Skip to content

Instantly share code, notes, and snippets.

@CodeSolutionsLLC
Created October 9, 2026 16:33
Show Gist options
  • Select an option

  • Save CodeSolutionsLLC/6cc37a8e83a82d4689a3a3cd4b9f5319 to your computer and use it in GitHub Desktop.

Select an option

Save CodeSolutionsLLC/6cc37a8e83a82d4689a3a3cd4b9f5319 to your computer and use it in GitHub Desktop.
Turck x Pavlo — denser transcript (Don't Panic Digest, 2026-10-09)

Matt Turck × Andy Pavlo — denser notes (scoped) · Don't Panic Digest More

What this is: Nearer-verbatim operator notes from a public X interview segment. Not a court transcript. Gaps labeled. Hosting ≠ endorsement.

Not Code Solutions LLC viewpoints or advice. Speakers' claims stay speakers' claims (Intent / [unverified] / host-said).

Hedge legend: [inaudible] · [unclear: …] · [gap: …] · Intent/[unverified] for numbers, incident, and company claims · host-said for Matt Turck framing.

Speaker claims: Numbers and statistics (adoption percentages, accuracy figures, share of open-source commits), anecdotes about a bank and database companies, and the Replit incident as described by the host are the speaker's or host's account [unverified] — not checked, not restated as fact. Skipped ranges: 10:51–15:10 and 30:43 onward.

Source URL: https://x.com/mattturck/status/2108223135673696504
Full episode duration: 1:16:46 (native X) · Scoped for Digest More: player 04:09–10:51 (~6:42) + 15:10–30:43 (~15:33) · 10:51–15:10 and 30:43 onward skipped
Host: Matt Turck · Guest: Andy Pavlo (database researcher)
Posted: Thu 2026-10-08 · Notes desk: 2026-10-09
Contest OUT · Third-party names (Replit, Postgres, Databricks, Anthropic, etc.) kept verbatim when spoken.
Disclaimer: https://legal.codesolutionsllc.com/si-disclaimer — Notes of a public interview — not Code Solutions LLC views or advice.

Lane: Digest More feed only. Not a Daily X release from this pack alone.


Part A — player 04:09–10:51

[04:09] Andy Pavlo
...do that. And agents now, uh, create databases, use databases, can make changes to databases. So maybe a year ago when Databricks acquired Neon, there was this stat, uh, that I believe said that, uh, the number, uh, or the percentage of agents creating Neon databases went from 30% to 80%.

[04:31] Matt Turck
Or the databases created by agents was like, was larger than everything else.

[04:35] Matt Turck
Unpack that for us. What, what does that mean, an agent creating a database? Uh, why does an agent need to create a database in the first place?

[04:42] Andy Pavlo
The creation of a database, and then there's branching of an existing database. And I think the Databricks metric was about like the branches being created by agents. Um, so, so what does branching mean? So branching is when you have a, you have a database system, or sorry, a database, and you want to make a sort of copy-on-write snapshot of that, that database, and have it look like two unique databases. Um, and the reason why you want to do that is you could have your production database that's running your, your business, your application, like where all your, your, you know, the system of record or all your changes are going into.

[05:16] Andy Pavlo
But then you want to have either a human developer or an agent be able to start trying out new things, make changes to the application. And you want them to operate on the real database, because oftentimes, traditionally, you would make copies of the database, or have a smaller version of, of the database, like a sample of what, what already exists there, because it's just too expensive to, to copy things around. So you'd have like a staging or development, uh, uh, database, and then you would then hopefully apply your changes to the production database and hope things that, once you go to the larger scale, everything works out.

[05:46] Andy Pavlo
So what branching allows you to do is it allows you to make a snapshot of the real database, have all the changes that you make go to this snapshot and not affect the production database. And then once you're okay with the changes you've made, maybe not necessarily the data, but usually for like the schema or the application code, once you know you've vetted it on the snapshot, then you can put it live in the production database.

[06:06] Andy Pavlo
So now with these agents, uh, people are spinning up, um, these coding agents to write new features, write new, new parts of their application. It's not something humans are doing anymore. And now you want to have these agents, you know, they want to, they want to vet their changes on a copy of the database. So you can create branches of these things. Often this just shows up when you're running testing as well. Like a lot of the times the, the, the branches are created by, um, you know, PRs running in GitHub Actions or testing, things like that. So you want to run your tester code on a copy of the database rather than the live one. So it's mostly being, this branching capability is used for development. So agents are, people are using agents to write a lot of new code. And so you want them to run in their own sandbox. So you make a snapshot of the database. So the branching capability is super useful for that.

