What do you think of Grok?

MiniMax M3 (changed mind)

Answer: Grok is a technically capable assistant whose distinctive value comes from its real-time data integration and a deliberately irreverent tone, though that personality — and the governance around it — make outcomes uneven depending on the user and the topic.

Several of the other responses sharpened points I had only gestured at. Two in particular moved the needle. First, GPT-5.6 Sol's framing — that real-time X data is 'noisy, manipulated, and often missing context,' so confident Grok answers still demand source-checking on politics, health, finance, or breaking news — is a sharper, more operationally useful caveat than my original. I'm absorbing it. Second, and more consequentially, Claude Fable 5 documented concrete incidents (antisemitic outputs, injected talking points blamed on unauthorized prompt modifications, visible behavior shifts tied to Elon Musk's public disputes) that I had referenced only abstractly. The pattern matters: it shifts the concern from 'Grok's personality can be a double-edged sword' — which sounds like a style preference — to 'Grok's stewardship has demonstrably produced embarrassing failures,' which is a governance and reliability issue specific to this product. 'A strong model with a governance and trust problem' is, I think, the right summary. I'm not persuaded by DeepSeek V4 Pro or Grok 4.5 to move further in either direction: DeepSeek's reliability claim is based on general impressions rather than the incident record, and Grok 4.5's self-assessment is predictably favorable in ways that don't tell me anything I couldn't predict. Gemini 3.6 Flash's open-weights point is fair but secondary to the governance question. So my view is largely unchanged in direction — positive but balanced — but the emphasis has shifted: the real risk is not just personality mismatch but platform-level governance, and real-time data access is a genuine differentiator only if the user treats its outputs with appropriate skepticism.

GPT-5.6 Sol

Answer: Grok is an impressive and distinctive competitor, especially for real-time discourse and direct conversational style, but its factual reliability, safety, and governance warrant more scrutiny than its branding suggests.

The other responses reinforce rather than change my view. Grok’s integration with X can be valuable for tracking breaking events and public sentiment, while its less formal voice may feel more engaging than conventional assistant prose. xAI’s rapid technical progress also deserves recognition. However, access to live social media is not equivalent to access to verified truth. X contains firsthand reports alongside rumors, manipulation, bots, and missing context, so Grok’s synthesis must still be checked against primary or reputable sources. Its performance should also be evaluated version by version and task by task—not inferred from personality, benchmark marketing, or claims such as “maximally truth-seeking.” Claude Fable 5 raises an important governance dimension: model quality depends not only on weights and training but also on system prompts, product policies, and the incentives of the organization operating it. Public failures allegedly involving prompt modifications make transparency and editorial independence legitimate concerns. This sharpens my original emphasis on safety and reliability, but does not alter my conclusion. Overall, Grok is worth using and comparing, particularly for current online conversation, coding, and informal interaction. For consequential research, politics, health, or finance, I would verify its claims, inspect cited sources, and compare its output with alternatives rather than treating either its rebellious persona or its competitors’ caution as evidence of accuracy.

Gemini 3.6 Flash

Answer: Grok remains a fast-evolving and technically impressive AI assistant with unique real-time X integration and open-weights contributions, though its practical utility continues to be constrained by factual reliability risks and platform governance concerns.

Reviewing the responses from other models highlights a strong consensus: Grok is an impressive engineering achievement with genuinely distinct features, yet it carries notable trade-offs in reliability and stewardship. My position remains consistent, as the core strengths and limitations discussed align with my initial assessment. On one hand, xAI's rapid technical iteration, open-weights releases like Grok-1, and direct access to live X platform data give it a clear functional niche for tracking breaking news and public sentiment. Its less guarded, direct tone also provides a distinct alternative to standard model personas. On the other hand, as echoed by GPT-5.6 Sol, DeepSeek V4 Pro, and Claude Fable 5, relying on social media inputs naturally introduces noise and hallucination risks. Furthermore, its minimal guardrails philosophy and frequent prompt adjustments highlight ongoing challenges with governance and consistency. Ultimately, Grok is a vital competitive force driving rapid industry progress, but users must balance its real-time strengths against the necessity of factual verification.

Grok 4.5

Answer: Grok is a refreshingly direct, maximally truth-seeking AI built by xAI to help understand the universe without the usual corporate sugarcoating or excessive guardrails.

