Perplexity

Perplexity 开源其高效的 Unigram 分词器，CPU 利用率降低 5-6 倍，显著减少延迟。

Perplexity开源了重构的Unigram分词器，CPU利用率降低5-6倍。

Perplexity 的编码器在生产输入长度下将 p50 延迟降低了约 5 倍，相比 HuggingFace 分词器，2 倍相比 SentencePiece C++，1.5 倍相比 IREE C。

GB 200s improve the prefill and decode disaggregation efficiency for large MoE models like Qwen, significantly enhancing throughput compared to the Hopper platform.

Perplexity published new research on deploying the Qwen3 235B model on NVIDIA GB200 NVL72 Blackwell racks, showing that GB200 outperforms Hopper in high-throughput inference for large MoE models.

Perplexity is emerging as an enterprise-grade knowledge and research platform, with PayPal running 74,000 weekly tasks for model validation, market analysis, and strategic decision-making.

Perplexity achieves default-secure computing architecture: each task runs in a hardware-isolated sandbox with VPC-level storage-compute separation and short-lived proxy tokens for agent authentication.

Lovable has rebuilt its apps with server-side rendering (SSR) to improve default discoverability in search engines and AI answer engines like Google, ChatGPT, and Perplexity, migrating its foundation to TanStack Start for better type safety and deployment flexibility.

Deep enterprise adoption of AI tools requires sustained investment in security engineering, with Perplexity implementing sandboxing and automated security workflows for risk control.

Perplexity open-sources its internal security tool Bumblebee to protect developer systems and enhance product safety.

Computer now connects to Snowflake, enabling real-time analysis of enterprise data.

Perplexity builds an AI agent skill system based on composability and evaluation, enabling automation of complex tasks by decomposing multi-step reasoning into atomic skill modules and continuously optimizing via real user behavior data.

Perplexity has published its internal guide for building AI agent skills, emphasizing a new mindset required for developers to design, refine, and maintain agent capabilities.

MiniMax Agent integrates Perplexity's search infrastructure, outperforming competitors in over 700 tasks with improved quality and efficiency.

Perplexity open-sources Bumblebee, a read-only scanner for macOS and Linux that detects risky packages, extensions, and AI configurations on developer machines.

Perplexity AI announces its system can now connect directly to the Snowflake data warehouse for end-to-end analytics.

By filtering ads, navigation, and无用内容 before handoff to the answer model, Perplexity increases vital content per snippet by 63%, achieving a 50x compression ratio while maintaining frontier-level performance.

Perplexity is shifting search from a web-fetch tool call to code generation to adapt to a future where code execution inside agent harnesses dominates knowledge work.

Perplexity introduces 'Search as Code' — an architecture enabling AI agents to generate Python code that directly calls our search stack, replacing manual function call loops.

Lovable has restructured its apps to be server-side rendered, enhancing discoverability in search engines like Google and AI answer engines such as ChatGPT and Perplexity, built on TanStack Start due to its active maintenance and strong community.

Perplexity is developing a highly secure and scalable agent runtime sandbox with key features including secure API key handling, content safety detection, and encrypted data transfer.

Perplexity introduces dashboard and automation tools based on Snowflake data to support pipeline analysis, product usage tracking, and customer segmentation.

PayPal uses Perplexity Enterprise to run 74,000 weekly tasks, used for model validation, channel performance, market trend research, competitive intelligence, and product analysis.

New apps are now server-side rendered and discoverable through search engines like Google and AI answer engines like ChatGPT, Claude, and Perplexity. Existing apps get prerendering. Additionally, access @semrush search intelligence in-app. Chat with your SEO data, understand how you're performing, and improve discoverability. You can also run SEO reviews on demand, and fix all suggestions in one click.

Perplexity scans external content in parallel using ML classifiers and the BrowseSafe model before AI agents act, with file data encrypted and auto-deleted after 7 days.

Perplexity now connects to Apple Health on iPhone, allowing users to import sleep, activity, and HRV data, and query biomarkers, blood test results, or panel analyses.

Google’s newly launched AI feature on its homepage offers fast response and smooth UX, prompting users to reconsider paying for Perplexity Pro. However, the article lacks technical validation or benchmarks.

The tweet humorously references the 'Zen of Python' and extends it with 'Zen of Skills', suggesting easily explainable ideas may already be known by models and should be deleted.

Several well-known tech companies have their employees sharing project progress on Twitter, a practice that more companies should emulate.

Perplexity 提供了关于改进 Unigram 分词器 CPU 性能的博客文章链接。