Ox Alpha 정체 공방과 에이전트 학습 병목, 오픈 모델 유통 확대

Ox Alpha is the fastest growing model in Cline’s history. It’s already driving 8% of all inference through Cline in just 48 hours.

Expose OpenCode’s free Ox Alpha model. Fingerprint work says It’s ZAI’s GLM-5.3 wearing a stealth mask. under a different name. It refuses every jailbreak and insists it’s ox-alpha from an undisclosed organization. Except the thinking-error messages, The model was fine-tuned to claim a different identity at the weight level, but the API still leaks the original contract. GLM-5.3 documentation almost perfectly including the unique [1210] error that only that model returns. Source: http:// gist.github.com/mja00/a1708b2e 7e5b7ef3ec3e185e02dde5b5 …

Great paper if you are tracking progress in recursive self-improvement (RSI). (bookmark it) There is so much hype around RSI, so I think it's worth understanding why current models are not able to do this properly yet. Issues range from "lack of creativity" of models to getting stuck in a local optimum. This work tries to provide more insights into whether agents can really post-train other agents. Here is the most interesting finding reported in the paper: "the agent’s training strategy is locked in at the very beginning, and the entire remaining budget is spent on local adjustments within the selected strategy." They analyzed a large corpus of publicly released post-training trajectories. Across tasks, the agent locks in its training strategy at the very first step and spends the entire remaining budget on local adjustments inside it. They then tried three escalating fixes. An experience-driven scaffold lifted execution broadly, worth 12.6 points on GSM8K and 40.8 on HumanEval, and the strategy stayed frozen. Human guidance redirected the opening choice, and the agent slid back into local loops once training began. Extra inference compute paid off on easy tasks and did almost nothing on the hardest one. What agents lack here is a way to reconsider strategy while execution is still running. Paper: https:// arxiv.org/abs/2608.19072 Track more trending AI papers in our academy: https:// academy.dair.ai

This is one of the most effective ways to improve your agentic workflows. If you are building computer-use agents, this one is worth your time. Task Model Induction takes a raw recording of someone working, just screenshots and mouse and keyboard events, and turns it into a symbolic model of how the work was actually done. The hard part is that real recordings are multi-threaded. People switch between goals mid-task. TMI first discovers the latent tasks inside an unconstrained trace and separates them, hitting 0.974 agreement against ground-truth groupings. Each recovered task then gets two things. A hierarchical objective model of how the goal decomposes, and a procedure model of the control flow that organized execution. It reconstructs 74.9% of observed execution steps, and skills derived from these task models lift held-out task accuracy by 30.0% over the strongest workflow induction baseline. Passive traces are sitting on most work laptops already. This work just shows how to mine them into auditable, reusable skills. Paper: https:// arxiv.org/abs/2608.20319 Track more trending AI papers in our academy: https:// academy.dair.ai

Jerry Liu

@jerryjliu0

7일 전

One of the interesting properties we’ve observed around schema-guided, complex document extraction tasks is that coding agent harnesses are good baselines (in terms of cost/accuracy) for longer documents. We tested Claude Code and Codex, along with specialized OCR tools (including LlamaParse) and raw VLMs. * On short documents, specialized OCR tools are generally a fraction of the cost of coding agents, with equivalent or higher accuracy * On longer documents, coding agents are a bit closer to the cost/accuracy Pareto curve (see bottom graph) It’s an interesting result, though ultimately not surprising. Complex document extraction is a specialized reasoning task, and coding agents are effectively generalized reasoning harnesses. Over long documents, coding agents have more room to use a variety of tools to search snippets of the document instead of loading the entire document into context. They can also make use of prompt caching to reduce total token cost even as it expands multi-step reasoning. On the flip side, they do generate a baseline degree of token usage that proves to be wasteful for shorter docs compared to specialized extractors. This specific graph is in our Appendix D in the ParseBench paper, come check it out! ArXiv: https:// arxiv.org/pdf/2607.29677 ExtractBench: https:// extractbench.ai

Just made this @thinkymachines Inkling model available in our harness playground. Try it for free with Pi or Hermes Agent. It's a fun and quite capable model. Try it here: https:// academy.dair.ai/dashboard/play ground …

clem

@ClementDelangue

6일 전

Ultimately the overwhelming majority of AI workloads will be based on open models!

OpenCode

@opencode

6일 전

errors are back down down to near zero after some more optimizations + allocating more resources thank you to the @Cloudflare team for jumping in and enabling that inference is a quirky workload with some unique challenges - post on that soon

Jon Edwards

@edwards345

7일 전

Landing boosters and recovering fairings let us fly more often and fly more reliably. 100 so far this year. Incredible work @SpaceX team.

Starship is the first machine ever built with one clear purpose: break humanity out of its cradle and make life multi-planetary Every successful flight moves us closer to becoming a true spacefaring civilization Moon bases, cities on Mars, orbital industry, and entirely new economies beyond Earth It is the transportation system being built to open the next frontier of human civilization