READING
Reading
Papers and links I think are cool.
Ankit Sonthalia, Haritz Puerto, Alexander Rubinstein, Martin Gubri, Seong Joon Oh · 2026 · Added October 13, 2026Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?(opens in new tab)The concept of bottling the agent in this paper is very interesting. Nowadays we're used to using llms by querying them
many times when solving related tasks, when they can be used as a tool to develop cheaper solutions for these workloads.
Also, I really appreciated seeing this paper compare bottling to Jev (a model that answers with structured values only,
never text) since that has become a hot topic in tech recently.
#llm#agents#quality-and-cost
Qirui Wu, Stan Birchfield, Hesam Rabeti, Angel X. Chang, Bowen Wen · 2024 · Added October 12, 2026GATOR: Generative and Agentic 3D Object Reconstruction From Casual Images(opens in new tab)I don't work in this field, but I'm always drawn to computer vision topics like object detection, and this one takes that a step further. GATOR turns a few casual photos into a complete, textured 3D object, even filling in the hidden parts, then uses an AI agent to refine it by comparing renders to the original photos.
#generative#agentic#object-reconstruction
Miao Rang, Zhenni Bi, Hang Zhou, Hanting Chen, An Xiao, Tianyu Guo, Kai Han, Xinghao Chen, Yunhe Wang · 2025 · Added September 28, 2026Revealing the Power of Post-Training for Small Language Models via Knowledge Distillation(opens in new tab)We are in an era where SLMs can be extremely powerful when used correctly. I'm in the process of learning how
to properly post-train SLMs to squeeze the maximum amount of value out of them, and came across this great article.
Utilizing offline on-policy knowledge distribution so these models can achieve superior performance is fascinating!
#Post-Training#SLM#Knowledge-Distribution
Connor O'Keefe, Yuri Volkov · 2026 · Added August 17, 2026How Cursor Router chooses the right model for the task(opens in new tab)Cursor's new router skips the guesswork: a classifier called Compass checks if a task is simple enough for a cheap model, and if not, a learned taxonomy picks whichever frontier model (Grok, Sol, Opus, or Fable) tends to win at that kind of work. Results so far: Auto Intelligence matches Fable-tier quality at 68% less cost, and Auto Balance beats Opus 4.8 while spending 41% less.
#frontier-models#routing#performance
Mohamad Abou Ali, Fadi Dornaika · 2025 · Added July 27, 2026Agentic AI: A Comprehensive Survey of Architectures, Applications, and Future Directions(opens in new tab)This survey untangles what people mean by “agentic AI” by framing systems as either symbolic/classical or neural/generative, then mapping the major architectures and applications across the literature. It also calls out governance and safety gaps and argues that the most practical path forward is intentional hybridization, not picking one paradigm and stopping there.
#agentic-ai#survey#governance