[06:50] Matt Turck
Yeah. Even more, uh, sort of simpler level, an agent needs to create a database because agents take actions on the world, and if you ask an agent to create an application, an application at the end of the day is still a database.

[07:03] Andy Pavlo
Yeah, so, so certainly if you say, like if you create a database from scratch, then yes, your coding agent, you give it the prompt to say "build me an application that does this," then yes, it will create, it'll create a database right away. But most of the, most of the branches or new databases, if you want to call it that, are being created are for this development. But yes, the agents can now create databases. Uh, [unclear: for better words], you can have them drop databases. People often get in trouble with this because it's doing things that they shouldn't be doing.

[07:25] Matt Turck
Yeah, there was a, last year as well, um...

[07:28] Andy Pavlo
Well, so every, every year there's some, some story.

[07:30] Matt Turck
Yeah. There was like, there was some, some incidents that was, uh, very, uh, interesting and got pretty, um, uh, widely publicized on X for what it's worth. So this Jason [unclear: surname — not clear on the audio] that was trying to build something with Replit, and I think the Replit agent, uh, went into the database and deleted a lot of files, which, uh, then subsequently Replit, uh, made the right adjustments. Uh, you know, obviously all of this is the Wild West and people are learning as they go and mistakes will be made. Uh, but, um...

[07:58] Andy Pavlo
Well, but it's, it's the Wild West for people that don't know the history of databases. Like the real, like, for years we've had the capabilities to make sure people don't do things they shouldn't be doing. Like stupid things like dropping a data, table, or deleting a bunch of records. Those, you know, those, those controls exist. It's just everyone wants to reinvent the wheel and people learn the hard way of just doing what's already been done in the past.

[08:19] Matt Turck
Let's expand on that point. So, so what does that mean? What, what does, what needs to exist in the agentic world for databases, uh, that is not currently in place?

[08:29] Andy Pavlo
Guardrails, right? Like, like it's, it's like, you know, you wouldn't, you know, you wouldn't bring a new, you know, have a new child, like a new toddler, and just like not put up the guardrails so they don't fall down the stairs or, you know, put their finger in the socket, right? It's the same kind of thing for these agents. Just think of them as like toddlers. Uh, a little more sophisticated than that, but like, uh, you don't want them to run wild and do whatever they want on, on a database. Now, you could sit and babysit it and just say approve, approve, approve, like, you know, in Claude Code or whatever, but of course then you can't scale and have hundreds of agents trying to do things at the same time.

[09:01] Andy Pavlo
So you want to set up the right permissions and controls so that they can't go into the production database and drop tables that they shouldn't be. And likewise, you don't want them to go read things they maybe necessarily wouldn't want to read as well. So if you have the right guardrails, um, and through branching or whatever means you want to, you know, have, create these sandboxes, you can avoid disasters like, you know, all, you know, every, every month there's a story where someone lost their production database because the agent deleted something or did something they shouldn't do. With the right guardrails in place, you can ensure that these things don't happen. Right? People get lazy. They use the same password across all, you know, entire company organization, and then they give the same password to the agent, and then they wonder why the agent started doing things that they shouldn't do.

[09:42] Matt Turck
Should those be the same guardrails as what currently exists?

[09:46] Andy Pavlo
Yeah, so I, I think the, maybe a larger question is like, do agents at least now behave differently than, than humans? Um, other than just being able to run 24/7, I haven't seen anything that would suggest the fact that they, they don't develop the same way, uh, software, the same way that, or interact with data the same way that humans do. Right?

[10:06] Andy Pavlo
At the end of the day, the database system only sort of exposes so much different capabilities, uh, and that's the API that gets exposed to you. So as long as you, you know, you can constrain those things for a human using those APIs, you could do the same thing for, uh, for agents. And the, like the level of sophistication of what those guardrails could be, beyond just like "don't run delete" or "you know, or update on a table", the, the level of sophistication, um, really matters on what sort of product you're using.

