tech
How technology works, and the ways it is impacting the world
Ben Thompson and Andrew Sharp analyze Anthropic's S1 filing, noting that while the core subscription and API business is extremely strong and profitable, R&D and training costs are exploding and remain out of the company's control due to intense frontier competition.
Ben Thompson breaks down the financial disclosures in Anthropic's S1, revealing that about 50% of revenue comes through the big three cloud providers (AWS, Azure, Google) and highlighting how massive convertible loans from Amazon resulted in significant repricing losses.
Andrew Sharp and Ben Thompson discuss Meta's recent quarterly earnings, where CEO Mark Zuckerberg announced aggressive increases in capital spending for AI infrastructure, accompanied by a new class of non-voting stock to fund future acquisitions and cash needs.
Ben Thompson posits that Netflix is a downstream casualty of the AI boom because Meta, under intense investor pressure to fund massive AI capital expenditures, is dialing up engagement and ad loads on Reels, capturing consumer eyeballs at Netflix's expense.
Netflix reported earnings showing fully saturated U.S. subscriber numbers with no additions this year, a 10% miss on engagement targets, and an advertising revenue miss down to $2.7 billion. Ben Thompson and Andrew Sharp discuss the company's reliance on continuous price hikes and how they face fundamental disadvantages in the ad-supported and engagement-driven internet landscape.
Media mogul John Malone predicted an upcoming global output deal and an $8 to $10 billion investment by SpaceX into David Ellison's Skydance to help deleverage the company's heavy debt pile. Ben Thompson and Andrew Sharp note the deep financial ties between Elon Musk and Larry Ellison.
The NSA issued an alert advising organizations using PHP for their systems to prepare for updates or take systems offline to address two critical vulnerabilities. Andrew Sharp and Ben Thompson discuss the severe cybersecurity risks these flaws pose to millions of websites running older PHP versions.
A certificate authority in India issued foundational 256-bit encryption certificates using extremely short keys due to a software error that misinterpreted the bit length. Ben Thompson and Andrew Sharp explain that while major browsers have updated to exclude them, deeply embedded Internet of Things devices remain permanently vulnerable.
After trying Meta's upcoming lightweight VR glasses in San Francisco, Ben Thompson described the resolution as breathtaking but expressed disappointment with the precarious fit and insubstantial feel. Andrew Sharp and Ben Thompson debate the tradeoffs of the headset compared to the Apple Vision Pro.
SpaceX AI has purchased Dot.com, which now redirects to the Grok bot download page. This acquisition occurred around the time of OpenAI's announcement of its new agent product, 'Dots'. The speakers find this move delightful and indicative of the 'wild west' nature of the current AI ecosystem.
Ben Thompson expresses disappointment with OpenAI's consumer-facing agent strategy, citing a lack of focus on practical everyday tasks and an aesthetic that feels dystopian. He believes companies like Meta and Google are better positioned in the consumer AI space.
Andrew Sharp argues that the increasing complexity and difficulty of using current technology, particularly mobile apps, creates a significant opportunity for AI agents. He suggests that agents should be marketed as 'pain relief' for everyday life, contrasting with the 'grind lifestyle' marketing often seen.
Andrew Sharp shares that his wife has become a user of the Muse agent after initially questioning its necessity. She found it useful after trying it for a couple of days, indicating potential for broader consumer adoption.
The podcast discusses the organizational structure of AI agents, comparing Codex's project-aligned approach to Claude's better context management for long threads. The speakers prefer organizing AI work into distinct threads per task, finding it more effective than massive, ongoing chat threads.
Despite the conceptual appeal of an overarching agent like Dots to manage other AI projects, the speaker found Dots to be poorly executed. It seemed unaware of his existing projects, focused on the wrong things like calendar conflicts, and requested unnecessary permissions, leading to a 'confused and confusing' user experience.
The speaker predicts that the AI agent paradigm will ultimately win over the current app-based system, not due to market forces or economic needs, but simply because it will offer a superior user experience. This improved experience is necessitated by the current poor state of technology and user interfaces.
Anthropic is stressing that Fable 5.1 and Mythos 5.1 are the exact same AI model. This clarification comes as companies face increasing pressure to justify and control AI usage costs, especially with the rise of more capable, lower-cost models from China.
Anthropic's Fable 5.1 model is reported to be more skilled in handling complex programming tasks, such as software projects and code reviews that span an entire application. The updated version also shows improved capabilities in scientific applications, including experiment design and data interpretation from dense diagrams.
Ben Thompson detailed his extensive efforts to reorganize his home's server rack and network infrastructure, involving the rewiring of 70-80 Ethernet runs. He emphasized the importance of network resilience, having recently incorporated a secondary internet source via Starlink.
Andrew Sharp expressed curiosity about Anthropic's strategy to address customer complaints regarding high token costs and frequent server downtime, especially given the improving performance of competitor models. Ben Thompson noted that Anthropic seems to lack sufficient compute resources, leading to these issues.
Ben Thompson humorously described Anthropic's release strategy for its Mythos model, suggesting a pattern of 'scaremongering' to gain attention from governments. He questioned the emphasis on the model's dangerous capabilities and noted the low adoption rate among customers.
Meta has agreed to a record $17.1 billion settlement with 29 states to implement stricter social media usage limits for children. This includes a two-hour daily limit on Instagram and Facebook, productive pauses, nighttime blocks, and restrictions during school hours. The settlement also mandates stronger age verification, safer content controls, and improved parental tools.
Ben Thompson expresses reservations about the Meta settlement, questioning whether it shifts parental responsibility to corporations and raises First Amendment concerns regarding content restrictions. He notes that settlements like this can have the force of law without going through the legislative process.
Andrew Sharp is taking a wait-and-see approach to the Meta settlement, pointing to Australia's ban on social media for children under 16 as a cautionary tale. Early results from Australia suggest that children found ways to circumvent the ban, with minimal change in platform usage.
Ben Thompson suggests that Meta's settlement could be a strategic move to set future regulatory standards. He refers to this as 'regulatory capture,' where a company influences the regulations that govern it, potentially to the detriment of competitors or the broader market.
NVIDIA announced a new initiative to establish independent financing platforms aimed at mobilizing over $500 billion in third-party capital for AI infrastructure development. This plan involves shifting GPU depreciation risk away from traditional lenders through financial engineering.
A listener questioned if NVIDIA's approach to financing AI infrastructure, which involves shifting GPU depreciation risk, mirrors the financial engineering that led to the 2008 subprime crisis. The discussion explored whether this could create preconditions for a cascading collapse or if it's the only way to expand capital for AI build-out.
The podcast discusses the increasing demand for capital to fund AI infrastructure, drawing a parallel to the historical "running out of money" issue from 1873. Ben Thompson notes the significant increase in debt raised by large companies for this purpose.
The discussion touches on how pension funds, designed for long-term liabilities, traditionally invest in infrastructure projects like toll roads due to their predictable, long-term payback. This type of 'patient capital' is contrasted with the rapid funding needs for AI infrastructure.