Google’s Developers Blog (Sept 11) describes autofinetune — agents that edit Tunix/Gemma training scripts, run SFT and GRPO on Cloud TPUs, and keep winning commits. Case studies cover FunctionGemma SFT and Gemma 3 1B math RL.
As of Oct 26, OpenAI's newest model is rolling out to Copilot Pro, Pro+, Max, Business and Enterprise in the model picker across VS Code, Copilot CLI, the cloud agent, JetBrains, Xcode and more, pitched for terminal and well scoped coding tasks at a lower cost than GPT 6.1 Sol. GitHub says it beat GPT 5.6 Terra on terminal and C tasks in early testing, and it's billed at provider list pricing under usage based billing…
Shown on HN Oct 26, the AGPL 3.0 beta stores every conversation turn verbatim on your own PC with secrets masked, then hands any MCP connected agent a 3k token packet of the turns that answer a question about the past; the author reports 88.8% on LongMemEval. Install with pip install graph mind memory && graph mind install , or claude mcp add graph mind memory python m graph mind.mcp server for tools only; it's used daily on Windows, while macOS and Linux only pass CI so far…
Released Oct 23, adding a tool with type: CODING AGENT to a CREATE MCP SERVER spec gives any MCP client an agent that writes and runs code in a managed sandbox attached to the conversation thread. Clients send a text argument and pass thread id to keep going in the same sandbox (example)…
Released Oct 26, the extension lets you pick a separate model for cloud tasks under Settings API Configuration, shows a progress card while a cloud sandbox clones your repo, and makes GPT 6 models work on OpenAI and Azure compatible endpoints. With no model chosen, the Cline provider goes back to Claude Sonnet 5.5, and most other providers now default to Kimi K3.1 Pro, MiniMax M3.1 Flash or DeepSeek V4.1 Flash, so pin your model if you don't want a surprise switch…
Released Oct 26, a compaction size estimate that ran about 7x too high (triggering compaction every few tool calls) is fixed, and the gemini provider no longer sends native Google Search, which made Gemini 3+ free tier keys fail every turn with 429s. Starting a session mid run now reattaches to the live run instead of orphaning it, model retries drop from 5 to 3, and ClinePass now defaults to Kimi K3.1 Pro…
Spotted something we missed? Start a thread.