Joejo Jiang @Joejojiang

Long Horizon Transformer Quantile Fault Prediction for Multi Site Industrial Predictive Maintenance David J Poland, Daniele Ravi, Na Helian arxiv.org/abs/2609.04840 [𝚌𝚜.𝙰𝙸]

The future of medical AI will require clinically grounded datasets, shared benchmarks, multimodal learning, and annotation systems that make expert knowledge easier to capture. Read the editorial by @ToddCHollon, MD: nejm.ai/44Nz5hT

Medical knowledge graphs may become MORE valuable in the LLM era — not less.

Four caches in LLM serving, clearly explained: Every LLM request reads the whole prompt and computes attention state for every token in it. This step is called prefill, and it impacts both the input bill and the time before the first token appears. In an agent loop, most of the prompt comprises text that the model already processed in the previous turn. There are four cache layers that prevent paying for the processed tokens at each turn. ↳ The KV cache holds the key and value tensors for every token at every layer, for one active request. ↳ Prefix caching keeps those tensors on the server instead. vLLM stores them in 16-token blocks and identifies each block by a hash that chains in the previous block's hash, so a block only matches if everything before it matched too. The scheduler stops at the first miss and prefills the suffix from there. ↳ Prompt caching is the same reuse that runs on a provider's hardware, with a price sheet attached. Anthropic charges 1.25x the base input rate to write an entry and 0.1x to read it. ↳ Semantic caching works differently. It embeds the incoming prompt, runs a similarity search over stored prompts, and returns a stored answer outright when the score is above a threshold. That's why it saves output tokens as well as input tokens. It's also why every request pays for an embedding round trip, including every miss. The first three match on exact tokens and cannot change what the model produces. This technique matches on similarity, which means it is quite susceptible to generating a wrong response since embeddings may match to a wrong prompt. The diagram below depicts all these techniques. To use these techniques, you don't need to build a custom serving stack. The transformers library already implements the cache as an object of KV vectors that you can preserve, so you can prefill a corpus once, retain the returned tensors, and reuse them across queries in about ten lines. And this KV cache is only one of four separate caching layers in an LLM stack. The other three are prefix caching on the server, prompt caching billed by a provider, and a semantic cache that skips the model entirely. I wrote a full breakdown of all four caches in LLM serving that you should know as an AI engineer, with code for each. Read it below.

Avi Chawla @_avichawla

2 weeks ago

x.com/i/article/2093…

23 177 1K 1.3M 2K

Natural-language input → The LLM acts as the Agent’s brain to understand intent → It autonomously selects and invokes the right tools, such as Cypher graph queries, vector-search APIs, or ranking algorithms → Processes the results → Generates a personalized response.

Avi Chawla @_avichawla

a week ago

Four caches in LLM serving, clearly explained: Every LLM request reads the whole prompt and computes attention state for every token in it. This step is called prefill, and it impacts both the input bill and the time before the first token appears. In an agent loop, most of

29 167 787 93K 1K

I’m still very curious about this: more and more clinical decision support systems are being adopted.

Irene Chen @irenetrampoline

a week ago

I get older, but my students stay the same age! And the buildings get newer. I'm teaching CPH100 again, an undergrad class tackling ML for healthcare. Follow along at irenechen.net/cph100 Our first class covered why we still have radiologists despite AI predictions in 2016 🤔

keyboard_arrow_left Previous keyboard_arrow_right Next

3 16 188 48K 135

NEW Health Policy - Recommendations for a national electronic health record in Spain: a Delphi study. Read it here: thelancet.com/journals/landi…

Real-world clinical settings are often more complex than we imagine, and defining meaningful tasks requires a solid methodological foundation. It is challenging, and I am still exploring the best way forward.

Stanford Center for Digital Health @StanfordCDH

2 weeks ago

Most LLM benchmarks miss the complexities of a real clinical practice. BRIDGE, from CDH-funded researcher Jonathan Chen and co-authors, introduces a multilingual evaluation framework built from 87 clinical tasks across 14 specialties. Read the study: nature.com/articles/s4155…

0 1 1 153 1

Find a one-click Instagram-style “plog” creation tool that uses computer vision. First, search GitHub to see whether such functionality already exists.

Anjula Dwivedi @HeyAnjula

a week ago

You installed Claude Code and stopped there. Here are 22 skills (actually) worth adding: If you're just getting started, check out these guides: Every one is a GitHub repo. Grouped by what they do: BUILD superpowers — plans and debugs before it writes Install →

8 62 346 21K 461

You installed Claude Code and stopped there. Here are 22 skills (actually) worth adding: If you're just getting started, check out these guides: Every one is a GitHub repo. Grouped by what they do: BUILD superpowers — plans and debugs before it writes Install → lnkd.in/eppbgRaK (259k★) gstack — 23 skills in one setup Install → lnkd.in/e8PYQ2z6 (124k★) caveman — no-fluff answers, no padding Install → lnkd.in/e4nxpEJi (92k★) ponytail — says nothing, writes one line Install → lnkd.in/eYzPcnAR (88k★) codex — runs Codex inside Claude Code Install → lnkd.in/eTweEPmw (30k★) i-have-adhd — stops it burying the answer Install → lnkd.in/eQEJZeMN (9k★) DESIGN ui-ux-pro-max — palettes and fonts that work Install → lnkd.in/eh5s-T6t (109k★) taste-skill — makes a page look expensive Install → lnkd.in/emKJrHrX (66k★) impeccable — the polish pass on any front end Install → lnkd.in/eMSX587R (49k★) hyperframes — turns a web page into a video Install → lnkd.in/ed-wSdsx (37k★) emil — beautiful interface animations Install → lnkd.in/eWebzpnh (20k★) gsap — smooth web animations Install → lnkd.in/e6avMtKZ (12k★) RESEARCH skill-creator — build your own skills Install → lnkd.in/er2tG4ZB (163k★) graphify — turns notes into a map Install → lnkd.in/eT9f97nW (94k★) last30days — what the internet said this month Install → lnkd.in/e6uRrEnT (53k★) agent-browser — lets Claude use any website Install → lnkd.in/eUS4cxjs (39k★) find-skills — finds any skill for you Install → lnkd.in/eFcmubhq (27k★) claude-hud — shows what Claude is doing live Install → lnkd.in/ecU6_AC9 (27k★) MARKETING remotion — build full videos in React Install → lnkd.in/ecX3gdvr (54k★) marketingskills — 40 marketing skills in one pack Install → lnkd.in/egt-7ZwM (41k★) humanizer — strips the AI tells from your writing Install → lnkd.in/eekWNVYm (31k★) social-media-skills — my own content pack, free Install → lnkd.in/emDvetxm (mine) Most install with npx skills add [repo]. I don't reach for all 22 daily. superpowers, caveman and last30days get the most use. Repost ♻️ to help someone in your network. P.S. Which one are you installing first?

Why does clinical AI still optimize for benchmark accuracy when the real bottleneck is whether patients can actually use it at home?

以结果为导向会带来什么？ 高效的决策？ 真正改变世界的能力是什么 一些思考