Escape Velocity: Why we don't need AGI

What happens when the trajectory of improvement increases so fast we don't care about AGI?

One of the presumed benefits of AGI is that it will lead to a super-human acceleration in intelligence, which will then unlock discoveries and advances across, well, everything. The basic argument is that if an AGI is just as smart as humans, but can “think” much faster, it will be able to find solutions to problems at previously unbelievable speeds. Trying to figure out how to make fusion work? A bunch of PhD brains can only think of new ideas and reason through them so fast, often in months or years. With an AGI, the speed limit is (theoretically) how much computing power we give it. Unlike a human, the AGI can work 24/7 and, again, we assume, come up with new ideas and test them out 1,000x faster. Suddenly, a lot of challenges we are facing at humanity-scale seem tractable because we have “zero cost” intelligence.

With AGI, we are building a universal hammer that can be great at everything. However, I have yet to come across use cases where someone actually wants AGI. Instead, they usually need a more specialized AI that has performance good enough that it feels smarter than the smartest-person-in-the-room.

Trying out combinations of proteins for new cancer therapies? Just task a few data centers. Test out many, many ideas of how to reduce carbon emissions during concrete manufacturing? Done. Design ultra-efficient antennas for global internet? Kick off the task on Friday, come back on Monday is the promise.

There’s just one problem- it’s not clear when we will get both AGI itself and AGI that is so cheap its inputs (novel data for a task, energy for computation, chips, etc) are a rounding error. However, it turns out we don’t need AGIs, or AGI that is universally cheap. We are on the verge of achieving a new type of AI improvement that I call Self-sustaining Escape Velocity (SEV). Once you have achieved escape velocity, having an AGI becomes irrelevant.

It will be easiest to understand SEV if we first talk about a few other ideas to help us frame our thinking.

The first is a classic lesson for startups: always hire someone who is smarter than the last person you hired. By following this rule, as your company grows it actually becomes more capable. It’s often assumed that this is very difficult because of the Peter principle, or basically “how can you actually know if someone is smarter than you?” However, this assumption is based on knowing how someone is smarter than you rather than merely establishing someone is smarter. The second scenario, establishing intelligence, is much easier. At its most basic level, give someone a challenge you failed to solve, and if they solve it, they’re smarter! The key lesson here is that we should, for the purposes of SEV, focus on evaluations rather than understanding AI performance.

Our second mental framework is to think about improving foundational models (and their scaling laws) as suffering from Tsiolkovsky’s rocket equation, also known as the “the tyranny of the rocket equation.” The rocket equation says that trying to launch larger and larger rockets becomes less and less efficient. This is due to a larger rocket needing even more fuel, which causes the rocket to weigh more, which in turn means you need more fuel to launch your (now heavier) rocket. Once you reach escape velocity however, the balance has tipped in favor of your rocket, and it is no longer at risk of crashing back down.

Currently, foundational model providers are struggling with a similar problem of more capable models requiring even more data. As they’ve begun to rely on synthetic data, generated by other AI models, it also becomes harder to build the larger model, because an even bigger synthetic data model is needed to generate more sophisticated data, which in turn requires… You can see where this is going.

When people talk about the benefits of compounding intelligence and breakthroughs made by AGI, they are primarily referring to the concept that an AGI has reached an intellectual escape velocity where all of the reasoning done by the model improves its answer or solution.

So foundational models are collapsing under their own weight and we don’t know how to know if they’re improving. What’s an AI company to do?

I think we should pursue a new strategy, Self-sustaining Escape Velocity (SEV). The promise of SEV is this: just keep dumping in some basic resource (compute, memory) and arrange your AI in a feedback loop to generate results that build on top of themselves. Once you have an AI in a setup where it can produce a better AI, your only constraint is how fast you can fuel the rocket engine.

The core of SEV is a hands-off feedback loop. Each time a new AI model is created, it is evaluated using a more sophisticated benchmark that is the result of the previous AI model, the baseline, pruning down the problem space into problems it cannot solve. The new model is a candidate to replace the old model if it makes meaningful progress on solving those problems. If the candidate proves it is indeed smarter than an old model, it becomes the baseline model. This new & improved baseline model is then used to challenge our synthetic data generation model in a critic/adversarial fashion to produce a higher quality model for synthetic data generation.

