T
@ThePrimeagen
This is also why people have bad opinions LLM just reinforce terrible ideas with lengthy, PhD sounding arguments that are useful as a child with a recorder
1.3K
36
63.6K
@karpathyposted on X
- Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. - Wow, feeling great, it’s so convincing! - Fun idea let’s ask it to argue the opposite. - LLM demolishes the entire argument and convinces me that the opposite is in fact true. - lol The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.
View original tweet on X →
Community Sentiment Analysis
Real-time analysis of public opinion and engagement
Sentiment Distribution
87% Engaged
84% Positive
Positive
84%
Negative
3%
Neutral
12%
Key Takeaways
What the community is saying — both sides
Supporting
1
elite adversarial sparring partners
the highest-value use is stress‑testing ideas—have the model demolish your draft, surface edge cases, and force you to patch weak spots before shipping.
2
persuasion and coherence, not truth
they assemble the most rhetorically convincing case from training patterns, so a fluent argument ≠ an accurate one.
3
arbitrator, not validator
you must judge which case has better evidence or keep your priors explicit.
4
outsourced conviction
widespread reliance on a single model or pipeline creates centrally steered consensus that feels like independent thought; that’s an architecture problem, not just an interface one.
5
“argue the opposite,” steelman + red team, multi‑agent councils, cross‑model reviews, and blind judges
isolate roles so no agent can simply echo what it helped build.
6
the argument that survives attack is the one worth building
.
7
humbling and erosion of facile conviction
useful for better thinking but disorienting for anyone who mistook polished prose for understanding.
8
Technical and design fixes suggested
embed falsification heuristics, reward early pushback, add verification layers (code/run checks, citations), and tune prompts (ban overly persuasive tokens) so models favor accuracy over smoothness.
9
democratizes dialectic thinking
; the exercise of iterated argument/counterargument is itself the epistemic gain.
Opposing
1
simulate argument
. Use critical thinking — don't outsource judgment to rhetorical skill.
2
weaknesses in your reasoning
, not omniscience on the model's part.
3
lack of genuine understanding
; true comprehension, they say, collapses artificial "sides."
4
persuasion drives real-world action
rhetorical power matters even if it isn't proof.
5
opinionated and take a stand
, preferring conviction over neutral simulacra.
6
conviction is just a prompt away
, and they haven't yet seen major downside risks.
7
intuition and inner knowing
as superior guides to truth.
8
mirror ideological or think-tank rhetoric
rather than inventing independent truths.
Top Reactions
Most popular replies, ranked by engagement
S
@santanu_ai
This is basically Socratic method on steroids. The goal was never to "be right" — it's to stress-test your thinking until only the solid parts remain. Most people use AI as a yes-machine. The smart ones use it as a sparring partner.
249
8
21.7K
S
@SHL0MS
working on a method called autoreason that is effectively autoresearch extended to subjective domains. autoresearch works because val_bpb gives you an objective fitness function. autoreason constructs a subjective one through independent blind evaluation, the same way science u
132
10
16.4K
F
@FakePsyho
> The post mistakes persuasiveness for truth: if an LLM can argue both sides well, that shows rhetorical skill, not reliable judgment. “It convinced me of the opposite” says more about weak standards of evidence than about what’s actually true.
131
8
11.3K
B
@bullrungenius
If an AI can flip your worldview in 5 minutes, your "meticulous" 4-hour argument was probably weak from the start.
2
0
132
S
@subcountability
Seem dangerous. There's some strong metaphysical assumption here, about one direction of an argument "actually" being true. We can probably reasonably talk away every good argument. People talking others into their view is probably a condition for action to be taken at all.
2
1
142
Report an Issue
Found something wrong with this article? Let us know and we'll look into it.