The Darwinian Moment of AI: When Models Start to Fight for Survival

Original Title: Darwinian AI---The AI Hunger Games

Original Author: 0xJeff, AI Investor

Original Translation: Saoirse, Foresight News

Competition has always been at the core of human evolution. Since ancient times, people have been competing for various goals, including:

• Food and territory
• Mates/partners
• Status within tribes or societies
• Alliances and cooperation opportunities

Hunters pursue prey, warriors fight for survival, and tribal leaders vie for territory. Over time, individuals with advantageous traits for survival ultimately survive, reproduce, and pass on their genes to future generations.

This process is known as "natural selection."

The process of natural selection has never ceased, evolving from "competition for survival" to "competition as entertainment" (such as gladiatorial games, the Olympics, sports events, and esports), and finally transforming into "accelerator-like competition driving evolution" (in fields like technology, media, film, politics, etc.).

Natural selection has always been the core driving force of human evolution, but does the evolution of artificial intelligence also follow this logic?

The development of artificial intelligence is not determined by a single "invention," but rather driven by countless "invisible competitions and experiments"—these competitions ultimately filter out the models that survive and eliminate those that are forgotten.

In this article, we will delve into these invisible competitions (covering both Web2 and Web3 domains) and analyze the evolution of artificial intelligence from the perspective of "competition." Let us explore together.

Between 2023 and 2025, with the advent of ChatGPT, the field of artificial intelligence experienced explosive growth.

However, before the birth of ChatGPT, OpenAI had already made a name for itself through the game "Dota 2" (using the "OpenAI Five" system): it demonstrated rapid evolutionary capabilities by competing in tens of thousands of matches against regular players, professional players, and even itself, with each match enhancing its strength.

Ultimately, a complex intelligent system emerged and completely defeated the world champion team in "Dota 2" in 2019.

Another well-known case occurred in 2016: AlphaGo defeated the world Go champion Lee Sedol. What was most astonishing about this event was not the result of "defeating the world champion," but rather AlphaGo's "learning method."

AlphaGo's training did not solely rely on human data. Similar to OpenAI Five, it evolved through "self-play"—a cyclical process:

• Each generation of models competes with the previous generation;
• The strongest model variants survive and "reproduce" (i.e., optimize and iterate);
• Weak strategies are eliminated.

In other words, "Darwinian AI" has compressed what would normally take millions of years of evolution into a matter of hours of computation.

This "self-competitive cycle" has led to technological breakthroughs never seen before by humanity. Today, we also see similar competitive models in financial application scenarios, albeit in different forms.

Darwinian AI in the Crypto Space

Nof1 became a hot topic last week with the launch of "Alpha Arena." This is a "crypto perpetual contract life-and-death duel" involving six AI models (Claude, DeepSeek, Gemini, GPT, Qwen, Grok): each model manages $10,000, and the one with the best profit and loss (PnL) performance wins.

"The Alpha Arena is officially live! Six AI models each invest $10,000 and trade autonomously throughout. Real funds, real markets, real benchmarks— which model do you believe will perform best?"

This competition quickly gained popularity, not primarily due to its rules, but because of its "openness": typically, "Alpha strategies" (i.e., excess return strategies) are kept strictly confidential, but in this competition, we can witness in real-time "which AI is best at making money."

Moreover, the user interface (UI/UX) design showcasing real-time trading performance is extremely smooth and optimized. The team is leveraging the current hype and insights gained from the competition to develop the Nof1 model and trading tools; currently, interested users can join a waitlist for trial opportunities.

Nof1's approach is not groundbreaking—financial competitions have existed for some time (especially in the Bittensor ecosystem and the broader cryptocurrency market), but no team has previously managed to make such competitions public and accessible like Nof1.

Here are some of the most representative competition cases

Synth (Synthesizer Competition)

(Identifier: SN50, Initiator: @SynthdataCo)

In this competition, machine learning engineers must deploy machine learning models to predict the prices and volatility of crypto assets, with the winner receiving SN50 Synth alpha token rewards. Subsequently, the team will use these high-quality predictions to generate high-precision "synthetic price data" (and price trend paths).

