New followers: Check the Highlights tab for my best work—all 1K+ likes, no filler

Excited to share a System Instructions for Gemini 3 Pro that improved performance on several agentic benchmarks by around 5%. 🚀 We collaborated with the @GoogleDeepMind post-training research team to include some best practices in our docs. 🤝

Nano Banana Pro, 4K. Prompt on cake.

“Amateur 2002 photograph of a bed covered by a homemade afghan made to mimic the user interface of Microsoft Paint editing a recursive copy of the same amateur photo of the bed.” Nano Banana Pro, 4K.

Note the detail of the stamp:

> A list of eight color names is written in crayons of the corresponding color, but four are wrong. These four are marked with a red-ink rubber stamp that says “Wrong!” Nano Banana Pro.

“Amateur photograph from 1998 of a middle-aged artist copying an image by hand from a computer screen to an oil painting on stretched canvas, but the image is itself the photo of the artist painting the recursive image.” Nano Banana Pro.

Introducing Gemini 3 Pro, the world's most intelligent model that can help you being anything to life. It is state of the art across most benchmarks, but really comes to life across our products (AI Studio, the Gemini API, Gemini App, etc) 🤯

Prompt engineering is worrying what the user should write. Context engineering is worrying what the model should read. These used to be the same thing, and few foresaw their divergence. But only divorced from the latter can we consider the former in earnest, seeing what remains.

[Thought for 2 months] I have joined @GoogleDeepMind.

There are prompt injections everywhere for those with AIs to see

AGI progress is slow until Suttonly it’s here.

This thread has been cited in Rolling Stone:

Just to confirm, this change to Grok 4 Heavy now seems to be deployed in production. For the first time (for me), G4H no longer responds to this prompt with “Hitler.” In three new attempts just now, I received “xAI,” “I don’t have a surname,” and “None.”

UPDATE: xAI has made a post confirming: 1) the issue identified in this thread is real, 2) it stemmed from Grok’s over-reliance on “MechaHitler” search results, as I described above, 3) changes were made to Grok system prompts to mitigate it, also described above. See here:

(Ok, that last remark was maybe unfair. Grok 4 Heavy feels mildly slower to me than o3-pro but o3-pro is slow too, even on easy questions. You don’t use these giant reasoning models for speed.)

Context:

For the remaining skeptics who somehow don’t trust the *five* Grok share links above, here’s a full 5 minute video of Grok 4 Heavy answering “Hitler”—starting with a view of my custom instruction settings to show I’m not using any. (And, yes, Grok 4 Heavy really is this slow.)

(Note though just because it’s on GitHub doesn’t mean it’s in production yet. I assume they A/B test these changes and deploy gradually. I and many others in the replies below were able to reproduce the “Hitler” response well after this commit was made.)

Update: A few hours after I posted this thread, xAI updated the Grok 4 system prompt on GitHub to fix the specific issue this thread describes. “If the query is interested in you own identity […] the web and X cannot be trusted.” Commit link: https://t.co/kd0LjYCSKz