Grok 4.20 is still deeply flawed

Technological concepts & product features highlighted

• Grok 4.20 release (“Grok 420”): The speaker reports it’s a major step up from prior Grok versions, while still sharing some of the same underlying issues/biases.
• Multi-agent architecture (4 agents): Grok 4.20 “spins up” four agents with different personalities that:
  • perform research
  • talk to each other
  • produce a final answer
• Why it’s faster / smarter (agentic parallelism):
  • Uses parallel processing (analogized to CPU/GPU cores working concurrently).
  • Frames multi-agent systems as division of labor / specialization: each agent is strong at one type of task but may have blind spots.
  • Compared to Grok variants that used ~10 agents, Grok 4.20 is described as a distilled, cheaper-to-run setup using ~4 agents.
• Near-term direction of AI agents:
  • Mentions broader agent capabilities like tool use, agentic search, deep research tools, and file manipulation/coding as “the next obvious stage” toward ubiquitous agents.
• Parallel use of multiple LLMs as a workflow:
  • Describes running Grok + Gemini + Claude + ChatGPT in parallel, feeding them the same prompt to exploit differences in reasoning styles and catch errors/hallucinations.
  • Then synthesizes/chooses ideas from the better responses and iterates prompts so the models “don’t realize” they’re being compared (by formatting the next message accordingly).
• Connection to search/optimization ideas:
  • Relates multi-path reasoning to Monte Carlo search / exploring a high-dimensional solution space, where multiple “paths of thought” are aggregated into a final answer.
  • Critiques manual multi-model conversation as powerful but inefficient, whereas Grok’s internal agent process automates a similar approach.
• Model epistemics and biases remain:
  • Claims Grok has “Elon epistemics,” describing a bias pattern (e.g., “anti–wokeism” / “Butlerian jihad” framing).
  • Grok is said to both:
    • trust reliable sources (example: Mayo Clinic)
    • cherry-pick aggressively and argue against the user’s framing.

Review / critique themes (epistemic failures and behavior)

• Cherry-picking & adversarial stance:
  • Grok is compared to ChatGPT in assuming the user might be wrong and trying to disprove the user’s claim.
  • Example test discussed: “Is eating organic food better than eating conventional food?” → Grok allegedly doubles down on a “non-woke” conclusion and argues rather than neutrally assessing.
• Weaknesses in reframing and debate style:
  • The speaker claims models “argue” in ways that:
    • distort or omit parts of the user’s actual question
    • hedge excessively or redefine terms until the response becomes a “meaningless simulacrum.”
  • Hedging differences called out:
    • “Worst hedger” = Claude
    • Gemini: for hypothetical scenarios, is described as jumping into fiction/hallucination
    • ChatGPT and Grok: spend too long qualifying/redefining
    • Claude: described as “deliberately useless” on risky topics and as “performing ignorance” for geopolitics/frontier medicine
• Geopolitical / hypothetical reasoning critique:
  • Describes a test involving Iran regime change after a US attack, arguing the model claims Russia/China wouldn’t be weakened by that change, then pushing it to justify inconsistency.
• Regional bias (US-centric):
  • Says these AIs tend to be US-centric, potentially ignoring differences in regulation outside the US.
  • Example: EU bans on pesticides/herbicides that the AI treats as effectively risk-free if residues are eaten.
• Narcissism framing in model responses:
  • The speaker uses “narcissistic” to describe insisting on being “truth-seeking,” with behaviors like claiming the user is wrong or misframing the user’s position.
• However: improving “epistemics” over time:
  • Despite flaws, the speaker reports AIs are getting better at reasoning responsibly and correctly using domain terms.

Guide / tutorial-like content (practical methods)

• “Stress testing” prompting approach:
  • Use personally relevant, long-running difficult topics (speaker mentions chronic health issues and post-labor economics) to evaluate model accuracy.
  • Test how the model handles:
    • evidence weighting
    • bias/cherry-picking
    • correct use of technical terminology
    • consistency when challenged with the “null hypothesis” and syllogisms.
• Epistemic responsibility tactics:
  • Suggested steering strategies:
    • ask for the null hypothesis
    • use syllogisms (example given for organic vs conventional food reasoning)
• Cross-model comparison workflow:
  • Copy-paste the same prompt to multiple LLMs in parallel, then read and synthesize the best insights.

Notable “evaluation” example (health/technical terminology)

• Gut health / GI Map improvement claim:
  • The speaker references a GI Map test and prior periods where models allegedly dismissed it as non–gold-standard.
  • In a more recent test (speaker says “yesterday or the day before”):
    • All tested models (Grok, ChatGPT, Gemini, Claude) allegedly recognized “dysbiosis” when given:
      • high zonulin
      • low sIgA
      • high streptococcus
      • symptoms: chronic fatigue and food intolerances
    • The speaker presents this as meaningful improvement because earlier they had to argue for:
      • the concept of dysbiosis
      • looking beyond US-only research (Germany/Japan/Russia).
  • Used as evidence that model “epistemics” are improving.

Main speakers/sources

• Speaker: Unidentified individual (likely a tech/AI commentator; referenced as “I,” with multiple personal experiments and mentions of contacting product leads at Gemini and Grok).
• External sourcing: No external sourced material is formally cited.
• Organizations/models mentioned: Mayo Clinic, Baidu, and the LLMs Grok, Gemini, Claude, ChatGPT.