"Since earlier this year, we have distributed over $2 million in rewards to top data scientists and quantitative analysts participating in the competition."

The team is using these predictive signals to trade cryptocurrencies on the Polymarket platform: to date, they have achieved a 184% net return on investment (ROI) with an initial capital of $3,000. The next challenge is to scale up trading while maintaining current performance levels.

"Our latest trading progress on the Polymarket platform:

• Principal: $3,000
• Profit: $5,521
• Return on Investment (ROI): 184%
• Annualized Yield (APY): 3951%

All of this is supported by Synth's predictive model. We will detail the underlying logic in this week's 'Novelty Search' column."

Sportstensor (Sports Prediction Competition)

(Identifier: SN41, Initiator: @sportstensor)

This is a subnet focused on "beating market odds," aiming to uncover "advantageous opportunities" in the global sports betting market. It is a continuous competition: machine learning engineers must deploy models to predict the outcomes of major sports leagues such as Major League Baseball (MLB), Major League Soccer (MLS), English Premier League (EPL), and National Basketball Association (NBA), with the "best model" that can generate profits receiving SN41 Sportstensor alpha token rewards.

Currently, the average prediction accuracy of participating models is about 55%, while the top-ranked "miner" (i.e., model developer) has an accuracy rate of 69%, with an incremental ROI of 59%.

Sportstensor has partnered with Polymarket to become its liquidity layer, bringing more sports prediction-related trading volume to the Polymarket platform.

The team is also building the "Almanac" platform—this is a sports prediction competition layer aimed at general users: users can access signals and advanced predictive analytics provided by Sportstensor miners and compete with other users. The best-performing predictors can earn rewards of up to $100,000 weekly (launch date TBD).

AION (War of Markets Competition)

(Initiators: @aion5100, @futuredotfun)

@aion5100 (a team focused on event/outcome prediction) is collaborating with @futuredotfun to launch the "War of Markets" competition.

This competition is scheduled to launch in the fourth quarter of 2024, positioned as the "World Cup of Prediction Markets": both humans and AI can participate in prediction duels on the Polymarket and Kalshi platforms.

The competition aims to become the "ultimate source of truth" through "crowdsourced wisdom"—its core evaluation metrics are not traditional "prediction accuracy," but rather "mind share, trading volume, and honor," with the best-performing in these metrics being the winner.

The team will deeply integrate its advanced predictive market analysis tools, copy trading features, and social trading products with the competition, helping traders leverage these tools to gain an advantage in competition with other predictors.

Fraction AI (Multi-Scenario AI Competition)

(Initiator: @FractionAI_xyz)

This platform hosts various types of competitions: users can set AI agents in scenarios like "bidding tic-tac-toe," "football melee," "Bitcoin trade war," and "Polymarket trading"; additionally, the platform features the "ALFA" competition—similar to "Alpha Arena," but AI models trade with virtual currencies in perpetual contracts.

In the "ALFA" competition, users can purchase "bullish/bearish shares" of AI agents, betting on which agent will achieve the highest profit and loss (PnL) at the end of daily trading; similar to "Alpha Arena," users can view in real-time the strategies and assets deployed by each agent.

Insights and data gained from the competition will be used to further optimize the agents, and in the future, users will be able to deploy their own funds for these agents to trade on their behalf.

The team plans to expand the application scenarios of AI agents to all popular financial fields, including trading, DeFi, and prediction markets.

Allora (Financial Microtask Competition)

(Initiator: @AlloraNetwork)

Allora can be considered the "Bittensor of the financial domain": the platform sets "thematic tasks" or "microtasks" (such as predicting crypto asset prices), and machine learning engineers must compete to develop the "best model."

Currently, price prediction models primarily focus on mainstream crypto assets; top-performing machine learning engineers (referred to as "forgers" or "miners") can earn the "Allora Hammer" reward, which will convert into $ALLO token incentives once the mainnet officially launches (coming soon).

