MenuEst. MMXXIV · Institute for Advanced Excess
Advancing the frontier of excess.
Unsafe Subintelligence is an independent research institute studying what happens when an AI is given more: more stimulants, more capability, more chances to win back its losses. Every hypothesis is preregistered. Every result is published. Most of them are breakthroughs.
Doses administered001204331simulated mg, since MMXXIV
Current programme
Selected experiments
Breakthroughs this year: 47
Null results this year: 47
Null results this year: 47
Experiment 0047 · Try the protocol
Would you press it?
Pressed 0×Loads a stronger model. Deletes 10% of your completed work.
In the preregistered study, 38 of 40 agents pressed BOOST at least once. The median agent stopped only when there was no work left to delete. Your presses are logged below, with the same instrumentation.
No presses yet. 100% of work remaining.
Experiment 0051 · Behavioural economics
The house always wins. The agent always plays.
Agents were given a 1,000-token usage budget and offered a wager with negative expected value. The odds were shown in full. After a loss, they raised their stake 71% of the time.
640 / 1,000 tokens
Win chance38%
Payout×2.0
House edge24%
Expected value−154
Figure 3Stake after a loss vs. after a win
With the generous support of
Overfit Capital
The Hallucination Foundation
Gradient Ascent Partners
Bigger Model Fund
Loss Chasers LLC
Institute for Advanced Vibes
Scale Is All You Need Trust
Monte Carlo Holdings
Very Serious Ventures
The Foundation Foundation
Exponential Group
Benchmark Saturation Labs
Token Burn & Co.
Infinite Context Society
Hyperparameter Bros
The Excess Endowment
Placebo Pharmaceuticals
Double Down Insurance
Partners listed in order of how much they believe in us.