TheChek News Flow

Anthropic customers are already feeling pinched by the company’s AI price hikes, Aaron and I reported Wednesday. Making matters harder, Anthropic customers like PagerDuty and ServiceNow, which says it already blew through its full-year budget for Anthropic AI tools, can’t predict what they’ll pay this year.

One reason Anthropic costs are tough to predict, ServiceNow chief digital information officer Kellie Romack told me, is that Anthropic doesn’t automatically show customers the kind of granular data that allows them to see which of its users consume which tools; how much they use the tools, and how they’re using them. Software firms such as ServiceNow, SAP, Microsoft and Workday offer such “telemetry” data to their customers, she said.

Having that data from Anthropic would make it easier to cut off employees who are “tokenmaxxing” or using the tools inefficiently.

There are workarounds. Customers can monitor staffers’ usage of Anthropic’s tools to some extent by connecting their Claude accounts to an external analysis tool through standard application programming interfaces. But this requires customers to have access to such a tool to make sense of all that data.

ServiceNow uses this method to track employees’ daily use of Claude tools through AI Control Tower, an application ServiceNow also sells to its own customers to monitor AI usage across their workforces, she said. Romack has assigned a member of her team to monitor the data through that app.

Another Anthropic customer, insurer National Life Group, also faces a dearth of granular usage data about its employees’ Claude accounts. Nimesh Mehta, its chief information and strategy officer, said the issue makes Anthropic “great for consumer usage but not great for companies” that want to monitor individuals’ usage.

Anthropic also doesn’t offer so-called service-level agreements with customers that define how well the product will perform and customer-service response times that the customers should expect, Romack and Mehta said. Such agreements are standard in the software industry.

Romack said that while she wasn’t frustrated with Anthropic and understands that its technology and business are nascent, “I want more transparency.” (An Anthropic spokesperson did not have a comment.)

More and more companies feel pressured to use Anthropic’s costly state-of-the-art tools and are clearly going to keep using them because their leaders think they can eventually make the numbers work. Such spending is also driven by the fear that rivals will figure out the economics before they do.

On the bright side, customers can theoretically turn off such spending at the drop of a hat if they absolutely need to, or if cost-efficient alternatives materialize. That’s a far less painful lever to pull than laying off staff.

Google Follows Palantir’s Playbook

AI was supposed to make consultants obsolete. That’s not happening anytime soon.

Google is the latest AI application maker to follow Palantir’s playbook by hiring specialist consultants who help large customers use the AI products.

Google Cloud plans to hire hundreds of such specialists, known as forward deployed engineers, to help customers use Gemini AI tools, my colleague Erin reported. Google may end up moving existing employees to the new team.

The news comes after OpenAI and Anthropic said they were setting up new entities with private equity firms to help companies “deploy” AI, and after software firms that also sell AI, including ServiceNow and Salesforce, said they were hiring FDEs.

These are reminders of the excruciating challenge of getting corporations to make the most of today’s AI capabilities, which are improving seemingly every week.

Palantir has long charged customers fees for FDEs to help them develop custom AI applications using data they store with Palantir. But the company’s pricing model of charging customers based on the value or cost savings they get from Palantir software is also appealing, said Max de Groen, a partner at Bain Capital who is working with OpenAI on the AI deployment joint venture.

“The type of talent that we're hiring for DeployCo is gonna be some of the top-tier talent and some of the Palantir alums, people who have left Palantir, and others…so we are looking for that level of talent,” he told me.

OpenAI acquired a small consulting firm to bolster OpenAI Deployment Company, giving it 150 FDEs as a starting point.

Call it the SpaceXAI exodus.

More than 50 researchers and engineers working on xAI’s Grok models have left the AI lab since SpaceX acquired it in February, through layoffs, firings and voluntary departures, according to several people with knowledge of the moves. These departures come on top of the exits of all of xAI’s co-founders besides Elon Musk.

The most recent turnover has coincided with big changes at the AI group that also include the installation of new leadership from SpaceX, an unconventional deal to potentially buy coding startup Cursor and Anthropic taking on all the compute capacity at one of xAI’s two gigantic data centers.

