Anthropic's Growth Is Slowing

The episode begins with a report that anthropic's Q3 revenue reached $16.8 billion, up from $11.5 billion in Q2. Despite being profitable on an adjusted basis for the second straight quarter, growth has slowed compared to earlier in the year. The company hosted a full day of investor presentations emphasizing safety, rapid growth, and efficiency to prepare for its upcoming IPO scheduled for November. Investors expressed some disappointment, as recent revenue figures, while impressive, fell short of the more bullish predictions earlier in the summer. Expected annualized revenues sit around $65-$74 billion depending on estimates, but some market watchers had hoped for figures exceeding $100 billion by year-end. The slowdown reflects the impact of a more challenging macroeconomic environment, rising interest rates, and concerns about AI infrastructure spending plans, which could see anthropic limiting their capital expenditure. Despite this, anthropic reported being cash flow positive, with an impressive ratio of revenue to expenses for compute, and an operating margin of roughly 11%. They also noted that a share of their revenue comes from distillation, with 5-7% from this activity, which involves other companies using anthropic's AI to train their models—a practice they expect to decline.

The episode shifts to a criminal case involving Abdul Rahman Al-Fazan, who, along with family members, booked one-way tickets from the US to Saudi Arabia. The FBI arrested him on October 9th, just days after he reportedly purchased tickets for October 12th. His attorneys sought to seal related filings to prevent him from fleeing, citing concerns that knowledge of the case could motivate him to escape. Additionally, there was a report from a San Francisco playground where a stranger tried to befriend an OpenAI employee, asking questions about their work. It was revealed this stranger was likely a spy attempting to steal corporate secrets, a scenario that illustrates the targeted risks faced by AI companies. OpenAI has taken measures, including training employees to recognize espionage tactics, and has previously arrested employees for planting listening devices. These threats are heightened by the industry's rapid hiring of foreign nationals, notably Chinese talent, which some see as a security vulnerability. The episode emphasizes the importance of insider risk management and the ongoing security challenges AI companies face against both external espionage and employee poaching.

Amazon announced the launch of two new AI wearables, the Bee Frontier wristband priced at $75 and the Bee Eclipse magnetic recorder at $100, with a subscription service costing $12 per month. The devices are Amazon’s first since its acquisition of the wearable maker B. The B Frontier can be worn on the wrist or pinned, while B Eclipse attaches magnetically to a phone, aiming to capture conversations and turn them into summaries, notes, and action items. Both devices leverage AI for note-taking, with the ability to generate summaries and suggested to-dos, and are designed with privacy in mind. Amazon asserts that conversations are processed in real time, with transcripts encrypted and audio recordings not stored, unless the user opts to connect to Alexa or other Amazon services. They also mentioned that users must give permission for recordings and can delete transcripts and summaries at any time. The devices are positioned as productivity tools but raise privacy questions amid growing surveillance concerns, especially considering recent backlash against such tech. Amazon emphasizes the security architecture used to protect data, asserting that audio is not stored, and that transcripts are encrypted, but acknowledges the limits of privacy depending on user settings.

The episode concludes with promising early reviews of Nvidia’s RTX Spark superchip, which aims to bring AI processing onto personal hardware, potentially revolutionizing how AI models are run. Tests show the RTX Spark outperforms Apple’s M5 Max MacBook Pro by roughly 25% in AI completion tasks with Gwen models. The platform supports AI workloads traditionally handled in cloud data centers, but now on local PCs, due to its powerful GPU and memory architecture. It’s positioned as ideal for privacy-sensitive applications, offline use, or those seeking tinkering freedom. The cost of these PCs, like Microsoft’s Surface Laptop Ultra at $2,600, reflects high-end specs, with some models supporting up to 128GB of unified memory—highlighting the trend of premium pricing for local AI hardware. AMD's latest chips push similar boundaries, but Nvidia currently offers a more accessible option to consumers wanting local AI. Experts suggest this hardware could significantly alter the future of personal computing and AI deployment, by reducing dependency on data center infrastructure and emphasizing on-device data security.