Now our baseline model and our synthetic generation model have both been leveled-up, so we can repeat the process, without human intervention!

If our process is truly self-sustaining then the only external input it needs is more compute power (and time, and memory) to improve itself.¹ And if our rate of self-improvement is fast enough, then our model improvement process will reach a point of escape velocity where improvements are not just linear or additive, but exponentially compounding. Compare this to our current scaling laws where we see foundational models have crossed over the tipping point and are achieving sub-linear gains in performance for their resource inputs- they’re going to teeter and effectively “fall back” to earth.

Imagine a chart that demonstrates how improvements in performance look for a model improvement process using SEV vs our standard “train, evaluate, iterate” approach now.

The startup (or tech giant) that cracks the code for exactly how to power that self-sustaining feedback loop will experience, literally, runaway success that is only limited by their resources.² I think this self-sustaining loop needs three fundamental pieces:

1. A solid RL policy and environment to steer the AI in its evaluations

2. Generation of synthetic data that focuses on quality rather than quantity

3. A highly optimized feedback loop to overcome drag in the system that will prevent achieving escape velocity.

Why does SEV mean we don’t have to care about AGI? With AGI, we are building a universal hammer that can be great at everything. However, I have yet to come across many (if any) use cases where someone actually wants AGI. Instead, they usually need a more specialized AI that has performance good enough that it feels smarter than the smartest-person-in-the-room. Pursuing AGI is one way, via boiling the ocean, to get to this. SEV, on the other hand, is a more targeted approach that focuses on setting up a system that can self-improve an AI in a limited domain.

This domain must be conducive to transitive improvements, meaning we can assume improvements to our AI can stack on top of each other. An example of a domain with good transitive properties is summarizing legal contracts. A domain like “contemporary performance art” is not. In my experience though, most problems that businesses care about solving are in transitive domains. Existing neural net models lend themselves to performing well in transitive domains, and the recent success of Test-Time Compute for reasoning models is another win in favor of transitive domains.

As an AI CEO/CTO looking for predictability in all this chaos, SEV is a very attractive approach. It is notoriously difficult to establish a stable trajectory in your AI improvements, which in turn means it is near impossible to predict where you’ll be in a few months, let alone a year, or the time horizon for your next funding round. With SEV, we’ve wrangled this chaos into a more predictable trajectory that is based more on resources than engineering sweat and AI researcher talent.

The good news is that many pieces of SEV already exist. We’re seeing massive leaps in making RL stable and easy to use. Likewise, with synthetic data generation we have reached large enough models to overcome earlier shortcomings. Although it hasn’t been widely appreciated, DeepSeek’s AI optimizations are the start of an avalanche of infra improvements we will see over the next several years.

So there you have it, SEV gives us a shortcut to get what we want out of AGI, without having to build AGI itself. This post lays out the strategy for SEV, but there are still many open questions in the tactics to implement SEV, and I would not be surprised to see many variants emerge that leverage hacks in specific areas. As an early reader of a draft put it, reaching SEV may reduce to a challenge of who can find the most impactful problem and solution space where the AI’s quality is (relatively cheaply) measurable.

Thanks for reading eDave.ai! Subscribe to be the first to hear about these upcoming posts:

• Why 10k is the big number for AI in 2025

• AI’s Moneymaker: WeChat meets fast fashion

• Agents vs Automons

• And more to come…

---

Thanks to Makoto Uchida for the original discussion that inspired this and feedback, as well as Alejandra Vergara for reading drafts and providing feedback.

1. Is Self-Sustaining the right way to think about this? What about iterative improvements an AI makes on itself? Self-sustaining is a form of iteration, but more tightly scoped to improvements that come from within the system itself.

2. Isn’t this just how typical training works w/ backprop and gradient descent? No, gradient descent and backprop are focused on the internal training of a model against its own training data. This keeps the evaluation data (problems to be solved) entirely separate until the outer evaluation phase, and uses more of an adversarial “challenge and strive to improve” approach that wraps around pure model training. SEV assumes many model candidates will be generated for each round in order to get to one that breaks out to a higher level of capability.