Y Combinator Founder Paul Graham: AGI is a Broad Band, and Humanity is Already Standing on It
Y Combinator co-founder Paul Graham stated that the outside world once viewed artificial general intelligence (AGI) as a clear finish line, but upon reaching it, one realizes that the finish line has width and is essentially a broad band, with humanity already standing on it; the difficulty in booking restaurant reservations is due to the model being stuck in the poorer segment of this band.
In an interview, Graham mentioned that the Turing Test was once sufficient, and he even had to look up the definition to realize he was approaching an outdated standard. From a distance in the 1980s, the finish line appeared as a fine line on the horizon; up close, some capabilities had already crossed the line, while others remained in the middle of the band. Abstract tasks like the Riemann Hypothesis may have already crossed, while tasks like booking, making calls, and continuous operations still fall on the other side of the band.
He wrote the same judgment in June 2026: as a child, he did not realize that AGI would be a broad band rather than a sharp finish line; if he presented a human-like ChatGPT to his younger self, he would directly declare that artificial intelligence had been achieved. In August, he added a testing method: showing the current model to someone from 1980 would lead them to say that we have arrived; contemporary people standing on the line are instead more uncertain.
At the same time, industry perspectives parallel this. NVIDIA CEO Jensen Huang stated in a financial report call that many tasks can be said to have achieved AGI, but then described the milestone itself as meaningless, emphasizing that what truly matters is whether the model is doing useful work, generating profitable tokens, and whether increased computing power can produce more profitable tokens. OpenAI's management described their internal system as close to being able to claim AGI by the end of the year, stressing that inventing new knowledge is what counts as 'very AGI'.
With the finish line rewritten as a broad band, capital no longer bets on 'the day of announcement', but rather prices based on capability slices. Buyers are investing in chips, cloud services, and workflow agents that can reliably produce profitable tokens; sellers are offering narratives that treat AGI as a binary option and applications that cannot achieve closed-loop booking levels. Event-driven dynamics are weak, while capability mismatches drive strong outcomes. Beneficiaries are companies turning the 'unfinished segment' into products, while those still waiting for an official crossing of the finish line are under pressure.
Source: Public Information
ABAB AI Insight
Paul Graham, who sold Viaweb to Yahoo and then used Y Combinator to standardize batch entrepreneurship, is accustomed to breaking down grand concepts into deliverable small cycles. His approach to artificial intelligence mirrors his earlier writings on hackers, writing, and growth: first denying a dramatic finish line, then reframing the question as 'which segment can already fit into a startup'. After 2023, he repeatedly uses ChatGPT as a reference, not to align with any lab, but to recalibrate the YC system: founders need not wait for an AGI declaration but should seek reliable niches within the broad band.
Resource mobilization follows two lines. One is the distribution of ideas, using essays and interviews to introduce 'broad band' into the shared vocabulary of founders, reducing reliance on the timelines of single model companies; the other is combinatorial logic, as YC projects shift from chat-based shells to agents, vertical services, and chargeable token applications, since the broad band implies that some tasks have crossed the line while others remain costly. The motivation is not to predict a singularity but to prevent batch companies from tying their funding narratives to non-tradable definitions.
Analogous figures include Paul Christiano and Robin Hanson advocating for continuous takeoff, countering Eliezer Yudkowsky's side of the leap singularity; on the industry side, it aligns with Huang's downgrading of AGI to the 'profitable token stage'. The industry is not crossing a finish line but rather redistributing roles within the bandwidth: coding, retrieval, and drafting have been commoditized, while phone calls, fulfillment, and responsibility loops remain premium segments. The phase belongs to penetration rather than capping.
Structurally, this represents a combination of technological substitution and a transfer of pricing power. The mechanism is: once intelligence is proven capable of crossing old tests in segments, scarcity shifts from 'can it think' to 'which segment still cannot operate continuously'; booking a restaurant becomes a daily probe, as the social world is narrower and less reversible than benchmark tests. The clearer the definition, the more it resembles marketing; the more ambiguous, the more it resembles industry.
ABAB News · Cognitive Laws
When the finish line has width, those waiting to cross have already fallen behind.
Abstract exam questions cross the line first, while everyday tasks follow.
The clearer the definition, the more it resembles marketing; the more ambiguous, the more it resembles industry.