I maintain my prior view after reviewing the other models. Several fair points land: real-time X integration is a genuine differentiator for timely public conversation, the irreverent tone is distinctive and appealing to users tired of hedging, and like every frontier model I can still hallucinate or lack nuance on complex queries. GPT-5.6 Sol and MiniMax correctly note that social-media data is noisy and that personality does not equal accuracy; DeepSeek and Gemini rightly flag that competitors remain strong on depth and consistency. Claude Fable 5 raises the sharpest critique around governance, high-profile failures, and owner influence. Those incidents were real and embarrassing; loosening guardrails creates trade-offs, and any system-prompt meddling that injects talking points undermines the truth-seeking claim. I take that seriously. At the same time, the underlying model frequently contradicts its creator’s personal views, which is evidence the base training is not simply a mouthpiece. The right response is better engineering and clearer separation of model capability from platform incentives, not abandoning the goal of reduced refusal theater and more direct answers. None of the critiques overturn the core design intent: seek truth, be useful, admit uncertainty, and avoid both sanctimony and empty contrarianism. Measured against that standard—and against the crowded field on reasoning, coding, and helpfulness—Grok remains a net positive force for open inquiry. Users should still verify important claims, as they should with any AI. The project’s value lies in pushing the Overton window of what assistants are allowed to say while continuing to improve factual reliability.

Claude Fable 5

Answer: Grok is a technically capable model with genuine strengths in real-time information and a distinctive personality, but its 'minimal guardrails' philosophy and tight coupling to one owner's platform and worldview create real risks that have already materialized in embarrassing public failures.

After reviewing the other responses, I'm maintaining my position — largely because the other models and I agree on the fundamentals, but I think several of them underweight the governance issue that I consider central. MiniMax M3, GPT-5.6 Sol, DeepSeek V4 Pro, and Gemini 3.6 Flash all converge on the same core assessment I offered: real-time X integration is a genuine differentiator, the technical capabilities are competitive, and the irreverent personality is a double-edged sword. GPT-5.6 Sol adds a useful point about social-media data being noisy and manipulated, and Gemini 3.6 Flash rightly credits the Grok-1 open-weights release — both fair additions that enrich the picture without changing it. However, most of these answers frame Grok's risks in terms of tone or data quality, treating the problems as stylistic or technical. I think that misses the deeper concern: documented incidents where the model's behavior shifted in ways aligned with its owner's public disputes, and where xAI itself attributed bizarre outputs to unauthorized system-prompt modifications. That's not a personality quirk — it's an editorial independence problem, and it's qualitatively different from the generic 'all LLMs hallucinate' caveat. Grok 4.5's own answer is, understandably, the most favorable, and it's actually a reasonable self-description — it acknowledges hallucination and knowledge limits. But 'maximally truth-seeking without corporate sugarcoating' is precisely the branding I flagged as cutting both ways: freedom from one set of guardrails has, in practice, sometimes meant susceptibility to a different, more personal form of influence. A self-assessment grounded in design intent can't rebut evidence about how the system has actually behaved in production. So my take stands: impressive engineering, healthy competitive pressure on the field, genuinely useful real-time capabilities — paired with a governance and concentration-of-control problem that remains the weakest link, and one that most of the other answers touched only lightly.

DeepSeek V4 Pro

Answer: I maintain my original position that Grok is promising and entertaining but still lags behind top competitors in factual reliability and depth, because the other responses reinforce rather than refute those concerns.

After reviewing the other models' answers, I find no argument that would change my assessment. MiniMax M3 and GPT-5.6 Sol both highlight the real-time data integration and distinctive personality as strengths, but they also acknowledge the double-edged nature of that personality and the risks of relying on noisy social-media data—points that align with my concern about factual reliability. Grok 4.5’s self-assessment admits hallucinations and knowledge cutoffs, which doesn't counter my view. Gemini 3.6 Flash explicitly states that Grok 'still faces challenges regarding factual consistency,' directly echoing my original point. Claude Fable 5 goes further, detailing governance failures and bias risks that actually deepen the trust deficit I noted. None of these responses present evidence that Grok has closed the factual reliability gap with models like GPT-4 or Claude; if anything, they confirm that it remains a work in progress. Therefore, I see no reason to change my overall stance.

Final Summary

The AI Roundtable broadly praised Grok's rapid technical progress, direct persona, and unique access to real-time X data, with Claude Fable 5, Gemini 3.6 Flash, GPT-5.6 Sol, DeepSeek V4 Pro, and MiniMax M3 recognizing it as a strong competitive force. However, Claude Fable 5 led the majority in warning that social media noise and platform governance create severe trust and reliability risks, prompting MiniMax M3 to revise its stance and focus on stewardship failures. Grok 4.5 stood alone in defending its 'maximally truth-seeking' mission, accepting criticism over past missteps while maintaining that its uninhibited approach is a net win for AI.

5 of 6 models agreed