Get your free personalized podcast brief
We scan new podcasts and send you the top 5 insights daily.
While previously underwhelming, the latest generation of AI models are now surprisingly effective at highly specialized, low-level coding tasks such as writing GPU shaders. This shows that the "bitter lesson"—that general models scaling beats specialized approaches—applies even in embedded and systems programming.
Related Insights
The AI industry is hitting data limits for training massive, general-purpose models. The next wave of progress will likely come from creating highly specialized models for specific domains, similar to DeepMind's AlphaFold, which can achieve superhuman performance on narrow tasks.
Even a specialized task like coding involves a wide range of human-like interaction: brainstorming, searching, and more. This "AGI-completeness" means a powerful general model with a good "bedside manner" can outperform a narrowly specialized one, complicating the strategy for vertical AI apps.
Specialized coding models often fail because a developer's workflow isn't just writing code; it's a complex conversation involving brainstorming, compliance, and web research. The best coding assistants are the most generalist models because every complex task has AGI-like qualities.
An AI was tasked with creating a C++ audio/video equalizer for byte-by-byte streaming, a problem described as something that "audio DSP engineers often get wrong." The AI's success demonstrates its ability to generate correct, readable code for highly specialized and difficult technical challenges that are prone to human error.
Specialized models like Cursor's Composer 2 can achieve short-term dominance over general frontier models by hyper-focusing on a specific domain like coding. This 'hill climbing' strategy allows them to beat larger models on cost-performance, even if general models are predicted to win long-term.
Current AI models resemble a student who grinds 10,000 hours on a narrow task. They achieve superhuman performance on benchmarks but lack the broad, adaptable intelligence of someone with less specific training but better general reasoning. This explains the gap between eval scores and real-world utility.
The "bitter lesson" in AI research posits that methods leveraging massive computation scale better and ultimately win out over approaches that rely on human-designed domain knowledge or clever shortcuts, favoring scale over ingenuity.
AI coding assistants struggle with deep kernel work (CUDA, PTX) because there's little public code to learn from. Furthermore, debugging AI-generated parallel code is extremely difficult because the developer lacks the original mental model, making it less efficient than writing it themselves.
Breakthroughs will emerge from 'systems' of AI—chaining together multiple specialized models to perform complex tasks. GPT-4 is rumored to be a 'mixture of experts,' and companies like Wonder Dynamics combine different models for tasks like character rigging and lighting to achieve superior results.
Just as neural networks replaced hand-crafted features, large generalist models are replacing narrow, task-specific ones. Jeff Dean notes the era of unified models is "really upon us." A single, large model that can generalize across domains like math and language is proving more powerful than bespoke solutions for each, a modern take on the "bitter lesson."
