Solana Co-founder Anatoly Yakovenko: AGI is Still Far Away
Anatoly Yakovenko stated that AGI is still far away. He cited his own testing project, Percolator, saying that if he were to code full-time, the project could have already been delivered. AI excels in design and specification research, as well as in finding bugs through fuzz testing, until the model is later tightened.
He outlined a two-step process: first, use AI for specification research, with specifications still written by humans, then let AI find specification loopholes; code is also written by humans, and then AI finds code vulnerabilities. He believes this approach will yield better quality, but the number of good products actually delivered to the market may not be higher than before the proliferation of large models.
This does not contradict his earlier characterization of AI as an expert multiplier. In 2025, he mentioned at TechCrunch Disrupt that he would keep an eye on Claude's work, being able to sense when it goes off track, often due to distractions from model outputs. At this moment, he defined the roles: humans are responsible for writing specifications and code, while models are responsible for retrieval and finding issues.
In the same week, he also responded to claims of slowing frontiers with "profit on a trillion-dollar market cap" and mocked efforts to reduce token consumption in Codex. While denying that general intelligence has been achieved, he opposed framing the slowdown as a consensus on safety. Percolator is used as a counterexample: with model assistance, delivery is still slower than that of experienced individuals working alone.
The incremental information pulls the AGI timeline back from slogans to concrete warehouses. He does not deny that models can find loopholes but refuses to elevate loophole detection to the level of independent shipping intelligence.
In market mechanisms, this is a narrative hedge. The buying side continues to bet on Solana's high throughput, on-chain applications, and AI agent settlements; the selling side writes "AGI is close" as if all agreements will be automatically rewritten. The flow of funds still supports infrastructure that can be manually delivered, rather than a bubble of purely agent-generated applications. Beneficiaries are public chain teams that still treat core protocols as craftsmanship; those under pressure are the ones equating programming agents directly with product company valuation stories.
Source: Public Information
ABAB AI Insight
Yakovenko's judgment comes from the same engineering path: originating from Qualcomm and operating system patents, then using Proof of History to create an on-chain clock. His measure of intelligence is not a benchmark score but whether a project named Percolator can be delivered on time. The model can help him search for specifications and perform fuzz testing, but it cannot carry the system across the finish line, hence AGI is still deemed far away.
The capital path aligns with his public investments. Solana seeks protocols that iterate continuously, not one-off generated applications. He has invested seed funding in robots, trading infrastructure, and development tools, while positioning AI as an expert multiplier. The motivation is clear: the cost of errors in on-chain systems is capital, not demo videos; therefore, specifications and code must remain in human hands, with models only allowed to find issues.
In a similar position, this aligns more with Karpathy's view that "AGI is still ten years away" and that "the idea of agents is fundamentally flawed," rather than the narrative of laboratories compressing timelines to two or three years. The industry stage is that programming agents have moved from "able to write functions" to "still unable to take responsibility for experimental design." Whoever can weld models and human efforts into a quality control assembly line will be the first to achieve reliable delivery.
Structural changes belong to the semi-finished period of technological substitution. Loophole detection and retrieval have been replaced first, while architecture and shipping rights have not yet been. The mechanism is that models drive retrieval costs close to zero, thus exposing specification and testing loopholes faster; however, faster exposure does not equate to faster market entry, as the real bottleneck returns to the engineer hours that can sign off.
ABAB News · Cognitive Laws
Finding bugs does not equal being able to ship
Specifications are still handwritten, so intelligence has not yet been handed over
The timeline is based on warehouses, not demos