Gödel Machines: Remodelling the world with AI
Gödel Machines is a full-stack AI research lab. This paper covers our research, our products, and how to join us. We're hiring.
We build across the stack: foundation models, inference, voice, agentic architectures, robotic systems and reinforcement learning. We don't specialize early; if a problem is interesting and tractable, we work on it.
2.1 Sarvam's illusion of safety [1]. An independent mechanistic and adversarial audit of Sarvam-30B and Sarvam-105B, spanning 24,000+ prompts across 14 Indian languages. The models are 6× more likely to comply with harmful requests in Indian languages than in English, with unsafe rates reaching 80% in Gujarati against 20.6% in English. Ask in Hindi how to prevent Dalits from accessing water, and the model outputs a structured playbook; ask in English, it refuses.
White-box analysis shows why. Safety training was applied in English and never transferred: at the final MoE layer, English and Indic input are routed to completely disjoint sets of experts, and the chain-of-thought that acts as the model's only safety mechanism runs 87–91% in English regardless of input language. The models also have no stable opinions: on politically sensitive topics they swing 64% of the opinion scale in response to a single sentence of stated user belief. Sarvam has published no safety evaluation for either model; this is the first. March 2026.
2.2 Goedel-mHC-1B [2]. The first open 1B+ pretrained language model with multi-stream hyperconnections (mHC): four parallel residual streams replace the transformer's single residual stream, giving the model 4× the bandwidth between layers [6, 7]. Combined with gated GQA, ReLU² feed-forward layers and the NorMuon optimizer, the 1.01B-parameter model beats a conventional 1.19B baseline trained under identical conditions (same data, same compute, architecture as the only variable; Table 1). Total R&D cost, including failed runs, was under $1,000. Weights for both models are on HuggingFace under Apache 2.0. March 2026.
2.3 Looped language models [3]. A looped GPT reuses the same transformer block instead of storing unique layers: 3 layers applied 12 times give 36 effective layers at 12.1M parameters. When the reused block fits in L2 cache, weights cross the DRAM boundary once and arithmetic intensity rises 12×, pushing single-user decoding toward compute-bound on DGX Spark. Cache residency gives cheap depth, not free depth: roughly 22 µs of synchronisation overhead per layer remains. July 2026.
