X open-sources For You recommendation algorithm

Grok 4.6 is now live in GitHub Copilot across CLI, IDE, and cloud, and Musk says a 48-hour autonomous build of a shooter game proves it can handle extended Gauntlet Loops.
X has open-sourced its For You recommendation algorithm with ranking weights and election-related filters, making government-requested censorship visible; Musk also clarified that ranking uses personalized probability predictions, not raw counts.
Perplexity released an agent-first Search SDK for any agentic harness, while levelsio argues Grok Bot is the "Claude Code moment" for ordinary knowledge work—a market potentially larger than coding agents.
Basecamp added CLI/agent access, which Jason Fried calls the biggest change in Basecamp 5 history, while Omarchy Linux 4 "Quattro" is being touted as a fully malleable agent-era OS.
YC recent batches saw solo-founder share double from 9% to 18%, while levelsio reported a record ~$25K month on X from ~73M views.

Grok 4.6 is now available in GitHub Copilot across the Copilot CLI, IDE, and cloud products. This is a notable expansion of xAI's model into one of the most widely used AI coding environments, giving developers another frontier-model option directly in their workflow. — via 1
Musk says Grok 4.6 worked continuously for 48 hours to build a shooter game, calling it proof the model can run "Gauntlet Loops" with sustained autonomy. He recommends running such flows through the Grok Build harness or in Grok Bot, positioning the release as a milestone for long-running agent tasks. — via 1 2
Perplexity's Search SDK is now available for any agentic harness, creating an agent-first Python tool that lets agents search in parallel and filter, deduplicate, and rank results in code. Perplexity frames this as making "Perplexity Computer" the best option for broad and deep research; Srinivas adds that its Agent API should be hard to beat for frontier web search. — via 1 2
levelsio argues Grok Bot represents a "Claude Code moment" for ordinary knowledge work, not just code. He contends every day OpenAI, Anthropic, and Google wait to ship a comparable product is a day they cede a market that could be larger than coding AI agents. — via 1

X open-sourced its For You recommendation algorithm, including the actual ranking weights and the filtering logic required for Brazil's 2026 election. Musk says this makes any government-requested censorship on the platform clearly visible to the public. — via 1
Musk clarified how the algorithm weighs actions: ranking is based on personalized probability predictions of behavior such as likes or reports, not simple counts, and he pushed back on the idea that a high report weight means one report outweighs hundreds of likes. The explanation is intended to show that recommendation scores are calibrated to user behavior models. — via 1
Users can now inspect and verify their own account reach and visibility, which Musk says undermines earlier claims of hidden account restrictions. This adds visibility into how posts perform and whether accounts are being quietly limited. — via 1

Basecamp has introduced CLI and agent access, and Jason Fried calls it the biggest change in Basecamp 5 — even all of Basecamp — history. Batch edits and template creation/editing now take seconds, making the product programmable by agents and power users. — via 1
Omarchy Linux 4 "Quattro" is earning outsized praise as an agent-era OS, with Tobi calling it one of the greatest software releases of his career and a fully malleable operating system for agents. Jason Fried also announced a Basecamp plugin for Quattro that surfaces account notifications in the system bar, built with Rob Zolkos. — via 1 2 3
Greg Isenberg proposed a five-part test for when a task should become an AI agent: recurring trigger, stable input shape, defined toolset, measurable completion criteria, and real-time judgment in the middle. The first four determine automability, while the fifth is what separates an agent from simple automation. — via 1
Replit's CEO offered more evidence of agentic coding speed: he says he "vibecoded" $450 in under two hours and took a Replit project from first prompt to TestFlight in less than 50 hours. These anecdotes suggest AI-assisted development is compressing the path from idea to shipped app. — via 1 2

The share of solo founders in recent YC batches doubled from 9% last summer to 18% this summer. This is a significant shift for the accelerator's portfolio composition and points to AI tools lowering the need for a co-founder. — via 1
Paul Graham advised a startup whose idea isn't working to reboot with a new idea and wait until the next Demo Day to present it. He also amplified the view that young startups shouldn't adopt big-company process early — the priority should be speed and users, with OKRs able to wait. — via 1 2
levelsio reported a personal best month on X with roughly $25,000 in total revenue, including about $19,000 from X's ad revenue share and subscriptions, on roughly 73 million views. He credits X's shift to sharing revenue only on original content and says it's welcome news for original writers and citizen journalists; he also plans to put around $4,000 into the Omarchy ecosystem as a plugin contest prize pool. — via 1 2

Discover