All published work
25 articles, newest first.
A peer-reviewed Nature study trains on 21 million RNA sequence–structure pairs and tests whether structural supervision improves RNA prediction. The results are promising, but benchmark success is not experimental validation.
An experimental Qwen3.5-4B derivative embeds an exact RISC-V processor in constructed neural layers and uses recurrent state as RAM. The design is inspectable, but not a replacement for ordinary tools.
A new H100 and H200 measurement study finds that speculative decoding's energy advantage depends on batch size, draft acceptance, and the cost of verifying rejected tokens.
The October 29 weight release makes Mistral's large multimodal MoE inspectable and self-hostable, while memory requirements, benchmark caveats, and a hosting-license restriction shape its practical use.
A synthetic protocell study links DNA mutations to vacuoles, movement, and cargo storage in RNA droplets. The missing step is an autonomous cycle of inheritance and selection.
A Nature study combines multi-ancestry genetics, transcriptomics, and computational gene prioritization to map 127 Alzheimer's risk loci. Its microglial signal is stronger than its neuron-specific claims.
Liquid AI's six new open-weight models reuse a 1.7B-parameter backbone across modalities, with practical local deployment paths and a consequential commercial license limit.
The new open-source CUDA kernel library combines quantization formats and GPU-specific execution paths; vLLM benchmarks show useful gains with important precision caveats.
A high-access review of OpenAI's internal monitoring system and two later agent incidents show why oversight needs its own security boundary.
A newly published network model finds persistent cooperative structure even when harmful interactions dominate, but its public code does not exactly match the latest paper's interaction rules.
I think understanding the system, testing it properly, managing agent context, and keeping strong computer science fundamentals matter more than manually inspecting every generated line.
I compared scGPT and Geneformer with PCA, Harmony, and scVI on two single-cell RNA-seq datasets. The results were useful, but not the sweeping foundation-model win one might expect.
The new Apache-2.0 PyTorch backend exposes Trainium as a native device for eager execution, compilation, distributed training, and custom kernels, but API gaps and hardware constraints still matter.
Clef and Clef-flash put Apache-2.0 weights, multimodal inputs, a compatible typed API, and local deployment behind models built to score bounded decisions instead of generate text.
A Nature study from Google DeepMind shows that detectable watermarks can be embedded in designed protein sequences and predicted structures while preserving tested function, but robustness and deployment remain open problems.
OpenShell 0.1 puts filesystem, process, network, and credential controls outside the agent itself, with an Apache-2.0 runtime that can span containers, Kubernetes, and microVMs.
My DecPort experiments found label-free, input-specific decision-behavior transfer across three frozen LLM backbones, while a preregistered final gate found no evidence that the learned shared DecisionCore itself was portable.
Anthropic reports that autonomous Claude agents surfaced ART, a repeat-associated reverse-transcriptase system in phages, but the preprint also exposes important limits in reproducibility and biological validation.
Qwen-Image-2.1 is a 7B image generator and editor with native RGBA output and day-one support across major inference stacks, but its research-only license sharply limits how open the release really is.
OpenAI’s September flagship improved long-horizon agentic performance and became the first model the company classified at its Critical cybersecurity threshold, while ARC Prize showed how strongly peak performance depends on the surrounding harness.
A September preprint separates phenotypic from genotypic diversity across self-replicating neural cellular automata, including 24 long 200 × 200 runs.
A Nature study turns papers and their code into MCP-backed agents; 74 of 100 computational-biology papers were successfully agentified, showing both the promise and the dependency problem.
Cactus Compute’s September 16 release uses nested 2-to-20-layer subnetworks so one open-weight checkpoint can target different edge budgets without maintaining separate models.
The September public beta moves context management, tool orchestration, subagents, and long-running session state behind a managed API while leaving execution environments configurable.
An independent reproduction finds exact Orthrus trajectories in every tested FP32 case, but only 43–45% under BF16.
