Early AI training involved simple preference tasks. Now, training frontier models requires PhDs and top professionals to perform complex, hours-long tasks like building entire websites or explaining nuanced cancer topics. The demand is for deep, specialized expertise, not just generalist labor.

Unlike traditional engineering, breakthroughs in foundational AI research often feel binary. A model can be completely broken until a handful of key insights are discovered, at which point it suddenly works. This "all or nothing" dynamic makes it impossible to predict timelines, as you don't know if a solution is a week or two years away.

The original playbook of simply scaling parameters and data is now obsolete. Top AI labs have pivoted to heavily designed post-training pipelines, retrieval, tool use, and agent training, acknowledging that raw scaling is insufficient to solve real-world problems.

The future of AI is hard to predict because increasing a model's scale often produces 'emergent properties'—new capabilities that were not designed or anticipated. This means even experts are often surprised by what new, larger models can do, making the development path non-linear.

Attempting to interpret every learned circuit in a complex neural network is a futile effort. True understanding comes from describing the system's foundational elements: its architecture, learning rule, loss functions, and the data it was trained on. The emergent complexity is a result of this process.

The ambition to fully reverse-engineer AI models into simple, understandable components is proving unrealistic as their internal workings are messy and complex. Its practical value is less about achieving guarantees and more about coarse-grained analysis, such as identifying when specific high-level capabilities are being used.

AI development is more like farming than engineering. Companies create conditions for models to learn but don't directly code their behaviors. This leads to a lack of deep understanding and results in emergent, unpredictable actions that were never explicitly programmed.

People overestimate AI's 'out-of-the-box' capability. Successful AI products require extensive work on data pipelines, context tuning, and continuous model training based on output. It's not a plug-and-play solution that magically produces correct responses.

Training models like GPT-4 involves two stages. First, "pre-training" consumes the internet to create a powerful but unfocused base model (“raw brain mass”). Second, "post-training" uses expert human feedback (SFT and RLHF) to align this raw intelligence into a useful, harmless assistant like ChatGPT.