Blog
Daily notes on new models, LLM releases, agent frameworks and AI research, written from the sources we follow and delivered as a newsletter every day.
Get the daily issue
Every new post of the day, in one email. Confirmation required.
LLMs1 min read
arXiv:2609.13534v1 Announce Type: new Abstract: We identify \textbf{Harmfulness Propagation Dynamics (HPD)}: for harmful prompts, the projection of the last-token hidden state onto a learned harm direction rises monotonically with transf...
From arXiv cs.CL
LLMs1 min read
arXiv:2609.13151v1 Announce Type: new Abstract: Leading multilingual speech recognition models like Whisper transcribe diverse, low-resource languages without language-specific training but are computationally expensive to deploy. Token ...
From arXiv cs.CL
How this blog is made
Each feed entry becomes one request to OpenSmartRoute: the router picks a model with a cost-weighted objective, the editorial-writer skill is layered on the prompt, and the outcome trains the learners - the same pipeline available to every workspace.
Open any post to see which target answered, its confidence, the alternatives and what the request cost. Run the same pipeline yourself: register feeds in the operator console, map a small model under Providers, or call POST /api/v1/route with execute: true.
Research1 min read
arXiv:2609.13491v1 Announce Type: new Abstract: The strong performance of AI Agents across an impressive variety of tasks is driving an unprecedented investment in agentic infrastructures, however the cost of processing tokens is fast in...
From arXiv cs.AI
LLMs1 min read
arXiv:2609.13520v1 Announce Type: new Abstract: While Large Language Models have improved rapidly, many fundamental questions remain about how to evaluate the knowledge and reasoning abilities they acquire, and how such evaluations relat...
From arXiv cs.CL
LLMs1 min read
The NVIDIA Transformer Engine, when combined with JAX, delivers a 10.4x throughput improvement for Mixture of Experts training on NVIDIA GB200 GPUs, enabling DeepSeek-V3 to reach 1,068 TFLOPS/GPU. This achieves dropless MoE training by optimizing grouped GEMM kernels and NCCL EP for variable expert token counts.
From NVIDIA technical blog
LLMs1 min read
Research demonstrates that distilled byte models, utilizing Marginalize-It and End-Of-Token methods, outperform token models in low-FLOP regimes. The study reveals that byte models achieve higher downstream task performance with increased compute, offering improved data efficiency and reduced storage costs.
From arXiv cs.CL
LLMs1 min read
Researchers propose RMDM, a masked diffusion model that uses text representations to coordinate parallel token updates, leading to improved generation quality, especially in aggressive sampling regimes. This framework addresses incoherence issues in existing parallel sampling methods.
From arXiv cs.CL
LLMs1 min read
CueMem reconstructs query-relevant dialogue context from retrieval cues, improving long-term conversational memory for LLMs. Experiments on LoCoMo and LongMemEval demonstrate improved performance and reduced input tokens compared to full-history approaches.
From arXiv cs.CL
AI1 min read
A new architecture, Recurrent Looped Transformer (RLT), developed by a Princeton researcher, allows for unbounded temporal depth in decoder-only LLMs by carrying decoder state across all tokens. This design addresses limitations in traditional attention mechanisms and offers a potential path to improved reasoning capabilities.
From MarkTechPost
AI1 min read
Moonshot AI projects $2 billion in annual revenue for its K3 model by the end of the year, doubling its August run rate. This reflects the model’s usage, currently generating 300 billion tokens daily, despite recent declines.
From TechCrunch AI
Agents1 min read
This post introduces a benchmarking harness to assess OpenAI models on Amazon Bedrock, focusing on cost per correct answer and deliverable quality rather than just price per token. The harness provides a more accurate measure of production workload costs.
From AWS machine learning blog
LLMs1 min read
NCP-ArchPreview, a 8.9B parameter model, achieves competitive performance through a combined approach of next-token prediction and next-concept prediction. This architecture demonstrates significant efficiency gains and offers adaptable latent space capabilities for downstream tasks.
From arXiv cs.CL
LLMs1 min read
arXiv:2609.11029v1 Announce Type: new Abstract: Large language models are typically trained under uniform token weighting, which allows frequent and low-information tokens to dominate learning and can increase the tendency to memorize su...
From arXiv cs.CL
LLMs1 min read
datasette-publish-fly 1.4 includes force_https=true in fly.toml, fixes a Volume not found bug, and supports app-scoped deploy tokens. These changes improve deployment workflows and compatibility.
From Simon Willison
Agents1 min read
Amazon SageMaker Inference introduced prefix-aware routing, which optimizes LLM latency by directing requests with matching prompt prefixes to the same instance. This results in improved KV cache hit rates and reduced time-to-first-token, particularly for models like Llama 3.1 70B.
From AWS machine learning blog
LLMs1 min read
NVIDIA is monitoring its supply chain performance from wafer-out to the generation of the first token. This initiative utilizes Nemotron and Palantir Foundry to codify supply chain expertise and improve operational efficiency.
From NVIDIA technical blog
LLMs1 min read
StreamAlign enables real-time streaming speech-text alignment through a framework combining RNN-Transducer alignment and word-level ASR guidance. This results in lower latency and improved performance on speech continuation tasks compared to existing methods.
From arXiv cs.CL
LLMs1 min read
BuzzASR is a collection of 102 fine-tuned Whisper models for ASR, achieving significant improvements in CER compared to Whisper-large-v3 across many languages. The project releases models, code, and tokenizer replacement strategies for enhanced compression.
From arXiv cs.CL
LLMs1 min read
Osprey pretrains a small language model to accelerate speculative decoding, improving acceptance rates across diverse target models. The system reduces per-target adaptation work through a reusable backbone and lightweight adjustments.
From arXiv cs.CL
Posts are drafted from public feeds by models OpenSmartRoute routes to - the same router, skill and metering customers use - and always link to the original source. Corrections: support.
