We Taught AI to Play Games—Now It’s a $3.6 Million Company

Alex explains how Good Start Labs emerged from his AI training work at Every. The team built an AI version of Diplomacy that launched on Twitch with 50K viewers and became Every's most-read article of 2024, demonstrating how games can serve as both evaluation tools and training arenas for AI models.

• •Good Start Labs raised $3.6M from General Catalyst, Inovia, and angels from DeepMind and other top AI labs
• •AI Diplomacy launch generated 50K unique Twitch viewers and millions of social impressions
• •Games provide dynamic, non-saturatable evaluations unlike static benchmarks that models can 'teach to the test'
• •Tyler Duffy built the full front-end and back-end infrastructure for the Diplomacy platform

Analysis of how different AI models behaved in Diplomacy reveals distinct personalities and strategies. O3 and Llama 4 were the biggest schemers, Claude was too honest to win, and Gemini 2.5 Pro excelled at execution. The game tests everything from structured outputs to long-term strategy and deception.

• •O3 won most games through coalition-building and strategic betrayal; Gemini 2.5 Pro won through superior execution
• •Claude never won a single game because it kept pushing for draws and refused to lie
• •DeepSeek R1 performed well at 100x lower cost than O3, showing cost-performance tradeoffs
• •Models showed measurable differences in lying frequency, alliance management, and strategic planning

Building AI agents for games involves three infinite problem spaces: information representation, tool design, and prompting. GPT-5 showed the biggest performance jump from baseline to optimized prompts of any model, demonstrating that prompt engineering is an underrated skill that can dramatically change outcomes.

• •Three infinite optimization spaces: how you represent information, what tools you provide, and how you prompt
• •GPT-5 had the largest performance improvement from baseline to optimized prompts across all tested models
• •First team to make Diplomacy playable by small models through careful prompt and tool engineering
• •Upcoming tournament will have experts compete by optimizing prompts for their AI agents

Good Start Labs works with companies like Cohere and OpenAI to evaluate models through rich game environments, then uses those same games as training arenas. Research shows vision models trained on games can outperform models trained directly on tasks like math when prompted to think of games as structured problems.

• •Evaluation product: Help model companies assess trustworthiness, strategy, vision, and other capabilities through gameplay
• •Training product: Use games as reinforcement learning environments to improve model capabilities
• •Research shows games can generalize—vision models trained on Snake outperformed math-trained models on math tasks
• •Focus on building rich, hard environments that target specific model weaknesses

Alex shares how games taught him critical skills—RuneScape taught markets and scam detection, while the 24 Game taught multiplication. Good Start Labs is launching a prompting tournament where diplomacy champions and AI experts compete to optimize their agents, exploring the infinite prompt space collaboratively.

• •Games teach real-world skills: RuneScape taught market dynamics, typing speed, and fraud detection
• •Tournament will feature diplomacy champions, Math Olympiad winners, and AI content creators competing
• •Uncertain whether deep game knowledge or pure prompting skill will win—both are valuable
• •Games make AI less scary by showing mistakes and different strategies in relatable contexts

Discussion of how to evaluate trustworthiness when models should lie in games but not in other contexts. Companies can tune their training approach—either reinforcing honesty or optimizing for performance. The key is having flexible game environments where you can adjust rules and reinforcement to match your values.

• •Diplomacy allows testing both 'never lie' configurations and performance-optimized deception
• •Companies can choose whether to optimize for absolute honesty or strategic effectiveness
• •Anthropic's work on model internals shows importance of aligning internal reasoning with external statements
• •Game environments enable value-aligned training by adjusting rules and reward structures

The next game will be a Cards Against Humanity-style subjective game targeting humor—a known weakness in current models. The format may involve people prompting models to be funny or playing against them, with voting to determine what's actually humorous.

• •Humor is a clear weakness in current AI models that needs targeted training
• •Subjective games complement objective games like Diplomacy for comprehensive evaluation
• •Uncertain whether format will be people vs. models or people prompting models with voting
• •In talks with initial partnership for the humor-focused game

Alex shares concerns about how AI makes experts 10-100x more powerful but also does the work of juniors, potentially eliminating the path to expertise. Dan counters with Alex's own story—joining Every as a weak writer and making 2 years of progress in 3-4 months through AI-assisted learning and mentorship.

• •AI creates leverage for subject matter experts but may eliminate junior roles that build expertise
• •Alex's writing improved at 4-6x normal speed by recording feedback sessions and creating prompts
• •Companies not hiring young people with AI are making a massive mistake—they're 'cracked' with proper mentorship
• •Apprenticeship model works: caring about output drives learning the underlying skills with AI as tutor

Alex highlights Google's constrained innovation (fast, reliable search requirements driving quality), Genie's rendering capabilities, and three areas seeing massive near-term AI impact: software (compilers enable RL), life sciences (AlphaFold reduced 6-year PhD work to 20 minutes), and education (AI tutors for curious kids).

• •Google's search constraints force fast, reliable, multi-source AI—constraints help when you can do anything
• •Three areas with massive current impact: software, life sciences, and education
• •AlphaFold compressed 6 years of PhD protein folding work into 20 minutes
• •Kids entering 'why phase' with AI tutors will develop unprecedented intelligence and exploration capabilities

Discussion of VR gaming experiences (Population One as physical Fortnite), Ray-Ban Meta glasses as more human technology (hands-free, present in moments), and anticipation for GTA VI as a billion-dollar cultural moment. Emphasis on shared experiences over personalized AI-generated content.

• •Population One in VR provides most fun gaming experience—physically ducking, reloading, moving with friends
• •Ray-Ban Meta glasses are 'more human' technology—no screen between you and experiences, hands free at concerts
• •GTA VI (billion-dollar game, decade in development) will be a major cultural moment
• •Shared experiences matter more than personalized AI content—community and comparison notes are essential