Astra 에이전트 실행, 문체 개인화, 추론 메모리 최적화와 Physical AI 데이터 병목

This is insane! GPT-6 Astra built this beautiful math animation in one go! (🔉 sound on) "Jaw-on-the-floor" moment. I've not been able to get anything close to this with previous AI models. True personalized learning has arrived!

Teknium 🪽

@Teknium

10시간 전

Okay hermes only does #2 out of those 3 Our subagents can't talk to each other, can only submit final work to the orchestrator, and context is fresh, provided by the orchestrator + skill refs for what the orchestrator had loaded going into it's creation

ChatGPT

@ChatGPT

8시간 전

ChatGPT Work can now pick up on what makes your writing sound like… you. Your favorite phrases. Your very specific sign-off. your capitalizations quirks. Connect the tools you use every day, like Gmail, Google Drive, Slack, and SharePoint, and ChatGPT Work will learn your writing style from your emails, messages, and files—and carry it into whatever you write next.

ChatGPT

@ChatGPT

8시간 전

Ready to sound more like yourself? To get started, go to https:// chatgpt.com/?surface=work& writing_style=setup … on web or you can also set it up in Settings → Personalization → Writing style and it will apply to all your messages across ChatGPT Work on web and mobile. Available on web for all paid plans with access to ChatGPT Work.

PagedAttention, clearly explained. (how vLLM manages KV cache like an operating system) every request a serving engine handles needs GPU memory for its KV cache, the key and value vectors stored per token so later tokens can attend over the prompt without recomputing it. the engine has no way to know how many tokens the request will generate. the simple approach is to reserve one contiguous slab per request based sized for the maximum allowed length (max_token_size). for example, a request with max_token_size = 4,096 tokens gets 4,096 slots the moment it arrives and holds them until it finishes, even if it stops at 40. the left side of the graphic below shows what that does to memory. → a large part of every slab is reserved but unused. those slots belong to one request until it finishes, and most of them never get written. → the grey pieces between slabs are fragmentation gaps. as requests of different sizes come and go, the free memory breaks into fragments that are each too small for a new slab, even when the total free space would fit several requests. a new request needs a contiguous slab, none of the gaps fit one, and the reserved slots cannot be taken back, so it waits while the GPU holds mostly empty space. the vLLM team measured 60 to 80% of KV cache memory wasted this way in existing serving systems. they introduced PagedAttention which is inspired from how operating system uses RAM. instead of one slab per request, the KV cache is handed out in blocks. a block is a fixed-size chunk of GPU memory that holds the keys and values for a fixed number of consecutive tokens, in vLLM the default value is 16. inside it sits the full KV state for those 16 tokens across every layer and every KV head. at startup the engine measures the GPU memory left after loading the weights and carves all of it into a pool of numbered physical blocks. that pool is the grid in the graphic below, and it is the only KV memory that will ever exist. the bottom row shows how a request uses that pool. → a request arrives and its sequence is split into logical blocks, L1 for tokens 1 to 16, L2 for 17 to 32, and so on. a logical block is only an index into the sequence, it holds nothing itself. → each logical block is assigned a physical block from the pool and the pairing is recorded in the request's block table. the assignment happens only when the previous block fills, so no request holds memory it has not written yet. → at each decode step the attention kernel reads the block table and gathers keys and values from whichever physical blocks they landed in. a request's tokens no longer need to be next to each other in memory, which is why consecutive logical blocks can sit at scattered grid positions. every physical block is the same size, so any free one serves any request and fragmentation gaps stop forming. when a request finishes, its block numbers go straight back to the pool for the next request to take. waste can never be more than the unfilled tail of a single 16-token block, which is how it drops under 4%. the reclaimed memory becomes batch size, meaning more requests generating at the same time on one GPU. that is where the 2 to 4x throughput gain comes from. each generation step is limited by how fast the model weights can be read from memory, and a bigger batch spreads that read across more tokens. the model itself is untouched. whatever attention variant it was trained with runs the same way under PagedAttention, because the change lives entirely in how the serving engine hands out memory. i wrote the full breakdown of every attention mechanism, from multi-head attention through FlashAttention and sparse attention, up to PagedAttention and RadixAttention. the article is quoted below.