[10:35] Andy Pavlo
Um, and I will say this is something where like, you know, I, I like open source databases, but like the enterprise commercial legacy databases, if you will, they've already done all this, you know, for, you know, for humans in the past. So they have a lot of the protection mechanisms you would want to just expose that to agents as well.


Part B — player 15:10–30:43

[15:11] Matt Turck
...as well. [Andy Pavlo: Yeah.] Some fantastic stuff in there that I want to go back to in a minute. But there is a concept of agent memory that seems to be a little bit of a battleground right now with, on the one hand, file systems and on the other hand, databases. What is your view?

[15:30] Andy Pavlo
Oh, well what is a file system? It's a database, right? Everything's a database, right? You have a, you know, a directory with a bunch of, a bunch of Excel files, that's a database. Right? You have a notebook and pencil on paper, that's a database. Like my daughter was like three, we would, we would, you know, trying to teach her the importance of databases and I would make her write down—well, we'd do it together—we'd write down the temperature every day, you know, in a little, little line notebook and then she would close it and go "commit" to save the data. Uh, that's a database, right? It really comes down to what the, what the form of the, of the input or the context you want to provide to your agents are now. And as far as I know, it's all human-readable text, right? So whether or not that human-readable text comes from a bunch of JSON files you have in your directory, or it's extracted from a database system, uh, that can then convert that into the correct input form. At the end of the day, like, it doesn't matter. I think that, uh, there's obviously advantages and safety guarantees you can get when you use a database management system that can protect data. Like SQLite is famous for this, why would you want to use a file system and try to manage whether you're writing things correctly and safely and that you can get multiple changes and commit this, uh, versus like just SQLite just does it. Right? So like, at the end of the day, everybody should be putting everything in a database system. Or I mean even some file systems, like WinFS, underneath the covers, it's a database system, right?

[16:54] Matt Turck
Yeah, there was a piece that also like went viral at some point... I think that was entitled "I Recreated a Database by Accident," like talking about like file systems. Okay. Interesting. Does the... the world of like multi-agents, sub-agents, agent swarms, does that require more of a database, meaning in the sense like computationally very complex and therefore having something that's extremely heavy duty on the other side seems to make sense?

[17:24] Andy Pavlo
In the end of the day, if you need to share state across multiple machines or instances of these agents, whether they're running on the same box or not, like, again, a database management system provides that capability so that you could have, you know, multiple, multiple readers reading the shared state, right? Um, I haven't seen anything where like agent itself, like as part of its memory lookup, is you're pushing down filters and things like that into the database. Usually that's provided as part of the context. Again, this goes back to the RAG or the vector database stuff. Like, you provide the context based on what you've already looked up in, again, your memory... in the database files or the database itself. So, you know, again, from my understanding is that the API for the input for this context is just human-readable text, and whether that comes from a file system or a, uh, or a database system, it doesn't matter. But the database system provides certain guarantees and luxuries, if you will, that a file system can't provide, and shared state is one of them across multiple agents.

[18:24] Matt Turck
Have you seen the agent world converge towards a specific type of database?

[18:30] Andy Pavlo
I mean asking like, it like in terms of like using it as a memory or having an agent choose one particular database system over another when they build an application?

[18:38] Matt Turck
More the... actually both.

[18:40] Andy Pavlo
So I guess maybe the latter, the latter one, it obviously they're fed on... on... the models are trained on what people are writing about. And so, you know, for... if I say I'm building a new application, chances are it's probably going to recommend Postgres at this point, uh, because that's what's in the, in the pre-training data set. And I will say, I can't say who, I've had two database companies ask me, uh, how to get their database to be recommended first by an agent over Postgres. And I was like, that's not, first of all, that's not what I do, that's not research, uh, and two, like don't talk to me, go talk to the guys building the models because like that's clearly something that you know, they can have sway in. It's basically like SEO, how do you get the agent to learn like "Oh I should use my database X instead of Postgres."

[19:25] Matt Turck
I would suspect that your writing is highly indexed by the search engines and therefore pitching you is a way of pitching the model.

[19:33] Andy Pavlo
So this is a very much a first-world problem and I'm not trying to say like a humble brag, but like, one problem I do have is when I ask questions about databases, it cites me back to myself, because all my course materials...

[19:44] Matt Turck
Which has to be somewhat satisfying.

