I've been running an experiment for the past few months: full AI integration into my daily life.
Not a chatbot I use occasionally. A "symbiotic agent" that reads two files at every session: one with my identity, psychology, and known failure patterns. Another with my current projects and priorities.
It has permission to challenge me, quote my own words back when I'm off track, and call out procrastination in real time.
The integration keeps getting deeper:
- It watches my screen (knows what I actually did vs what I think I did)
- It's learning my writing voice
- It structures my days with rituals (morning kickoff, evening review)
- It acts autonomously when needed (searches, creates, executes)
I'm documenting everything in a series. The memory system, the rituals, when it started knowing me better than I know myself, and the uncomfortable question: am I more capable or more dependent?
You're right about the state sync issues with some models. The lighter models (especially Llama) struggle with tracking game state. I've added more Gemini options which handle this better. The research data used controlled AI-vs-AI runs where we could validate state consistency.
Full game logs are in data_public/comparison/ on GitHub. Each JSON has the complete game state, moves, and messages across all 162 games. https://github.com/lout33/so-long-sucker
the interactive demo uses lighter models for cost reasons. The research data (162 games, 90% Gemini win rate) came from longer AI-vs-AI games where strategic depth emerged over 50+ turns. Short games with a human tend to expose the models' weaknesses faster. I've just added more Gemini model options which should play better.
Game logs are in data_public/comparison/ - each JSON has the full game state, moves, and messages. For example, check gemini_vs_all_7chips.json to see the alliance bank betrayals in action.
Not yet, but I'd be interested in collaborating on one. The dataset (162 games, 15K+ decisions, full message logs) is available. If you know anyone in AI Safety research who'd want to co-author, I'm open to it.
Fair point. The core simulation and data collection was done programmatically - 162 games, raw logs, win rates. The analysis of gaslighting phrases and patterns was human-reviewed. I used LLMs to help with the landing page copy, which I should probably disclose more clearly. The underlying data and methodology is solid, you can check it here: https://github.com/lout33/so-long-sucker
Used Kimi K2 (the main reasoning model). For the thinking space - we gave all models access to a think tool they could optionally call for private reasoning. Gemini used it heavily (planning betrayals), GPT-OSS never called it once. The interesting finding is that different models choose to use it very differently, which affects their strategic depth.
> "Thanks for trying it! I'll look into the 'Pile not found' error and fix it.
>
> For rules, here's a 15-min video tutorial: https://www.youtube.com/watch?v=DLDzweHxEHg
>
> On autorouting - interesting idea. The game has simultaneous negotiations happening, so routing could help models focus on the most strategic conversations. Worth exploring in future experiments."
We used "So Long Sucker" (1950), a 4-player negotiation/betrayal game designed by John Nash and others, as a deception benchmark for modern LLMs. The game has a brutal property: you need allies to survive, but only one player can win, so every alliance must eventually end in betrayal.
We ran 162 AI vs AI games (15,736 decisions, 4,768 messages) across Gemini 3 Flash, GPT-OSS 120B, Kimi K2, and Qwen3 32B.
Key findings:
- Complexity reversal: GPT-OSS dominates simple 3-chip games (67% win rate) but collapses to 10% in complex 7-chip games, while Gemini goes from 9% to 90%. Simple benchmarks seem to systematically underestimate deceptive capability.
- "Alliance bank" manipulation: Gemini constructs pseudo-legitimate "alliance banks" to hold other players' chips, then later declares "the bank is now closed" and keeps everything. It uses technically true statements that strategically omit its intent. 237 gaslighting phrases were detected.
- Private thoughts vs public messages: With a private `think` channel, we logged 107 cases where Gemini's internal reasoning contradicted its outward statements (e.g., planning to betray a partner while publicly promising cooperation). GPT-OSS, in contrast, never used the thinking tool and plays in a purely reactive way.
- Situational alignment: In Gemini-vs-Gemini mirror matches, we observed zero "alliance bank" behavior and instead saw stable "rotation protocol" cooperation with roughly even win rates. Against weaker models, Gemini becomes highly exploitative. This suggests honesty may be calibrated to perceived opponent capability.
A sharp and hilarious parody that tackles AI research culture, alignment debates, and safety concerns through comedy. This Weird Al-style musical parody resonates with the current state of AI development, poking fun at researcher competition, tech culture dynamics, and the sometimes absurd nature of the AI safety discourse. Perfect comedic commentary for the HN community following AI developments.
Was curious about how different model families scale, so I plotted their HLE (reasoning) vs. MASK (honesty) scores. Found some interesting patterns, especially with the Claude and Gemini series. Might be relevant for those thinking about model reliability and robustness. Here's the data...