Gödel Machines
Gödel Machines is an AI research lab working on language models, inference, voice interfaces, agents, robotics and reinforcement learning. This overview brings together our published research, open models and projects.
Our work spans model architecture, the systems used to run models, and their applications. We pursue problems across these areas when we see a useful question we can investigate and test.
2.1 Sarvam's illusion of safety [1]. An audit of multilingual AI safety, examining why Sarvam models respond differently to harmful requests across languages.
2.2 Goedel-mHC-1B [2]. An open language model exploring how hyperconnections can improve training efficiency and performance.
2.3 Looped language models [3]. Experiments in reusing model layers to add depth and make inference more efficient.
2.4 On systems of intelligence [4]. An essay on how AI changes software and services, and where human accountability belongs.
2.5 Physical AI: Quadruped control architecture [6]. Part 1 of the Physical AI series: how a quadruped's control software coordinates balance and movement.
2.6 Physical AI: Quadruped mechanics [7]. Part 2 of the Physical AI series: why robots use four legs, and the mechanics behind walking.
2.7 Inference: From noise to words [8]. Part 1: how text diffusion became practical.
2.8 Inference: Making every denoising pass fast [9]. Part 2: making each denoising pass cheaper with W4A16 expert kernels.
2.9 Inference: Fewer, smaller forwards [10]. Part 3: D2F and Funnel decoding schedules.
2.10 Physical AI: Quadruped market survey [11]. Part 3: quadruped manufacturers, applications, and the Indian market.
2.11 Challenges [5]. Open machine-learning problems with public baselines and a leaderboard.
Overhear. A voice interface for email, web search, tasks and workflows, designed to work through conversation without a screen. Development has ended and the project is no longer maintained.
Table 1 lists the status of our released projects as of July 2026.
We are hiring. To apply, send a project you would like us to review, such as a repository, paper or demo, to hi@goedelmachines.com. A résumé is optional.
[1] Gödel Machines. Sarvam's illusion of safety. March 2026.
[2] Gödel Machines. Releasing goedel-mHC-1B. March 2026.
[3] Gödel Machines. Looped language models. July 2026.
[4] Gödel Machines. On systems of intelligence. July 2026.
[5] Gödel Machines. Challenges are now live. April 2026.
[6] Gödel Machines. Physical AI: Quadruped control architecture. September 2026.
[7] Gödel Machines. Physical AI: Quadruped mechanics. September 2026.
[8] Gödel Machines. Inference: From noise to words. September 2026.
[9] Gödel Machines. Inference: Making every denoising pass fast. September 2026.
[10] Gödel Machines. Inference: Fewer, smaller forwards. September 2026.
[11] Gödel Machines. Physical AI: Quadruped market survey. September 2026.
Correspondence: hi@goedelmachines.com. 6th Floor, Ilyas Mohammed Khan Estate, Road No. 1, Mithila Nagar, Banjara Hills, Hyderabad, Telangana 500034.