# Good AI Labs > Good AI Labs pursues intelligence that learns from experience and acts in the world through decision models, robot learning, spatial memory and persistent agents. An independent AI research lab founded in 2024. Contact: research@goodailabs.com. ## What we’re building ### Decisions and robot control Local decisions for software agents and experimental robot control. - [Reflex-1](https://www.goodailabs.com/research/reflex-1/) - [Samsara-VLA](https://www.goodailabs.com/research/samsara-vla/) ### Spatial memory Video retrieval and scene reconstruction. - [re1](https://www.goodailabs.com/products/re1/) ### Agents for ongoing work Use tools, remember prior work and return to an ongoing project. - [Colony](https://www.goodailabs.com/products/colony/) ## Systems - [Colony](https://www.goodailabs.com/products/colony/): Personal agents that keep working across conversations. Give them projects, tools and memory. They can write code, use your Mac, build small apps and return to scheduled work on a computer you control. - [re1](https://www.goodailabs.com/products/re1/): Spatial memory for physical AI. Return to an observation, inspect its source video and reconstruct the place for simulation. ## Current research - [Reflex-1](https://www.goodailabs.com/research/reflex-1/): a 421M-parameter local decision model for tool routing, intent classification and action selection. October 5 release: 92.93% on a reserved 495-example Banking77 subset; 21/30 native MiniWoB++ episodes with a fixed controller. These are not full benchmark scores. Banking77 CPU inference: 375.2 ms median, 452.1 ms p95 on Intel Xeon Platinum 8558, FP32, four compute threads; 2.17 GiB peak process RSS. Current GPU latency is unmeasured. Conditions and limitations: https://www.goodailabs.com/research/reflex-1/#current-results. Measurements: https://www.goodailabs.com/research/reflex-1/current-evaluation.json. - [Samsara-VLA](https://www.goodailabs.com/research/samsara-vla/): experimental recurrent robot control, 217M parameters. The released checkpoint reached 94.50% native LIBERO success across 2,000 episodes. Samsara-VLA Tiny reached 89.30%. Private model downloads and a playground with inference on the user's device. Measurements: https://www.goodailabs.com/research/samsara-vla/release-evaluation.json. - [VIO-lens-2](https://www.goodailabs.com/research/vio-lens-2/): research in progress. Reflex-1 and Samsara VLA include evaluation conditions and downloadable JSON measurement summaries. reflex-1 model and weights: https://huggingface.co/gai-labs/reflex-1. Current runtime release: 1.0.0. Public FP32 weights, tokenizers and Transformers inference code; no account or API key required. ## Writing - [Introducing Reflex-1](https://www.goodailabs.com/blog/reflex-1-fast-decisions/): A 421M-parameter model for tool routing, intent classification and action selection, with public weights and measured CPU performance. - [Introducing Samsara-VLA](https://www.goodailabs.com/blog/samsara-capability-and-compute/): Robot control with memory, in two compact models. - [Selecting Samsara's checkpoint](https://www.goodailabs.com/blog/a-number-you-cannot-tune-for/): The selected 80,000-step checkpoint finished two episodes ahead of the 34,000-step candidate. All five results from the same 2,000-episode evaluation. - [Introducing re1: spatial memory for physical AI](https://www.goodailabs.com/blog/re1-video-memory/): Search recorded events, check the source video and revisit captured places in 3D. - [Introducing Colony: personal agents with memory and tools](https://www.goodailabs.com/blog/colony-runtime-and-model-budget/): Give an agent an ongoing job, a place to work and memory it can use next time. Colony is opening for technical previews. - [Why we build local models](https://www.goodailabs.com/blog/goodness-first/): We want people to be able to run and adapt the models we build. That puts hardware requirements and operating control in the research brief. - [Four tests for a robot placement](https://www.goodailabs.com/blog/four-ways-to-grade-the-same-robot/): A placement can reach the goal and fall over afterward. Four historical Samsara protocols tested when success should count. - [Choosing hardware for a local model](https://www.goodailabs.com/blog/sovereign-intelligence/): reflex-1 runs on a laptop CPU as well as an H200. The request, latency target and runtime memory determine which machine fits the application. - [Budgeting a local decision service](https://www.goodailabs.com/blog/security-review-is-the-market/): Our reflex-1 timings start with the model loaded. A deployed service also has startup, queues and tool calls to account for. - [Adapting 9M parameters inside a 421M model](https://www.goodailabs.com/blog/debugging-by-decomposition/): reflex-1's main adaptation run took 75 minutes on one H200. Where the trainable parameters sit and what that timing includes. ## Follow the lab - [X](https://x.com/goodailabs) - [LinkedIn](https://www.linkedin.com/company/goodailabs/) - [Hugging Face](https://huggingface.co/gai-labs) - [YouTube](https://www.youtube.com/@goodailabs) - [DEV](https://dev.to/good_ai_labs) - [Zenodo](https://zenodo.org/communities/goodailabs/) - [Hacker News](https://news.ycombinator.com/user?id=goodailabs) - [Reddit](https://www.reddit.com/user/goodailabs/) - [Instagram](https://www.instagram.com/goodailabs/) - [Threads](https://www.threads.com/@goodailabs) - [Kaggle](https://www.kaggle.com/goodailabs) ## Pages - [About the lab](https://www.goodailabs.com/team/) - [Updates](https://www.goodailabs.com/updates/) - [Careers](https://www.goodailabs.com/careers/) - [Contact](mailto:research@goodailabs.com) Every content page has a plain-text twin at its own URL plus llms.txt. [Full public reference](https://www.goodailabs.com/llms-full.txt): all public page text, model comparisons and evaluation conditions.