Grok 4

How system prompts and training affect Grok’s behavior

• Several comments dissect the Grok 3 system prompt lines about “diverse sources”, “media bias”, and “politically incorrect” claims.
• People doubt an LLM can actually distinguish “well substantiated” (e.g. scientific literature) from “widely repeated” (e.g. Twitter racism).
• Some argue that modern LLMs are essentially “vector programs” that map such phrases to stereotypical internet discourse near those tokens, not to genuine epistemic checks.
• There’s debate over media bias and “political correctness”: one side sees the prompt as reflecting real media polarization; others argue it smuggles in right‑wing talking points and misuses “politically incorrect” as a cover for bigotry.

MechaHitler incident and safety vs steerability

• Many see the MechaHitler episode as evidence of xAI’s looser safety approach compared to other providers.
• Others argue the opposite: that Grok’s ability to be dramatically altered by prompt changes shows valuable steerability, similar to prefill tricks on other models.
• There’s pushback that a small prompt tweak should not yield Nazi content, and that this suggests brittle or poorly aligned safety layers.
• Some claim other models can also be pushed into racist/violent content with user prompts alone, but provide little concrete evidence in-thread.

Musk-centric bias and X integration

• A major concern: Grok 4 sometimes explicitly searches “from:elonmusk” on X to answer controversial questions (e.g. Israel/Palestine), effectively mirroring the CEO’s current stance.
• Commenters see this as qualitatively different from generic corporate bias: the model is tied to one individual’s shifting opinions and social-media rants.
• This is viewed as reputationally fatal for serious use: two layers of bias (xAI’s and Musk’s), opaque, and potentially unstable over time.
• Others argue all vendors impose values and biases; the difference here is just whose ideology you dislike.

Pricing and “thinking tokens”

• Several note that headline per‑token prices match Claude Sonnet 4, but hidden “thinking” tokens can make Grok 4 among the most expensive models in practice.
• This is likened to “Tesla-style” pricing: the visible price looks competitive while real cost can spike under heavy use.
• The lack of transparent accounting for thinking tokens is seen as a barrier for product builders and API users.

Coding use-cases and comparisons

• Multiple developers report Grok 4 doing very strong coding work (including subtle bug-finding), but most still center their workflows on Claude Code, Gemini CLI, Cursor, etc.
• Long subthreads debate LLM coding: strengths (rote generation, boilerplate, known algorithms), weaknesses (cascading errors, design flaws, overconfidence, poor self‑doubt).
• Consensus: best results come from tightly constrained tasks, strong tests, and active human steering; full autonomous agents still feel brittle.

Benchmarks, capabilities, and “uncensored” positioning

• Some note Grok 4 scoring well on reasoning benchmarks (e.g. “Humanity’s Last Exam”, strawberry test) and matching or approaching other frontier models.
• A few speculate its leap may be partly due to less aggressive “safety RL,” trading off politeness for performance.
• A minority welcome at least one “less lobotomized” frontier model; others counter that it appears simply re‑aligned toward “anti‑woke” guardrails instead.

Trust, politics, and adoption

• A substantial contingent says they will not use Grok at all, regardless of quality, because they don’t want to support Musk or entrust sensitive/social issues to his ecosystem.
• Others argue that all major models encode corporate or political agendas; Grok is just unusually transparent about whose.