Princeton's "CEO-Bench": AI Agents Simulated Running a Company for 500 Days — Most Go Bankrupt

A research team at Princeton University has built "CEO-Bench," a benchmark in which AI agents run a fictional software company for 500 days (in simulation). The results were grim: many of today's leading models fell below their starting capital, and only three models finished in the black.

Even more interesting, a simple rule-based heuristic that uses no AI at all outperformed nearly every AI model. This suggests there are domains where the autonomous decision-making of AI agents still falls short of simple rules.

These results offer an important perspective for evaluating the practical utility of AI agents. Against the expectation that agents can "autonomously handle complex tasks," it turns out that human-designed rules can still be more effective for long-term strategic decision-making.

A Qwen 3 Inference Engine Written in C from Scratch Goes Public

A CPU-only inference engine for Qwen 3 (4B and below), written from scratch in C, has been released in the Reddit LocalLLaMA community. Its only external dependencies are libc, libm, and cJSON (with optional parallelization via OpenMP). It loads directly from HuggingFace safetensors and performs 4-bit quantization on the fly.

The developer is a C programmer with no machine learning background who built the engine while learning how LLMs work from the ground up. They leaned on ChatGPT as a "teacher," but wrote all the code themselves and prioritized readability for learning purposes. The result runs at just 1 token per second, but it's an implementation with high educational value.

Compared with existing fast C implementations (qwen3.c), it stands out for aiming to be "easy to read and suitable for learning" — making it a valuable resource for developers who want to understand the internals of LLMs.

DuckDuckGo Starts Serving CAPTCHAs to Local LLM Agents, Impacting Search Integration

Reports are piling up from users in the LocalLLaMA community that local LLM agents based on llama.cpp have started running into CAPTCHA blocks when using DuckDuckGo search. Users have observed agents outputting messages like "DuckDuckGo is blocking with a CAPTCHA. Let me try other approaches."

This suggests that search portals are stepping up countermeasures against the use of search tools by AI agents. It's a practical lesson: in automated workflows built on local LLMs, the availability of search APIs can become a real constraint.

role-model: A Routing Protocol Unifying Local and Cloud AI

A hybrid local/cloud AI router called "role-model" was announced on Hacker News. It combines deterministic routing with fallback to a controller model, selecting the optimal model (local or cloud) based on task type.

What's distinctive is its design philosophy: the greater the differences between models, the more value routing provides. Rather than splitting traffic among frontier models, it shines when dividing work across a lineup of models with clear gaps in speed, cost, and quality. The project also expects a secondary effect: encouraging the emergence of a market for specialized models.

From the Japanese-Speaking Community: GEO, AI Crawler Specs, and Strands Agents

On Qiita, practical Japanese-language articles were published covering "GEO (Generative Engine Optimization)," a content optimization technique for the AI search era; a roundup of web crawler specs from major AI services; and tool-integration implementations with AWS Strands Agents. GEO in particular — a method for making your content more likely to be cited by AI search engines like ChatGPT and Perplexity — looks set to become a new guiding principle for content creation.

---