Rollouts finds offending PRs; one click starts cloud agent

#1 𝕏

Cursor announced that Rollouts identifies the offending PR when it catches a regression and opens an issue, with one click starting a cloud agent to fix it. Rollouts usage credits are included through Oct 3.

#2 📝 OpenAI News

A practical guide to building with GPT-6 - OpenAI shared “A model guide for the GPT‑6 family” on October 2, 2026, offering practical tips for getting the best results from GPT‑6 models while managing time and cost. OpenAI describes GPT‑6 as its most advanced suite of models yet, with options for different kinds of work.

#3 𝕏

Google Research announced a next-generation Federated Learning system that uses Trusted Execution Environments (TEEs) to deliver verifiable differential privacy. The system moves computation server-side to cut training times.

#4 𝕏

@adithya_s_k and the Hugging Face team released an open guide and resources for multi-harness RL using a proxy that supports four API formats and captures vLLM token IDs and log probabilities without modifying agent harnesses. Training Liquid AI’s LFM2.5-2.6B across four harnesses raised its score from 42% to 54% with 31% fewer tool calls, while imitation training on 3,189 successful Qwen3.8-27B rollouts plateaued at 47.5%.

#5 𝕏

NVIDIA AI shared a tutorial on fine-tuning NVIDIA Nemotron 3.5 ASR for dialects and languages after fine-tuning reduced word error rate on Najdi and Hijazi Saudi Arabic from 55% to 30%.

#6 𝕏

Harrison Chase commented that agent teams want their own trace review UI, and that LangChain’s LangSmith Custom Apps lets users build one on their own data and ship it directly into the workspace.

#7 in

Vercel shared that Rogo engineers build the company’s internal apps on Vercel, deploying over 73,000 times last month. Coding agents ship to production in 5 minutes and run workflows from churn risk analysis to the sales team’s deal desk.

#8 𝕏

Google Research announced that Kaggle’s Gemma 4 Developer Agent Competition is live, challenging entrants to post-train an open model into a coding agent for everyday hardware. The competition offers a $100K prize pool and lists November 25, 2026, as the deadline.

#9 𝕏

LlamaParse’s agentic parsing tier extracted the BEA’s new Q2 GDP-by-state data from an infographic without a text layer and turned it into a 52-row table.

#10 𝕏

Clement Delangue of Hugging Face announced that decision models can now run on-device in llama.cpp—described as free, fast, and private—and shared the command `llama serve -hf ggml-org/Kev-4B-GGUF`.

#11 𝕏

Harrison Chase commented on a Google Research multi-agent paper’s pattern of giving verifiers clean context, preventing explorers’ reasoning from swaying them toward bad proofs while still allowing unconventional ideas. He said Deepagents subagents receive separate context for the same reason.

#12 𝕏

Guillermo Rauch demonstrated an app that uses “quines”—programs that output their own source code—to have models “teach me back” what they did. The demo lets users reveal the Svelte code powering it step by step.

#13 𝕏

Google AI announced that a Project Suncatcher prototype satellite, built with Planet, launched aboard SpaceX’s Transporter-18 rideshare mission to test how Google TPUs withstand spaceflight. The findings will inform potential links between multiple satellite constellations for scaled machine learning, using near-constant sunlight in low Earth orbit to generate up to 8x more solar power than on Earth.

#14 𝕏

Aravind Srinivas recapped seven recent open-source contributions from Perplexity spanning models, benchmarks, and tools, including Perplexity’s pplx-decider-v1-27b, which he described as a state-of-the-art multimodal decision model. He added that more open-source contributions are coming soon.

#15 𝕏

Mustafa Suleyman announced that a Voice and Transcribe Streaming model is now available on Vercel, describing it as #1 for quality and speed. He claimed it is cheaper than any other hyperscaler and 60% cheaper than Eleven Labs.

#16 𝕏

Guillermo Rauch praised svelte.dev’s SvelteKit 3 and expressed excitement about Async Svelte and Remote Functions. He said a tiny app built and deployed end-to-end in 15 seconds, describing it as instant and saying “fx + opus” figured it out effortlessly.

#17 𝕏

Thariq demonstrated an animation editor he asked Claude to make for iterating on a jump animation in his game prototype. He said he used Claude for teaching and finding references to improve animation quality and shared a side-by-side video.

#18 𝕏

An interactive 3D jet engine made with Opus 5.5 was shared; users can cut it away and pull it apart.

#19 𝕏

Garry Tan highlighted Capy’s unexpected, model-agnostic cross-session coordination, saying his GBrain collaborator Sina started a fix wave that steered around Tan’s existing work across multiple Capy threads.

#20 𝕏

Lenny Rachitsky recapped how, five minutes before OpenAI’s DevDay live demo, @thsottiaux’s Dot noticed production was down and asked to fix it—prompting the response, “I don’t think you’re there yet, little Dot, but thank you for trying.”

Also covered by: @Lenny Rachitsky