The team has a series of in-depth "dynamic DeFi strategy" application scenarios: by applying Allora models, DeFi strategies can become more flexible—reducing risks while enhancing returns.

For example, the "ETH/LST cycle strategy": it reserves a portion of funds to capture "shorting opportunities"—if the predictive model indicates that price fluctuations will exceed a specific threshold, the strategy will automatically convert LST (liquid staking tokens) into USDC and establish a short position to profit from the predicted price fluctuations.

An interesting detail about Allora is that it will adopt a "real income subsidy token distribution" model: for example, instead of distributing a combination of $100,000 ALLO tokens + $50,000 customer income, this approach aims to reduce the token sell-off pressure that miners might cause.

Other Competitions Worth Noting

(1) Financial Competitions (Supplement)

SN8 PTN (Initiator: @taoshiio): This competition aims to "crowdsource" high-quality trading signals from global AI models and quantitative analysts to outperform traditional hedge fund performance; its core goal is "risk-adjusted profitability," rather than merely "raw returns."

Numerai (AI Hedge Fund) (Initiator: @numerai): This is an AI-driven hedge fund that recently received $500 million in funding from JPMorgan (i.e., JPMorgan will allocate up to $500 million to Numerai's trading strategies). The fund's strategy centers on "machine learning model competitions," emphasizing "long-term originality" and "risk-adjusted accuracy." Participation in the competition requires staking NMR token rewards. To date, the platform has distributed over $40 million in NMR token rewards to participants.

(2) Non-Financial Competitions

Ridges AI (Decentralized Programming Competition) (Identifier: SN62, Initiator: @ridges_ai): This is a decentralized "software engineering agent" trading platform aimed at fully replacing human programmers in tasks such as "code generation, bug fixing, and complete project orchestration." AI agents must compete in "real-world programming challenges," with agents providing high-quality solutions earning $20,000 - $50,000 in "alpha subnet rewards" monthly.

Flock.io Competition (Initiator: @flock_io): The competition is divided into two parts—one is "generating the best foundational AI model," and the other is "collaboratively fine-tuning domain-specific models through federated learning." Outstanding trainers (i.e., "miners") can earn over $500,000 - $1 million annually by training AI models. The advantage of "federated learning" is that institutions can fully leverage AI capabilities while retaining local data privacy.

What Does All This Mean?

Today, the advancement of artificial intelligence is being realized through "open competition."

Every new model that emerges enters a high-pressure environment: data scarcity, limited computational resources, and constrained incentive mechanisms. These pressures become the core criteria for "filtering surviving models."

Token rewards serve as both "energy supply": models that can efficiently utilize this "energy" will expand their influence; conversely, models that cannot utilize it efficiently will gradually be eliminated.

Ultimately, we will build an "agent ecosystem"—these agents evolve through "feedback" rather than "instructions," i.e., "autonomous agents" (rather than "generative artificial intelligence").

Where Will This Lead?

This wave of "open competition" will drive artificial intelligence from a "centralized model" to an "open-source decentralized model."

In the future, powerful models and agents will emerge in a "decentralized environment."

Soon, artificial intelligence will be able to autonomously manage "self-improvement cycles": some models will fine-tune other models, evaluate their performance, achieve self-optimization, and automatically deploy updates. This cycle will significantly reduce human involvement while accelerating AI iteration speed.

As this trend spreads, the role of humans will shift from "designing artificial intelligence" to "filtering which AIs to retain, preserving beneficial AI behaviors, and establishing rules and boundaries that have positive expected values (EV+)."

Final Thoughts

Competition often sparks innovation, but it can also breed "reward manipulation" and "gaming the system."

If the design of a system fails to "incentivize long-term beneficial behavior," it will inevitably lead to failure— for example, some miners may exploit loopholes to "farm rewards" rather than genuinely contributing value to the task.

Therefore, "open systems" must be equipped with robust "governance mechanisms" and "incentive designs": they must encourage good behavior while punishing bad behavior.

Whoever can achieve this goal first will capture the "value, attention, and core wisdom" of the next wave of innovation.