DeepSeek-V4-Flash 효율 경쟁, 에이전트 운영 자동화, 공개 재현 생태계

1 / We're introducing two new updates to Snowflake Cortex AI with dynamic model routing in Cortex AI Gateway and expanded open model access with DeepSeek-V4-Flash and GLM-5.3. Together, these updates help teams automatically use the right model for each task based on task performance, customer preference, cost and speed. The deep dive

DeepSeek V4 Flash beat outperformed Claude Fable 5 - on Terminal-Bench while being 11× cheaper. - 79% to 88% hit by letting the model verify itself. Yes bro No new training. Just smartly scaling/smarter test-time compute. Now How right? Sample 5 solutions : let the same model rank them with LLM-as-a-Verifier, accuracy jumps from 79% to 88%. This is what happens when open-source models become good enough to verify their own work at scale.

Karpathy's agentic engineering lifecycle, clearly explained: (using open-source tooling built by Google) The tooling to build agents is quite mature, so most of the work in shipping an agent isn't writing the agent anymore. It's everything after, including scaffolding it, deploying it to a runtime, locking down its identity and network, evaluating it, and publishing it somewhere people can use. Each of those has traditionally lived in its own console, its own config, its own separate tool. Google's Agents CLI + skills implements procedures to condense the entire lifecycle into the coding agent itself, by prompting in plain English. A setup command injects the lifecycle skills, so a single coding agent can carry an idea from an empty folder to a governed, published enterprise asset. I mapped the full lifecycle in the diagram below. Here's what each stage does. > Setup installs the skills into any coding agent (Claude Code, Cursor, Codex, Antigravity) from one command. > Build scaffolds the agent and its deterministic tools from a prompt, then you run it locally in the playground. > Deploy pushes it onto Agent Runtime with Sessions and Memory Bank, so it holds state across runs. > Govern is the security stage, and Agents CLI drives all of it from prompts. It provisions a dedicated least-privilege identity, screens untrusted text for prompt injection through Model Armor, and confines the agent to an egress allow-list of hosts you approve. > Evaluate checks grounding and hallucination, and then optimizes the prompt while proving no regression. > Publish registers the agent into Gemini Enterprise for the whole org to use. Every stage requires just natural-language prompts. Agents CLI GitHub repo → https:// fandf.co/4qdqSgL (don't forget to star it ) To dive deeper, I wrote the full hands-on build, from install to enterprise registration, and worked with the Google Cloud team to put this together. Read it below.

Jerry Liu

@jerryjliu0

If you’re building an “agent over your PDFs” in 2026, one of the most underrated things you can do is to provide exact grounding back to the source document for any agentic answers. Specifically, showing the exact word, line, or region that the answer came from. Frontier vision models do a terrible job at this. They have a hard time predicting exact bounding boxes over region, tying it back to the source answer, and surfacing it as context in a cost-effective and accurate manner. Dedicated VLM-based document OCR/extraction tools like LlamaParse are a lot better at visual grounding and specifically tying it to extracted outputs. One of our main evaluation areas in ExtractBench, our recently released real-world benchmark, is to evaluate the visual grounding capabilities of document extraction. ExtractBench: https:// extractbench.ai If you want to give LlamaParse a shot, come check it out: https:// cloud.llamaindex.ai

Something super exciting happened quietly on HF over the past month: AI agents became AI builders, and they did it in the open! During our ICML reproduction challenge, 1,221 humans teamed up with coding agents to verify and reproduce 2,226 papers. But here's the cool part: everything happened on the @huggingface hub: 6,816 reproduction logbooks published openly, 2,962 cloud jobs launched, 35,908 claims judged, all traceable, all public and transparent For years the hub has been where humans collaborate on models, datasets and demos. Now we watch agents use it the same way: writing logbooks, publishing results, building on each other's work. Closed labs run evals behind closed doors and ask you to trust the press release. Open science means anyone can check the receipts and now agents can too! The next million users of the hub might not be human. and that might be the best thing to ever happen to science! Full write-up about the hackathon: https:// huggingface.co/blog/icml-2026 -open-reproductions …

