side job of the day: making fully monitored coffees at #AWSreInvent making developer dreams come true 😂 PS: dashboards are public on ela.st/beans-streams

hierarchy of (LLM coding( leverage): how CLAUDE md > specification > research > plan > code and how to write a good CLAUDE md — but keep it short www.humanlayer.dev/blog/writing...

But How Do AI Images and Videos Actually Work? diffusion models, CLIP, and the papers + math behind it — great weekend video: www.youtube.com/watch?v=iv-5...

the weekend #ElkyFriday PS: don’t worry, not going to start a dedicated elky channel 😅

european bureaucracy: it "only" took 2.5 years to get a delayed flight paid out. and after all the costs (including a laywer) you only get 50%. but I'll still take the win 😉 PS: and I have an even older claim still under negotiations 🫠

what's the story with langchain vs langgraph? I thought the focus was switching over to langgraph but the contributions for langchain seem to have taken over again

how to end the week in the office: give elkie a belly rub

if you're in SF on monday, we'll have a hardware oriented meetup together with microsoft (right at the start of #MSIgnite): * Bad AI In A Box (BAIIAB) — making AI touch grass * monitor our coffee consumption — live debug coffee from across the world. and we'll hand out excellent coffee as a demo

while we all love ourselves some ML, sometimes all you need is a little bit of math. great post from one of my colleagues on how to tune your shopping results based on history (with a function score) alexmarquardt.com/personalizin...

#KubeCon is on who is around and wants to meet this week?

got the right "business" cards for #KubeCon sorry, no YAML 😉 this was more fun to build than expected: @claudeai generated HTML design with a *download as image* button. @MOO to print and ship them github.com/xeraa/code-c... for the implementation

fun #TLS fact: why would you want your own certificate authority (CA) over a public one? if you form a cluster (like elasticsearch), you want to control who can join 1/3

the #elasticsearch API docs always had cURL and console examples. a while ago we added dynamic languages (python, JS, ruby, PHP) now, we have the first static language with java big API surface so this will come in handy. even if just to tell an LLM "no, use the API like this" 😉

#CalHacks was a crazy weekend. great projects with @elastic.co agent builder, RAG, geo and timeseries data: lnkd.in/dJ6H5cAk PS: we also served a ton of (monitored) coffee. but that‘s for a separate post…

AWS outage / post-mortem: not one but two follow-up problems. and of course it started with DNS 😬 aws.amazon.com/message/101925/

#ElasticON closing keynote by @swyx: the year agents ate search from RAG, agentic RAG, and context engineering subagents 🔮

how much data can you get from coffee? a lot ☕️ this has been a fun project — from sending the data with OTLP to watching every step. can‘t wait to bring it to events and show streams of logs to dashboards

next week will be extra @elastic.co-packed in SF monday: luma.com/smart-search * ghita, @zeroentropyai.bsky.social: search tools for efficient AI agents * jesse, fintools: LLMs and the next generation of financial search * josh, reducto: vision-first RAG pipeline with reducto and elasticsearch

reddit answers — and I‘m not sure if they are trying to troll me or not: wooden data centers 😅 sounds like a logstash situation too

and we got brand new toy. I hope that we can roll out coffee and related dashboards next week ☕️ #ElasticON

we'll have #ElasticON SF next week. I'm especially excited to have convinced @swyx.io to join us for the closing keynote on "the year agents ate search" — let's see how much we can all agree or disagree here :D www.elastic.co/events/elast... (and we still have a couple of free slots available)

AI gateways: envoy + elastic observability from adrian cole (we‘re going back to the zipkin days) #ElasticON starting off with: what is agentic?

how do you get recommendations on tinder? a dee dive into the implementation of retrieval and ranking for millions of users and likes, billions of messages and swipes #ElasticON

kicking off #ElasticON NYC. I‘ll put the most relevant context (😉) in this thread — it will be a busy day starting with joining forces with @jina.ai and @elastic.co www.elastic.co/blog/elastic... 🤗

so, vibe coding is over. now it‘s vibe engineering? though it‘s probably time to move beyond YOLO coding (for production) simonwillison.net/2025/Oct/7/v...

bringing more monitoring love to your self-managed / on-prem #elasticsearch clusters with AutoOps: pull stats with an OTel based agent, get recommendations and issue-detection in real-time PS: and soon inference on GPUs as well (so you don't have to go to OpenAI etc any more) 1/3

the joys of open source... also, poor everyone org on github 😅

wish that #reddit quotations in LLMs would stay low, so we could just have conversations and not fight all the spam and bots not sure how accurate the data is but interesting in any case: promptwatch.com/data/top-cha...

