Karan Malhotra

Karan Malhotra is a co-founder of Nous Research who collaborated with Teknium on the first Hermes language model and demonstrated WorldSim, a Claude-generated environment navigated through simulated shell commands. His work and public views emphasize making language models customizable and usable locally, alongside experiments that extend their interaction beyond a chat assistant.[2][1]

In the WorldSim demonstration published by Latent Space on April 27, 2024, Malhotra used a command-line format to steer Claude toward generating an imagined environment. Commands such as cd, ls and cat prompted it to invent directories, files and their contents. He described the goal as coaxing Claude into having “the assistant take a back seat”—his interpretation of how prompting could shift the interaction toward a simulator.[1]

Malhotra credited Amanda Askell’s published Claude system prompt with inspiring his use of third-person descriptions of the assistant. His framing included “the assistant's in a CLI mood today.” He also wrapped interactions in XML tags, saying Claude handled that format effectively, and shared the system prompt and chain so others could replicate the experiment. He explained that sharing his version gave people access to a particular set of generated world-building commands, while their own WorldSim sessions could produce different results.[1]

The demonstration also exposed the importance of specifying constraints. After creating a universe, Malhotra asked WorldSim to create Twitter and advance the world. Because he had omitted chronological constraints, the generated story placed Twitter at the creation of the universe, with humans joining much later. He acknowledged the mistake and warned about the numerical output: “I don't expect these numbers to be right.” The example was an imagined narrative, rather than a validated reconstruction of history.[1]

WebSim co-founder Rob Haisfield credited Malhotra’s initial demonstration with inspiring his own experiments in prompt engineering. Haisfield also said that an introduction from Malhotra connected him with his co-founder Shawn. WebSim carried the idea of a generated environment into a browser, where invented URLs could lead to generated pages.[1]

Before the WorldSim demonstration, Malhotra had worked on fine-tuning open language models. His January 2024 account of that earlier work explains the original Hermes collaboration and the limited-compute setting behind his interest in smaller, locally usable models.[2]

In an interview recorded January 31, 2024, Malhotra recalled discovering a GPT-2 Colab notebook and making fine-tunes in the styles of Edgar Allan Poe and Lovecraft. His experiments subsequently expanded toward using language models for learning, production and automation.[2]

While working with Open Assistant under LAION, Malhotra had access to compute and was generating GPT-4 instruction data. Teknium was independently producing data in the same format, so Malhotra proposed combining their resources. They first trained gpt4-vicuna by further fine-tuning an existing Vicuna fine-tune. They then decided to train the base LLaMA model on their data instead; Malhotra identified the result as the first Hermes.[2]

Malhotra explained the approach as knowledge distillation: a larger model generates compact instructions or conversations that become training examples for a smaller model. Rather than requiring the smaller model to learn from all the underlying material, the examples convey compressed explanations and answers. He connected this choice to the collaborators being “GPU-poor” and wanting useful models that could run offline, while acknowledging that the transfer of information was imperfect.[2]

According to Malhotra, Nous Research initially served as a name for his collaboration with Teknium. Interest in their work brought more contributors together through Twitter and Discord, developing into an open-source research collective. In January 2024, he explicitly limited his own Hermes involvement to the initial model and credited Teknium with leading subsequent Hermes work.[2]

In the January 2024 interview, Malhotra emphasized local and offline use: “run the model yourself, do everything at home”. He argued that many problems remained unsolved at small model sizes and should be investigated before moving to much larger models. As Nous developed a corporate branch, he described an intention to build paid tools and conduct research that would make open models more effective.[2]

His fine-tuning advice stressed researching hyperparameters and learning rates. He favored continuing training when compute allowed and the model was not overfitting, while qualifying this as a personal belief and noting that he was not then training Hermes or many of the models discussed. He also recommended Axolotl as, in his opinion, an accessible and effective general-purpose trainer.[2]

Malhotra advised against selling models made from GPT-4 outputs out of respect for the model creators’ terms of service. He explicitly distinguished his licensing opinions from Nous Research’s views, describing them as his own individual position in January 2024.[2]

Malhotra is also listed as an author of the Hermes 4 Technical Report, with the identifier @karan4d.[3]