August 2025 Summaries

Perplexity has announced Comet Plus, a $5 subscription service intended to give users access to premium content from participating publishers while creating a new compensation model for journalism in an AI-driven web. The company argues that traditional publisher revenue models based on clicks and pageviews do not account for AI-mediated uses of content, such as search citations and agent-driven tasks, and says Comet Plus will distribute subscription revenue to partners after deducting a small amount for compute costs. Subscribers will be able to access participating publishers’ sites and content, use AI assistants to perform tasks on those sites, and receive answers informed by premium sources, while Pro and Max members will receive the service as part of their existing plans. Perplexity plans to announce its initial publishing partners when the Comet browser becomes freely available and is inviting interested publishers to join the program.

Perplexity has expanded its Comet AI browser to all U.S. Pro subscribers and Enterprise Pro users, positioning it as an enterprise-ready assistant that automates routine online work such as scheduling meetings, drafting emails, researching topics, managing tabs, and interacting with web-based business tools. Built on Chromium, Comet supports Chrome extensions, single sign-on, existing web applications, and enterprise security policies, while offering administrator controls for permissions, adoption insights, and file transfers. Perplexity states that Comet maintains SOC 2 Type II certification and GDPR and HIPAA compliance, stores browsing and AI interaction data locally with end-to-end encryption, and accesses limited purpose-specific data only when users request personalized assistance. Its agentic capabilities are intended to handle multistep processes including vendor research, content creation, customer analysis, and market intelligence, with integrations for Gmail and calendars and compatibility with web interfaces such as Slack and CRM systems. Perplexity plans further integrations, automation, collaboration features, and AI model options, and announced a live demonstration scheduled for August 28, 2025.

Perplexity has partnered with The Roost Podcast Network and Theo Von’s “This Past Weekend” to integrate its AI-powered answer engine directly into the show’s spontaneous conversations. The tool will support Theo Von, his guests, and production team by providing on-screen answers to unscripted questions as they arise, replacing traditional web searching with immediate responses. Perplexity and the show describe the arrangement as an embedded conversational feature rather than a conventional sponsorship, designed to match the podcast’s fast-paced and curiosity-driven format. Theo Von said the platform helps him get answers while maintaining the flow of discussion, while Perplexity Chief Business Officer Dmitry Shevelenko said the partnership reflects the value of asking questions and exploring the world.

Perplexity describes how it achieved immediate support for OpenAI’s GPT-OSS 20B and 120B open-weight models on NVIDIA H200 Hopper GPUs through its in-house ROSE inference engine, choosing FP8 rather than the models’ native MXFP4 quantization because Hopper lacks dedicated FP4 tensor cores. The effort required model-specific changes for grouped-query attention with attention sinks, YaRN positional encoding, biased projections, and sparse mixture-of-experts layers using SwiGLU, Top-K-only softmax routing, and bias-enabled DeepGEMM kernels. Perplexity adapted FlashInfer for sink attention, extended MoE execution while preserving bfloat16 precision for routing accuracy, and optimized GPU kernel launch configurations for the models’ smaller hidden dimensions. Harmony chat formatting, including role, channel, and tool-call metadata, was handled in a frontend formatter and parser so ROSE’s token-based backend could remain unchanged. Benchmarking found that single-node deployments avoid inter-node overhead, with different tensor and expert parallelism choices balancing latency and cost for input processing, while EP4/DP1/TP4 generally performed best for output decoding; using the 20B model to speculatively decode for the 120B model yielded only limited gains because their active parameter counts are relatively similar.

Perplexity has integrated OpenTable’s network of 60,000 global restaurant partners to help users search for and reserve restaurants through conversational AI. Users can describe specific preferences such as cuisine, ambiance, dietary needs, views, noise level, décor, or happy-hour offerings, and Perplexity will provide tailored options based on information aggregated from across the web. The integration is designed to reduce the need to switch between apps, browse extensive reviews, or coordinate reservations manually, allowing diners to book directly through an OpenTable “Reserve” button and receive confirmation. OpenTable says the partnership connects its restaurant network with Perplexity’s users as more people use AI to discover dining options.

Perplexity argues that modern AI assistants should be distinguished from traditional web crawlers because they retrieve current web content only in response to specific user requests, use it immediately to answer questions, and do not store it or use it for model training. It contends that blocking such user-driven agents can restrict access to information for tasks such as medical research, product comparisons, and news discovery, potentially giving infrastructure providers undue control over which tools users may employ. The passage strongly disputes Cloudflare’s claims that Perplexity conducted stealth crawling or impersonated other traffic, asserting that Cloudflare likely misattributed millions of requests from BrowserBase, a third-party cloud-browser service that Perplexity says it uses only minimally. Perplexity characterizes Cloudflare’s analysis and technical diagrams as inaccurate and argues that the dispute exposes weaknesses in Cloudflare’s ability to distinguish legitimate AI-assisted browsing from malicious scraping.

Large language model inference consists of a compute-intensive prefill phase that processes input tokens and builds KV caches, followed by a memory-bound decode phase that generates output tokens one at a time; combining both workloads on the same hardware can increase decode latency and cause output stutter. Perplexity addresses this through disaggregated prefill and decode, assigning separate nodes to prefill requests and decoder nodes to generate responses after receiving transferred KV caches, improving throughput and latency consistency at the cost of additional time to first token from network transfers. Its KV Messenger coordinates these transfers over RDMA-capable networks using libfabric and fabric-lib, beginning layer-level KV page transfers as soon as they are produced and transferring needed metadata, logits, or hidden states as well. The system supports tensor-parallel and data-parallel deployments, including KV-cache sharding, replication, topology-aware NIC selection, and an HND memory layout that makes sharded cache transfers more efficient. For speculative decoding, the final input token is handled on the decoder to avoid synchronizing sampling and structured-output logic across prefillers. In tests with DeepSeek-R1, separating prefill reportedly raised throughput from inconsistent rates below 50 tokens per second to more than 90 tokens per second with one prefiller supporting three decoder nodes, while deployments of the 480B-parameter Qwen3-Coder model use eight-GPU shards and parallel transfers across associated network interfaces.