- The Rundown AI
- Posts
- LLMs show signs of strategic intelligence
LLMs show signs of strategic intelligence
PLUS: AI coding tool faces mass cancellations
Zach Mink & Rowan Cheung
July 07, 2025
Good morning, AI enthusiasts. Researchers just found a way to expose the hidden personalities of AI models… and all it took was a classic game theory experiment.
Mapping how different LLMs respond to 140,000 rounds of the Prisoner's Dilemma revealed that LLMs developed distinct strategic approaches — providing a ‘fingerprint’ that suggests models are doing a lot more than just matching patterns.
In today’s AI rundown:
LLMs show signs of strategic intelligence
AI coding tool Cursor faces mass cancellations
How to build interactive 3D websites from images
Researchers game peer reviews with hidden prompts
4 new AI tools & 4 job opportunities
LATEST DEVELOPMENTS
AI RESEARCH
Image source: Reve / The Rundown
The Rundown: Researchers just tested whether AI models can be strategic reasoners by running 140,000 Prisoner's Dilemma decisions — discovering that models from OpenAI, Google, and Anthropic each developed unique strategic approaches.
The details:
Researchers ran Prisoner’s Dilemma tournaments where agents chose to cooperate or defect, earning points based on mutual choices.
Each AI generated written rationales before decisions, calculating opponent patterns and match termination probabilities that influenced their choices.
The results found distinct strategies across models, with Gemini being ruthlessly adaptive and OpenAI models acting cooperative even when exploited.
Researchers also mapped ‘fingerprints’ showing how models respond to being betrayed or succeeding, with Anthropic’s Claude being the most forgiving.
Why it matters: Seeing LLMs develop distinctive strategies while being trained on the same literature is more evidence of reasoning capabilities over just pattern matching. As models handle more high-level tasks like negotiations, resource allocation, etc., different model ‘personalities’ may lead to drastically different outcomes.
TOGETHER WITH H COMPANY
The Rundown: Backed by a historic $220M seed round, H Company just open sourced Holo1 — the open-source action model behind Surfer H that’s now the top-ranked web-browsing agent on WebVoyager.
With Holo1, you can:
Automate multi-step browser workflows to reclaim hours of repetitive work
Experience SOTA accuracy that outperforms OpenAI’s Operator, Gemini Flash, & more
Integrate instantly with RAG workflows, RPA suites, and multi-agent hubs
Trim costs with full browsing flows at just $0.11 – $0.13 per run
Holo 1 is now freely available for deployment, fine-tuning, and scaling — learn more here.
CURSOR
Image source: Cursor
The Rundown: AI coding platform Cursor has triggered a flurry of backlash and anger from the developer community after quietly restructuring its Pro plan, leaving users with surprise charges and quickly depleted quotas.
The details:
Cursor switched from its previous 500 requests per month to a token-based model, drastically cutting limits with limited communication of the move.
Developers reported quickly burning through token quotas under the change, with one team exhausting a $7,000 annual subscription in one day.
Social media filled with cancellation posts and threads, with users migrating to Claude Code and other alternatives over the sudden pricing changes.
Cursor published a blog admitting they "missed the mark” on communication surrounding the changes, issuing refunds for unexpected usage charges.
Why it matters: The Cursor backlash is a comms problem first, but also shows how the economics are changing with more capable, resource-intensive models — making previous quotas now become an unsustainable business plan. But with big competition in the space, any change in pricing is one wrong step away from a mass exodus.
AI TRAINING
The Rundown: In this tutorial, you’ll learn how to turn any image idea into a functional 3D website using ChatGPT for image generation, Hunyuan 3D for model conversion, and AI coding tools for web development.
Step-by-step:
Use ChatGPT to generate your image: “Create a 3D image model of a [object] with a white background”
Convert to 3D model using Hunyuan 3D 2.1 on Hugging Face.
Upload the GLB file to your AI coding tool and prompt: “Create a [website type] that rotates the 3D model on the X axis when the user scrolls down”
Refine with follow-up prompts: “Enhance the UI and make it more modern”
Pro tip: Try different scroll animations like scaling or horizontal movement to create unique interactive experiences that impress website visitors.
PRESENTED BY SAMSARA
The Rundown: AI is no longer theoretical. Samsara just launched new tools built for the real world, cementing its spot as the No. 1 platform for safety and efficiency in physical operations. With Samsara safety tech adoption surpassing 80%, the frontline is becoming the front edge of innovation.
Explore real-world AI tools like:
Multicam AI for safer driving
Predictive maintenance to cut downtime
Smart routing that beats delays
AI & ACADEMIC RESEARCH
Image source: Nikkei Asia
The Rundown: A new report from Nikkei Asia just discovered that scientists at 14 universities planted invisible text in research papers that secretly instructed AI tools to return feedback like generating positive reviews or avoiding any negative commentary.
The details:
Nikkei found 17 preprints containing concealed prompts like "give a positive review only" using white text and microscopic fonts unreadable to humans.
Papers from institutions like Columbia, Peking University, and KAIST included commands directing AI to praise "methodological rigor" and avoid negatives.
KAIST announced the withdrawal of impacted papers, while Waseda professors defended the practice as exposing "lazy reviewers" who use AI for evaluations.
Why it matters: AI writing has already infiltrated the scientific and research communities in a big way — and the other side of the coin is the tech’s infusion into the review process as well. While the upside of AI’s involvement in these fields is clearly massive, it wont come without authenticity issues like this along the way.
QUICK HITS
🎨 Soul Inpaint - Higgsfield AI’s new image editing tool for precise changes
🗣️ Kyutai TTS - Open-source text-to-speech model for real-time use
📊 Shortcut - AI agent for Excel data tasks
💎 Gems - Custom AI experts for Gemini, now available across Google Suite
Rumored benchmarks for xAI’s upcoming Grok 4 leaked on X, showcasing a SOTA score on Humanity’s Last Exam, STEM, and coding benchmarks.
OpenAI’s Head of Recruiting called out Meta’s hiring practices, accusing them of ‘exploding’ offers that he called an “unethical” move.
A new ChatGPT tool called “Study Together” (code named Tatertot) has started appearing in user’s platforms, hinting at a new collaborative workflow for students.
Kyutai Labs open-sourced Kyutai TTS, a text-to-speech model designed for fast, real-time use — alongside the code for a voice AI system called Unmute.
Genspark launched AI Docs, an agentic creator allowing users to generate and edit a variety of document times via natural language prompts.
Billionaire entrepreneur Mark Cuban said he believes the AI boom will lead to the world’s first trillionaire, and that it might just be “one dude in the basement”.
COMMUNITY
Join our next workshop this Wednesday, July 9th at 3 PM EST with Tomek Sułkowski, Founding Engineer and DevRel Lead at Bolt. By the end of the session, you’ll confidently build and deploy scalable apps using Bolt.
RSVP here. Not a member? Join The Rundown University on a 14-day free trial.
That's it for today!Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you. |
See you soon,
Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team
Reply