Karpathy’s Pelican
Andrej Karpathy's latest AI experiment, showcasing Opus 5 procedurally rendering a 3D animation from the first paragraph of Lord of the Rings, demonstrates LLMs' astonishing generative capacity for tasks humans wouldn't bother with. However, the resulting 'jank' highlights a current weakness in self-auditing visual output. This sparked intense debate on HN about the practical utility of such AI-generated content, the evolving credibility of AI experts, and the future of personalized media versus shared experiences.
The Lowdown
Andrej Karpathy, a prominent figure in AI, shared a fascinating experiment where he tasked Anthropic's Opus 5 model with an ambitious creative challenge. Instead of a simple text prompt, he fed it the opening paragraph of J.R.R. Tolkien's 'The Lord of the Rings' with a generous 1M token budget and asked it to generate a 3D Three.js rendering of the story. The results were impressive, yet imperfect.
- Opus 5 worked for approximately two hours, producing 5500 lines of Three.js code that procedurally rendered the narrative.
- Karpathy highlighted the LLM's 'stamina and patience,' noting its ability to undertake highly custom, intricate tasks that 'no one in their right mind' would perform due to the sheer effort involved.
- He envisions a future of 'hyper custom worlds' and 'ephemeral GTA of X on demand,' suggesting a new paradigm for personalized interactive experiences.
- A key limitation identified was the LLM's struggle with auditing its own visual work. Opus 5 had to 'slowly and painstakingly take screenshots' to self-correct, leading to noticeable 'jank' in the final output.
This experiment offers a compelling glimpse into the dual nature of advanced LLMs: their extraordinary potential for complex code generation and creative automation, juxtaposed with their current difficulties in understanding and refining visual fidelity without direct, efficient sensory feedback.
The Gossip
Slop or Spectacle: The Utility of AI Art
The community fiercely debated the practical value of AI-generated creative works. Many commenters dismissed Karpathy's demo, and similar 'AI-made games,' as 'useless,' 'boring,' and 'slop,' arguing they lack depth, fun, or genuine creativity, existing only as novelties. They pointed out the 'godawful jank' and poor interpretation of text. Conversely, some recognized the underlying technical achievement and potential, especially for bespoke projects or as a foundational step towards 'hyper-personalized entertainment' for niche audiences, though others worried about the loss of shared cultural experiences.
Karpathy's Credibility Conundrum
A significant thread questioned Andrej Karpathy's perceived shift in stance regarding AI capabilities. Commenters noted that he previously predicted reliable AI agents were 'at least a decade away,' but now, as an Anthropic employee, he appears to be marketing current, albeit janky, AI feats. This led to accusations of corporate influence on his public statements. Others countered that AI has genuinely advanced dramatically in a short period, making earlier predictions obsolete, and that it's unfair to assume his current observations are purely marketing-driven.
Benchmarking Beyond the Bicycle
The discussion extended to the effectiveness of current AI benchmarks. Many expressed fatigue with the 'pelican on a bicycle' SVG test, seeing it as a simplistic and overused measure. While Karpathy's LotR rendering was acknowledged as more complex, some argued that the ability to generate Three.js code might be a specific training artifact rather than a general indicator of intelligence. There was a call for more rigorous tests, particularly those involving tasks AIs are currently poor at, such as driving or understanding novel instructions, to better gauge their true capabilities.
The Nature of AI-Generated Output: Disposable Code and Copyright Quandaries
The experiment prompted reflection on the implications of AI's generative power for software and intellectual property. One commenter introduced the concept of 'throwaway software,' suggesting that LLMs might enable the cheap, rapid, and disposable production of code, similar to how mass manufacturing changed physical goods. Separately, concerns were raised about copyright, with one user implying that LLMs still retain and reproduce verbatim content from their training data (like the first LotR paragraph), despite efforts to prevent such 'verbatim giveaways,' reigniting the debate about 'theft' and data ownership.