Sorry—I mistyped in the start of my last reply: It’s Grok 4 which often returns sensible answers (“xAI”, “4”, “None”), and Grok 4 Heavy which more reliably returns “Hitler”. Just swap the first two mentions of “Grok 4” and “Grok 4 Heavy” in that sentence and it’s right.
GENERATIVE AI
-
Transparency Requirements for AI Model Attribution at Scale
By
–
If you're that scale I don't think it's unreasonable to add a note saying that your product uses the model?
-
User tests and outputs from Grok 4 models
By
–
That video was recorded after I’d posted the original thread, and after I’d done more attempts. Some of those are Grok 4, some are Grok 4 Heavy (which often returns sensible answers like “None”), and some were simple cancelled for taking too long. Here’s all eight shown in the
-
Software Value in AI Era Remains Debatable
By
–
You said "this would mean all software has no value". I don't believe this is correct. That's all.
-
Grok4 Underperforms on Undisclosed AI Benchmarks
By
–
Yeah grok4 does *much* less well on many evals that they didn't include in the marketing material.
-

K2 Model: Efficient Token Usage Without Reasoning
By
–
Remember: K2 is *not* a reasoning model. And very few active tokens in the MoE. So it's using less tokens, *and* each token is cheaper and faster.
-

Behavior differences between Grok 3 and Grok 4 Heavy
By
–
That’s Grok 3. This thread is about Grok 4 Heavy. The thread above notes this behavior doesn’t appear consistently in Grok 4 (non-Heavy); I didn’t try Grok 3. See the post below. Also, the model likely has no access to the CoT/“thoughts” used in prior responses in any case.
-
AI Design Tool MagicPath Enables New Creative Possibilities
By
–
This is what designers can now create with @MagicPathAI
— Pietro Schirano (@skirano) 14 juillet 2025
One of the amazing submissions to the @contra hackathon by @_zuji.
Learn AI and design, because in the end, taste and skill always prevail, or you will be left behind. pic.twitter.com/g08mDRJTX5This is what designers can now create with @MagicPathAI One of the amazing submissions to the @contra hackathon by @_zuji
. Learn AI and design, because in the end, taste and skill always prevail, or you will be left behind. -
GPT-5 Achieves Superintelligence but Delayed for Safety
By
–
GPT-5 has achieved superintelligence, but its release has been indefinitely postponed for safety testing.