36. Fable/Mythos, Loop Engineering [Persian]

Overview

The video covers recent developments in AI model capabilities, focusing on the shift toward “loop” / agentic coding systems. It also discusses competitive and safety implications for major labs—especially Entropic and OpenAI—framed through ideas like “Mythos/Fable” and “Loop Engineering.”

1) IPO hype and “dangerous capability” claims

• The narrator argues that when major AI-related companies pursue IPOs (e.g., mentions of SpaceX, Entropic, OpenAI), public discussion can become hype-heavy, with narratives from investors and government potentially exaggerating risks.
• He claims Entropic’s “Mythos” was presented to government as potentially “very dangerous”, such as rapidly finding cybersecurity vulnerabilities.
• A proposed executive order (attributed to Trump-era advisors) to restrict or delay something is said to have been initiated due to danger-and-geopolitical-competition claims, then canceled.
• The narrator implies that these political swings influence which capabilities become broadly available later.

2) “Loop Engineering” as the key technical shift

The main technical concept is Loop Engineering, contrasted with prompt engineering.

Instead of sending a single prompt, loop-based systems define:

• a goal
• iterative steps/iterations
• a final condition / stopping criterion

The narrator describes loops as systems where the model handles prompting, verification, and iteration, rather than users manually crafting multi-step prompt sequences.

This creates a closed system (a loop of agents/tools/verifiers), which is argued to:

• improve reliability of outputs
• enable major productivity gains

3) Productivity and code-quality claims for “Mythos/Fable”

The narrator reports Entropic’s claims that:

• AI output/code volume increased dramatically (e.g., “8 times as many lines” per time period)
• AI-written code is more accurate and increasingly readable
• AI can find bugs humans miss and improve code review

A key concern raised is that if capability improves via a closed loop, faster AI-assisted development could accelerate rapidly—potentially outpacing human oversight.

4) Safety debate: global governance vs. restricting access

After discussing fast capability growth, the narrator highlights a policy risk:

• the loop/capability feedback cycle might keep increasing productivity without a natural plateau, escalating impact over time.

He argues (via commentary attributed to a referenced “brand”/article) that:

• an international/global organization should coordinate and oversee research
• this oversight might even include pausing research to assess impacts

The narrator suggests this outcome may not occur automatically, but frames it as a need for global governance.

5) OpenAI’s stance and “safety net” (blocking certain capabilities)

The video claims OpenAI published an article advocating:

• broad access to advanced AI assistants
• automation of work and research
• an eventual aim to “automate humans” (framed by the narrator as strategic alongside economic growth)

However, the narrator contrasts this with what he describes as a competitor-targeted approach:

• OpenAI uses a “safety net” / safeguard that refuses certain outputs (e.g., specific biology questions or cybersecurity misuse)
• The narrator interprets this as potentially breaking competitors’ loops by restricting what others can derive from the models, which could reduce:
  • interoperability
  • innovation

6) Demonstrations and the narrator’s own “loop” experiments

The narrator provides examples of loop-based systems that generate:

• apps and code “from feature requests”
• game creation (including Minecraft-like and horror game examples)
• 3D mapping/visualizations
• networking/packet-behavior visualizations (e.g., HTTP/UDP/TCP/ping)
• video/animation clock sequences
• technical optimization tasks (e.g., a matrix multiplication benchmark)

He also describes an experiment in which the system:

• plans optimizations step-by-step
• uses a verifier (possibly another model such as Haiku as a cheaper judge)
• achieves large speedups (from seconds to milliseconds), including:
  • GPU usage
  • compiler/library optimizations

7) Critique of model comparisons and token-cost fairness

The narrator discusses a referenced “brand” article arguing that comparisons should consider cost, not only benchmark scores.

He further claims:

• some looping approaches can generate many tokens internally (“burn tokens”) to keep searching for a better answer
• this may be acceptable for large labs (less token-sensitive) but could be problematic for:
  • open-source users
  • cost-limited users

He concludes that benchmarking without accounting for cost can be misleading.

Overall takeaway

The video’s central message is that “loop engineering”—goal → iterations → verification → completion—is a major driver of AI progress in coding and automation.

At the same time, the narrator emphasizes the double-edged nature:

• productivity gains may accelerate dangerous feedback loops
• safety measures and governance proposals remain contested, especially regarding:
  • access restrictions vs. open development
  • competitor constraints
  • cost/benchmark transparency

Presenters or contributors

• The video narrator/host (name not provided in the subtitles)
• Entropic
• OpenAI
• SpaceX (mentioned in the context of IPO/hype)
• Trump (mentioned in relation to an executive-order discussion; not otherwise credited)
• Haiku (mentioned as a “cheaper model” used as a verifier)