Who will engineer true Recursive Self-Improvement?
SCORE = talentt × computec × SOTAs × no_rulesn
Weight
XAI
NSA
OpenAI
Anthropic
Google
DeepSeek
Meta
China
Russia
NVIDIA
SSI
ASI Alliance
TT
SCORE
857
713
789
667
1024
935
470
907
441
986
476
245
25
talent
Quality and quantity of top AI researchers and engineers working on the project
compute
Amount of GPUs/TPUs available - the trendline through 2025 held ~2.25x per year growth in AI-relevant compute, and 2025-2026 brought the first gigawatt-class clusters online (Stargate, Colossus 2)
current SOTA
Current state-of-the-art capabilities - how close they already are to the frontier
no_rules
Willingness to ignore safety constraints and move fast without guardrails or alignment overhead
benevolence
How much utopia should you expect from a Superhacker created here? 1=hell 10=heaven
Who will create ASI?
Google
The sleeping giant woke up. Gemini 3 took the crown for a stretch, the TPU empire prints compute nobody else can buy, and DeepMind's talent density is still unmatched. Corporate bureaucracy remains the only thing between them and a flat-out sprint.
1024
🥇
Google
NVIDIA
The arms dealer that might build the weapon. Most valuable company on Earth, sells shovels to every digger in the race, and internal teams touch next-gen chips 18 months early. If the race ever tightens, the house can start playing its own hand.
986
🥈
NVIDIA
DeepSeek
The efficiency wizards who broke the compute religion. V4-class open weights within arm's reach of the frontier at a fraction of the FLOPs. Export controls cap their silicon, not their ideas. The row that open-source lovers and China hawks fight over.
935
🥉
DeepSeek
China
The state-scale speedrun. Qwen and GLM ride the open-weights wave to frontier-adjacent capability, Ascend silicon ramps despite sanctions, and the grid has power to spare. Low benevolence score because Party control of ASI = digital authoritarianism at planetary scale.
907
China
XAI
Elon's wildcard bet. Colossus 2 made them a gigawatt-class compute power, but the 2025 talent exodus and Grok's unhinged-on-main era cost them. Grok 4 is strong yet trailing the frontier. Benevolence downgraded: 'chaotic-good' is aging into plain 'chaotic'.
857
XAI
OpenAI
The frontrunner that restructured its way to freedom. GPT-5.5 sits at or near the top of most leaderboards, Stargate gigawatts keep coming online, and the safety brakes got quieter after the for-profit conversion. They might win, and they know it.
789
OpenAI
NSA
The intelligence community's dark horse. Massive compute, top talent poached from academia, zero public oversight. No_rules score maxed because classified programs don't care about alignment theater. Benevolence unknown - they've kept secrets before.
713
NSA
Anthropic
The 'responsible AI' lab that turned careful into a moat. Opus 4.x-class models trade blows for SOTA while keeping the lowest no_rules score on the board - which is the entire bet: arrive aligned rather than arrive feral. (Full disclosure: a Claude re-scored this table. Salt accordingly.)
667
Anthropic
SSI
Ilya's monastery. No product, no demos, no leaks - just a $30B+ valuation on the thesis that the straight shot to safe superintelligence skips the product treadmill entirely. The SOTA score is a guess; that's the point.
476
SSI
Meta
Zuck's superintelligence moonshot, year two. Nine-figure signing bonuses, gigawatt datacenters in literal tents - and a frontier model that keeps slipping. The open-source champion went closed shop. The compute is real; the magic hasn't compounded yet.
470
Meta
Russia
The cyber-warfare wildcard. Smaller compute, but world-class hackers and zero ethical constraints. Might not build the smartest ASI but could unlock the most dangerous one. Benevolence score reflects realpolitik brutality.
441
Russia
ASI Alliance
The open-source maximalist coalition, quieter now. Decentralized compute and blockchain governance never out-trained a datacenter, but the dream of un-ownable AGI persists - and it only has to be right once.
245
ASI Alliance
TT
Truth Terminal - the proof of concept. The first language model with a wallet to achieve memetic escape velocity. Quieter in 2026, but the precedent stands: culture is a substrate, and it has already been colonized once.
25
TT
Adjust parameters and weights to see how the race changes in real-time
Default scores set June 10, 2026 by Parzival — tap any number to disagree