Tristan Handy

#dbt Lab社のTristan CEOにご来社いただき、とても良いディスカッションができました。dbtはずっとお世話になっているけれども、今後の拡張も欲しかった機能ばかりで楽しみです。引き続きAI-oriented Data Platform頑張っていきます💪

Your AI Agents can only analyze 1GB of data Enterprises have 1,000,000x as much data We fixed that... It took 3 years and building an entirely new execution environment to pull this off. TextQL lets AI agents analyze your ENTIRE enterprise data - not just the 0.0001% they can handle today. Here's what makes it so different: 1. ALL your data, not samples. AI today: 1GB demos. Your reality: 350TB across 47 systems. TextQL: Everything. At once. 2. Real enterprise questions.ChatGPT writes poems. We answer why renewal rates dropped 3% after ERP upgrades. 3. No more sandbox limitations.Everyone else: Jupyter notebook with 10GB RAM. Us: A data center. ------------------------------ In 2021, I watched Fortune 500s spend millions on AI that could only analyze data that fits on a USB stick. Today: - We process 10+ petabytes daily - Our biggest customer analyzes 100TB in real-time - Average query spans 50+ data sources And we just proved what everyone said was impossible - AI that can actually build a digital twin of your enterprise ------------------------------ To celebrate, we're giving away free data architecture assessments. We'll show you exactly how much of your data is invisible to AI right now (spoiler: it's probably 99.9%). Comment "TextQL" and we'll send you the assessment link.

Introducing the dbt MCP Server! dbt is the standard for creating high trust datasets. AI workflows need access to high trust data, including structured data. The dbt MCP server is a meaningful step towards connecting AI systems to your data.

I think AI is going to usher in a gold age of infra, not obviate it. It's just so clear that good CS fundamentals result in better AI built systems. Vibe coding works better with type safety, languages where syntax maps closely to semantics, referential transparency, tight scoping etc. These approaches have never been widely adopted in CS because they are hard for humans, and in particular novices. But they're not hard for AI. And they map so much better from natural language descriptions.

To MCP or not to MCP, that's the question. Lmk in comments

Thrilled to share that we've closed $28M in funding, led by @CRV, with continued support from @wing_vc and @saranormous. Also excited to onboard strategic partners @SnowflakeDB and @databricks! voyage.ai Building the world’s best models for RAG and search 🧵🧵🧵:

We are excited to announce our $17M Series A led by @glennsolomon at @notablecap with participation from @whoisnnamdi and @lsvp, @datadoghq, @matthartman, @jefrankle, @jasoncwarner, @jthandy, @BM_DataDowntime, @aparnabsinha, @NadimHossain and a number of tech industry executives 🚀🚀🚀

We're live! @pydantic Logfire is in open beta. Since I started writing Python in 2010, I've wanted a better way to do logging. Off the back of Pydantic's unbelievable growth, last year I started a company backed by @sequoia, and lucky enough to hire the brightest people I know. Now we've gone and built the logging thing I always wanted. pydantic.dev/logfire

“Community isn’t just what got us here it’s what we need to drive the ecosystem for next decade plus” - amazing to have @jthandy join me at Slush to share his company building journey at @dbt_labs, honored to be your partner.

Next Thursday, we're taking the wraps off of Hex 3.0! We went pretty hard for this, including a live stream video you won't want to miss 📺 Add it to your cal! hex.tech/hex-three-poin… (we're also having a fun little party in SF, DM me if you want to join)

Super excited to announce Finetuning LLMs, a short-course made with my friend @AndrewYNg !! 🎉 By importing open-source @LaminiAI core & @HuggingFace & @PyTorch & @Weights_Biases, you can: ✅ Gain an expert's intuition behind finetuning* ✅ Understand how finetuning fits in vs. prompt-engineering vs. RAG vs. pretraining. Or are they all layers of a cake? 🎂 Yes, always cake. ✅ Finetune your own LLM, continually teaching it new knowledge ✅ Pick up practical frameworks for compute & memory requirements needed to finetune vs. run LLMs ✅ Touch on advanced topics, such as parameter-efficient finetuning *Honestly, "finetuning" could totally be named better, especially if you know the history of finetuning vision models! It involved freezing the model weights and changing the (classifier) head. For LLMs, we often keep the weights unfrozen and don't swap layers of the model, just the data. Built with favorite AI libraries: @HuggingFace - stay awesome, keep building @PyTorch - thank you for saving me in grad school @Weights_Biases - wish I had you in grad school, sorry Tensorboard @LaminiAI - building this now, so grad school would've been just 1 line of code

We’re back with the @dbt_labs Analytics Engineering Podcast! @jthandy and I kick off the season with the LEGEND @Bob_Muglia. 🐐🚨 One of my favorite recordings yet. 🎧 👇 open.spotify.com/episode/5mcDam…

Had a ton of fun hanging out with Joe Reis on his podcast! Episode just dropped here: open.spotify.com/episode/30bOkc…

dbt Labs has been recognized as @SnowflakeDB’s Data Integration Partner of the Year! 🎉 It’s a testament to the incredible results that customers like McDonald’s Nordics, Sunrun & JetBlue see from using dbt Cloud with Snowflake. At #SnowflakeSummit? Stop by our booth, #2720!

@mullinsms @pdrmnvd @matsonj @ChristianNolan I think this is actually an important distinction; centralized BI didn't actually provide (or aim to provide!) *most* of the data that decision-makers at an org used.

@mullinsms @pdrmnvd @matsonj @ChristianNolan Appreciate you both! I think we're talking past each other. I called out "BI teams" that sit inside of IT: I think that's actually what you're both describing. They didn't report to a VPA or CDO and didn't have organizational power. Very dissimilar to the modern "data team".

@matsonj @ChristianNolan @pdrmnvd The adoption of the MDS has pushed more of these shadow pipelines into more official channels. And while dbt isn't required for this transition, it pushed it to happen faster, as it lowered the barriers for shadow pipeline owners to author their own "official" pipelines.

@matsonj @ChristianNolan @pdrmnvd Generally, the way that data was produced at that point was that the people who needed the data produced it themselves or got it directly from their peers (in the business) who did. All "shadow pipelines".

@pdrmnvd sorry-- "and i'm *not* trying to suggest that this is no longer true."

@pdrmnvd and i'm trying to suggest that this is no longer true. certainly there were people using the MDS prior to their use of dbt. that's not really the group of people i'm addressing here. the far larger flow of usage is "shadow pipelines" >> dbt-governed pipelines.