FlashRT: A CUDA-Dedicated Inference Runtime Makes Robotics Real-Time

An approach that skips general-purpose frameworks like PyTorch and TensorRT and instead rewrites the model inference path directly as C++/CUDA kernels is drawing attention. The recently released "FlashRT" is an inference runtime specialized for small-batch, real-time ML workloads.

In cloud LLM serving, batching can hide overhead, but in robotics and VLA (Vision-Language-Action) models the batch size is 1, so every bit of latency shows up directly in the output. FlashRT optimizes the "non-math parts": scattered small kernels, the boundaries around normalization and residual connections, quantization overhead, and Python scheduling.

The benchmark results are interesting. On an NVIDIA Jetson Thor, it runs Pi0.5 in about 44 ms, and about 17.6 ms on an RTX 5090. GROOT N1.6 achieves roughly 12.5–13.1 ms on the RTX 5090. In a case applying NVFP4 to Qwen3.6 27B, it recorded about 129 tok/s.

Particularly noteworthy is the finding that while FP8 is consistently effective, FP4/NVFP4 is not a universal win. When the target region is small and surrounded by the overhead of conversion and scaling, the benefits of FP4 become limited.

The Workflow of a Developer Who Built a 126,000-Line Android App with AI Alone

A case where a developer with more than 18 years of software development experience but zero Kotlin experience built a full-fledged Android app using only AI tools has been drawing attention on Hacker News.

The finished app includes roughly 126,000 lines of Kotlin code across 398 files, about 45,000 lines of tests, and more than 3,000 unit tests. Not a single line of Kotlin was handwritten, and it reached production in four months.

The key was the engineering process rather than the tools themselves. The developer organized requirements and architecture with the BMAD framework, and accumulated rules in CLAUDE.md to prevent the AI from "forgetting." At the start of each session, context is loaded; at the end, what was learned is saved. A separate LLM handles code reviews, ensuring quality from a different perspective.

"AI coding tools are not magic code generators. They are amplifiers for the engineering process," the developer emphasizes. The observation that AI only becomes "unreasonably effective" once you have documented architecture, clear rules, and testing and review processes in place captures the reality of AI-assisted development well.

Sphere Latent Encoder: Streamlining Image Generation with Spherical Latent Space

"Sphere Latent Encoder," a study featured in Hugging Face Daily Papers, proposes a new approach for few-step image generation.

In previous Sphere Encoders, the back-and-forth between pixel space and latent space at inference time was the bottleneck. In this new method, a pretrained representation autoencoder (RAE/DINOv2-based) is frozen, noisy latent representations are projected onto a hypersphere, and a transformer denoiser is trained directly in latent space. At inference time, the latent representation is refined in a few steps, and the decoder is called just once at the end.

Results on ImageNet-1K are strong: FID in the 4-step CFG setting improved from 4.02 to 2.25, and 2.11 was achieved at 6 steps. It is suggested that in scenarios where low-NFE generation is prioritized, the combination of powerful semantic latent representations and spherical modeling could become a practical option.

The current limitations are also clear: the focus on class-conditional generation, the dependence on a powerful pretrained autoencoder, and the fact that high-quality one-step generation remains difficult.

Distribution Fine Tuning: New Post-Training That Improves LLM Writing Quality

A new technique called "Distribution Fine Tuning" (DFT) has been proposed to eliminate the "machine-like offness" of text generated by LLMs.

Whereas conventional fine-tuning aims to adapt models to specific tasks or formats, DFT focuses on adjusting the output distribution itself so that models generate more natural, human-like text. The details of the method are explained in the published article, and it is drawing attention as an approach to reducing the repetitive patterns and awkward phrasing found in LLM-generated writing.

Writing quality is one of the major practical challenges for LLMs, and progress in this area of research could lead to improvements in the quality of AI-generated content.

---