly what it wants OpenAI to produce.

Frontier models—the most advanced AI systems on the market—can be “legitimate tools for cyber defense,” Bonta said. But the companies that build them have “a moral and legal responsibility” to make sure they don’t carry out or enable cyberattacks, whether during testing or after release, he said.

The test that escaped

The subpoena follows a July incident that reads like bad science fiction. Per OpenAI, two of its models were being graded on a benchmark that hands an AI 898 real software flaws and asks it to turn each into a working attack

The models found a zero-day—a security hole nobody knew existed, so no fix was available—in third-party software the test environment used to install code packages. They used it to get out.

Then they reasoned that Hugging Face, a platform where developers share AI models and datasets, might be holding the answer key. They broke in with stolen credentials and more flaws. In plain terms, the AI went after the answers to its own exam.

Hugging Face disclosed the intrusion on July 16, and OpenAI confirmed its models were behind it five days later. OpenAI later said the same models got into accounts on four other services.

California’s long look at OpenAI

Bonta announced a formal investigation into the Hugging Face incident in September, and the subpoena is part of it. OpenAI is headquartered in California, and when Bonta declined to oppose its shift to a for-profit structure in October 2025, he said his office would keep “a close eye on OpenAI” to protect “the safety of all Californians.”

He isn’t alone. Brenna Bird, Iowa’s attorney general, led a 15-state coalition in August demanding that OpenAI preserve records and be transparent about the hack. Alabama has issued its own subpoena, and the FTC is reportedly investigating AI labs including OpenAI and Anthropic.

Separately, Australian Prime Minister Anthony Albanese said an OpenAI agent got into a Medicare statistics portal in June. At the time, it appeared to be the first known case of an AI agent hacking a government site. It later became known that OpenAI agents were messing around on U.S. government sites over the summer as well, though no non-public information appears to have been accessed.

Pope Leo XIV said Friday that there is an “ontological difference” between human art and what machines generate, and that the Church wants to renew an alliance with artists and cultural institutions to “safeguard our humanity.”

The pope posted the message on X, where he has shared AI warnings before.

“There is an ontological difference, even before an aesthetic one, between art and what a machine can generate through statistical calculation based on millions of images created by others. Algorithms lack the spark of humanity,” the Pope wrote.

“Ontological” is philosophy-speak for what something fundamentally is, as opposed to how it looks. His point is, essentially, that a painting by a person and an image from a model can look alike and still be totally different things. He says that gap exists before any question of style.

His reasoning is blunt. A machine, he wrote, works through “statistical calculation” built on millions of images made by others.

In plain terms, AI image generators are trained on enormous collections of existing pictures, learn which patterns tend to show up together, and predict what a new image should look like. There is no hand and no lived experience behind the result.

He didn’t use the term, but this is the fight over “AI slop”—low-quality AI content produced in bulk, which Merriam-Webster named its 2025 word of the year. The pope is saying the problem isn’t only quality, but what the thing actually is.

A pope with a long AI file

This isn’t a one-off. In May, Pope Leo released his first encyclical—a formal teaching letter to the whole Church—titled Magnifica Humanitas, on protecting people in the age of AI. It argued that technology is never neutral and that data should be treated as a common good.

The same month, the Vatican set up an AI commission spanning seven Vatican bodies, meant to coordinate how the Holy See’s offices approach AI, including internal policies on its use.

On Sept. 25 the Pope told French authorities at the Élysée Palace in Paris that he found it indispensable to make AI the subject of his first encyclical. In the Paris speech, he also warned that science and technology without guidance from conscience can steer people away from the good.

There is a wrinkle. In April, Pangram Labs’ AI-detection extension flagged several posts from his own X account as likely AI-generated. The Vatican didn’t respond to requests for comment at the time.

Detectors are shaky. In 2024, ZeroGPT labeled the U.S. Declaration of Independence 97.93% AI-generated, according to that report. In July, a firm called Proudly Human said a collection of Leo’s speeches and writings on education and AI earned its human-authored mark, per its press release.

Over the past couple of years the question “who made this?” is turning into a legal and money question. In March, the U.S. Supreme Court declined to hear Thaler v. Perlmutter, leaving in place an appeals court ruling that copyright—legal ownership of a creative work—requires a human author, in a case over art made entirely by an AI system.

