They didn't write: "95% of business analytics is now automated" they wrote: "95% of business analytics queries are automated" small difference in words, big difference in meaning. There's hype in the post, but I think this is true: With a well-built ontology / semantic layer, 95% of questions can be answered by an AI agent Building a semantic layer is a good amount of work (so not 95% automation of analytics work), but not nearly as much work as answering every question. We see the same at @definiteapp. Converting a question to SQL is just not that hard for frontier models with the right context and guardrails definite.app/blog/ontology-a…

we're 1% of this and have the devastating anthropic bills to prove it

we're building our agent (@definiteapp) with claude code. We use Skills to tell Claude how to use features within our product. This keeps the main CLAUDE md small, while still giving the agent all the details it needs. e.g. we have a data-lineage skill and claude knows how and when to use it based on context in the app

you're right, text-to-sql doesn't get you near an insight a real business user would want. there's simply too much ambiguity and missing context. e.g. Business user: give me churn by # of employees A b2b SaaS company might have the churn data in Stripe and the # of employees in Hubspot. The LLM would need to: 1. Take the raw stripe data, create cohorts and calculate churn 2. Find the best "# of employees" to use from the 5 options in Hubspot 3. Figure out how to join Stripe and Hubspot 1, 2, and 3 are all tricky and open for interpretation (e.g. should we join on this UUID that is in both Hubspot and Stripe but called _meta_cust_id in Stripe and customer_id in Hubspot?) Here's just one record from the Stripe subscriptions table. Questions from just this one record: 1. Is amount in dollar or cents (it's cents) 2. Do we need to join to the subscription items too (yes)? 3. How do we know if the subscription is past due (you don't from the subscription)? 4. Is interval always monthly (no)?

zoom out a bit. I've been working with LLMs since GPT-3. We were trying to get it to write SQL to answer business questions (i.e. "self serve analytics"). GPT-3 could sort of do this for really basic questions with the right prompt. We were < 70% accurate. GPT-3.5 we got to 80%. GPT-4 we got to 90%. We continued adding guardrails (things that make life easier for the LLM) and a more agentic approach (i.e. the agent can explore the database and correct errors). The models kept getting better too: Smaller models: better / faster for certain tasks (e.g. searching a schema). Context windows: GPT-3 = 2k tokens -> GPT-5 400k tokens Reasoning: run a query -> get an error -> show the LLM the error -> LLM self-corrects with GPT-5 / Sonnet 4.5, we added another 9 and we'll keep adding them for a decade.

"12 weeks total" -> 12 minutes later

anthropic has built-in tools (e.g. text_editor_20250429) that claude4 was **trained** to use. Claude Code uses these tools very well. Cursor built their own tools which claude is not as good at using. docs.anthropic.com/en/docs/a…

opus-5.x is going to be fucking nuts. The training set from Claude Code is pure gold.

cursor for powerpoint is just cursor + `npm install reveal.js`

added s3 support to smallpond on whim [link in comments]. I don't think smallpond gives me anything I don't already get from @duckdb, so I'm bailing on it.

@definiteapp is cursor for data (backed with the best open source data infra like @ApacheIceberg and @duckdb)

check the dev console, do you see this? unknown tool research_kickoff_tool.clarify_with_text

the amount spent creating a model is a fact. if it costs pennies or billions, by definition, changes the facts. if you run a mile in 15 minutes, good for you, you ran a mile. if you run a mile in under 4 minutes, the world takes notice.

tools: ["human"]