The Labs

The Institutional Shift

For most of its history, AI research happened primarily at universities. MIT, Stanford, CMU, Toronto, Montreal, Berkeley — these were the places that mattered. Faculty published papers, trained PhD students, and set the research agenda. Industry labs existed (AT&T Bell Labs, Xerox PARC, IBM Research) but generally operated on long time horizons and published openly.

That changed in the 2010s. As deep learning began producing commercially valuable results — image recognition, machine translation, recommendation systems — talent and compute migrated from universities to industry at an accelerating rate. Today, the most consequential AI research comes from a handful of private labs, and the compute required for frontier work is beyond the reach of any university. A single training run for a model like GPT-4 is estimated to cost over $100 million in compute alone.¹

Understanding these labs — their histories, incentives, and philosophies — is essential context for reading the field. Their decisions about what to build, what to release, and what to keep private shape what AI looks like for everyone.

OpenAI

Founded in 2015 as a nonprofit research lab with the stated mission to "ensure that artificial general intelligence benefits all of humanity." Early backers included Elon Musk, Sam Altman, Peter Thiel, Reid Hoffman, and others, who collectively pledged $1 billion. The original thesis was that concentrating AGI capabilities in a single company would be dangerous, so an open, nonprofit alternative was needed.

The lab's trajectory since then has been a case study in how compute costs shape organizational structure:

• 2015-2018: Nonprofit era. Published influential work on reinforcement learning (OpenAI Gym, PPO algorithm), robotics (dexterous manipulation), and language (early GPT). Most work was open-source.
• 2019: Created "OpenAI LP," a "capped profit" subsidiary, arguing that the nonprofit couldn't raise the capital needed for frontier research. Investors can earn up to 100x their investment, after which excess returns go to the nonprofit. Microsoft invested $1 billion.
• 2020: Released GPT-3 (175 billion parameters) but only via API — no open weights. This marked the shift from "open" to "controlled access."
• 2022: ChatGPT launched in November. Reached 100 million users in two months — the fastest consumer technology adoption in history. This was not a new model; it was GPT-3.5 with RLHF (reinforcement learning from human feedback) fine-tuning and a chat interface. The product innovation mattered more than the model innovation.
• 2023: GPT-4 released. Microsoft invested another $10 billion, becoming a 49% owner of OpenAI LP. The paper provided almost no architectural or training details — a sharp departure from the lab's original ethos.
• 2023 (November): The board fired Sam Altman as CEO over concerns about the pace of commercialization versus safety. Most employees threatened to quit. Altman was reinstated within days, and the board was restructured. This episode revealed the tension between the nonprofit mission and the commercial reality.
• 2024-2025: Continued scaling. o1 and o3 reasoning models. DALL-E for image generation. Sora for video generation. Ongoing restructuring of the nonprofit/for-profit relationship.

OpenAI's significance is less about any single model and more about proving that scaling language models could produce commercially viable products. They didn't invent transformers (that was Google Brain), they didn't invent RLHF (that was DeepMind), and they didn't invent large language models (GPT-1 was one of several contemporary efforts). What they did was push the scaling curve harder than anyone else and build the first product that made LLMs legible to non-technical users.

Anthropic

Founded in 2021 by Dario Amodei (former VP of Research at OpenAI) and Daniela Amodei (former VP of Operations at OpenAI), along with several other ex-OpenAI researchers including Tom Brown (first author on the GPT-3 paper) and Chris Olah (known for mechanistic interpretability work).

The founding thesis was that AI development needed to be safety-focused from the organizational level, not just the research level. Key distinctions from OpenAI:

• Constitutional AI (CAI): Anthropic's training approach. Instead of using human feedback for every response, they provide the model with a set of principles (a "constitution") and have the model critique and revise its own outputs against those principles. The goal is to make safety training more principled and less dependent on individual human annotators' preferences.²
• Interpretability research: Significant investment in understanding what happens inside neural networks. Chris Olah's work on mechanistic interpretability — understanding individual neurons and circuits in trained models — is some of the most cited safety research.
• Responsible Scaling Policy: A public framework that ties capability development to safety evaluations. If a model crosses certain capability thresholds, additional safety measures are required before deployment.
• Public Benefit Corporation (PBC): Anthropic is structured as a PBC, which legally allows the board to weigh public benefit alongside shareholder returns. This is different from both OpenAI's nonprofit/capped-profit hybrid and a traditional corporation.