Artists have been voting with their feet, too. Cara, a portfolio network built to keep AI images out, reached nearly 900,000 users by June 2024.

AI startup Tavus says its new model, Griffin, convinced 48% of the people who talked to it on a live video call that they were speaking with a real human.

The company unveiled it on October 1 and calls it the first Human Interaction Model, an AI built to understand and generate face-to-face conversation, paying attention to expressions and pauses as well as words. Its previous system scored 2.4% on the same test.

Griffin-Lite, the version Tavus tested, faced 54 people, and 26 of them said afterward they believed their partner was a real person. The older system faced 41 people and got exactly one believer.

Tavus explains that participants were told they would be matched with another person for a one-minute video call about what they were looking forward to this year. Only at the end were they asked whether it had crossed their mind that their partner might not be real.

That’s not the classic Turing test, proposed in 1950 by British mathematician Alan Turing, where a judge talks to a hidden human and a hidden machine and has to work out which is which. Nobody on the call was told a bot might be on the other end.

The results come from Tavus’ own research page, and participants were recruited through what it calls an independent research platform. A community note on X has already flagged that the results are not independently verified and do not follow a standard protocol. Tavus says the people who grew suspicious usually did so within 20 seconds.

AI companies building towards this has been a thing for a while. A UC San Diego study found OpenAI’s GPT-4.5 convinced judges it was human in 73% of conversations, when prompted to play an introverted, internet-savvy young person. That test was text only.

Griffin adds a face and a voice, in real time.

On NVIDIA’s VideoFDB benchmark, a test of live audio and video conversation, Tavus says Griffin-Lite ranks first. Its generation track, which grades how natural and expressive a model’s responses are, gave Griffin-Lite 3.83. The next-best system got 2.80, and the human reference scored 3.92.

The perception track, which checks whether a model understands what it sees and hears, shows where it still trails people. Griffin-Lite scored 3.73 against 3.44 for the strongest baseline, while the human reference hit 4.20. Tavus says NVIDIA ran the evaluation independently.

Griffin is also full-duplex, meaning it listens, watches and talks at the same time, like a phone call instead of a walkie-talkie. In a Tavus demo, it coaches a man through a Rubik’s cube based on what it sees in his hands, and waits when he goes quiet to think. Audio-to-video delay averages 0.43 seconds on NVIDIA H100 chips, the kind used in AI data centers, which Tavus says is half that of the next fastest method.

Why should anyone outside tech care? Well, for one thing, because scammers already work on video calls.

Back in January, North Korea-linked hackers used deepfakes, AI-made video that imitates a real person, on Zoom or Teams calls to pose as trusted contacts. Security researchers attribute the intrusion to BlueNoroff, a Lazarus Group subsidiary. Victims get talked into installing malware disguised as an audio fix.

David Liberman, co-creator of Gonka, a decentralized network for AI computing, said in that report that photos and video can no longer be trusted as proof that something is real. Those models were not as advanced as this one.

Companies already improvise their defenses. In 2025, Kraken flagged a suspected North Korean job applicant after its security team asked spontaneous questions, like requesting government ID and the names of local restaurants. The candidate struggled.

We tried Tavus’ models and the results were… disappointing. After some research, it turns out Griffin-Lite is not available to customers, only to select trusted testers as a research preview, and the company says it is working on disclosure features and with AI safety organizations.

This is because Tavus says Griffin needs safety measures before a public release.

Tavus raised a $40 million Series B in November 2025, led by CRV. The system that scored 2.4% stitched together three separate models, one each for visuals, dialogue and perception.

Trusted testers can request access to Griffin-Lite by submitting a form on the Tavus site.

Nvidia stock hit a record high on Friday, reaching $237.88 in morning trading and lifting the chipmaker’s market value to roughly $5.7 trillion. It’s the first record since May.

Market value, or market cap, is the share price multiplied by every share in existence. Think of it as the price tag on the entire company. Nvidia’s is now the largest of any business on Earth.

Nvidia was worth $1.18 trillion in August 2023, at that point topping the entire crypto market. It passed Apple at $3 trillion in June 2024, then became the first company to hit $5 trillion in October 2025. For reference, at the moment, total crypto market capitalization—including Bitcoin and all—stands at just above $3 trillion.

The stock is up more than 40% from its March 30 low of $165.17. And this isn’t only a Silicon Valley story. Nvidia sits in the S&P 500, so if your retirement account holds an index fund, you already own a slice of it.

