Issue #25·Thursday, January 22, 2026·20 min read·10 stories

Perplexity's Police AI Deal Sparks Warnings

OpenAI's 5-device hardware roadmap, Apple's chatbot Siri, plus 100B vector search in 200ms.

Yesterday, Perplexity's deal to provide AI to police departments surfaced, prompting immediate warnings from experts about public safety implications. Separately, OpenAI's internal hardware roadmap details a five-device strategy, including a 2026 earbud launch, signaling a push beyond pure software. For builders focused on search, a new ANN v3 paper achieves 200ms p99 query latency over 100 billion vectors.

▲NEWS

8 stories

AI Earbuds Lead OpenAI's 5-Device Hardware Push

OpenAI is reportedly pivoting into consumer hardware, with plans for five devices by 2028. The first product, 'Sweetpea' earbuds, is slated for a September 2026 launch, aiming for 40-50 million units in its first year. This signals OpenAI's move toward vertical integration and direct consumer relationships, suggesting future AI features may be tied to specific OpenAI hardware and influencing product strategy for builders creating AI-powered devices.

Read full story→

Experts Warn on Perplexity's 30% Error Rate in Police AI

Perplexity is offering free AI chatbots to law enforcement for tasks like crime scene analysis, but independent research shows a 30% error rate. Experts warn that deploying general-purpose AI with known error rates in sensitive public safety contexts is risky. Expect severe scrutiny and legal exposure if your AI product's errors impact high-stakes decisions, given the current lack of regulatory safeguards.

AI Cloud Market Shift: ByteDance Takes 13% Share, Targets Alibaba

ByteDance is aggressively expanding its Volcano Engine cloud offering in China, targeting Alibaba and Tencent with AI agents built on proprietary models. The company captured 13% of China's AI cloud services revenue in H1 2025 and was Nvidia's largest customer in China in 2024.

55 US AI Startups Raised $100M+ in 2025

55 US AI startups raised $100 million or more in 2025, including several multi-billion dollar rounds. The report details specific funding, valuations, and lead investors across AI infrastructure, healthcare, and legal tech. This signals a strong fundraising environment for founders in the AI market.

LLM Comparison Tool Tracks Price, Performance, Speed

whatllm.org is a side-by-side comparison tool for large language models. The platform tracks price, performance, and speed across the AI ecosystem, with daily updates to its benchmarks.

Gurman: Apple Siri Chatbot Coming, Uses Google TPUs

Mark Gurman reports Apple will transform Siri into a generative AI chatbot, codenamed 'Campos', for iOS 27. This update will leverage Google Foundation Models (AFM v11, comparable to Gemini 3). A preceding iOS 26.4 update will use AFM v10. Apple is reportedly discussing hosting the advanced chatbot on Google's TPU servers, signaling a potential policy shift from its usual Private Cloud Compute.

AI Reverse Engineers Molecules 10x Faster

Researchers from NYU and the University of Florida developed a generative AI method for molecule reverse engineering, published in Nature. Their neural network proposes viable molecular structures 10 times faster than existing methods, maintaining accuracy. This AI-driven approach allows scientists to specify desired properties and quickly generate potential structures, accelerating drug and battery development.

Zoomlion Deploys Humanoids in 6-Minute Excavator Production

Chinese heavy equipment manufacturer Zoomlion produces an excavator every six minutes, with humanoid robots deployed since 2024 to support production. These humanoids handle logistics, pre-assembly, and quality inspection within a broader robotic manufacturing system that has been in place since at least 2022. The company plans to commercialize its robotics platform.

⚙TECHNICAL

2 stories

100B Vector Search Hits 200ms P99 Latency with Turbopuffer ANN v3

Turbopuffer's ANN v3 achieves 200ms p99 latency over 100 billion vectors (200TiB) by optimizing memory hierarchy. It uses hierarchical clustering and RaBitQ binary quantization, which compresses vectors 16-32x and shifts the bottleneck from bandwidth to compute. This deep dive demonstrates how to achieve 100B vector scale with low latency and cost by balancing memory hierarchy and compute bottlenecks through techniques like hierarchical clustering and RaBitQ quantization.

Rust Powers S3's Core, Formal Methods Ensure Consistency, New AI Vectors Arrive

An interview with Mai-Lan Tomsen Bukovec reveals AWS S3's engineering principles, including its extreme scale, the challenge of strong consistency, and the strategic use of Rust for critical code. S3 also introduced S3 Vectors, a new data primitive optimized for high-dimensional search, directly supporting AI/ML applications like RAG.

◈ANALYSIS

2 stories

“AiPhone” Commentary Projects Siri’s Google TPU Future

A recent commentary projects Apple's Siri will evolve into an "AiPhone" through a two-stage AI rollout. The author envisions a Gemini-powered Siri in iOS 26.4, followed by the conversational chatbot Campos in iOS 27. This commentary highlights Apple's potential use of Google's TPUs for the advanced model, noting privacy implications.

Value Share, Not Tokens: OpenAI's Enterprise Pivot

One analysis suggests OpenAI's 2026 enterprise strategy will pivot from selling raw API tokens to value-sharing models. This shift, signaled by strategic hires and CFO posts, aims to maximize revenue and reduce customer churn. The author predicts OpenAI will seek a fraction of the value created by customer applications.

⚒TOOLS

1 story

Open-Source Tool Sandboxes AI Agents, Blocks Secret Exfiltration

yolo-cage is an open-source tool that creates secure sandboxes for AI coding agents to prevent secret exfiltration and dangerous actions. It enforces Git branch isolation, scans and blocks secrets in egress HTTP/HTTPS traffic, and restricts GitHub CLI/API commands. Builders can thus use AI agents with a limited blast radius, deferring security decisions to PR review.