The Claude model family is Anthropic's product. Claude 3.5 Sonnet (mid-2024) demonstrated that safety-focused training didn't have to come at the cost of capability — it performed competitively with GPT-4 on most benchmarks while generally refusing more categories of harmful requests. This is the tool you build with daily.

Google DeepMind

The result of a 2023 merger between DeepMind (founded 2010 in London, acquired by Google in 2014 for ~$500 million) and Google Brain (Google's internal AI research division). The merged entity is led by Demis Hassabis, DeepMind's co-founder.

DeepMind's reputation was built on reinforcement learning and game-playing AI:

• AlphaGo (2016): Defeated Lee Sedol at Go, a game previously thought to be decades away from AI mastery. This was the moment that convinced many people deep learning could solve problems beyond pattern recognition.
• AlphaFold (2020): Solved the protein structure prediction problem — a grand challenge in biology for 50 years. AlphaFold 2 predicted protein structures with accuracy rivaling experimental methods. This may be the single most impactful AI application to date, affecting drug discovery, biology, and materials science.³
• Gemini (2023-present): Google's multimodal foundation model family, competing directly with GPT-4 and Claude. Gemini benefits from Google's proprietary TPU hardware (more on that in Chapter 18) and from integration across Google's product ecosystem.

Google Brain's contribution was different: fundamental research that the field built on. The transformer architecture came from Google Brain (Vaswani et al., 2017). BERT came from Google AI Language. TensorFlow, the first major deep learning framework, came from Google Brain. Word2Vec, which kicked off the word embedding era, came from Google Research.

The merged entity has the deepest research talent, the most compute (TPU pods that no other lab can match), and the broadest product surface (Search, Gmail, YouTube, Android, Cloud). Their challenge is organizational: translating research breakthroughs into products within a large corporate bureaucracy. OpenAI shipped ChatGPT while Google was still debating internally how to release a chatbot without embarrassment.

Meta AI (FAIR)

Meta's AI lab, originally called FAIR (Facebook AI Research), was founded in 2013 and led by Yann LeCun, one of the three "godfathers of deep learning" (with Geoffrey Hinton and Yoshua Bengio). LeCun's leadership gives FAIR a distinctive character: more emphasis on self-supervised learning and less on the autoregressive language model paradigm that dominates the field.

Meta's defining strategic choice is open weights:

• LLaMA (2023): Initially leaked, then officially released with open weights. LLaMA 2 followed with a permissive license. LLaMA 3 (2024) pushed further, with the 405B parameter model being the largest openly available foundation model.
• PyTorch: Meta developed the PyTorch deep learning framework, which became the standard for research (displacing TensorFlow) and is now the foundation for most ML development. PyTorch is arguably Meta's most important contribution to the field — more impactful than any single model.
• Research publications: FAIR publishes prolifically and openly. Their papers on vision transformers (DINOv2), self-supervised learning, and multilingual models are foundational.

Why open weights? Meta's business model doesn't depend on selling API access to models. It depends on social media advertising. For Meta, making AI models freely available commoditizes the model layer, prevents competitors (OpenAI, Google) from building a moat around proprietary models, and creates an ecosystem of developers who build on Meta's stack. It's strategy, not charity — but the effect is that Meta has done more to democratize frontier AI than any other company.

Mistral

Founded in 2023 in Paris by Arthur Mensch, Guillaume Lample, and Timothee Lacroix — all formerly at Meta AI/FAIR and Google DeepMind. Mistral represents the European entry into frontier model development.

