Hi HN, I'm Hongyin. I built Vibivibi (https://vibivibi.com) because I kept starting a coding agent session on my laptop and wanting to continue it on a server, or hand it to a colleague. I want an easier way to do it but won’t trust anyone hosting my traces who promises they won’t sell it (market price per coding trace is $50 to $100).What Vibivibi does: a small CLI, `vibi`, finds the sessions...
https://www.subconscious.dev/blog/subconscious-glm-5-2-makes-compact-obsolete
GLM-5.2 is a turning point for coding agents. It's the first model a business would actually pay to replace Claude Opus with. We gave GLM-5.2 the ability to control its own context during agentic tasks, and paired it with Subconscious Cache — ensuring reliable long-term memory far beyond the context window.This blog post walks through our approach to long-horizon agents, long-context LLMs,...
https://huggingface.co/papers/2507.16784
- We build the Thread Inference Model (TIM) based on the transformer architecture, and its dedicated runtime TIMRUN.- TIM + TIMRUN = Intelligent workflow generation, context engineering, and multi-hop tool use happens at the runtime level- TIM + TIMRUN supports virtually unlimited reasoning enabled by context pruning, significantly improves the efficiency for long-horizon reasoning tasks-...
- TIM + TIMRUN = Intelligent workflow generation, context engineering, and multi-hop tool use happens at the runtime level
- TIM + TIMRUN supports virtually unlimited reasoning enabled by context pruning, significantly improves the efficiency for long-horizon reasoning tasks
- Inference API is live at https://subconscious.dev/
- More details: https://github.com/subconscious-systems/TIMRUN
Paper: https://arxiv.org/abs/2309.10814 An automatic NLEP generation toolkit is opensourced: https://github.com/luohongyin/langcode
Example Colab notebook is included in the Github repo.
This work introduces the following features of NLEP
1. NLEP is a full python program that prints the target response of LLMs. 2. Task-general NLEP prompting outperforms task-specific chain-of-thought prompting on math, symbolic, and natural language. 3. Enable the chain-of-thought reasoning ability of small models (RoBERTa) on text classification 4. Hierarchical instructing via program completion.
The interesting aspect is that the ChatGPT response quality is not improved with search results, and it can be misled by noisy search.
Our slurm is not fixed yet, but I've brought the model back online for now - it will be temporally diabled again when our slurm is fixed.
Gradio does not allow tasks more than a minute, so the demo might fail for long texts.