Brilliant paper from LinkedIn. (bookmark it) I have been saying that memory is one of the most challenging things to get right when building agents. It's an are you want to invest time in optimizing. This paper focuses on whether memory is portable, which is crucial given that more users are using agents across models, agents, and providers. If you have ever upgraded the model behind an agent and watched it start forgetting things it used to know, this one is for you. In this work, the same agent history is stored four ways, kept verbatim for long context reading, chunked for retrieval, compressed into model-written notes, and normalized into a fixed-schema knowledge graph. Then the model reading it gets swapped. Fixed-schema knowledge graphs barely move, 0.0004 accuracy change across a writer swap. Model-written notes move by +9.91 or -13.28 points depending on which direction you migrate, so the same pair of models gives you opposite results based on which one wrote the memory. Partial embedding migration is also expensive. A 50/50 mixed index captures only 4.96 of the 11.90 points that full re-embedding gives you. The two failure modes have different causes. 80% of the notes deficit comes from information lost when the note was written. 81% of the retrieval deficit comes from retrieval missing the right chunk. Repair needs the raw history. Store-only repair of notes failed to reach 90% recovery in all 48 test cases, while keeping the source history recovered 34 of 48. Paper: https:// academy.dair.ai/papers/does-yo ur-agents-memory-survive-a-model-upgrade-a-controlled-study-of-memory-por-2609.05339 …

There are chapters of the humanoid race you can't buy with unlimited money: the deep engineering of the hardware and designing the right AI recipe And then there are chapters money can buy. Figure has entered the chapter of scaling intelligence through data and compute

“Qwen-Drive-1.0” Qwen’s first autonomous driving foundation model! Most driving VLMs learn to talk about scenes, but not explicitly model 3D geometry or plan motion. Qwen-Drive-1.0 unifies 3D perception, driving VQA, and trajectory planning on top of a pretrained VLM, using a BEV perception head and flow-based Planning Expert. So they created one model that can understand the scene, reason about it, and drive, while largely preserving general vision-language ability. https:// alphaxiv.org/abs/2609.00111

Guillermo Rauch

@rauchg

10시간 전

The case for *total* Linux victory (desktop, mobile, server, iot, robots, embedded…): 1) Most agent-programmable People will want fully personalized computing experiences. macOS makes it hard to even ship a tiny app to your menubar. 2) Open source training data Blender beats all its competitors because it's optimal in the most important dimension: agents can train on it and use it efficiently. OSS wins. 3) Design being democratized macOS has historically had a 'monopoly' on beautiful OS design. The global design community is now huge, they're motivated, they have tokens. 4) Computing moving to cloud sandboxes More and more of our personal and professional computing experiences will come from agents that orchestrate work in the cloud, using sandboxes. Those sandboxes will be Linux. Bonus: seamless continuum between local & remote 5) Security Every piece of software is now at major risk, Linux included. But Linux has the collective investment of everyone on the planet throwing tokens at it.

Guillermo Rauch

@rauchg

10시간 전

"Where do you want to go today?" In 1994 Microsoft launched an ad campaign under this motto. I think it represented the potential of computing very well, but it was quite aspirational. The internet & www was in its infancy, software was hard to obtain (CDs and long downloads), programming your own computer was tricky. In fact, it was so tricky I switched to Linux in order to teach myself how to code and get true 'root' on my PC 😁 Now, I look at what AI is enabling, the limits on what you can ship vanishing, and the new worlds being created, and I realize that question is perfect for 2026.

Tired of choice overload? With so many AI models available, it’s easy to just default to the most capable (and most expensive) option for every coding task. Smart Routing in Unity Gateway takes the guesswork out by automatically matching your tasks to the right model based on complexity. It uses lightweight models for simple edits and reserves frontier models for the heavy lifting. Paired with Omnigent, it automatically matches your task with the best model and coding harness so developers can stay focused on writing code. High-quality results. 30%+ lower task costs. https:// databricks.com/blog/smart-rou ting-unity-ai-gateway-match-frontier-quality-30-lower-cost-task?utm_source=twitter&utm_medium=organic-social …