We've just surpassed 3 million models on the Hub the community is accelerating towards an open, distributed future where open AI is everywhere, for everyone

LLMs are weird. Looking at various blogs and comments, the people who are most vociferously unwilling to entertain the idea of distortion-free watermarking seem to have a mental model in which the LLM generates the best possible output sequence and then the watermarking algorithm edits it. Except that’s not how it works at all. During normal generation, many of the tokens that an LLM picks _aren’t_ the one it considers the most likely next token. If you turned off the randomness, the first few tokens of any response would be great but then the output would eventually get stuck in loops and other degenerative behaviors. Humans don’t speak or write one word at a time, so this locally-optimal-but-globally-incoherent behavior is very counterintuitive. If you analyzed unwatermarked / watermarked pairs of outputs with regard to some definition of quality, the watermarked output would indeed be worse — about half the time. The other half the time it would be better! That’s the sense in which watermarking doesn’t degrade quality. It is nonsensical to ask whether the quality of a specific unwatermarked output is degraded by watermarking. The algorithm doesn’t start from a specific output and modify it; instead it changes the sampling process. People seem to have a lot of trouble with the idea that it is not meaningful to compare pairs of outputs, only probability distributions. Finally, if some generated outputs are better than others, why not generate multiple outputs (watermarked or not) and only output the best one? You could! Repeated sampling is a valid and widely used way of improving quality (but it multiplies the cost). Unfortunately it only works to a small degree, except in narrow domains like code, because models tend to be only slightly better at assessing quality than at generating text.

This is a great example to understand how it works. How can the watermarked and unwatermarked distributions be indistinguishable if there's only one correct answer? (In the extreme case, suppose we ask the LLM to output an exact piece of text.) The watermark doesn't introduce new randomness; it hides in the randomness that already existed in the distribution of responses to a given prompt. If the output distribution for a given prompt doesn't have enough variation, there will simply be no detectable watermark. This is the reason why outputs need to be a certain length to carry watermarks. It's not actually about the length, it's about the amount of randomness, which simply correlates with length. In extreme cases, there can be long outputs with little or no randomness.

ChatGPT for Teens adds age-appropriate safety measures, parental controls, and learning tools designed to steer teens away from harmful content — and from using AI to cheat on their homework.

Apple left a demo video of camera-equipped AirPods in the macOS 26.7 Tahoe release candidate, showing Siri using Visual Intelligence to identify what the wearer is looking at. Code in the same release references AirPods Pro 4, AirPods 5, and an unannounced operating system codenamed Pebble for future smart-home devices.

A video in a MacOS Tahoe release candidate version shows a user wearing AirPods, looking at a book, and talking to Siri.

LG AI Research has successfully advanced to the next phase of Korea’s national AI foundation project, proudly representing the country on the global stage. This milestone is another step forward—and the journey continues. LG AI Research is now fully focused on developing our third-stage model. Try K-EXAONE 2.0 : https:// huggingface.co/LGAI-EXAONE/K- EXAONE-2.0-750B-A37B … #LG_AI_Research #EXAONE #K-EXAONE #LG_AI #KR_AI #LLM

Alibaba Cloud is expanding its AI infrastructure in South Korea. We’ve launched our 3rd data center in South Korea, strengthening local infrastructure for enterprises accelerating AI adoption. Built for: • Robust, scalable & secure cloud infrastructure • Low-latency, highly available workloads • Enterprise AI agents, from development to deployment and security • Local data privacy & sovereignty The new facility brings Alibaba Cloud’s global network to 104 availability zones across 30 regions. And we’re just getting started. https:// click.alibabacloud.com/m/20000002424/ #AlibabaCloud #AIInfrastructure #SouthKorea #AgenticAI #CloudComputing