Anton Troynikov

Anton Troynikov is a cofounder of Chroma, an open-source embedding database built for AI applications. With his cofounder Huber, he helped turn tools for internal embedding experiments into a storage layer for developers using language models. He also worked on Stable Attribution, an attempt to estimate training-image influence on Stable Diffusion outputs, and coauthored the 2025 report Context Rot with Kelly Hong and Jeff Huber.[3][4]

Chroma's early development supplies context for Troynikov's later retrieval and evaluation work. In his March 2023 interview, he recalled that the team initially built it for experiments involving algorithms, visualization and manipulation of embeddings—a model's numerical representation of its input. As document question answering attracted developers, the team packaged its internal tool to make it easier for others to use.[3]

Troynikov described a practical constraint behind the design: for a long time, Chroma was just him and his cofounder Huber, and they could not dedicate a full engineer to operating a vector store. They prioritized easy deployment and rapid experimentation, alongside inserts, updates, deletes, embedding-API integration and results that an application could place in a language model's context. Chroma was a day-one LangChain integration partner at launch.[3]

The retrieval workflow he explained embeds both a document corpus and an incoming question, searches for nearby document vectors, and supplies the retrieved text with the question to a language model. This lets an application bring in private or other information unavailable during model training. He also described Chroma's vector index working with a metadata store, allowing conventional filters to constrain similarity searches—for example, returning matches only from documents containing a specified phrase.[3]

Troynikov said the difference between static academic benchmarks and changing production data informed Chroma's founding: deployed systems need ongoing decisions about training, monitoring and whether performance is improving. During the July 2023 Llama 2 discussion, he described Chroma's ongoing evaluation of the model for retrieval-augmented generation. He used SciQ multiple-choice questions with supporting paragraphs as a quick test, while considering community-sourced benchmarks closer to actual production data. His hypothesis was that smaller models could work well with retrieval because external documents could supply facts, leaving the model to reason over them.[3][1]

Earlier that year, Troynikov had described another application of the team's embedding work: Stable Attribution, an attempt to identify training images that influenced a particular Stable Diffusion generation. Removing examples one at a time and retraining would be computationally impractical, so the team pursued similarity in latent space. Its method incorporated diffusion-model properties and attention maps, treating those maps as transformations of the space in which similarity searches were performed.[3]

He emphasized that this similarity reflected how the model represented images, rather than how alike they looked to a person. The method leaned on reproduction and did not account for model generalization; its rankings were ordinal, and its cutoff was heuristic. The team had not performed the proposed comparison against a gold standard using a toy diffusion model. These limits left the method's accuracy as causal attribution unresolved.[3]

The team originally planned to reject images not generated by Stable Diffusion using its watermark, but Troynikov said the watermark was absent or undisclosed in many hosted versions. They chose to release for feedback. He also gave a copyright-policy rationale: he opposed strengthening derivative-work restrictions and feared that a backlash could give large rights holders broader control. As he put it, “we didn't want Disney to own everything forever”.[3]

In March 2023, Troynikov rejected the existential-risk arguments he had encountered, but expressed concern about individually targeted propaganda and easier access to dangerous knowledge. He emphasized the difficulty of acting in the physical world, drawing on what he described as his background as a roboticist. His preferred engineering approach involved measurable task outcomes and internal feedback mechanisms; he compared text-only prompt engineering to “trying to pick a lock with a wet noodle”.[3]

He also reported that ChatGPT, Copilot and a recent trial of Cursor had accelerated his programming work by helping with unfamiliar libraries, bugs and documentation. He acknowledged ChatGPT's hallucinations, but found it useful for pointing him toward an answer and clearing unfamiliar tasks out of the way so he could focus on work he knew well.[3]

In July 2023, Troynikov argued that access to model internals could enable research and applications unavailable through hosted APIs alone, particularly through examining embeddings and internal states. He was optimistic about agents while explicitly calling them “nothing but toys today”, and hoped models trained for particular agent tasks would make them more useful. He also argued that safety work benefited from examining actual model internals rather than reasoning only about hypothetical systems.[1]

Troynikov, Kelly Hong and Jeff Huber jointly authored Chroma's July 14, 2025 report, Context Rot: How Increasing Input Tokens Impacts LLM Performance. It evaluated 18 language models across modified Needle in a Haystack retrieval, conversational question answering and repeated-word reproduction tasks, although not every model appeared in every experiment. The retrieval experiments held task complexity constant while varying input length; the repeated-word task increased output length with input length.[4]

The authors reported increasingly unreliable, nonuniform performance as context grew. In their cleaned LongMemEval evaluation of 306 prompts, focused inputs averaging about 300 tokens produced higher performance than full conversation histories averaging about 113,000 tokens. They argued for context engineering—careful construction and management of the information supplied to a model. The study did not exhaust real-world use cases or explain the mechanism behind the observed degradation.[4]