Looped Transformers Paper: Turning Transformers into Programmable Computers
max_paperclips · x · 2026-09-02
Dimitris Papail shares the backstory of his group's paper, "Looped Transformers as Programmable Computers." He recalls being "AGI pilled" by davinci in 2022, leading his group to pivot to understanding language models. The paper attempts to explain how Transformers can implement general-purpose algorithms of arbitrary depth via prompts. He notes that while he still doesn't fully understand Transformers, the rise of agents—where the object of study becomes the instrument—makes the field incredibly exciting.
Related event: Researcher Reflects on Journey from Transformers to Agents(2 posts)→
More from Research
- CrowdStrike fine-tunes Nemotron 3 for security triage; Cyber Defense Benchmark shows zero passes — joshua_saxe · 2026-09-02
- Discussion on Looped Transformer efficiency and interpretability — aryaman2020 · 2026-09-02
- View: Transformers May Not Gain Much by Reasoning in Latent Space — xuanalogue · 2026-09-02
- Opinion: Abandoning CoT Monitoring for Latent Reasoning is Inevitable — Darpinian · 2026-09-02
- Generation is cheap, but physical verification remains the bottleneck in science — shyamalanadkat · 2026-09-02
- A Primer on Flow Matching: Transporting Probability Distributions via Vector Fields — ariG23498 · 2026-09-02