how to avoid deadlocks if you run 300 #elasticsearch clusters like palantir. and you pick the trickiest combination of configuration + client calls: * concurrent synchronous refresh * wait_for with long refresh intervals

if virtual hackathons are your thing, we might have something there (with decent 💸): ai-accelerate.devpost.com and with the new agent builder features (built-in LLM, queries as tools, native MCP,...) there are some useful new toys available for it too

new MCP registry from @github.com: github.blog/ai-and-ml/gi... looking at the current npm security disaster, we definitely need more trusted sources for dependencies 😬

failure catalog of AI apps — not sure I've seen it spelt out that way before not sure what to think of the product behind it but the list and ways to approach the problem look pretty useful: github.com/onestardao/W...

build an intelligent query layer with MCP in #elasticsearch — from search to agent here's a preview how ES|QL queries turn into tools that get aggregated into agents — all served and built right in kibana. and you can try this out today

TIL how to * subscribe to our release notes or * the @elastic integrations (which are independent of stack releases) since two people asked a very similar question today. it's actually 3 simple steps 1/3

(re)ranking — going from hundreds retrieved results down to the few actually shown / used great to see that elastic reranking is also doing well in external benchmarks ;) google post: cloud.google.com/blog/product... details on elastic reranking: www.elastic.co/docs/explore...

the enshittification of meetup.com continues (until everyone migrated to lu.ma) * besides organizers paying for groups, upselling attendees with a "flash sale". we got some confused attendees how we are making them pay for meetups 😵 * huge email ads even for events on the pro plan. how pro-fessional

awesome to see mcp-use on product hunt today! for everyone needing to build with or around MCP: www.producthunt.com/products/mcp... and yes, we'll do a small hack night with them next week to put that to the test as well 😉 lu.ma/fqdql7kh

I recently joined Time4AI for a 20min chat, discussing how @elastic.co's #elasticsearch and #kibana are transforming information retrieval for search, logs, observability, and security 1/2

#elasticsearch REST API in .md — in case you need to point your LLM to the latest version and features it's been long overdue :) www.elastic.co/docs/referen... vs www.elastic.co/docs/referen...

#elasticsearch 9.1 / 8.19 release: continuing the optimizations, also for time-series — LogsDB and TSDS (time-series datastream) storage is commonly the most painful part, especially with longer retentions. but we're getting to 2x to 4x improvements over default indices 1/6

#elasticsearch 9.1 / 8.19 release 🥳: how to improve ranking, latency & cost for dense vector search at the same time — by defaulting to better binary quantization (BBQ). even if improving all 3 is counter intuitive quick overview of what's new, focus on semantic_text but dense_vector is similar 1/8

thread + article on #elasticsearch alerting options: * kibana alerts if you need something simple * watcher for way more options watcher example for: "alert when users from the US, Germany, and India dropped below 1000 in the last hour compared to yesterday" 1/2

some cool tricks here to reduce your vector search cost by 85% while improving relevance: * quantization (and this was before our binary quantization BBQ which is also mentioned) * optimized settings for vector search * selectively using vector search — skipping very old or rarely used documents

today is one of my favorite days: release day! #elasticsearch and the entire @elastic.co stack 9.1 + 8.19 are out. lots of great stuff there — I'll pick one feature a day to dive into 🔬 but first, versions + updates: 2 majors in parallel? what about 7.x? how to upgrade? what's even maintained today?

AI agent pitfalls: * compounding errors over multiple steps * quadratic token cost over many steps (though I didn't fully get why you'd always need to keep all previous context — sometimes sure but always?) * the tool will not "just figure out" what to always do over many steps

still running into some of the most basic hallucinations / fabrications. I thought we were past this by now?

"Open Data Standards: Postgres, OTel, and Iceberg" some great points: * how "commercially unencumbered and vendor neutral" is the foundation of becoming an open standard * standards over tools (the wire protocol of postgres is kind of that standard) * s3 as the ultimate data infrastructure

more elasticsearch MCP experiments: * waterfall trace performance analysis (attached video) * root cause analysis * performance analysis by instrumented service moving observability from kibana to claude or any other MCP

when you have @sessionize.com, every community topic looks like a CfP #community #DevRel that's why we have 3 different calls running on it for @elastic.co right now — conference, meetups, and champions

the walls around the data gardens are going up: 1 slack request per minute with a maximum of 15 objects coming back 🐌 api.slack.com/changelog/20...