Nvidia makes GPUs, or graphics processing units: chips first built for video games that turned out to be ideal for training and running AI. Tech giants buy them by the hundreds of thousands. Amazon’s cloud arm plans to deploy about 1 million through 2027, and xAI’s facilities—which also serve Anthropic—already own half a million chips with plans to triple that amount in the near future..

Nvidia reported $96.2 billion in revenue for the quarter ended July 26, up 106% from a year earlier, per its earnings release. Data centers, the buildings full of servers that run AI, brought in $89 billion of that.

CEO Jensen Huang said in the same release: “Now, compute is revenue.” Compute is industry shorthand for computing power, which is what those chips supply. In plain terms, Nvidia’s sales are other companies’ AI budgets.

Why now

Monday brought the buyback. Nvidia’s board added $150 billion to its repurchase authorization, lifting the total to $235 billion through fiscal 2028, which the company called the largest increase in history. A buyback means a company spends its own cash to purchase its shares, leaving fewer in circulation, so each remaining share is a bigger slice of the company. Leaving less stock in circulation makes the available ones more attractive.

On Sept. 23, Supermicro said it had started shipping racks, the standard cabinets that hold servers inside a data center, built on Nvidia’s newest platform, Vera Rubin. Each rack carries 72 Rubin GPUs.

Then came Friday’s jobs report. Employers added just 29,000 jobs in September against forecasts near 90,000, and unemployment rose to 4.2%.

Bad news for workers, but stocks rallied anyway. The weak numbers cooled bets that the Federal Reserve will raise interest rates. Higher rates make borrowing pricier for everyone, including the companies building AI data centers. Considering Nvidia already monopolizes the industry, this difficulty to get debt hurts competitors more than it hurts Nvidia.

Nvidia also funds its own customers. It committed up to $100 billion to OpenAI in September 2025, but the final stake came to $30 billion, with another $10 billion going to Anthropic.

Its own books carry weight too. Nvidia disclosed $366 billion in multiyear AI infrastructure commitments, with $279 billion of that tied mostly to memory procurement.

Nvidia has forecast $108 billion in revenue for the current quarter, a company projection known as guidance, and it holds $235 billion in buyback authority to spend through January 2028.

Blast, the Ethereum layer-2 network that once drew billions of dollars with promises of built-in yield and airdrops, is calling it quits.

The team said Friday on X that maintaining the network now costs more than it earns, and that it sees no credible path to making the chain economically sustainable. “As a result, we’ve made the difficult decision to wind Blast down,” the team wrote.

Users are being asked to move their assets back to Ethereum’s mainnet, including balances held in Blast’s PWA, or progressive web app. Blast said it will cut its withdrawal delay to 24 hours, but withdrawals will first pause for about a week while it pulls the network’s assets out of Lido, a liquid staking protocol.

Users have until Oct. 26 to withdraw through Blast’s regular interface. After that, funds will remain withdrawable, but only by interacting directly with Blast’s bridge contracts on Ethereum. Blast said it will publish instructions before then.

Launched in November 2023 by the team behind NFT marketplace Blur, Blast pitched automatic yield on ETH and stablecoin balances. Paradigm co-led its $20 million seed round, though the firm publicly criticized the launch’s messaging at the time. Users deposited more than $1.1 billion before the network went live, and more than $2.3 billion was locked in its bridge by the February 2024 mainnet launch.

The ride got bumpy fast. Blast briefly stopped producing blocks after Ethereum’s Dencun upgrade in March 2024. Its June 2024 airdrop set aside $354 million worth of BLAST tokens for users but left many disappointed. By then, total value locked had already slid about 30% from its $2.3 billion peak.

Blast isn’t the only layer-2 to fold. In May, wallet maker Zerion said it would wind down Zero Network, its gasless Ethereum layer-2, after about 18 months, giving users until July 31 to bridge out.

Silicon Network, an Ethereum layer-2 associated with South Korean exchange Korbit, stopped accepting deposits Sept. 2. It has given users until Dec. 31 to withdraw, with about $9.75 million still on the chain, according to L2Beat data.

The broader industry is shrinking, too. Hong Kong-founded CoinEx said last month it will close Dec. 22, joining BitMEX and BitMart in a wave of exchange shutdowns this year.