Doug Turnbull

For the Season 01 finale of Show Us Your (Agent) Skills, I wanted three people whose workflows will immediately improve how you use agents: @HamelHusain has done more than almost anyone to drag LLM evals from vibes into discipline. Christopher Fonnesbeck builds Bayesian models for the Mets, Brewers and Yankees at @pymc_labs, on what production causal inference looks like when the stakes are real. @softwaredoug spent years on search at Shopify and Reddit, and writes about relevance with more clarity than almost anyone in the field. Each of them will show the skills and workflows they've built to encode that hard-won expertise into things an agent can actually run. The real setups, not the slide versions (or LinkedIn posts). @twiecki and I co-host. ​Register to join live or get the recording afterwards → luma.com/ltpzpqgw

@macastel3 Most work at these companies is about investments in future products / directions, not keeping the lights on

How is it even possible that a company like Meta for example lays off thousands of employees and the day to day business is not affected? Very weird dynamics in those huge tech companies

It's easy for an agent to reason+debug why bm25 / grep produce results they do, even when they're weird

@helloiamleonie 90s Inspiration :)

all of our personal portfolio pages look the same: minimalist + white/beige background + black font i have the urge to change mine to a maximalist, full of color vibe. something that looks like a kid drew a rainbow. but i'm also a basic millenial. the struggle...

@bryonkuchML @maxrumpf Nothing eliminates the challenges of getting good training data from clicks, etc :) But it massively simplifies by having a single unifying model that can match AND rank. And customizable via just a prompt

@bryonkuchML We should hear from @maxrumpf The hypothesis might be that if you train an agentic model on open e-commerce training set Then later in-context, you share training data from YOUR specific e-com problem... Then an agentic model could adapt using in-context guidance

Bullish on agentic search models turbopuffer.com/blog/reinforce…

dangerously close to the last few features needed to replace even the trickiest queries from gen-1 search engines...

@thinkingshivers I never look at AI edits or suggestions for this reason. I just have it point out copy-editing issues A human editor wouldn't rewrite anything, they'd just point out the issues sto me. I expect the same of AI.

I submitted a draft of my short story to Claude for copy editing. Sometimes he’ll suggest a re-write of a particular sentence. My version: “She was just a rich girl, with a carelessness about her that could only come from being born into privilege.” Claude’s suggestion: “She was just a rich girl, careless in the way only privilege allows.” It's a matter of taste, but I personally think Claude's version is better. It's saying the same thing but more deftly. But when I swap his sentence in then plug the paragraph into Pangram, it goes from being high confidence that it's human to low confidence that it's human. If I keep doing this, will it start to read like AI slop? If I keep doing this, is it even my writing anymore? So I'm keeping my version, the one I think is worse, and I'm disquieted by the fact that there could be a better version of this story that I now need to specifically avoid.

@JnBrymn Is day coding better than day drinking?

I’ve recently heard that the mid-tier “day coders” are the ones that are most screwed. These are the ones who code for their job only and aren’t excitedly exploring the technology just because it’s fun. - Juniors are fine because they’ll adapt quickly. - Seniors are fine because they became seniors by excitedly exploring. - people that career changed into dev roles are fine because they understand how to adapt.

At a tech group last night, people didn't seem worried about junior devs and AI coding Some junior devs thrive. Others create slop. Some senior devs thrive. Others don't want to adopt.

I made a poem

wait until you hear the annualized revenue of turbopuffer.supply (please do not ask the gross margins please do not ask th

turbopuffer x SID An easy way to tell a good from a great AI researcher: how much do they think about infrastructure. Infra extends beyond what’s running on the GPUs: Slow environments will bottleneck your training steps. More parallel and powerful models make this problem worse. RL environment specifics are usually secret, but we shared some details in a recent post with our friends at @turbopuffer Training great models requires great infrastructure and we’re excited to be working with the best.

@Sirupsen Congrats man! It's been great to watch

turbopuffer crossed $100M run-rate in March. 19mo after $1M. Profitable & <$1M raised. Cursor・Anthropic・Notion・Cognition・Harvey・Bridgewater・Ramp・Linear・Legora・Superhuman・Atlassian・Granola We’d be nowhere without them. We work like hell to exceed their expectations.

we are seeing more and more RL puffin', here's a model specifically trained for it on tpuf from @SID_AI

So many low cost / in between steps: use your existing search? Chunk then use lexical? Snippet out the info you need at search time?

Teams jump from "I have existing search" to "I need to embed billions of chunks and this will take 1 year / cost $10 million" Meanwhile nobody's sure the RAG product will even pan out.

If you want to understand why BM25 and ColBERT are very strong on deep research, look at the traces of agent on BCP

Won't be a cheaper time than now to buy a ticket maven.com/softwaredoug/c…

Next Cheat at Search w/ Agents: Oct 2026