The Long Leash and the Narrow Fences: Eric Ries and Jeremy Howard on Building AI’s Edison Lab

Eric Ries responds to the AI industry’s pattern of raising large rounds and spending heavily on models and compute before anything touches market by pointing to the seduction of the demo. AI demos are “unbelievably good” and “magical,” he says, so founders easily convince themselves they are the exception and do not need to test with customers. But the same principle still holds: you do not know in advance what customers will want, and you have to discover it through experiments and revealed actions rather than asking electrons what they want.

Ries argues that many AI companies have copy-pasted the SaaS stack into AI, crossing out labels and assuming everything works the same. That pushes the product-market-fit question to another layer: some API companies see tremendous usage and assume their customers will find product-market fit with end users, even though the value chain can be two, three, or four layers deep. If anything about this stack is assembled differently, Ries expects “a lot of carnage” in applications that never understand the end customer.

Because AI has real infrastructure and operating costs, Ries says he draws more on lessons from physical manufacturing, deep-sea oil drilling, and nuclear power plants than from pure software. On moats, he invokes the parable of picking up dimes in front of a steamroller: OpenAI could in theory “nuke you from orbit,” but it can only focus on so many things at once. The strategic response is to move fast, grab use cases, and build in a way that permits rapid iteration and pivoting if assumptions prove wrong.

Jeremy Howard says fast.ai and Answer.AI share the same goal: maximizing the public benefit of AI for as many people as possible. Answer.AI is a public benefit corporation, and Howard sees language as such a powerful modality that it deserves particular focus now that computers have begun to unlock it. When he and Rachel started fast.ai, the idea of a broad online deep-learning course for non-PhDs, or of building applications in a few lines of code, was considered ludicrous; now, with venture and angel money and a team, he feels it is time to invest more deeply in that mission.

Howard traces Answer.AI’s origin to years of conversations with Ries and to a limitation in fast.ai: everything it created—research, courses, software—required a strong coding background, restricting it to less than 1% of the world’s population. Natural language and other natural modalities seemed like a way to make AI far more accessible. Howard also worried that AI would accelerate centralization of power and reduce opportunities for most people, which made fast.ai’s mission feel like it was failing; Ries rejects the idea that fast.ai was a failure and credits the goodwill Jeremy and Rachel built over years of helping people enter the field.

Funding an R&D lab was not a normal startup exercise. Answer.AI is a for-profit R&D lab with no proprietary technology, no obvious product-market fit, no defined market, and no five-year financial plan, so Ries says the usual investor questions did not apply. He argues that industrial R&D has drifted into hyper-specialization, separating research from development, while the best research often happens when the researcher is coupled to the application. He recalls a data-center project where researchers optimized energy efficiency, only to learn from the data-center builder that the builder did not operate the facility and first cared about footprint; making the thing energy-efficient required a 300% size increase, and once that fact reached the lab, a junior PhD saw a different scientific problem in about five seconds.

Answer.AI’s funding came through a mix of luck and network: after a podcast appearance, Howard asked interviewer Alessio for advice, and Alessio said he could give them a term sheet tomorrow. When they brought in angels, Howard says every AI insider he invited except one said yes, while some traditional investors still asked about proprietary technology, market, and position in the SaaS stack. Ries describes the operating model as “a long leash with narrow fences”: leaders set the research thesis, but researchers choose what to pursue.

Ries warns that abstraction layers are useful only when you understand how the layers interact, and that prematurely deciding where the boundaries sit in a nascent field leaves opportunities out of the consideration set. Full-stack engineers are prized precisely because they break abstraction boundaries when needed. In AI, he says, the research and development sides should inform each other rather than being separated too early.

Howard applies that to agents: there are places where agents line up well with language-model capabilities, but most places they do not. A deep understanding of the “linearized subgraph matching” operations underlying LLMs reveals that agents requiring genuinely novel planning—not Nobel Prize-level novelty, just a sequence not almost exactly present in training data—will struggle. Agents that mix and match patterns seen in training data will be amazingly good: labeling emails, adding CSS to HTML, or iteratively debugging common Python problems. Agents that try to independently create a research breakthrough are a different matter.

Howard adds a researcher’s example: much of his code creates new algorithms, often only four or five lines long, and language models have never helped him create those algorithms because by definition they are not in the training data. If the algorithm were already widely known, he says, he probably would not need to ask the model for help. That low-level research experience shapes what development Answer.AI tries to do.

Answer.AI’s efficient Llama 3 fine-tuning work came from Kerem, a former master’s student living in Turkey, who decided to start the project; the company has no managers, and colleagues joined in to help where they could. Howard says this fits the lab’s Edison-style model and the historian Eric Gilliam’s description of Answer.AI as “a long leash with narrow fences.” The narrow fences came from a thesis: there is overinvestment in training foundation models from scratch and in gold-plated hardware, and underinvestment in the resource-constrained real world.

When Howard visited a big Bay Area lab and explained the efficiency work, someone asked what it was for and summarized it as bringing cost down. Howard’s answer is that this is the definition of accessibility: a 10X cost reduction means many more people can use the technology. Ries adds that cost can be a difference in degree that becomes a difference in kind, especially because software now faces physical supply-chain constraints such as power access and how much HBM memory can be manufactured worldwide. Many things people want from generative AI are impossible today simply because of cost.

One consequence of cheap fine-tuning could be continuous personalization. Today’s models are “amnesiac,” Howard says, like the protagonist of Memento writing scraps in a context window to remember who they are talking to. If fine-tuning became cheap enough to continuously train an individual agent on an inexpensive virtual machine, builders could dedicate resources to each customer or use case and unlock hyper-personalization that context windows cannot sustain.