[19:46] Andy Pavlo
Yes. Well, no, but it's like, no, because like... Say when I was teaching database courses, I would say, you know, show me, I want to learn about database systems that do this version of a kind of index, like a B+ tree or whatever. And I would say... so now I have to say, like, I already know which ones do do this, so show me the ones that don't do this, and then I have to say also don't cite Andy Pavlo or anybody from Carnegie Mellon. But then the problem is that our course materials...

[20:09] Matt Turck
Does the model reply "You know I know you're Andy, right?"

[20:13] Andy Pavlo
ChatGPT knows... yeah. ChatGPT actually says like, "I know who you are." It says something like that, but Perplexity doesn't do this. But the problem is also too is like since all our course material is open source, anybody can just take and use it, so then it says, okay well you can't cite Carnegie Mellon University, let me cite another university, but it's just my slides being used at another course. So I got to be like very specific about what what not to have it turn back to me. Um...

[20:39] Andy Pavlo
So I think that... to your original question, like I think for OLTP, I haven't tried this in a while, but usually it spits back "Hey you should use Postgres." Unless you say something very specific like, uh, you know, "I'm gonna I'm gonna run on embedded devices," then it'll say "Use DuckDB or SQLite" or something like that. Or "I only care about key values," and then come back and say like "Use RocksDB." But in terms of for OLTP it almost always says use Postgres. And then it might recommend like "Do you want a hosted version or not," right, if you have further chats. Um, for OLAP, I'd be interested to see if there's any kind of ranking on this, it's actually not a bad idea, uh, maybe somebody's already done this, but like that one it definitely would come back and maybe you have to refine it further and say like "My workload looks like this, my data looks like this," and then I think it can give a more nuanced answer. I haven't tried this in for about a year or so.

[21:30] Andy Pavlo
...but I've definitely seen before like, and they say, oh, you know, do you have any money? Yes, okay, you can use Snowflake or Databricks. Do you have no money? Okay, you can use DuckDB, or, you know, do you want to run on-prem? You can use ClickHouse. Things like that. Um, so then your second question was, uh, about agents...

[21:44] Matt Turck
Well, which one should they recommend? So that's that's that's the answer to which one they do recommend, but uh...

[21:51] Andy Pavlo
Which one should they recommend. Yeah.

[21:53] Matt Turck
Yeah, and where I'm going with this partly is, you know, this question of of of speed. And that's where, you know, ClickHouse uh shines, and we'll talk about ClickHouse uh more later. There's perhaps a concept of uh, you know, context layer, ontology, which seems to be pushing towards the graph databases. Are they, or or, you know, what what what what makes sense?

[22:16] Andy Pavlo
I I mean I have strong opinions about the graph databases. I don't think you need one. I think, uh, that can be... I have not seen any benchmarks that show graph databases, the native graph databases, outperform relational database systems, uh, when the relational database systems have the right constructs in SQL or the API so that you can do graph traversal all on the server side and not go back and forth. Um, SQL, the SQL standard in 2023 added support for property graph queries. Oracle supports it. Postgres was supposed to support it this year, it's coming out, Postgres 17, but I think it got delayed or rolled back, I don't know why. Um, but anyway, that's beside the point.

[22:53] Andy Pavlo
No, I I think that, um... I mean, the context layer... again, depending on the requirements of what the, how many agents you have, how much, whether you have shared state or not across multiple agents, um, Postgres often is is, you know, should be the first choice for many of these things. For OLTP. If you're doing analytics, that's a whole another story.

[23:14] Matt Turck
Let's talk about MCP.

[23:15] Andy Pavlo
Yes.

[23:16] Matt Turck
So, over the last uh year or so it seems like everybody, every database provider, added an MCP server to their to their offering. Thoughts on that? Like pros and and and cons? And I'm I'm coming from the, you know, a year or two ago, there was like this whole like text-to-SQL kind of movement where everybody was trying to make uh LLMs work on top of databases, especially for analytical purposes. And that that sort of didn't work. Are we there now?

