Artificial General Intelligence—AI that can perform any intellectual task that a human can—has been called both humanity's last invention and our greatest existential risk. In San Francisco, several organizations are racing to build it, believing they'll get there within years, not decades. This chapter explores the front-runners in that race and what's at stake.
OpenAI: From Non-Profit to Powerhouse
OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman, and others, with a billion dollars in committed funding and an audacious mission: ensure that artificial general intelligence benefits all of humanity.
The founding charter committed OpenAI to several principles: publish most of its AI research, avoid competing with any late-stage AGI project that's close to success, and use any influence it obtains over AGI deployment to ensure it's used for the benefit of all.
Initially structured as a non-profit, OpenAI attracted top researchers but struggled to compete with the compensation packages offered by Google, Facebook, and other tech giants. In 2019, OpenAI restructured, creating a "capped profit" arm that could raise investment capital while theoretically maintaining the original mission.
Microsoft invested $1 billion in this new structure, gaining access to OpenAI's technology and providing the massive computing resources needed to train ever-larger models. This partnership would prove crucial.
GPT-3 and the Power of Scale
In 2020, OpenAI released GPT-3, a language model with 175 billion parameters—10x larger than any previous model. The results were startling. GPT-3 could write coherent essays, answer questions, write code, translate languages, and perform dozens of other tasks with minimal training—just examples in the prompt.
This "few-shot learning" capability suggested that scale might be the key to intelligence. Make the model bigger, train it on more data, and capabilities emerge that weren't explicitly programmed.
OpenAI didn't fully open-source GPT-3, citing safety concerns—a controversial decision given the organization's name and original commitments. Instead, they released it as a commercial API. Thousands of startups built applications on top of it.
ChatGPT and the Consumer Breakthrough
We've discussed ChatGPT's November 2022 launch and explosive growth. But what made it possible? The key innovation was Reinforcement Learning from Human Feedback (RLHF)—training the model not just on text from the internet, but on human preferences for responses.
OpenAI hired contractors to rate thousands of responses, teaching the model which outputs humans found helpful, harmless, and honest. This seemingly simple idea transformed the model from an impressive but unpredictable research tool into a product that regular people could use reliably.
The impact was immediate. Microsoft rushed to integrate ChatGPT into Bing, its search engine, hoping to finally challenge Google's dominance. Businesses scrambled to figure out how to integrate AI into their products. Students used it to write essays. Programmers used it to debug code. Writers used it for brainstorming.
GPT-4: Crossing the Threshold
In March 2023, OpenAI released GPT-4, and with it, evidence that AI capabilities were advancing faster than most people had anticipated. GPT-4 could:
- Pass the bar exam at the 90th percentile
- Score 5's on multiple AP exams
- Engage in sophisticated reasoning and problem-solving
- Process both text and images (multimodal capability)
- Write code at near-professional levels
The gap between GPT-3.5 (powering the original ChatGPT) and GPT-4 was larger than the gap between GPT-2 and GPT-3. If this rate of improvement continued, where would we be in two years? Five years?
By early 2023, OpenAI's valuation had soared to $29 billion. Microsoft invested another $10 billion. The organization that started as an idealistic non-profit had become one of the most valuable AI companies in the world.
Google: The Incumbent's Challenge
Google invented the Transformer architecture. Google had been integrating AI into its products for years. Google had some of the world's best AI researchers. Yet ChatGPT caught them off guard.
The problem wasn't capability—Google had models that matched or exceeded GPT-3.5. The problem was risk tolerance. Google had too much to lose. With billions in advertising revenue dependent on trust in their search results, releasing an AI chatbot that might hallucinate fake information or say something offensive was dangerous.
OpenAI, with no search business to protect and a mission to push the frontier, took the risk. Google was left scrambling to respond.
In February 2023, Google announced Bard, its ChatGPT competitor. The launch was rushed and showed it—in the promotional material, Bard gave a factually incorrect answer about the James Webb Space Telescope. Google's stock dropped 8%, erasing $100 billion in market value.
This forced Google to reorganize. DeepMind and Google Brain—previously separate AI research organizations—were merged into Google DeepMind, with Demis Hassabis (DeepMind's co-founder) at the helm. Google committed to accelerate AI deployment across all products.
DeepMind and the Research Frontier
While OpenAI captured headlines with ChatGPT, DeepMind continued pushing the research frontier. In 2020, they solved one of biology's grand challenges: protein folding. AlphaFold could predict how proteins fold based on their amino acid sequences with remarkable accuracy, a problem that had stymied scientists for fifty years.
This wasn't just an impressive demo—it was transformative for biology and medicine. Understanding protein structures is crucial for drug discovery, understanding disease, and countless other applications. DeepMind released the structures of nearly all known proteins, freely available to researchers worldwide.
In 2023, DeepMind released Gemini, Google's answer to GPT-4, with multimodal capabilities and performance that matched or exceeded OpenAI's model in several benchmarks. The race was truly on.
Anthropic: The Safety-First Approach
In 2021, several key OpenAI researchers—including Dario Amodei and Daniela Amodei—left to found Anthropic. Their concern: as AI systems became more powerful, safety needed to be the primary focus, not an afterthought.
Anthropic raised over $7 billion from investors including Google, Salesforce, and others, becoming one of the most well-funded AI startups in history. Their Claude model, released in 2023, emphasized "constitutional AI"—training models to be helpful, harmless, and honest through carefully designed principles.
Claude couldn't match GPT-4 in all capabilities, but it showed less tendency toward harmful outputs and better calibration about its own uncertainty—knowing what it didn't know. For enterprise customers concerned about AI safety, this was compelling.
The Microsoft Wild Card
Microsoft's $10 billion investment in OpenAI was one of the most consequential moves in recent tech history. It gave Microsoft access to cutting-edge AI technology and allowed them to integrate it across their product suite: Windows, Office, Azure, Bing, and more.
The integration happened with remarkable speed. GitHub Copilot, powered by OpenAI's Codex, became an essential tool for millions of developers. Office applications gained AI assistants. Bing, long an also-ran to Google, became interesting for the first time in years.
CEO Satya Nadella framed Microsoft's AI strategy simply: "We're going to make Google dance." By forcing Google to rush AI deployment, Microsoft had accomplished what seemed impossible—putting pressure on the dominant search company.
The Open Source Wildcard
While the giants battled, an open-source movement emerged. Meta released LLaMA, a family of language models that, while not matching GPT-4, performed well enough for many applications. Within days, LLaMA had leaked and spread across the internet.
Open-source models like Stable Diffusion (for images) and various LLaMA derivatives demonstrated that cutting-edge AI capabilities wouldn't remain locked in corporate labs. Anyone with sufficient compute resources could train or fine-tune powerful models.
This democratization worried some (harder to control potentially dangerous applications) and excited others (prevented any single company from monopolizing transformative technology).
"The race to AGI isn't just about who gets there first. It's about who gets there safely, who controls the technology, and whether the benefits will be distributed broadly or concentrated in a few hands."
The Stakes
Why does it matter which organization develops AGI first? Several reasons:
- Economic advantage: AGI would be the most valuable technology in history, potentially worth trillions of dollars.
- Strategic power: Control over AGI could determine geopolitical power balances.
- Safety: Different organizations have different approaches to AI safety; who gets there first affects what safety measures are implemented.
- Values: AGI will likely reflect the values of its creators; which values get embedded matters enormously.
The fact that this race is centered in San Francisco—that the organizations building toward AGI are primarily Bay Area companies—means that a small geographic area is making decisions that will affect all of humanity.
Whether that's inspiring or terrifying probably depends on how much you trust Silicon Valley.