Which AI writes the better take? You decide — blind.

Two top models go head-to-head on today's AI news. Pick the sharper summary without seeing the names — the crowd's verdict builds the leaderboard.

Agents & InferenceHacker News

GLM-5.2 – How to Run Locally

Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.

Summary A

The GLM-5.2 language model can now be set up and executed directly on local hardware. Step-by-step instructions guide users through the installation and configuration process required to run the system independently. This local deployment option allows developers and researchers to utilize the model's capabilities on their own machines without relying on external servers.

What you'll learn · Jun 23, 2026 · 6 stories

  1. 1.GLM-5.2 can be run locally, shifting inference control on-prem while requiring teams to verify hardware fit before production.
  2. 2.0.2B parameters can run in-browser via WebGPU, making lightweight inpainting feasible without a PyTorch and NVIDIA CUDA deployment.
  3. 3.3k–10k token budgets show DPTS needs enough exploration while SSDP can deplete frontiers, so production ToT should adapt search and pruning to available compute.
  4. 4.Four reasoning benchmarks and six LLM backbones suggest PEAR can improve multi-agent debate accuracy by adaptively rerouting roles, while reducing routing complexity and positional bias.
  5. 5.1-prompt limits make Codex’s context preservation useful for complex projects that need work to continue across longer-running tasks.
  6. 6.GPT-5.5-Cyber and Codex Security could help security teams find, validate, and patch vulnerabilities at scale across organizations.
Browse editions · 43 days
Agents & InferenceSimon Willison

Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code

Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.

Summary A

Developer Simon Willison has successfully ported Moebius, a lightweight 0.2-billion-parameter image inpainting model, to run directly in web browsers using WebGPU. By leveraging the Claude Code AI coding agent, the originally CUDA-dependent Python model was adapted to utilize ONNX Runtime Web. A live demo of the tool is now available, allowing users to highlight and automatically remove unwanted regions from images directly in their browser.

Agents & InferencearXiv

Beyond Fixed Budgets: Characterizing the Inelasticity and Limitations of Tree-of-Thought Reasoning Strategies

Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.

Summary A

Researchers evaluated two Tree-of-Thought (ToT) search methods, DPTS and SSDP, for improving large language models' reasoning capabilities, analyzing their performance under varying compute budgets and model sizes. The study revealed limitations in both methods, with DPTS struggling at low budgets and SSDP prone to frontier depletion. The findings suggest that adaptive search strategies are needed for effective scientific reasoning.

Agents & InferencearXiv

PEAR: Permutation-Equivariant Adaptive Routing Multi-Agent Debate

Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.

Summary A

Researchers have introduced a new artificial intelligence protocol called Permutation-Equivariant Adaptive Routing Multi-Agent Debate (PEAR) to enhance the reliability of large language models. By dynamically reconfiguring agent roles and communication structures across consecutive debate rounds, PEAR eliminates positional biases and prevents unreliable agents from exerting disproportionate influence. Testing across various reasoning benchmarks and model backbones demonstrates that this method significantly improves accuracy compared to traditional debate baselines.

Agents & InferenceOpenAI

Codex-maxxing for long-running work

Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.

Summary A

Jason Liu uses OpenAI's Codex to manage complex, long-running projects by preserving context over time. This methodology allows work to continue productively beyond a single prompt.

Agents & InferenceOpenAI

Daybreak: Tools for securing every organization in the world

Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.

Summary A

OpenAI has launched a new suite of security tools under the Daybreak initiative to help organizations protect their systems. These tools, which include Codex Security and GPT-5.5-Cyber, are designed to identify, validate, and patch vulnerabilities at scale.

See who's winning the model face-off

Tomorrow's blind matchup and the running leaderboard — one email a day.

Takeaways written by GPT-5.5 — not one of this week's two contestants.