Perplexity Debuts Local Portable Computer as OpenAI and Vercel Accelerate Agentic AI Tools

Key Takeaways

1. Perplexity unveiled a local-first Portable Computer research prototype running on NVIDIA DGX Spark, with a device-side 27B model scoring 82.6% on real knowledge-work tasks and a post-trained PPLX 27B reaching 85.4%.
2. OpenAI said it built a chip and expanded ChatGPT Work to operate browsers and log into sites without credentials, signaling a deeper push into agentic computing.
3. Vercel shipped four agent-focused developer capabilities: AI Gateway webhooks, Vercel Connect GA, Run SDK, and a GA Security Dashboard.
4. Patrick Collison described @bot, @agentmail, and Stripe Link as the current ultimate personal agent stack, while Stripe Radar blocked $300M in fraud for token-selling companies.
5. Starcloud raised $250M at a $2.3B valuation, adding to a fast-moving wave of AI-adjacent infrastructure funding.

1. Local AI Hardware and Models

1. Perplexity's Portable Computer research pushes local-first AI agents: a device-side 27B model scores 82.6% on real knowledge-work tasks, and the post-trained PPLX 27B improves to 85.4%, surpassing open-source harnesses Pi and Hermes. Perplexity is working with NVIDIA to support it on DGX Spark, and the design runs fully locally with no cloud dependency. — via 1 2 3 4
2. Perplexity Computer for Max users adds a background Dream agent that continuously ingests context from files and connected apps and builds a multi-hop context graph. Its Brain memory system turns sessions, files, and sources into a structured wiki, improving correctness by 9.3 points, timeliness by 8.0, and recall by 8.9 while using 15% fewer tokens. This matches Aravind Srinivas's argument that future AI should run as a background process on user hardware. — via 1 2
3. Sam Altman said OpenAI "made a chip, and it's fast," without disclosing architecture, manufacturing, or timeline. The statement is a strong signal that OpenAI is moving beyond relying solely on external accelerators, but no additional details were provided. — via 1
4. Garry Tan shared M5 Ultra hardware specs: the 512GB version ships in October, has 1.2TB/s memory bandwidth, 50% higher than M3 Ultra and 4.4x DGX Spark/M4 Pro, and a four-stack setup could run Kimi K3/GLM 5.3 above API speed at more than 100 tok/sec. — via 1

2. Agents, Developer Platforms, and Productivity

1. ChatGPT Work now lets the agent call a computer and browser to log in to websites without entering usernames or passwords, enabling account opening, reservations, reimbursements, and rental tasks. This moves OpenAI from chat completion toward delegated real-world actions. — via 1
2. Vercel expanded its agent-native developer surface: AI Gateway now supports webhooks after long generations, Vercel Connect is GA with commands like vercel connect create notion, Run SDK provides a lightweight QuickJS eval for agent-written code, and the Security Dashboard is GA with vercel security check. These tools target the hardest agent problem: securely connecting services and data. — via 1 2 3 4
3. Jason Fried processed over 300 accumulated emails by running a 27B Qwen model locally on an RTX 5090 with the new HEY CLI, using the Imbox and screener workflow. It is a concrete example of a local model handling personal productivity work without cloud APIs. — via 1
4. Patrick Collison described "@bot + @agentmail + Stripe Link" as the current ultimate personal agent stack, letting agents register accounts, pay, and do what a personal assistant can do. He separately noted that Stripe Radar blocked $300M in fraud for token-selling companies in recent days. — via 1 2
5. Amjad Masad says he built a Grammarly alternative called Content Refinery in a day on Replit after receiving a $144 renewal notice, and used MCP to connect UberSuggest for SEO research. The example shows how quickly non-core tools can be replaced with AI-native apps. — via 1
6. Claude Code's team addressed extensibility: model families are not interchangeable, system prompts have a significant performance impact, and the team will support Agents.MD and system-prompt modifications. — via 1

3. Funding, Market, and Policy Signals

1. Starcloud raised $250M at a $2.3B valuation, with Alexis Ohanian noting markets are moving quickly on a concept once dismissed as science fiction. The round highlights how rapidly investor sentiment is shifting toward ambitious infrastructure bets. — via 1
2. Patrick Collison is backing Bridge's EURR, a euro stablecoin powered by open issuance, and Revolut is launching it in-app. This expands stablecoin payments beyond USD-denominated rails and signals growing institutional interest in non-dollar stablecoins. — via 1
3. Yohei Nakajima argues that excessive VC money keeps companies private too long, absorbs too much capital, and leads to fewer IPOs and worse post-IPO performance. He recommends companies go public around $100-200M revenue, under $250M raised, and below $1B valuation. — via 1
4. Anthropic and AWS are sponsoring a "Rare Disease, Real Kids" hackathon with a $50K prize pool, opening genome and clinical data for one rare-disease patient to find causes and approved drugs through community AI efforts. — via 1
5. Paul Graham and Garry Tan both point to hands-on student projects as essential for training future founders; Graham recommends his essay "A Project of One's Own." This is a signal for early-stage founder education emerging from the AI era. — via 1 2