PLUS: The end of online anonymity and how Claude can now do your taxes for free
Happy reading
Apple is making a significant push into on-device processing with the launch of its new M5 chip family. The new MacBook Pros are built to handle intensive AI workflows locally, marking a major hardware offensive in the AI race.
This move equips developers and creators with the tools to build the next generation of AI applications right from their laptops. But is this bet on powerful, private, on-device AI enough to challenge the dominance of cloud-based giants?
In today’s Next in AI:
Apple's M5 chip AI offensive
The end of online anonymity via AI
The rise of China's open-source models
How Claude can do your taxes for free
Apple's AI Offensive
Next in AI: Apple is making a major push into on-device AI, unveiling new MacBook Pros powered by M5 Pro and M5 Max chips designed to handle intensive AI workflows locally.
Explained:
The new chips use a Fusion Architecture with a Neural Accelerator in each GPU core, enabling up to 4x faster AI performance compared to the previous generation for tasks like processing LLM prompts.
M5 Max models now support up to 128GB of unified memory with up to 614GB/s of bandwidth, empowering developers to train complex AI models directly on their laptops.
The upgrade extends beyond the processor, with up to doubled SSD speeds of 14.5GB/s, increased starting storage, and next-generation connectivity including Thunderbolt 5 and Wi-Fi 7.
Why It Matters: Apple is betting on a future where powerful AI runs directly on your device, offering greater privacy and speed by reducing reliance on the cloud. This move equips developers and creators with the hardware to build and use next-generation AI applications right from their laptops.
The End of Anonymity
Next in AI: New research demonstrates that large language models can deanonymize users on platforms like Reddit and Hacker News with high accuracy. The models link pseudonymous accounts to real-world identities simply by analyzing unstructured text posts.
Explained:
The AI-powered method extracts identity-relevant features from your posts—like hobbies, professional jargon, or city mentions—and uses them to find matching profiles across different platforms.
In tests matching Hacker News accounts to LinkedIn profiles, the system identified users with surprising accuracy, achieving recall rates as high as 67% with up to 90 percent precision.
This process is both fast and cheap, costing as little as $1.41 per target using commercially available APIs, making large-scale deanonymization feasible for the first time.
Why It Matters: The long-held assumption that burner accounts provide a reliable privacy shield is no longer valid. This development fundamentally changes the threat model for online expression, requiring a more cautious approach to what you share publicly.
The Eastern Model
Next in AI: Chinese open-source AI models from firms like DeepSeek, Alibaba, and Moonshot are surging, offering performance that rivals top Western models at a fraction of the cost. This shift is providing startups and developers worldwide with powerful new foundational tools.
Explained:
Top-tier performance is now accessible and affordable. Models from firms like Moonshot AI and MiniMax are not just cheaper alternatives; they are topping global usage charts, surpassing established US models in total text processed for some applications.
They are becoming the default foundation for builders worldwide. Alibaba's Qwen series has overtaken Meta's Llama in cumulative downloads, and an estimated 80% of Silicon Valley startups using open-source AI are now building on Chinese models.
The pace is accelerating with increasingly specialized models. DeepSeek is set to release its first major multimodal model this week, while other companies are training models on non-US hardware like Huawei's Ascend chips, signaling a push for greater independence.
Why It Matters: This open-source push gives developers and startups unprecedented access to powerful AI tools, leveling the playing field for innovation. It also positions Chinese firms to define the underlying infrastructure and standards for the next generation of AI applications globally.
Your New Accountant
Next in AI: A detailed case study shows how Anthropic's Claude can successfully prepare a complex, 42-page federal tax return for free. This offers a powerful alternative to paid software like TurboTax for those with complicated financial situations.
Explained:
Claude successfully processed a 42-page tax return that included multiple schedules and forms for inherited IRAs, capital loss carryforwards, and foreign tax credits.
The most effective workflow involved uploading all source documents and having Claude fill out current IRS form PDFs directly, a process the author has since packaged into a reusable skill.
This experiment was driven by a desire to avoid funding Intuit, whose documented lobbying efforts have consistently worked against government-led free tax filing initiatives.
Why It Matters: This real-world example shows AI democratizing access to specialized knowledge, bypassing expensive software built to solve problems that incumbents helped create. It provides a practical blueprint for using AI to manage complex administrative tasks far beyond simple text generation.
AI Pulse
Warwick warns that many AI cancer pathology tools may be taking "shortcuts," learning to identify statistical correlations in tissue samples rather than true biological signals, raising concerns about their reliability in patient care.
Cekura launched on Hacker News as a YC-backed platform for testing and monitoring conversational AI agents, using synthetic users and LLM-based judges to find regressions across full conversations before they hit production.
Judges threatened legal consequences in India's Supreme Court after a junior judge used fake, AI-generated orders and legal citations in a property dispute ruling, calling it a matter of "institutional concern."
Experience protects older workers from AI-related job displacement, according to a Dallas Fed analysis, which found that the pay premium for tacit knowledge gained from hands-on work is increasing in AI-exposed industries.