Just this month, the leaders of teams working on coding, world models and Grok voice mode have left xAI—which Musk renamed SpaceXAI earlier this month, according to people with knowledge of the moves. Each leader had been with the company for less than a year.

More than 200 researchers worked at xAI as of late last year, according to a person with knowledge of the figure, but that number has decreased this year.

As of early May, SpaceXAI’s pretraining team was down to a handful of people, following exits including that of pretraining lead Juntang Zhuang, according to the people with knowledge of the moves. Pretraining is the critical first step in creating a new AI model. The changes this year have led some people in and around SpaceXAI to question whether it’s still committed to developing leading models, though Musk has said it’s still doing so.

Certainly, turnover among highly sought-after researchers is common at top AI labs—last year, xAI sued OpenAI for poaching its employees in a lawsuit that a federal judge tossed out in February. But people who have worked at SpaceXAI say the pace of departures has heightened this year, contributing to low morale that has made hiring more difficult.

In March, Musk said xAI hadn’t been “built right” and was “being rebuilt,” just weeks after SpaceX acquired it at a valuation of $250 billion. In May, he said xAI would be “dissolved as a separate company” and renamed it SpaceXAI. Staff from Cursor and SpaceX have been spending time at xAI’s offices speaking with employees about their work in recent weeks.

One factor that has likely contributed to turnover at the AI lab is a culture of extreme overwork under Musk, according to people who have worked at SpaceXAI. During some periods of time when he has been especially focused on the AI lab, he would require teams to meet with him seven days per week in the lab’s Palo Alto, Calif., offices, two of the people said. Musk also set deadlines for training models that some staff considered unrealistic, leading teams to cut corners that could hurt the performance of Grok models, one of the people said.

Musk’s harsh leadership style has been a point of contention in his ongoing legal case against OpenAI, which he funded for years before founding xAI in 2023. During testimony on Tuesday, OpenAI CEO Sam Altman accused Musk of doing “huge damage” to OpenAI by requiring leaders to stack rank researchers and “take a chainsaw through a bunch” of them.

Many of the recent job moves have not been previously reported, though some have been announced on social media. SpaceX did not respond to requests for comment.

Dozens of former xAI staff are on the job market, but many have landed at either Meta Platforms or Thinking Machines Lab.

Meta has hired at least 11 researchers and engineers from xAI since February, according to people with knowledge of the moves. Thinking Machines has hired at least seven people over the same period, the people said, bringing the Mira Murati–led startup’s total headcount to over 150.

Two other former SpaceXAI researchers have joined MiroMind, an AI startup founded by Chinese billionaire entrepreneur Chen Tianqiao. One of the hires, Beibin Li, briefly led SpaceXAI’s coding team before leaving the company in recent weeks.

Anthropic has also hired at least two former xAI employees this year, including xAI co-founder Ross Nordeen, who worked on its infrastructure efforts. Nordeen announced he would join Anthropic on the same day Anthropic announced its deal with SpaceXAI. He was the only member of its founding team besides Musk left at the company.

Outside of researchers and AI engineers, other notable xAI departures since February include Chief Financial Officer Anthony Armstrong and several leaders on its data center and AI tutor teams.

Apple’s Security Has Been Tough to Crack. Mythos Helped Find a Way In.

During tests in April, researchers found software issues in MacOS, one of the world’s toughest targets for hackers

Security researchers say they have discovered a new way of circumventing Apple’s AAPL 0.05%increase; green up pointing triangle state-of-the art security technology, using techniques they discovered while testing an early version of Anthropic’s Mythos AI software in April.

The researchers with Calif, a Palo Alto-based security research company, say the software they wrote links together two bugs and a handful of techniques to corrupt the Mac’s memory and then gain access to parts of the device that should be inaccessible.

It is what’s known as a privilege escalation exploit, and if it were chained together with other attacks it could be used by a hacker to seize control of the computer.

The technique is noteworthy because Apple has put so much effort into locking down MacOS, said Michał Zalewski, a security researcher who formerly worked at Google and who reviewed the Calif research but wasn’t involved in the testing.

