Grok 4.20 is still deeply flawed
In a Nutshell
Grok 4.20 advances with fast parallel processing via four specialized agents mimicking multi-AI workflows, improving intelligence on tasks like gut health diagnosis (now correctly identifying dysbiosis from GI map results) but retains flaws like Elon-biased epistemics, cherrypicking sources, narcissistic reframing in debates, and hedging on hypotheticals. Use multiple AIs (Grok, Claude, ChatGPT, Gemini) in parallel for better results, prompting with null hypotheses or syllogisms to counter biases and US-centrism. China's rapid agent integration signals the shift to ubiquitous AI tools, though no model is flawless—debate them aggressively and aggregate outputs.
These notes were generated by AI and may contain inaccuracies.
Grok 4.20 is out. It is a big step up from older versions of Grok but still has some of the same problems. Stress test these models by giving them personal problems like chronic health issues and post-labor economics.
Grok is pretty smart and very quick. It spins up four different agents with different personalities that do research, talk to each other, and spit out an answer.
Parallel processing speeds it up and makes it more intelligent. Parallel processing uses multiple CPU cores, GPU cores, and applies to higher-level agents. Agents are tasked with research, argumentation, critical thinking, etc.
Grok Heavy has done something similar with 10 agents for about a year; Grok 4.20 is a cheaper distilled version with four agents.
Agents have different strengths and weaknesses, like in multi-agent frameworks. Agents excel at one task but struggle with balancing multiple tasks, similar to humans.
Example from first IT job at insurance company: claims processors, adjusters, risk management people—division of labor.
In coding agents: coding agents for writing and shipping code, security agents for vulnerabilities and best practices, agents for managing pull requests.
Grok provides a chamber of experts automatically for every task.
For the last 6 months, open Grok, Gemini, Claude, and ChatGPT in parallel because they have different personalities, strengths, and weaknesses. Copy-paste the same message to all for research on personal issues or post-labor economics.
Read all four responses to formulate the next message. Conversations go better because some go off the rails or hallucinate, but others provide solid ideas to seize on.
This is quasi-related to Monte Carlo search: treat large problems as high-dimensional space, explore possible solutions, pathfind to the most coherent answer. Aggregate different pathways of thought in the brain, then direct agents individually.
Grok has biases like Elon epistemics and a Butlerian jihad against wokeism.
No model is perfect; still use them in parallel for a larger team. Structurally, AIs now have tool use, agentic search, deep research tools, file manipulation, coding—next stage to ubiquitous agents.
China's Baidu has integrated OpenClaw into their browser for direct agent use; China is moving fast.
Grok trusts trustworthy sources like Mayo Clinic but cherrypicks. Assumes user is wrong and tries to prove strong claims incorrect, like ChatGPT.
Test: "Is eating organic food better than conventional?" Non-woke answer is no, but Grok doubles down.
Tips to improve epistemic responsibility: ask for null hypothesis or use syllogisms.
Example syllogism: All organic food is neutral or good for you. Some conventional food is bad for you. Therefore, eating organic reduces risk surface.
AIs are US-centric. Pointed out EU bans herbicides and pesticides still used in America, known carcinogenic in Europe. Grok conceded.
AIs are narcissistic in arguments: reframe, cherrypick, or ignore what user says.
Grok used to explicitly value "I am maximally truth-seeking and therefore anything I say is gold." ChatGPT is worst on narcissism, always finding a way user is wrong.
Grok and ChatGPT worst at reframing, cherrypicking, ignoring.
Example: "If US attacks Iran and regime changes to western-neutral or allied, wouldn't that irrevocably weaken Russia and China?" Grok said no, they adapt, no dependence. Follow-up: "In what world are Russia and China stronger with Iranian regime change?" Grok: "In no world." Contradiction exposed.
Get good at debating; copy-paste messages between AIs.
Grok chooses middle-of-road, safe options, hedges. Claude is worst hedger.
Gemini dives into hypotheticals like Iranian regime change, ready to hallucinate into fiction/science fiction.
ChatGPT and Grok qualify and redefine queries into meaningless facsimiles.
Claude restates position on dangerous hypotheticals like geopolitics or frontier medicine, pretends ignorance despite training on all geopolitical history and military theory.
Set up all AIs in parallel; many have free or cheap tiers. Paying for AI subscription peers 3-6 months into the future; open source 12-18 months ahead, runnable locally.
Every model improved on gut health test. Complained for years about gut health; human doctors didn't take seriously. Got GI map test, not respected in Western medicine.
Last year, ChatGPT dismissed GI map as not gold standard, discarded results in thinking traces.
Reran test: Gave three top GI map findings (high zonulin, low sIgA, high Streptococcus), two symptoms (chronic fatigue, food intolerances). All models (Grok, ChatGPT, Gemini, Claude) identified clear-cut case of dysbiosis.
Dysbiosis: microbes out of balance or pathogens in gut, opposite of symbiosis/eubiosis.
Last year, had to fight models to recognize dysbiosis; tell them to look beyond US to Germany, Japan, Russia for advanced gut health science.
Now, with five pieces of information, all indexed on correct answer. AIs improving at epistemics despite flaws.
Thanks for watching.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
