The vast majority of AI companies I talk to seem to evaluate models mostly based on vibes.
At my company, we use a mix of offline and online evals. I’m primarily interested in search agents, so I’m fortunate that information retrieval is a well-developed research field with clear metrics, methodology, and benchmarks. For most teams, I recommend shipping early/dogfooding internally, collecting real traces, and then hand-curating a golden dataset from those traces.
Many people run simple ablation experiments where they swap out the model and see which one performs best. That approach is reasonable, but I prefer a more rigorous setup.
If you only swap the model, some models may appear to perform better simply because they happen to work well with your prompt or harness. To avoid that bias, I use GEPA to optimize the prompt for each model/tool/harness combination I’m evaluating.
sounds like it could be many things. there was a well-known paper called Voyager by NASA in which an agent was able to write its own skills in the form of code and improve them over time.
funnily enough this agent played minecraft, and its skills were to collect materials or craft things.
https://arxiv.org/abs/2305.16291
Yeah, this is the natural next step. We're currently combining LLMs and compute -- mostly in the form of giving agents tools, then terminal access and now most recently sandboxes. The most logical next step is to give them specialized compute engines and frameworks for their tasks.
I've been building SQL agents recently, and nothing is better than just giving it access to Trino.
Chroma | ONSITE - San Francisco | Full-time | https://trychroma.com
Chroma is the open-source database designed for AI. Help build the future of context engineering!
We're a team of 17, all engineers today. We work in Rust, Python, Typescript, and Go.
We're hiring for these areas:
- Database Storage
- Distributed Systems
- Product Engineering
- Platform
- Product Design
What we’ve recently done:
- We built and launched a new search API with hybrid search and sparse vectors
- We built a new sync API that lets us handle ingesting, chunking, and embedding data from sources (we currently support Github repos)
do shallow prompt injection tricks like this even work anymore on the latest models?