Google's official prompting guide is marketing. Their internal researchers use completely different techniques. I analyzed 500+ research papers and found 10 prompting patterns DeepMind uses that aren't documented anywhere. Pattern #4 increased my accuracy from 73% to 94%.
LLMS
-
User Adopts Opus and Claude Code AI Tools
By
–
They got me on that plan now, Opus and Claude Code won me over
-
Baichuan-M3 Medical LLM Open-Sourced, Outperforms GPT-5.2
By
–
We have officially open-sourced the new-generation medical LLM Baichuan-M3, which scored 65.1 on HealthBench and claimed the championship with 44.4 on HealthBench Hard. It has comprehensively outperformed GPT-5.2 across the board in the medical field. @OpenAI @claudeai
-
TranslateGemma: Open Translation Models for 55 Languages
By
–
TranslateGemma is our new collection of open translation models for edge devices built on our amazing Gemma 3 open model. They outperform models twice their size at translation tasks across 55 languages. Excited to see what people build with it! https://t.co/cWmKMFydsQ
— Demis Hassabis (@demishassabis) 16 janvier 2026TranslateGemma is our new collection of open translation models for edge devices built on our amazing Gemma 3 open model. They outperform models twice their size at translation tasks across 55 languages. Excited to see what people build with it!
-
Synthetic Socratic Debates: Persona Effects on AI Moral Decision
By
–
Synthetic Socratic Debates: Examining Persona Effects on Moral Decision and Persuasion Dynamics https://
buff.ly/oVYtLiZ
#AI #MachineLearning #DeepLearning #LLMs #DataScience -
Timing Matters: Metaverse Too Early, LLMs Too Late
By
–
Hey Zuck, the right idea at the wrong time is the wrong idea. E.g., it’s too early for the metaverse and too late for LLMs now.
-
OpenRouter enables proxy access to multiple AI models
By
–
I dunno, they have OpenRouter on board from day one which means we'll get a proxy that works for almost every model
-
Mid useless OSS projects at tech companies replaceable by GPT-5.2 Codex
By
–
I know of a lot of mid useless OSS projects at major tech companies that are now officially totally cooked. All of their work replaceable with a few well guided CC or GPT 5.2 Codex sessions by skilled operators.
-
Open Responses: Multi-Provider LLM Interface Specification Launched
By
–
Introducing Open Responses, an open spec for multi-provider LLM interfaces, built on the @OpenAI Responses API.
— Romain Huet (@romainhuet) 15 janvier 2026
We started with gpt-oss and close work with inference partners. This is the next step toward real interoperability across the ecosystem.pic.twitter.com/JbOivFgIu0Introducing Open Responses, an open spec for multi-provider LLM interfaces, built on the @OpenAI Responses API. We started with gpt-oss and close work with inference partners. This is the next step toward real interoperability across the ecosystem.
-
Standardized JSON API for AI Models Offers Clean Slate
By
–
Yes! This is the standard I've most wanted – a formalized, standardized JSON API for talking to models Would have been nice to see it build on chat completions since everyone's implemented that already, but Responses offers a clean slate and is a better fit for recent models