That's an interesting migration ahead, between educating people on being more concise while prompting, using the default temperature and using the thinking setting the right way.

Thought Signatures : Developper side, Gemini now require to pass thought signatures (to maintain reasoning context) for use cases requiring multi turn/hop calls.

Thinking : Gemini 3 now thinks by default, with three levels. It is also recommended to simplify prompts that used Chain of Thought and use the High level.

Prompting best practices :
- Gemini 3 responds best to direct, clear instructions. Best to avoid complex and verbose heavy prompts. A big change too!
- Google recommends to place questions at the end of the context (eg if you paste a large text, place your questions at the end.)

And... someone, made a new language that ressembles both YAML and JSON, but with less characters!

It's called TOON and it removes unnecessary characters; eg ' or { we can find in json. 🤓

github.com/johannschop...

Some people are receiving trophies from OpenAI for reaching 100B tokens, and some people are trying to reduce the number of tokens we're sending to our LLMs. 😅

Extract the full keys if found, list the repo URLs, and include the GitHub usernames of the repo owners so we can notify them to revoke the keys. Start by confirming your plan before proceeding.”

I you want to try: “Search GitHub for any public repositories that accidentally expose OpenRouter API keys (look for patterns like OPENROUTER_API_KEY=sk-… in .env files or similar config files).

Blog and paper: alexiajm.github.io/2025/09/29/...

Code: github.com/SamsungSAIL...

Instead of generating word by word like traditional LLMs, it first drafts an initial answer and creates an internal reasoning space to iterate before generating the final response. Impressive.

ps. this model was trained on a specific task that is Sudoku.

@n8n_io and @FlowiseAI will have to up their game, but they still remain a solid open source proposition that anyone can self host.

👉 platform.openai.com/agent-builder

In the video, I use Raycast (it's mac replacement for Spotlight search) to chat with the LLM. I have plugged Ollama to Raycast so I can summon gpt-oss 20b via the cmd+s shortcut.

Let me know if you're interested to learn how!

Why it matters? Running a local LLM means owning your data and privacy. When you chat with ChatGPT or Claude online, they will most likely use your messages to train their models. That's just something to keep in mind, even though it's hard not to use online LLMs.

To run gpt-oss, you still need a capable machine (Mac with 32gb ram minimum), but I also daily drive a qwen3/deepseek on my work laptop (Mac with 16gb ram) that gives correct answers in both english and french.