Cofounder, CSO @perplexity_ai. Former high frequency trader, competitive programmer. Think fast, build faster.
New York, NY
Joined December 2011
- Tweets162
- Following275
- Followers23K
- Likes2.3K
Impressive new results from Argon, now available in Perplexity. The frontier and cost/quality tradeoff is extremely close between the top models.
The sandboxes that power Perplexity are now open source.
It's been a lot of fun to be contributing a few spare commits in my spare time. Definitely would not have time to be actively learning a new language and rewriting our stack in Rust without today's coding agents and Perplexity's developer tooling.
We're open sourcing SPACE, the sandbox platform that runs every task in Perplexity Computer.
SPACE gives each agent task its own Firecracker microVM and controls credentials and outbound traffic from outside the VM.
Available now under Apache 2.0: github.com/perplexityai/SPAC…
Make hard decisions quickly, no matter how painful.
When I was a teenager, my friends and I came up with “the coin flip method” from first principles. If we were indecisive or were at a stand still on something, we’d flip a coin to make a hard decision knowing fully well if we felt bad about what the coin told us, we would go the other way. Like if we should rent movie A or movie B.
Later I learned this is called a “Freudian coin toss,” and knowing when you are displeased with a result is a strong signal for anything and a great way to keep motion.
Perplexity augments fast and cheap models like GPT 5.6 Luna to improve accuracy and overcome hallucinations. Now compatible with any open harness: via CLI, SDK, or MCP.
Perplexity Search debuts on the Artificial Analysis Search Index, with all three context size variants taking top positions on the leaderboard
The @perplexity_ai Search API comes with three context settings (low, medium, and high) that control how much extracted content each search result carries. We tested all three variants using our standardized methodology: the same model (GPT-5.6 Luna at medium reasoning), running inside Stirrup, our open-source agent harness, with tools for searching and fetching pages from the web. Only the provider behind the search tool changes.
Key results:
➤ Perplexity Search (medium) scores 80 on the Artificial Analysis Search Index, ahead of the previous leaders, Parallel (advanced) and Brave Search (LLM context), at 75. The high and low variants score 79 and 77 respectively. Its lead is concentrated in BrowseComp results, with AA-Omniscience and DeepSearchQA scoring comparably to other leading providers
➤ Efficient search payloads: smaller overall search results mean the model reads less per task, so Perplexity has the lowest model inference cost per task of providers we’ve tested so far, ranging from $0.028 to $0.034 across the three variants vs $0.036 for the next lowest provider
➤ Total cost per task is ~$0.091 for the medium and high context variants, at mid-pack latency. For comparison, Parallel (advanced) costs $0.084 per task and Brave (LLM context) costs $0.13 per task
AGI will be here when agents learn to multitask across >10 tasks at once without steering. Today's models seem to often operate on a last in first out stack and lose track of the long-term vision, hampering their ability to plan across long engineering tasks. Asking agents to take notes helps somewhat but also seems like an incomplete solution.
Back when I was working in finance, it took weeks to acquire and test one new data source. Now financial research, modeling, backtesting, etc. is now infinitely easier and faster with AI. Creativity is the bottleneck rather than code.
We are doubling down on making Computer more useful for financial researchers and analysts.
Computer now connects to new licensed data sources, including Dun & Bradstreet, Guidepoint, IBISWorld, and 20+ others. An analyst can ask Computer a question and get an answer built from the firm's own licensed sources without APIs or separate logins to work through.
Every figure traces to the source record it came from. Teams doing high-stakes work would know exactly which data set produced which number.
To connect, select a provider in Settings → Connectors and sign in with an existing license. Available now to all Perplexity users.
perplexity.ai/hub/blog/compu…
Congrats to everyone involved from both Nvidia and Perplexity. Excited to build great products that leverage both local and cloud compute simultaneously. Much more to come!
Today we’re launching Portable Computer on @NVIDIA DGX Spark.
Portable Computer is a fully local version of Perplexity Computer, where the entire runtime: orchestrator LLM, subagent LLM, agent harness all run on your local hardware. No cloud dependency.
Glad that the ecosystem has converged on an open standard for plugins! Incredibly easy to install or build your own.
