Grok 4.6 Launches in GitHub Copilot, Perplexity Ships Agent SDK, and Anthropic Expands Safety Disclosures

Key Takeaways

1. Grok 4.6 is now available in GitHub Copilot across CLI, IDE, and cloud; Musk also demonstrated a 48-hour autonomous game build. — via 1 2 3
2. X open-sourced its For You recommendation algorithm, adding Brazil 2026 election filtering and making government censorship requests visible; Musk explained ranking uses personalized probability predictions rather than raw like/report counts. — via 1 2
3. Perplexity released an agent-first Search SDK for any agentic harness, while NVIDIA open-sourced NeMo Switchyard for model routing in agent workflows. — via 1 2
4. Anthropic published its second Responsible Scaling Policy risk report and a watermarking FAQ tied to the EU AI Act, saying quality and cost are unaffected. — via 1 2
5. Qwen3.8-27B is now Hugging Face's fourth most popular model ever, and DeepSeek-V4-Pro-0813 is available on Novita with a 1M-token context. — via 1 2
6. Databricks closed an $18.8B Series M, and swyx highlighted that Genie agents now generate more than 70% of platform queries. — via 1

1. Model and Product Releases

1. Grok 4.6 is integrated into GitHub Copilot for CLI, IDE, and cloud products. Musk also said Grok 4.6 worked for 48 continuous hours to build a shooter game, claiming it can run Gauntlet Loops and recommending the Grok Build harness for autonomous work. — via 1 2 3
2. Gemini 3.7 Flash is now available in the Gemini app, per Demis Hassabis. — via 1
3. Runway launched 1080p early access for Seedance 2.5, promising clearer detail at higher resolution. — via 1
4. Hugging Face says Qwen3.8-27B has become its fourth most-popular model of all time and shared a simple llama serve command with a draft model for local inference. — via 1 2
5. Novita now supports DeepSeek-V4-Pro-0813 on Hugging Face, offering a 1M-token context window aimed at reasoning, coding, and agent workflows. — via 1

2. Agents, Open Source, and Developer Tooling

1. Perplexity's Search SDK is now available for any agentic harness; the agent-first Python SDK lets agents search in parallel and filter, dedupe, and sort results in code, while Aravind Srinivas argues the Agent API is hard to beat for frontier web search. — via 1 2
2. NVIDIA open-sourced NeMo Switchyard, a library for model routing in agent workflows, enabling frontier-model use for complex reasoning/planning and high-capacity specialized execution with Nemotron Lightning. — via 1
3. X's For You algorithm is now open source; Musk says the release exposes how ranking weights work, adds Brazil 2026 election filtering, and makes government-required censorship visible. Ranking is based on personalized probability predictions for user actions, not simple like/report counts. — via 1 2
4. ChatGPT updates this week include auto-generated quizzes, booking search, Google Drive file support, and home suggestions; Greg Brockman says the team is moving toward not manually choosing models, and multi-agent v2 can now delegate to any supported model including Luna. — via 1 2
5. Hamel Husain's hands-on review of Grok Bot highlights its autonomous multi-thread/task capabilities and clean iOS experience, while computer use still lags Codex. — via 1

3. Safety, Policy, and Regulation

1. Anthropic released its second risk report under the Responsible Scaling Policy, detailing system risks and preparedness. — via 1
2. In a new FAQ, Anthropic says it is adding text watermarking to comply with the EU AI Act; other major model developers signed the same code of conduct. The method reportedly has no practical effect on Claude output quality, adds no hidden characters or token cost, and cannot be traced to specific individuals. Separate commentary from Hamel Husain argues no blog post will change developer views on watermarks, keeping the topic contested. — via 1 2
3. Elon Musk says government-owned airports must not favor any religion and called DFW's planned Islamic wudu ablution facilities illegal; he directed review of state funding to DFW and IAH and referred both airports to the U.S. Department of Transportation. — via 1

4. Market and Community Signals

1. Databricks closed an $18.8B Series M. swyx joked the round keeps funding the "delay all meetings" era, adding that Databricks' Genie Ontology auto-captures enterprise context, allowing Genie agents to answer directly and power over 70% of platform queries, driving consumption and revenue. — via 1
2. AI Engineer NYC 2026 is scheduled for Oct 12-14 with a dedicated AI + Finance track covering investment banking, hedge funds, private equity, and payments; speaker applications are open. — via 1
3. Ethan Mollick flags that independent game developers face harsher AI scrutiny than large studios despite being resource-limited, and says adoption friction rather than capability may determine AI's economic impact. — via 1 2
4. Hamel Husain's fall AI evals course adds agentic evaluation methods and safety/risk/privacy considerations for deploying custom agents, with the next cohort starting Sept 5. — via 1