Mind Lab puts continual learning to the test with Macaron-V1 (11 minute read)
Mind Lab claims its Macaron-V1 model surpasses GLM-5.2 in its benchmarks. The model was built by attaching five LoRA expert modules, each with about one billion parameters, to GLM-5.1. The system dynamically switches to the expert model best suited to the task it is given. The accumulated data from model use can be distilled into a dedicated LoRA adapted that is continually updated as the model is called.
|
|
How OpenAI Built GPT-Live (8 minute read)
OpenAI rebuilt its voice architecture around a full-duplex model that listened and spoke simultaneously. The system combined stateful inference, asynchronous delegation, dynamic context management, and low-latency media transport to keep conversations responsive while supporting advanced reasoning and tool use.
|
One agent, every surface: how we built the Kiro agent harness (20 minute read)
Kiro is an agentic IDE with features such as specs, steering, and hooks. The Kiro agent harness is a lightweight server-side process that runs alongside codebases, starts quickly, and owns everything on the agent side. The IDE, CLI, and Web clients own how the user interacts with the agent and how it presents the agent's work. The only way to cross that boundary is through the defined protocol interface. The well-defined interface between server and client means the agent code evolves independently of the clients.
|
OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems (34 minute read)
OpenAI's recently released solutions for ten major open mathematics problems show that AI is now superhumanly capable at cyber and coding and superhuman at advanced math, the same way non-AI computers have been superhuman at basic math for a long time. These problems were well-defined, formalized problems where the solution could be easily verified. A lot of what constitutes AI R&D is verifiable. The lab that gets traction on true AI R&D self-improvement loops will find themselves in an overwhelmingly strong position.
|
|
Orchard (GitHub Repo)
Orchard is an open-source agentic modeling framework. Its foundation is a thin, Kubernetes-native environment service that exposes generic primitives with no assumptions about the harness, trainer, inference backend, or task domain sitting above it. This foundation allows for recipes that use the same substrate for trajectory distillation, on-policy RL rollouts, and evaluations. Datasets, training recipes, and evaluation protocols stay portable across harnesses, domains, and projects instead of being rebuilt for each new study.
|
MirrorCode (8 minute read)
MirrorCode is a benchmark that tests AI models on long-horizon tasks. It contains tasks where AI models have to reimplement entire programs end-to-end without access to the original source code. AI-generated solutions must match the original program's output exactly on end-to-end tests. The benchmark's 25 target programs span different areas of computing, including Unix utilities, data serialization and query tools, bioinformatics, interpreters, static analysis, cryptography, and compression.
|
From RLVR to RLSVR (GitHub Repo)
RLSVR expanded RLVR beyond inherently verifiable problems by transforming open-ended tasks into proxy environments with rules and outcomes that generated their own reward signals. SpyRL demonstrated the approach through multi-agent self-play, where predetermined roles and voting made evaluation automatic.
|
|
GPT-5.6 Sol Uses Twice the Tokens of GPT-5.5 (2 minute read)
GPT-5.6 Sol xhigh now uses more than twice as many tokens per session as GPT-5.5 xhigh in Codex workflows. This cuts the effective value of a token-based quota by more than half. At the same token price, 2.25x the tokens means roughly 2.25x the cost for a similar token mix. GPT-5.6 Sol also adds a cache-write charge that GPT-5.5 didn't have.
|
|
| | |
