Perplexity Computer now generates interactive 3D learning visualizations

#1 𝕏

Aravind Srinivas demonstrated inline 3D visualizations in Perplexity Computer, where “Visualize…” can generate educational, interactive widgets and animations for a requested learning topic. He says the feature is best experienced on Standard or High effort.

#2 in

🥞 Carl Vellotti says Claude Code can be modified with files built from four elements—trigger, action, place, and data—and that Claude can write and load these mods from a user’s description. He also created a free resource with five templates and 12 prompts for experimenting with Claude Code mods.

Also covered by: @Thariq, @Jason Zhou

#3 𝕏

Harrison Chase shared ModelRouterMiddleware, a LangChain project by @dbreunig, in the context of a DSPy and Jev model-router example. Jev reads the first message to select a model for the entire run, with a custom hook allowing reselection after each tool result.

#4 𝕏

Teresa Torres recapped how comparing an input tree with an output tree can yield multiple change sets, though only one may tell the real story. Her fix: give the agent rules for meaningful moves—distinguishing a semantic merge from deletion and source reassignment—and log each move as it builds the tree.

#5 𝕏

Guillermo Rauch announced that a KVM zero-day was confirmed through Vercel’s Sandbox bounty program, affecting what he called the industry’s gold-standard solution for Linux virtualization. He thanked Paulos and other researchers for helping improve sandbox security for agents and said a full writeup was coming.

#6 𝕏

Garry Tan said he had just deleted about a thousand lines of markdown from gstack because frontier models are now good enough that the old tricks generally aren’t needed anymore.

#7 𝕏

Garry Tan commented that harnesses can use the same frontier models with better coordination to produce better code outcomes. He noted a tension for frontier model companies, which want their native harnesses to use more tokens rather than fewer.

#8 𝕏

Madhu Guru commented that AI adoption is largely a product problem, with shallow usage even among the 2% paying for AI because products resemble “airplane cockpits with 100 levers,” including connectors, permissions, model selection, and token usage. Guru predicted product teams will make AI products easier to use over the next 12 months.

#9 𝕏

Guillermo Rauch predicted security will become an increasingly large function in software companies, spanning verification engineering and capital allocation. He framed this as both a trust challenge for a “2-person-and-a-dog company” and an opportunity for small teams to disrupt areas shaped by poor global cybersecurity and increasing reliance on centralization—areas he’s excited to invest in.

#10 𝕏

Aravind Srinivas said the goal is to vertically integrate agentic infrastructure by owning and optimizing sandboxes for the best silicon, describing NVIDIA’s Vera CPU as “far better than x86.” He added that deployment of Perplexity Computer on Vera is beginning to roll out, with more details coming soon.

#11 𝕏

Peter Yang shared a free recording of his live session—attended by 100+ people—on building an AI system to proactively take work off your plate. His paid newsletter includes a self-paced AI course with 25+ lessons that is designed to be kept up to date as AI changes, plus $600+ in AI tool credits and an invitation to his private Slack community. It was offered at $150/year for the next 5 days before rising to $200/year on October 8.

Also covered by: @Peter Yang