Mistral's approach is notable for efficiency:

• Mistral 7B (2023): Outperformed LLaMA 2 13B on most benchmarks despite being nearly half the size. Demonstrated that architectural innovations (sliding window attention, grouped-query attention) could close the gap between small and large models.
• Mixtral 8x7B (2024): A Mixture of Experts (MoE) architecture that activates only a subset of parameters for each token. 47B total parameters but only ~13B active during inference, making it much cheaper to run than its full parameter count suggests.
• Open weight release: Like Meta, Mistral has released models with open weights, though their licensing has become more restrictive over time as commercial pressure grows.

Mistral raised over $600 million within its first year, reaching a $6 billion valuation. Its existence matters for AI geopolitics: the EU's AI Act creates a regulatory environment where European-developed models may have strategic advantages in European markets.

The Rest of the Field

Beyond the five labs above, several others are worth tracking:

The Open vs. Closed Debate

This is the defining disagreement in AI governance, and it cuts across technical, economic, and ethical lines.

The Case for Closed

Proponents (primarily OpenAI, Anthropic, and some safety researchers) argue:

• Frontier models are dual-use technology. A model that can write helpful code can also write malware. A model that can explain chemistry can help synthesize dangerous substances. Releasing weights makes these capabilities permanently and irrevocably available.
• API-gated access allows the developer to monitor and restrict harmful use. You can filter outputs, rate-limit users, and shut off access. None of this is possible once weights are in the wild.
• Safety alignment is fragile. RLHF and constitutional AI training can be removed through fine-tuning on a relatively small dataset. Open weights + commodity hardware + a few hundred dollars of compute = an unaligned model.⁴

The Case for Open

Proponents (primarily Meta, Mistral, and the broader open-source community) argue:

• Concentration of power is itself a risk. If only three companies control frontier AI, those companies effectively govern what AI can and cannot do for everyone. History suggests this kind of concentration is bad for society, innovation, and competition.
• Transparency requires access. You can't audit a model you can't examine. Safety research requires the ability to probe, test, and stress models — which requires open weights. The closed labs publish safety research, but external researchers can't verify it.
• Innovation happens at the edges. Most applications of LLMs — fine-tuning for specific domains, building agentic systems, combining models with tools — happen outside the labs that built the base model. Open weights enable this far more than API access.
• The cat is already out of the bag. Once one lab releases open weights (Meta), the argument that other labs should stay closed weakens. The capabilities are available regardless.

The honest answer is that both sides are right about the risks they identify and wrong about the risks they minimize. Concentrated power is genuinely dangerous. Unrestricted access to frontier models is also genuinely dangerous. There's no clean solution, and anyone who tells you otherwise is selling something or hasn't thought it through.

How the Labs Make Money

This matters because it determines what gets built and what gets prioritized.

The economics are important: as of early 2025, none of the pure-play AI labs (OpenAI, Anthropic) are profitable. They're subsidized by investor capital, burning cash to build capability and market share in the hope that the market will be enormous. This is the classic Silicon Valley playbook (Amazon lost money for years), but it means the current AI landscape is shaped more by investor confidence than by sustainable business models.

The University Response

Universities haven't disappeared from AI research, but their role has changed. They're less likely to produce frontier models (they can't afford the compute) and more likely to contribute foundational theory, evaluation methodology, safety research, and applications to non-commercial domains.

A few university labs remain highly influential:

• Stanford HAI (Human-Centered AI) — Produces the annual AI Index Report, the most comprehensive survey of the field's state.
• Mila (Montreal) — Yoshua Bengio's lab. Strong on theory, responsible AI, and climate applications.
• Berkeley AI Research (BAIR) — Robotics, RL, systems. Produced several key researchers who went on to start or join labs.
• University of Washington — NLP strength. Produced key work on instruction tuning and evaluation.

The brain drain from universities to industry is real and well-documented. The solution isn't obvious: you can't train a frontier model on a professor's grant budget. Some labs (Google DeepMind, Meta) address this by maintaining research partnerships with universities and allowing dual appointments, but the gravitational pull of industry resources is strong.

The labs build the models, but they don't build the chips. That distinction matters enormously, because the entire AI industry runs on hardware produced by a surprisingly small number of companies. The hardware stack — who builds what, why Nvidia dominates, and what the supply chain looks like — is the subject of the next chapter.