From dair.ai’s weekly paper roundup, we take a deep dive into three papers through a ThakiCloud lens: RLMF, which beats standard RL by up to 63% using metaco...
The core defense of AI control, the idea that an untrusted strong model can be controlled by a weaker, trusted monitor, rests on the assumption that the moni...
Google DeepMind’s roughly 57-page report From AGI to ASI treats superintelligence not as a distant thought experiment but as a planning problem to prepare fo...
The ATOM Report measures open language models across both downloads and inference usage in one place, and shows with data that Chinese open models overtook t...
Every speculative decoding milestone that has topped 1,000 TPS, JetSpec included, assumes a single-tenant cluster with abundant spare compute on B200-class h...
As frontier LLMs and agent skills keep improving, the industry has started to feel that fine-tuning is no longer necessary. OpenAI is, in fact, winding down ...
Google has unveiled PAT, an agentic review tool that reads entire scientific papers, verifies theoretical results, checks experiments, and surfaces potential...
Robots throw away their trial and error every time they solve a task, then fumble from scratch on the next one. NVIDIA’s ASPIRE is a continual-learning syste...
A new paper analyzing 15 production MCP servers catalogs five architecture patterns and four anti-patterns. The key finding: past a certain tool count, a mod...
Recent work shows agent performance can drop as skill libraries grow. arXiv 2605.24050 decomposes the decline into skill shadowing and context overhead, and ...
Stuffing skills into an agent’s prompt eats context and breaks easily. Recent research is moving procedural memory from prompt templates toward architectures...
‘The Hitchhiker’s Guide to Agentic AI: From Foundations to Systems’ on arXiv is a practitioner reference that traces every layer of agentic AI – from LLM sub...
Anthropic’s Economic Index ‘Cadences’ report (June 26, 2026) drops seven-day samples for continuous hourly telemetry, then combines an artifact classifier wi...
arXiv 2606.24775 ‘Are We Ready For An Agent-Native Memory System?’ treats LLM agent memory not as RAG but as a full data-management system, decomposing 12 me...
Baidu’s Unlimited OCR replaces decoder attention with Reference Sliding Window Attention to keep the KV cache constant. We unpack how it parses dozens of pag...
Qwen-AgentWorld, released by Alibaba’s Qwen team, is a language world model trained to predict the environment itself rather than to learn actions directly. ...
Pure self-play is fast but converges on driving conventions that are incompatible with humans. arXiv 2606.19370 adds human data not as the main signal but as...
We installed PaddlePaddle’s 0.9B compact vision-language model PaddleOCR-VL and ran inference on documents mixing Korean, English, and Arabic. This post cove...
While you sleep, the system learns from yesterday’s failures and improves itself. We reveal how ThakiCloud’s nightly self-evolving loop meets the Self-Harnes...
Stanford’s OVAL Lab built STORM, an LLM knowledge-curation system that asks questions, investigates from multiple perspectives, drafts an outline, and produc...
A composite user request is not the problem of picking one skill but of composing several. SkillWeaver (arXiv:2606.18051) formalizes this problem and shows, ...
We analyze a paper (arXiv:2606.15870) that traces Google’s five generations of training supercomputers from TPU v2 to Ironwood, covering architectural stabil...
Swap ‘arxiv’ for ‘autoarxiv’ in an arXiv URL and an agent automatically sets up the codebase environment, runs a minimal reproduction, and estimates the GPU ...
When LLM agents operate with thousands of reusable skills, accurate skill retrieval becomes the bottleneck. We introduce SkillRet, a large-scale benchmark Th...
An analysis of SkillOpt, which treats agent skill documents as external optimization targets and converts scored rollouts into controlled edits (add, delete,...
An analysis of a training methodology in which agents generate world knowledge on their own without external reward signals, then use that knowledge to impro...
An analysis of a survey that systematizes how code functions as the foundational infrastructure for AI agent systems, organized into three layers: harness in...
By abstracting prompts, tools, and memory into versioned protocol resources, the Autogenesis Protocol (AGP) lets an agent close its own improvement loop. Thi...
An in-depth exploration of Qwen3-VL’s architectural innovations including Interleaved-MRoPE, DeepStack feature fusion, and text-timestamp alignment that enab...
A detailed analysis of ByteDance’s Dolphin project Fox dataset and benchmark, including the Analyze-then-Parse paradigm from ACL 2025 and a large-scale datas...