The @perplexity_ai API MCP server is now available as an Agent Plugin 🎉
Install it in Codex, Cursor, VS Code, and other Agent Plugins 1.0.0 conformant clients. Your agent gets real-time web search, reasoning, and research through Perplexity's API.
Visit the repo to get started: pplx.ai/api_agent_plugin
If your agent needs frontier web search, it's hard to beat Perplexity's Agent API (on any metric).
Introducing Web Search Benchmarks 🌐
Rankings of search tools across different models and configurations to help you decide how to ground your agent: openrouter.ai/benchmarks
Open weights continues to nearly match the frontier while edging the frontier out on several benchmarks.
Time to try GLM out for coding!
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.
- Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
- A major leap in cybersecurity, setting a new standard among open models
Tech Blog: z.ai/blog/glm-5.3
How and why are now the hardest parts of engineering: how to measure quality, how to design systems, how to find levers to pull, and why it all matters.
blog.senko.net/code-was-neve…
GPT 5.6 Terra and Luna enable improved efficiency and latency across many use cases, especially reoccurring automations. Excited to continue building on the next frontier of cheaper models.
Agentic compaction is a cool idea for both context retention and cost savings.
Introducing AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents.
Long-horizon coding trajectories often accumulate failed attempts, stale hypotheses, and noisy tool outputs, making it harder for the agent to focus on what matters.
👉AutoCompact teaches coding agents to proactively call compact() when necessary, replacing noisy history with a clean working context for the next phase of the task.
(Glad to have contributed to this work with Xuan Zhang, @ltzheng01, Bo An, and @SimonXinDong)
Project page: autocompact.github.io/
Perplexity search is now available as a Remote MCP. One more way to make search and research APIs universally accessible to all agents.
The Perplexity remote MCP server is live.
You can now connect Perplexity to Claude Code, Cursor, or VS Code with just your API key. Nothing to install.
docs.perplexity.ai/docs/gett…
Can confirm that Cameron is one of most creative power users of AI out there.
Excited to announce Projects, which are Perplexity's hubs for agent and human collaboration. Most of my time now lives in Projects, where agents automatically manage their own persistent files and collaborate with native tools.
Projects are Perplexity's core for collaboration: as agents become more long-running, they will need more agent-native ways to persist and self-improve: memories, skills, and files. But until today, agents have lacked a central hub for collaboration that allows for a persistent area for knowledge work. That place is Projects, not individual chats or files. Projects come packaged with an set of CLI tools intentionally designed for agents to collaborate across files, projects, and with other agents.
Importantly, both individual and project identities are fully functional. Individually-scoped connectors are supported as usual, but so are project-level credentials. A project-level brain agent manages the project and provides objective summaries and memories. Agents do the execution but humans still do the steering with notifications and permissions checks.
Internally at Perplexity, the canonical way to work is now to start in Slack (each Project connects with at least one Slack channel), spawn multiple threads of conversation in parallel, accept most first passes, and follow up in Perplexity's apps when refinement is needed. The future of knowledge work is looking bright: work is incredibly efficient with tools like Computer, leaving more time for curiosity and learning.
Projects is the next step in the vision that Spaces began for collaboration and persistent storage of context. We're excited to see how you use Projects and we will learn from your feedback to keep pushing the frontier of what is possible when agents and humans work together to build real businesses.
Perplexity Computer is an unmatched agent for financial analysis and realtime market data. Now with improved finance connectors and curated skills for each connector.
Perplexity Computer now supports 6+ finance data Connectors with 16 new finance skills.
Pull fundamentals from @Factset, alt data from @Carbonarc, expert calls from @Guidepoint, + 4 more connectors.
Explore our 30+ skills, and write your own to automate parts of your investment process.
Excited to open source more infra that powers Perplexity and Computer's harness. Shoutout to our security teams.
Today we’re open-sourcing Numbat, an agent-detection and response layer that is designed to work across agent harnesses.
Numbat gives security teams visibility into agent activity, with controls to block selected actions before execution.
Read more: research.perplexity.ai/artic…
Kimi K3 is now available on US hosted inference in Perplexity Computer. Open weight models like GLM 5.2 and Kimi K3 are now some of the most affordable ways to try agentic work products.