Ries compares this to Edison’s focus on manufacturability: many applications of electricity did not require genius to imagine, but making them practical required hundreds of thousands of painstaking experiments. Edison’s lab produced both “light bulb” inventions—obvious applications enabled by a new capability—and “phonograph” inventions that required a bigger conceptual step. Answer.AI gives every team member a $500-a-month credit card for AI tools, and Howard says the purchases are mostly familiar ones: coding helpers, Copilot, Perplexity, and products built on models such as Cohere’s Command R Plus.

Howard identifies law and education as broad areas where computing with language is an obvious fit, calling them heavily language-based domains. One lawyer they work with describes a law firm as “just a very large language model”: you feed in text and get out text. Both fields also offer large opportunities to help society.

In law, Howard argues that wealthy people and organizations use the law as a weapon to stop less wealthy people from doing things they have every right to do, and society mostly accepts that injustice. Bringing down the cost of high-quality legal advice would mean the law is used as a weapon less often. In countries like the US, law is also used for gatekeeping in regulated markets, and Howard says he wants to knock down barriers that stop people from bringing their ideas into the world.

In education, Howard speaks as a homeschooling father who has spent hundreds or thousands of hours researching the academic literature, listening to teacher interviews, and teaching his own child. Teachers are constrained by a system in which many children must go through the same path at the same time in the same way, with limited customization; he has seen how delighted his daughter is when that constraint is removed. Making law and education more accessible would help more people build what they want to build and become who they want to be.

Howard says California’s proposed SB 1047 was created with good intentions and has some good features, but he focuses on the part that tries to ensure the safety of AI models. About a year ago, he studied that question when Governance AI and groups such as OpenAI were developing a framework for such laws, and he red-teamed a draft. He interviewed more than 70 experts across relevant fields and concluded, counterintuitively, that such a policy would likely be ineffective and could produce the opposite result: a less safe situation.

The core problem is that a model is a purely dual-use technology, like a pen, paper, calculator, or the internet. A developer can test a model extensively and claim it never produces bioweapon recipes, but a user can fine-tune or prompt it to do whatever they like; the developer cannot ensure otherwise, just as a calculator seller cannot ensure the buyer will not use it for nuclear calculations. If regulation requires ensuring safety, the practical result is that raw models are never released—only layers on top, like ChatGPT versus downloadable Llama 3.

Raw models are more powerful because users can fine-tune them, study their weights, gradients, and activations, and do clever caching. If safety regulation makes raw model release impossible, models become extremely rivalrous goods, jealously guarded secrets available only to big states and big companies. Howard argues that this centralizes power, reduces transparency, and turns a dual-use technology that could improve cybersecurity or vaccines into something a few super-rich organizations race to control.

Ries says he does not object to foundation-model labs whose mission is to discover whether AGI is possible; commercialization there funds research. His concern is that fundraising gravity has pushed many entrepreneurs away from utility and into science fiction, leaving a reservoir of applications that do not require AGI unbuilt. He cites an Anthropic investor and AI-safety advocate who realized that a frontier-only world, where the closest thing to AGI is connected to real-world systems, invites unsafe deployments that could have been handled by smaller, properly fine-tuned, intrinsically safer models.

Ries also warns that some foundation labs have become “schizophrenic,” with a safety-research and AGI mission on one side and a huge commercial apparatus on the other. In some cases the commercial side barely knows or cares about the safety agenda, creating rival tensions that can rip an organization apart. If he ran such a lab, he would reconnect research to the customer, borrow Toyota’s “go and see” practice, and make customer success part of the lab’s responsibility.

In the quick-fire round, Howard calls agents overhyped and resource efficiency underhyped, repeating that most agent use cases are incompatible with the mathematical foundations of models. Asked about startups, he praises Perplexity for its speed and says Grok is fast and taking a foundationally sound different approach. He names two breakthroughs that would change his mind: a breakthrough in energy or resource requirements, and a breakthrough in planning and reasoning beyond subgraph matching, such as energy-based models or diffusion models for text.

Ries says LLMs have changed his perception of human intelligence by showing how much of it is encoded in language and linguistic processing; he jokes that obscure post-modernist literary theories are about to become empirically testable. The breakthrough that would shake him is a realization that the problem is not scaling LLMs but understanding cognition itself. He compares current LLM training to brute-forcing every combination of Minecraft blocks until one behaves like a calculator: it may work, but it may be the most inefficient possible way to discover a critical algorithm.

Jacob Effron and Jordan reflect that getting Jeremy Howard and Eric Ries together was a treat, saying they could have spent two hours with either one. They describe Howard as a polymath who is deep in AI regulation, founded a healthcare AI company, and has worked in education through fast.ai, which makes him well suited to run a research lab with many possible directions.

The hosts find the combination with Ries especially interesting: most research labs are hardcore AI people pursuing AGI, while Ries is focused on customers and learning what they need. They like the pair’s argument that today’s model capabilities already support many more applications if models become cheaper and more accessible, and they see 2024 as a year of applications across verticals. Legal and education stand out because the value proposition is clear and customer pain is easy to investigate.

On healthcare, the hosts agree with Ries that incentives can be warped so that a good product is not even used; the best entrepreneurs in the space understand end incentives and still build mission-driven products that fit them. They note that Answer.AI has three lawyers in-house working on the legal side, and they connect the “law is language in, language out” idea to their own work with data. They also flag Howard’s regulatory argument—that trying to make AI safer can make it less safe—and the irony that open source, with the most eyes on it, may end up safest.