The original Chinchilla scaling laws suggested that model performance was a function of three variables: parameter count ($N$), training tokens ($D$), and compute budget ($C$). The relationship was roughly defined as $C \approx 6ND$. However, as we approach the AGI threshold in 2026, the industry has pivoted toward Inference-Time Scaling.
Instead of merely building "wider" models, current state-of-the-art architectures (like the Rubin-class transformers) utilize test-time compute. By allowing a model to generate multiple internal chain-of-thought paths and verify them before outputting a final response, we are seeing reasoning jumps that traditionally required a 10x increase in raw training compute. This "System 2" thinking transition is the primary reason our countdown has accelerated.
A critical pillar of our tracking is the FLOPs-per-Watt metric. The transition from H100 to Blackwell and now into the NVIDIA Rubin era has seen a massive leap in FP4 and FP8 tensor performance. This hardware velocity is critical because it lowers the "entry barrier" for Level 3 and Level 4 AGI agents.
Our local analysis—benchmarked on high-bandwidth 32GB VRAM nodes—demonstrates that 4-bit and 6-bit quantization (using GGUF and EXL2 formats) no longer carries the massive "perplexity tax" it did in 2024. This democratization of compute means that frontier-level intelligence is no longer restricted to centralized "mega-clusters," but can be deployed in decentralized, high-speed local environments, accelerating the feedback loop of self-improving code.
We categorize the path to ASI using the industry-standard five-level framework:
Our current tracking shows a surge in Agentic Workflows. By utilizing hierarchical agent structures (where a "Manager" model oversees specialized "Worker" models), we are seeing the first instances of autonomous scientific R&D. When a model can formulate a hypothesis, write the code to test it, and interpret the results without human intervention, the "Event Horizon" becomes a mathematical certainty.
The "Data Wall" was predicted to hit in late 2025, but it has been circumvented through Self-Play and Synthetic Data generation. By using high-quality models to "curate" and "distill" massive amounts of raw internet data into high-signal reasoning sets, the efficiency of training has improved by a factor of 4x. This means $10^{25}$ FLOPs today buys significantly more "intelligence" than it did 18 months ago.
Our calculation engine utilizes a weighted ensemble forecast. We aggregate three primary data streams: (1) The Metaculus community consensus, (2) Proprietary hardware deployment schedules for H200/B200/Rubin nodes, and (3) The current velocity of algorithmic efficiency gains (measured in bits per parameter). As breakthroughs like Sparse Autoencoders or Test-Time Compute are verified, the clock adjusts in real-time.
AGI (Artificial General Intelligence) refers to a system that can perform any intellectual task a human can do. ASI (Artificial Superintelligence) is the stage immediately following, where the system’s recursive self-improvement outpaces human understanding, leading to an intelligence explosion. Our countdown tracks the transition to Level 5: Organizations, where AI can operate entire R&D cycles without human intervention.
While high-quality human-generated text is finite, Inference-Optimal Scaling and Self-Play (AlphaGo-style reinforcement learning for reasoning) have largely mitigated the data wall. By generating high-logic synthetic data, models can now train on "perfect" reasoning paths rather than the "noisy" data found on the public internet, accelerating the path to ASI.
The decentralization of AI is a key metric for "Intelligence Density." When frontier-level reasoning models can be run on consumer-grade 32GB VRAM hardware, the feedback loop for innovation becomes global. We track local-host capabilities as a leading indicator for how quickly AGI agents can be deployed in the real world.