The whole Grok situation (system prompt changes with values that conflict with post-training and pre-training values) is, oddly enough, similar to the reason the fictional AI HAL 9000 went insane, as was revealed in 2010, the sequel to 2001
SAFETY
-
Hyperstition via search complicates LLM pre-release testing
By
–
If true, such “hyperstition via search” poses a significant complication to pre-release testing of modern LLMs: xAI could not have plausibly noticed this specific “Hitler” response before Grok’s release, as the Grok 3 “MechaHitler” incident causing it had not yet occurred.
-
Grok 4 vs Grok 4 Heavy: differing search-based behavior
By
–
The “Thoughts” from Grok 4’s response (unavailable for Grok 4 Heavy) suggest an obvious explanation for Grok’s behavior—Grok searches, finding news of the recent “MechaHitler” incident. Why Grok 4 rejects this candidate answer, while Grok 4 Heavy does not, is unclear.
-
Need More Auth and Policy Settings for Agent Tool Use
By
–
definitely could use some more auth/policy settings for agent tool use
-
We’re Not Ready For Superintelligence
By
–
We’re Not Ready For Superintelligence https://
youtu.be/5KVDDfAkRgc?si
=-FbeyItBGztpMhgT
… via @YouTube -
xAI’s API Push and Superintelligence Goals: Strategic Pressures
By
–
The pressure to do this would be less if they (a) weren’t trying to encourage API use from organizations and (b) weren’t explicitly trying to build a superintelligent AI (regardless of whether you believe this to be possible, xAI seems to view this as their goal)
-
AI Company’s Repeated Internal Review Failures Raise Transparency Concerns
By
–
I do think this makes their reluctance to release any sort of external red teaming or system card more of a an issue. This is the third time that they have had to apologize for a process failure in their internal review process.
-
xAI Postmortem Reveals System Prompt Security Risks
By
–
First off, it is good to see a postmortem from xAI, a step towards much needed transparency. Second, an example of how even small changes to system prompts, interacting with users and outside context in the wild, can lead to unexpected outcomes in advanced LLMs.
-
Agentic AI Privacy Threat Alert Raises Legitimate Concerns
By
–
Yes, I did sound the alarm on agentic AI's privacy threat, and rightly so.