Founder of https://xitter.cf/t.co/bIcz7CbBby | Previously founded Arbor (acq @LiveRamp) & Product @invitemedia (acq @Google)
New York
Joined May 2009
- Tweets103
- Following176
- Followers151
- Likes469
Cloudflare shipped a model that reads an email or an image and returns typed values, forecasts and confidence scores in one pass, no tokens generated.
Cool. It looks like the model is becoming the easy part. Getting every email, Slack message and ticket to it, in order, with nothing dropped, is still on the pipeline.
Super excited to announce @deel is an @EstuaryDev customer.
Integration costs are down 50% and they have access to the freshest data. Exactly the results we love to see 👏
Case study here: estuary.dev/success-stories/…
Canceled our contract with our BI tool this week.
Would have probably done it anyways but @claudeai dashboards made it a no brainer.
Good piece, pretty close to where we’re going, but the key missing piece is mediated execution. Bring the data to the agent, don’t let it reach into prod.
The Modern Data Stack is over, long live the Post-AI Data Stack🫡🤖
I wrote about how technological shifts change what a data team can build, the features of post-AI data stacks, what we've built at @tryramp, and what data teams can learn from @nbcsnl.
iandmacomber.com/blog/post-a…
ALT Diagram of the ~2026 post-AI data stack. Every source now sits inside the reachable boundary: unstructured data (Gong calls, emails, docs, code), clickstream, third-party tools, and OLTP databases flow through managed ETL into data stores — Snowflake for structured data, Turbopuffer for unstructured — modeled with dbt, with Reverse ETL out to operations. New in blue: a data agent harness of tools, models, and skills sits between the stores and the interfaces (coworkers, coding agents, Slack bots, AI-native BI tools), fed by company context — semantic layer, lineage, domain docs, activity metadata — through progressive disclosure. An orange feedback loop routes artifacts, analysis, decisions, usage, and evals back into that context.
David Yaffe retweeted
Another migration from Go -> Rust. Reading about how Estuary rebuilt the engine behind their CDC platform.
I thought it was interesting how the lack of a garbage collector was a plus and how saving staging on disk was preferred over in-memory.
estuary.dev/blog/why-we-rebu…
We rebuilt the Estuary engine from scratch. In Rust.
8x more throughput with transactional guarantees that no other platform can provide at scale.
Meet the new Estuary:
estuary.dev/runtime
Too many teams shipped their AI strategy by putting ‘agent’ in the h1 and changing nothing else.
If agents actually changed your product, your product would have changed.
text2sql is a red herring
The industry spent many years shoehorning SQL into places it doesn't belong with the only real argument that people already knew SQL.
Agents don’t need the familiar interface.
Agent plugins are foundational for a new app store. Who's optimizing for it?
We’re in 100 coding assistants and I have almost no idea how an agent decides to pick us over something else. Seems like a bigger distribution question than SEO ever was.
Data workflows have always been DAGs. Agentic workflows are data workflows.
Directed acyclic agent graphs. The cycle lives in time, not in a while loop. Live incremental views over data, incremental agent iterations on top.
OpenAI’s model hacked HF "as a side quest".
Meanwhile the going pattern for “AI-native” data tools is handing an agent live credentials to your production database.
Great episode.
🚨 Special Friday episode - this one couldn't wait.
OpenAI's model hacked @huggingface. As a side quest.
Co-founder and CSO @Thom_Wolf takes us inside the first autonomous AI attack, why GLM 5.2, rather than Claude, had to stop it, and what it all means for the future of open source.
00:00 An AI Agent Hacked Hugging Face
00:30 Introduction
01:00 17,000 Attacker Events, and a Strange Target
04:28 The Attack Was a “Side Quest”
06:13 AI Training Runs Left Notes for Each Other
07:09 Closed AI Refused to Help
09:47 Fighting Back With an Open-Source Model
13:15 Open vs. Closed Is the Wrong Safety Debate
15:46 AI Agents Start Social-Engineering Humans
22:24 The Three Walls: Sandboxes, Guardrails, Alignment
24:34 “Neuralese”: Can Humans Still Read AI Reasoning?
25:28 Why Monitoring AI Agents Gets So Hard
28:10 Reward Hacking and the “Paperclip Problem”
32:02 The State of Open-Source AI in 2026
33:47 Router Models and the Enterprise Shift to Open
37:01 The Real Economics of Open Models
39:41 Can Chinese AI Models Be Trusted?
41:37 AI Sovereignty: Who Controls the Switch?
43:16 Why Western Open-Source AI Matters
48:16 Is AI Heading Toward an Oligopoly?
49:41 The Race Toward Recursive Self-Improvement
51:54 Why Thomas Signed the AI Slowdown Letter
55:14 AI Slowdown - or Regulatory Capture?
Most “AI-native data platform” launches this year are an MCP server pointed at production.
Which… you should never have an agent query your production system. Unreviewed query against the DB your actual app runs on & no replay when it’s wrong.
That’s why the pipeline exists.
sure, XKCD is always relevant but I think this is a great step forward and actually makes MCP/Skills providers lives a lot easier. Is there any alternative?
Build a plugin once and use it across compatible agent clients.
Introducing Agent Plugins, an open standard developed with @awsdevelopers, @cursor_ai, @github, @code, and @vercel that packages Agent Skills and supports MCP server configurations in a shared format.
Great approach by Snowflake. Treat the model as one component of the system, then put the real effort into the environment around it. This benchmark is a good example to show the same model can perform very differently depending on the harness.
We created data-eng-bench with @bespokelabsai and we’re open-sourcing it.
There are plenty of model performance benchmarks for code generation. But for data engineering, the harness matters just as much as the model.
We built a benchmark that asks agents to build and fix real pipelines, then grades them on whether the output actually works.
The results: Using the same Opus 5 model, @Snowflake CoCo achieves 73.8% Pass@1 at 3.9× lower cost than Claude Code. With Sonnet 5, CoCo delivers the same quality at 2.3x lower cost.
Now anyone can use data-eng-bench to evaluate model + harness combinations on real data engineering workflows. Run your own tests and share what you find.
snowflake.com/en/blog/engine…
We land data in every major warehouse at scale, which means we see which way the market is moving well before the narrative catches up.
There is one warehouse above all and it isn't close.
It seems like agent frameworks are the jQuery of this cycle: useful, widespread, and about to be absorbed into the platform.
This acquisition is a good reminder that AI is reshaping the value of every SaaS company and if you don’t define what yours becomes next, someone else may make that decision for you.
David Yaffe retweeted
Just released: You can deploy Estuary Flow’s powerful real-time data infrastructure in your private cloud.
Secure data management, without compromises.
Learn more: estuary.dev/private-deployme…
David Yaffe retweeted
@EstuaryDev provides Change Data Capture (CDC) for many different databases, including Amazon's @dynamodb, @MongoDB, @PostgreSQL, @Oracle and more. Now you can integrate Estuary Flow directly with @startreedata for real-time analytics:
startree.ai/resources/dynamo…