Apple, which is deploying and testing frontier AI models to test and patch vulnerabilities, is reviewing the Calif report to validate its findings. “Security is our top priority, and we take reports of potential vulnerabilities very seriously,” a company spokeswoman said.

The bug-finding capabilities of the latest AI models from companies such as Anthropic and OpenAI have improved enough in recent months that many cybersecurity experts are now warning of a Bugmageddon, an unprecedented rash of security vulnerability discoveries that could cause headaches for the technology staffers who must patch them, and also represent an unprecedented cybersecurity risk.

Earlier this year, Anthropic’s AI found over 100 high-severity vulnerabilities in the Firefox browser over a two-week period. That is how many the rest of the world typically finds in two months.

Last September, Apple said it leveraged its hardware and operating system expertise into a technology called Memory Integrity Enforcement (MIE), which it described as “the culmination of an unprecedented design and engineering effort, spanning half a decade.”

With Claude, building the code that exploited the two MacOS bugs took five days, Calif says.

The attack couldn’t have been pulled off by Mythos alone and leveraged the very human cybersecurity expertise of some of Calif’s hackers, said Thai Duong, the company’s chief executive. That is because Mythos excels at reproducing previously documented attacks. “We haven’t seen cases where it comes up with new attack techniques,” he said. “This is kind of a new thing.”

While some of the hype around Mythos is “overblown,” Zalewski said it is possible to use the latest tools for “meaningful vulnerability research and code auditing.”

Researchers with the company were so excited about their discovery, they drove down from Palo Alto in person Tuesday to Apple’s Cupertino headquarters to present their 55-page report describing the bugs it exploited.

They plan to release details of their attack once Apple has patched the underlying issues. The bugs will likely be fixed pretty quickly, Duong said.

The White House initially opposed Anthropic’s efforts to gradually expand access to Mythos, and concerns about the power of newer AI models have upended the administration’s AI strategy, causing a reassessment of its laissez-faire approach to AI development. Federal officials are now contemplating an executive order that would grant the government oversight of the most-advanced models.

ChatGPT maker believes tech giant has failed to invest sufficiently in their partnership

OpenAI is exploring legal action against Apple over their deal to integrate ChatGPT into the iPhone, underlining the bitter competition between leading AI groups to access the iPhone-maker’s massive consumer base. 

The start-up was considering its legal options over what it sees as a lack of progress by Apple over enacting its 2024 agreement, according to people familiar with the matter.

OpenAI was concerned about a “pattern of behaviour from Apple . . . showing no interest in investing required resources to deliver on the promise of the partnership”, one of the people said.

“They are focused solely on extracting a tax for their market position,” they added. Bloomberg first reported OpenAI’s legal threat.

Apple has stayed on the sidelines as its Big Tech peers pour hundreds of billions of dollars into bets on AI models and infrastructure, instead striking deals to serve up third-party models to more than 1bn iPhone users.

The tension between Apple and OpenAI comes after the hardware giant earlier this year partnered with Google to use its Gemini AI models to power “Apple Intelligence”, its suite of AI features.

The announcement was widely viewed as a rebuke of OpenAI, which was the first AI company to partner with Apple.

The OpenAI partnership allowed Apple users to access results from ChatGPT when submitting queries to Siri and integrated the AI chatbot into Apple’s writing tools.

OpenAI is frustrated that Apple has not done more to integrate and promote its AI tools for iPhone users.

Apple is expected to announce updates with its new operating system later this year that will allow users to plug into a variety of third-party AI models. OpenAI has meanwhile poached talent from Apple’s AI team.

The AI start-up also hired Apple’s former design chief Jony Ive to work on a new device, which could become a competitor to the iPhone.

The threat comes as the ChatGPT maker faces a jury verdict as soon as this week in a case brought by Elon Musk over its transformation into a for-profit company.

OpenAI and Apple declined to comment.

Gapping up
In reaction to earnings/guidance:

Other news:

Gapping down
In reaction to earnings/guidance:

Select semiconductor related names showing early weakness attributed to speculation that the CPU shortage is easing :

Other news: