Claude Code 사용 한도 조정, 에이전트 인프라 측정, 실시간 영상 생성과 장기 기억 연구

ClaudeDevs

@ClaudeDevs

1일 전

Starting Sept 14, we're permanently raising standard weekly limits by 25% for Pro, Max, Team, and seat-based Enterprise plans. Until then, the current 50% increase will stay in place until Sept 13.

ClaudeDevs

@ClaudeDevs

1일 전

Compared to today, this works out to a 17% reduction in weekly limits on Claude Code. We’re working on exciting changes that will make it feel like you’re getting more from Claude, while having more visibility and control of your usage. Can’t wait to share them.

Are you measuring AI infrastructure performance on how agents actually run? @SemiAnalysis_ built AgentX to do just that. Vera Rubin NVL72 performance measured by NVIDIA on the SemiAnalysis AgentX workload shows up to 30x better throughput per megawatt than GB300 NVL72. Learn more https:// nvda.ws/4x2WyXX

MiniMax (official)

@MiniMax_AI

1일 전

@rehan_shei may have just invented the best benchmark yet: can it generate faster than Twitch can stream? Incredible to see @fal pushing MiniMax H3 into entirely new experiences.

MiniMax H3 Max renders a 5-second clip in under 3 seconds. You wired MiniMax H3 Max into a live stream so it never stops. You can now watch infinite AI television a 24/7 stream

Banger paper from Apple. If you build MCP servers, this can help you turn your specification into an evaluation suite. (bookmark it) It's actually a really neat idea that's easy to implement. And it showcases the awesomeness of MCP. Agent Seer starts from a single MCP spec and synthesizes multi-turn agent test scenarios with no examples, no live tool access, and no domain-specific tuning. Function names, natural-language descriptions and typed parameter schemas already carry enough semantics to generate graded scenarios with synthetic tool outputs, which then expand into mock-data-grounded dialogues. Hand-built agent benchmarks demand deep domain expertise, do not scale across tool ecosystems, and go stale as soon as an API changes. Generating them from the live spec keeps pace with the ecosystem instead. They ran it on seven MCP specifications spanning different domains and suite sizes, with complete tool coverage on small and medium specs. Parameter schema complexity predicts quality variation far better than tool-suite size does. And argument value accuracy is the dominant failure mode, a sub-dimension that coarse name-match tool-calling metrics cannot see at all. Paper: https:// arxiv.org/abs/2608.26133 Chat with Paper: https:// academy.dair.ai/papers/agent-s eer-synthesizing-scenarios-from-specification-understanding-2608.26133 …

Agents reach websites in 6 ways. (Google and Microsoft are betting on the same one.) Ask an agent to buy something and watch it work. It screenshots the page, looks for something shaped like a button, clicks, and screenshots again to see what happened. You pay tokens every time it looks, and one redesign breaks the run. The site already knows what it can do. It has a search, a cart, a checkout, and none of that is written down anywhere a program can read. Here are the six, ordered from furthest from the interface to closest. 1) Raw API. Your script calls the backend with a key you manage. Precise and fast, though you had to find the endpoints yourself and the website never enters the picture. 2) Backend MCP server. The company describes its actions as named tools and your agent connects to them. Someone who understands the product wrote those definitions, so they name real actions instead of guessed ones. The interface is still skipped. 3) Computer use. The agent sees the live page as an image and clicks around it. Nothing to configure, but every look costs money and a layout change confuses it. 4) Browser automation over the DOM. The agent reads the page's underlying code instead of a picture of it. More reliable than pixels, except the tooling is generic, so the agent still infers meaning from anonymous divs and unlabeled buttons. 5) WebMCP, the one Chrome and Edge are building together. The page declares its own actions with names, plain-English descriptions, and typed inputs, and your agent calls them inside your live browser session. 6) The site's built-in assistant. The company ships a chat box, picks the model, and pays for the tokens. It works, and it is not your agent, so nothing it learns about you carries anywhere else. Three things vary across the six. Whose agent does the work, what you configure before anything happens, and what the agent receives when it arrives. Every option except one gives up at least one of the three. Raw API and backend MCP hand you typed actions and remove the website. Computer use asks for no setup and hands over pixels. The built-in assistant is precise and belongs to the vendor. WebMCP is the only one that keeps all three. Your own agent, nothing to configure, and named actions instead of guesswork. I wrote the full breakdown. The article is quoted below.

// Memory is what breaks long-horizon agents // It's undeniable how important memory/recall is for long-horizon tasks. If you are building for long-horizon tasks, this is a great read. (bookmark it) They set up an LLM agent to run a football club for 20 in-game years, through 26 tools and roughly 340 to 400 decision stops, scored by a deterministic engine with no LLM judge anywhere in the loop. Results: All 15 frontier models survive every horizon while the scripted baselines mostly die out. Neither scale, price, vendor, nor token spend predicts the ranking, and the order only settles late in the run. What separates the top models is managerial behavior, cutting slow-payoff investment near the end and opening contract renewals well before the deadline. Two universal failures were found. No model learns the market's hidden prices from hundreds of rejected bids, and self-managed memory collapses into either an archive that only grows or a plan rewritten every season. Paper: https:// arxiv.org/abs/2608.18423 Chat with Paper: https:// academy.dair.ai/papers/fm-benc h-a-benchmark-for-long-horizon-management-with-competing-agents-2608.18423 …

Deep RNN by hand ~ 12 steps walkthrough below A Recurrent Neural Network reads a sequence one step at a time, carrying a hidden state forward. Stack three of those hidden layers on top of each other and you have a Deep RNN: deep in time, and deep in layers. I drew and calculated one entirely by hand. Goal: run a three-layer RNN over a sequence of four inputs, carrying every hidden state forward yourself. 1. Given Four inputs X1 to X4, recurrent weights and biases for hidden layers a, b, c, and an output layer y. 2. Initialize Let us set the hidden states a0, b0, c0 to zeros. Nothing has been read yet. 3. First hidden layer (a) We build the transformation matrix by laying the input weights, the state weights and the biases side by side. We stack X1, the previous state a0, and an extra 1 underneath. Multiply the two, and a1 = [0, 1]. 4. Second hidden layer (b) Let us do it again, one layer up. Now a1 is the input, and b0 is the previous state. Multiply: b1 = [1, -1]. 5. Third hidden layer (c) Once more. b1 is the input, c0 is the previous state, and c1 = [1, 1]. 6. Output layer (y) Let us read the answer off the top of the stack. Weights and biases against [c1; 1], and Y1 = [3, 0, 3]. 7. Carry the states forward We copy a1, b1, c1 across. This is the whole trick of a recurrent network: the states are the only thing the next input gets to see. 8. Process X2 Repeat steps 3 to 6 for the second input: three hidden layers, then the output. Y2 = [5, 0, 4]. 9. Carry the states forward Let us copy a2, b2, c2 across, exactly as before. 10. Process X3 Same four moves, third input. Y3 = [13, -1, 9]. 11. Carry the states forward We copy a3, b3, c3 across, one last time. 12. Process X4 Repeat once more. Y4 = [15, 7, 2]. You have just run a Deep RNN over a whole sequence by hand. The outputs: Y1: [3, 0, 3] Y2: [5, 0, 4] Y3: [13, -1, 9] Y4: [15, 7, 2] The takeaway: the hidden states are the memory, and they are the only memory there is. Everything the network learns from X1 has to fit in those little two-cell columns and get handed forward, one step at a time. In my class I have students trace this with a finger. Follow the edges and you can see it plainly: where each memory cell is stored, how it gets passed to the next step, and how it shapes the encoding of the next input. Then count the hops X1's information has to make before it reaches the output. That is why long sequences forget. I also have them trace it backward, which is where you get a feel for backpropagation through time. Now imagine every one of those hops is a multiplication of the gradient. Chain enough of them together and the number runs away from you: a little above 1 and it explodes, a little below 1 and it vanishes. Save this post! #AIbyHand #RNN #DeepLearning

Hy4 preview runs in vLLM

You can run locally Qwen3.8-27B-Uncensored for cyber use it doesn’t refuse - it already 1M+ downloads. - Image + text. - Feed it a photo Get a direct answer. - Zero-filter replies. - FastMTP for up to 3x faster generation. If you want raw multimodal output locally, this is the one. - http:// huggingface.co/HauhauCS/Qwen3 .8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF …

The old card is faster on Qwen3.8-27B generation. A 9-year-old $650 eBay V100 just outperforming a 2025 Strix Halo on Qwen3.8-27B decode. - 56 tokens/sec. Price-to-performance is wild right now.

Ready to build models that support knee MRI interpretation? Check out this starter notebook by Pilkwang Kim. It is a great starting point to read the DICOM acquisitions, sample and normalise the MRI slices, and turn a pretrained vision backbone into twelve abnormality predictions.