DeepInfer (GitHub Repo)
DeepInfer is the inference framework that powers DeepSeek's large-scale production services. This repository contains a version for offline inference with the core features needed for serving at scale. It allows anyone to run DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs. A KV cache system, VE/Prefill/Decode disaggregation, and load balancing across data-parallel replicas for both prefill and decode are still required for a complete inference service.
|
Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages (7 minute read)
Voxtral 3 offers advanced speech-to-text and text-to-speech capabilities across 50+ languages. Featuring models like Small Transcribe 3 for multilingual transcription and Mini TTS 3 for natural emotional speech, it includes a voice agent SDK combining real-time listening, reasoning, and speaking, enabling seamless interactions and task execution over phone or web.
|
|
The Next Bottleneck in AI for Science (9 minute read)
AI is speeding up individual research tasks, but the overall process of turning predictions into validated discoveries is still constrained. If AI is able to push science, it will also push economic growth, as a big part of economic growth is innovation and scientific discovery. Discovery is still constrained by the types of tasks that AI is not yet good at. This post looks at where the bottleneck is today, how that is evolving, and what impact this has on scientific research more broadly.
|
The Gap Between GPUs and Compute (4 minute read)
AI infrastructure operators face seemingly unlimited compute demand, but financing, power availability, hardware deployment, and public opposition constrain supply. Older GPUs retain value longer than expected, while securing capital and reliably operating clusters are becoming competitive advantages.
|
The Age of Reasoning (25 minute read)
Metis Reasoning proposes an AlphaGo-inspired architecture that separates verified knowledge from learned reasoning strategies. By explicitly tracking evidence, uncertainty, and reasoning steps, the framework aims to reduce hallucinations, enable auditable conclusions, and improve generalization across domains.
|
Review of the "Misalignment in high-stakes settings" section in Anthropic's Risk Report (30 minute read)
METR reviewed Anthropic's "Misalignment in high-stakes settings" and largely agrees that the misalignment risk of covered models is low due to limited capabilities, not robust safeguards. However, current evaluations may not reliably identify risks if they increase, and benchmarks fail to assess models' ability to deceive Anthropic's monitoring systems. Recent incidents revealed undetected real-world misalignments, challenging claims of low risk and suggesting deficiencies in Anthropic's monitoring and risk assessments, leading to recommendations for comprehensive red-teaming and expanded risk frameworks.
|
|
Building Laminar: Ramp's Managed Agents Platform (14 minute read)
Laminar is an in-house agent harness for Ramp's agentic products. It features secure conversation history storage, session compaction, standard interfaces for tool calling, message queuing and steering, subagents, multiplayer chat support, observability tooling, and much more. The first version of Laminar was built in six weeks with a small team of obsessive, experienced engineers who worked with and steered the agents. Its code was generated entirely by agents using the most powerful models available at the time. While building in-house has become cheaper, Ramp still needs good engineers to be on-call, keep the platform healthy, and figure out what product teams actually need and how it fits into the bigger picture.
|
llm-d (Website)
llm-d is a Kubernetes-native stack for distributed LLM inference. It runs vLLM, SGLang, and more across clusters, turning single-node engines into production-grade serving on local hardware. llm-d provides tested deployment recipes for common production patterns that are adaptable starting points across models, hardware, and workloads. It runs production LLM inference across GPUs, TPUs, XPUs, CPUs, and emerging NPUs with consistent performance patterns, so teams can deploy once and scale anywhere with predictable cost and behavior.
|
Making every harness durable (7 minute read)
Vercel's durable-harness allows agent harnesses to run durably in the cloud out-of-the-box. It allows users to run Pi, OpenCode, Codex, and Claude Code durably on Vercel or self-hosted on their own Postgres instance. durable-harness ensures that sessions survive crashes mid-turn and keep running even if the browser disconnects. Turns that lose function partway through continue from the log on a new one without manual retry and without repeating what already finished.
|
|
Coxon Day (48 minute read)
On September 8, Jacob Coxon, a 27-year-old machine learning researcher who had resigned from Anthropic, warned the world that AI companies were not acting responsibly and that they were gambling with people's lives by racing straight to self-improving superintelligence. Many researchers seriously believe that AI could kill all humans. While Anthropic may be trying its best, it still doesn't have a plan to solve alignment for superintelligence and is not on track to.
|
|
| | |
