← Bonus Resources / The Papers That Built the Future
Bonus Resource · 01

The Papers That Built the Future

Four landmark AI research papers — what they actually said, why they mattered, and what they got quietly wrong.

Most people have heard of "the transformer paper." Very few have read it. These summaries are not simplifications — they are translations. Each one tells you what the researchers found, why it changed everything, and where the cracks later showed.

Paper 01 · 2017

"Attention Is All You Need"

Vaswani et al., Google Brain — the paper that introduced the Transformer architecture

In 2017, eight researchers at Google published a modest-sounding paper about improving machine translation. It turned out to be the foundation of almost everything that has happened in AI since — ChatGPT, Gemini, Claude, Stable Diffusion, AlphaFold. All of it rests on this 15-page document.

Before this paper, AI systems processed language sequentially — word by word, like reading one character at a time. The core idea of the Transformer was "attention": instead of processing a sentence in order, let the model look at the entire sentence at once and decide which words matter most in relation to each other. When the model reads "The animal didn't cross the street because it was too tired," attention lets it figure out that "it" refers to "animal," not "street" — and do so by comparing every word to every other word simultaneously.

The key insight: You don't need to process language in sequence. You can attend to all of it at once, weigh what matters, and represent meaning through relationships — not positions.

This made training dramatically faster (parallelizable, meaning it runs on many chips at once) and dramatically more scalable. The paper itself was about translation. Within two years, the same architecture was running on hundreds of times more data and producing systems that could write, reason, and code.

What it got wrong (or didn't anticipate): The authors had no idea what they had built. The paper's final line thanks Google Brain's infrastructure team. There is no mention of AGI, consciousness, risk, or societal impact. The architecture they designed for better translation became the engine of a technology that would reshape economies. The gap between what the paper intended and what it produced is one of the cleanest illustrations of why "we just wanted to improve benchmarks" is not a sufficient safety framework.

Architecture NLP Self-Attention Unintended consequences
Paper 02 · 2020

"Scaling Laws for Neural Language Models"

Kaplan et al., OpenAI — the paper that made bigger feel inevitable

If the Transformer paper told the world how to build powerful AI, the Scaling Laws paper told it why to keep making it bigger. Published by OpenAI in 2020, this paper found something startling: AI performance improves predictably as you scale up model size, dataset size, and compute — and it does so across a smooth curve, like a mathematical law.

This was not obvious before. Researchers had assumed there would be diminishing returns, that models would plateau, that bigger wasn't reliably better. The scaling laws paper showed that, over many orders of magnitude, you could simply graph "how big is the model" against "how well does it perform" and get a clean, predictable relationship.

The key insight: Capability is not random. It scales. If you can predict performance from compute, then AI progress becomes, in some sense, an engineering problem — not a research breakthrough problem. You just need more resources.

This paper arguably did more to accelerate the AI race than any technical innovation. If you can predict the curve, investors and labs can fund along it. "We know it will work — we just need more chips" is an easier pitch than "we think this research direction might pan out."

What it got wrong: The paper focused on loss — a technical measure of prediction accuracy — rather than on whether loss translates to useful, safe, or aligned behavior. It also didn't capture "emergent capabilities": abilities that appear suddenly at scale and weren't predicted by the smooth curve. Scaling laws said capability grows smoothly; reality showed it sometimes jumps. That gap between predicted smooth growth and actual discontinuous leaps is one of the harder problems in AI safety today.

Scaling Compute Empirical ML Emergent capabilities
Paper 03 · 2022

"Constitutional AI: Harmlessness from AI Feedback"

Bai et al., Anthropic — the paper that tried to build values into AI

By 2022, it was clear that large language models were powerful — and that they would happily produce harmful, misleading, or manipulative content if asked in the right way. The dominant approach to fixing this was human feedback: hire people to rate outputs as good or bad, and train the model to prefer the good ones. This worked, but it was slow, expensive, and inconsistent. What one rater found acceptable, another found objectionable.

Anthropic's Constitutional AI paper proposed something different: give the model a set of principles — a constitution — and have it critique and revise its own outputs against those principles. Instead of thousands of human ratings, you give the model a ruleset and let it argue with itself. The AI becomes, in effect, its own safety evaluator.

The key insight: You can encode values as explicit, readable principles and use AI feedback — rather than only human feedback — to reinforce them. This makes the alignment process more transparent and more scalable.

The practical output of this research was Claude — Anthropic's AI assistant, which was built using this method. Constitutional AI is why Claude will decline some requests while explaining its reasoning, rather than simply refusing or silently complying.

What remains open: The hard question is whether a constitution is sufficient. Rules can be gamed, interpreted narrowly, or followed in letter while violated in spirit. A model trained to appear harmless is not necessarily harmless — it is trained to appear so. Constitutional AI was a genuine advance in making AI alignment legible and scalable. Whether it solves the deeper problem of building AI that is reliably good — not just trained to seem good — remains one of the field's central open questions.

Alignment RLHF Safety Value specification
Paper 04 · 2023

GPT-4 Technical Report

OpenAI — the paper that revealed almost nothing, and why that matters

The GPT-4 Technical Report is one of the most-read and least-informative papers in AI history — deliberately. Published in March 2023, it describes a system that passed the bar exam, outperformed most humans on standardized tests, and could process both text and images. It contains almost no details about how GPT-4 was built, what data it was trained on, how large it is, or how its safety measures work.

OpenAI justified this secrecy on competitive and safety grounds: publishing full details would help adversaries reproduce dangerous capabilities. Critics, including many AI researchers, argued the opposite — that without transparency, there is no accountability, and "trust us" is not a safety standard.

The key insight the paper refuses to give: We don't know what GPT-4 is. We know what it can do on benchmarks. We do not know what it is doing internally, what its failure modes look like under distribution, or whether its safety properties hold outside of tested conditions.

The report is important not just as a technical document but as a policy document. It marks the moment when frontier AI development became explicitly closed — a reversal from OpenAI's founding mission of open research. It established a precedent that powerful AI systems need not be scientifically transparent, and that capability demonstrations could substitute for mechanistic understanding.

Why it matters beyond the technical: The GPT-4 Technical Report is a signal flare for AI governance. It asks: when a private company builds something genuinely powerful and potentially dangerous, what are its disclosure obligations? Who gets to audit it? What does "responsible deployment" mean without external verification? These questions still don't have good answers. The paper didn't cause this problem — but it made it impossible to ignore.

Multimodal Benchmarks Transparency AI governance