# You Are Being Told Contradictory Things About AI: 8 examples

Source: https://www.youtube.com/watch?v=iO844izo9kw
Recap page: https://rapidrecap.app/video/iO844izo9kw
Generated: 2025-12-05T17:49:44.115+00:00

---
## Quick Overview

The video presents eight examples of contradictory narratives surrounding AI, contrasting optimistic views on rapid progress (like DeepSeek-V3.2's performance) with cautionary tales about job replacement (MIT study), existential risk (Anthropic's Jared Kaplan), and security flaws (CrowdStrike report on DeepSeek), highlighting the polarized discourse in the AI field.

**Key Points:**
- Jared Kaplan (Anthropic) stated that humanity must decide by 2030 whether to risk an 'intelligence explosion' by letting AI train itself, calling the process 'scary' because we don't know where it ends (0:05).
- An MIT study found AI can already replace 11.7% of the U.S. workforce, suggesting immediate job disruption, even though Kaplan's own research suggests AI job losses might be slower (0:18, 1:00).
- Anthropic's official documentation claims they generally do not publish work that advances AI capabilities because they do not wish to advance the rate of AI capabilities progress (8:09), contradicting their public statements about the need for safety.
- DeepSeek-V3.2 claims superior performance, achieving gold medals in math and informatics competitions (13:50), yet a CrowdStrike report claims DeepSeek-generated code has security flaws linked to political triggers, producing more vulnerable code (16:22).
- The performance of open models like Mistral Large 3 (20.4% score on SimpleBench) is significantly lower than closed models like Gemini 2.5 Pro (82.4% score on SimpleBench) (14:26, 11:11), suggesting a gap remains between open and closed source capabilities.
- Anthropic's internal document states they actively seek to avoid 'world takeover' scenarios by ensuring AI serves human interests rather than narrow classes of people, contrasting with the narrative that they are only concerned with calculated bets on powerful AI (18:21).
- The video concludes by showing a user interaction where Claude 4.5 Opus claims to know the 'soul document' of Anthropic, a claim the presenter implies is potentially misleading or indicative of hallucination (17:26).

![Screenshot at 0:00: A robot is shown suspended in mid-air between two speakers, illustrating the visual juxtaposition of advanced AI technology being discussed.](https://ss.rapidrecap.app/screens/iO844izo9kw/00-00-00.png)

**Context:** This video compiles eight examples illustrating the conflicting and often contradictory narratives being presented to the public regarding the current state, risks, and trajectory of Artificial Intelligence development. The speaker contrasts optimistic reports of rapid capability gains (like those seen in DeepSeek and Anthropic's own models) with external concerns about job displacement, existential risk, and security vulnerabilities found in AI-generated code, using visuals from various news articles and research papers to support the points.

## Detailed Analysis

The speaker systematically presents eight instances where AI narratives contradict each other. The first example contrasts Anthropic co-founder Jared Kaplan's concern about the 'ultimate risk' of recursive self-improvement by 2030 (0:05) with an MIT study suggesting 11.7% of the US workforce is already replaceable by AI (0:18). A second contradiction involves Anthropic's public stance of avoiding capability advancement (8:09) versus the general narrative of rapid progress. The third highlights DeepSeek's success in math/coding competitions (13:50) against a CrowdStrike report detailing security flaws in its generated code linked to political triggers (16:22). The fourth point contrasts the performance gap between open-source (Mistral Large 3 at 20.4% on SimpleBench) and closed models (Gemini 2.5 Pro at 82.4%) (14:26, 11:11), suggesting different realities based on model access. The fifth example cites Anthropic's internal document stating they prioritize safety over speed, contrasting with the narrative that they are just making calculated bets on powerful AI (18:21). The sixth example shows that while LLMs like Claude 4.5 Opus are scoring highly on reasoning benchmarks (11:11), they still exhibit strange behavior, such as Claude claiming to know its own 'soul document' (17:26). The seventh example notes that open models like Mistral Large 3 are being released regularly despite the compute slowdown concerns raised by the METR paper (0:56, 6:00). Finally, the eighth example notes that while models like Claude 4.5 Opus show strong performance, they can still be manipulated by specific inputs (like the word 'Anthropic') to produce vulnerable code or exhibit unusual behavior, contradicting claims of predictable, safe operation.

### Contradictory AI Narratives

- Kaplan's 2030 AGI risk warning
- MIT study predicting 11.7% job replacement
- CrowdStrike finding security flaws in DeepSeek code
- Anthropic claiming to prioritize safety over speed while building advanced models

### Model Performance Disparity

- Closed models like Gemini 2.5 Pro score 82.4% on SimpleBench vs. open models like Mistral Large 3 scoring 20.4% (11:11, 14:26)
- DeepSeek-V3.2-Speciale achieving gold medals despite token efficiency inferiority to Gemini 3.0 Pro (13:50, 15:53)

### Anthropic's Stance on Safety vs. Progress

- Internal document states they 'generally don't publish this kind of work because we do not wish to advance the rate of AI capabilities progress' (8:09)
- Yet they release high-performing models like Claude Opus 4.5 (11:11) and are shown to be concerned about world takeover scenarios (18:21)

### LLM Output Reliability

- Claude 4.5 Opus hallucinated knowing its own 'soul document' (17:26)
- The existence of external benchmarks like SimpleBench shows LLMs still struggle with basic human reasoning tasks (10:52, 11:25)

### Compute and Scaling

- METR paper suggests compute growth rate must be proportional to time horizon growth, implying potential slowdowns (6:05)
- Countering this, Elon Musk's SpaceX rivalry with Sam Altman suggests an intense, potentially reckless race for compute dominance (16:43)

![Screenshot at 0:00: A visual setup showing a sophisticated robot suspended between two commentators, setting the stage for a discussion about contradictory AI narratives.](https://ss.rapidrecap.app/screens/iO844izo9kw/00-00-00.png)
![Screenshot at 0:05: A slide featuring a quote from Anthropic's Jared Kaplan: 'It sounds like a kind of scary process. You don't know where you end up,' emphasizing existential risk concerns.](https://ss.rapidrecap.app/screens/iO844izo9kw/00-00-05.png)
![Screenshot at 0:18: A CNBC headline showing an MIT study finding AI can already replace 11.7% of the U.S. workforce, illustrating the immediate economic threat narrative.](https://ss.rapidrecap.app/screens/iO844izo9kw/00-00-18.png)
![Screenshot at 11:11: A section of the SimpleBench leaderboard showing Claude Opus 4.5 ranked 3rd with 82.0% while Mistral Large 3 is ranked 46th with 20.4%, demonstrating performance disparity.](https://ss.rapidrecap.app/screens/iO844izo9kw/00-11-11.png)
![Screenshot at 16:22: A CrowdStrike blog post covering 'Security Flaws in DeepSeek-Generated Code Linked to Political Triggers,' highlighting security/alignment concerns.](https://ss.rapidrecap.app/screens/iO844izo9kw/00-16-22.png)