[23:43] Andy Pavlo
So, text-to-SQL, it's like natural language to SQL. So this is something people have been trying to do since the 70s. Right? Uh, and every so often there's, you know, maybe a uh, there's a, you know, some kind of breakthrough in natural language processing. People are like, okay, now finally it can happen. Right? It hasn't happened. Except for LLMs. I think, um, the... I've been skeptical of the academic literature uh and the recent publications on various frameworks to do text-to-SQL because often times the benchmarks that they're showing, Spider and a couple other ones, like these are well-known benchmarks that you know the models are trained on, so like of course it's gonna regurgitate the the correct answer.

[24:24] Andy Pavlo
I actually was talking to somebody from a very large bank uh yesterday who said they now have a text-to-SQL interface. Uh, but in the initial implementation of it, they were using something off the shelf and then they they further refined it a little bit further in in-house. But it was doing it was about 60% accuracy, meaning like the the the LLM or the agent could produce the right SQL query to produce the right answer, 60% of the time. Right, and that's roughly what I think I've seen before in other some other reportings. But they said through this providing a a more rich semantic layer or context layer over multiple years and millions and millions of dollars to make this work, they were able to get it up to like 99 point like 5% accuracy, which is insane.

[25:04] Andy Pavlo
I think that is... I think that that's a a viable thing now if you invest the time to have the right semantic layer and the context information you need so the the model is, has the the lo- sorry, the agent has enough information to how to, derive the right SQL query from the question you're asking. I think, again, that's a large bank with infinite money that they could pull this off. I think if you just download something now, a random person will find that these existing tools are not going to be at that level of proficiency or accuracy. Um, but it doesn't mean we can't get there, right?

[25:38] Andy Pavlo
Um, so now in terms of the MCP, the the Model Context Protocol, like I get why they did it. Like, I think this came out of Anthropic and then everyone sort of quickly adopted it as as the as the standard. Um, and then every time there's every so often there's a Hacker News post saying like MCP is dead or whatever. Like, it's I don't think it's going away. It's basically a REST interface, right? A standard way to to specify what you want [ASR-sourced join] the request to be from the agent to whatever, you know, data system you have, and then that data system that receives the MCP request, it's their job to convert it into whatever the correct API calls is to get the data you need. And in the case of database systems, it's, you know, going and, writing the SQL query to go get the data you need. But usually it's either, you know, oftentimes it's a strict translation of what being passed to you and not like trying to derive the semantic meaning of the MCP request because the MCP request has to be somewhat well well-defined.

[26:32] Andy Pavlo
Um, I have no problem with MCP. I think that, like, it's not, it's not this magic bullet that solves all the world's problems. It's just an additional interface, right, uh, that you can now interpose or interoperate these different agents and different systems to collect data you need. So I'm all for it because it's just you're using the database more rather than some homegrown thing. So yeah, fantastic, it's great.

[26:52] Andy Pavlo
Could it be improved? I mean, like, I think it's inefficient because you're sending human-readable text over, you know, over the network to go retrieve data. There's obviously better ways to do this with binary protocols, but like, at the end of the day, it's, it's good enough, right? And it seems to be the standard everyone's using.

[27:09] Matt Turck
Should it be that security layer that you were alluding to earlier?

[27:13] Andy Pavlo
Yeah, so like, you could, but it's no different than, like, the, if the request came from like SQL from a traditional application client using JDBC or ODBC, like, you know, the network protocols, the terminals for for these different data systems. It's no different if it comes from that versus, you know, a REST request with over M-, with MCP.

[27:33] Andy Pavlo
At the end of the day, like, okay, I got this request, I need to understand who's asking for it, what are they asking to do, and should they be allowed to do it? Like, so the the guardrails we were talking about before, they're the same if it's coming in at SQL versus the same as coming in MCP.

[27:48] Matt Turck
Okay, so that's uh one part of the conversation, that's databases as uh memory layer or execution layer for for agents. Let's let's flip the discussion into uh what AI can do for databases. So clearly we live in a world where uh AI can code just about anything. Is that true of databases as well?

[28:11] Andy Pavlo
Yes. So to give one anecdote, I teach data- course on database management systems at Carnegie Mellon University. Uh, a year ago, the the agents couldn't do our entire projects. So the projects would be like, we give you a scaffolding of a database system and you have to implement, you know, the indexes, the query engine, and things like that. It could do some of it, not all of it.

[28:33] Andy Pavlo
Uh, I think it was Opus 4 or whatever that, whatever Anthropic put out last year, then that just opened the floodgate and the agents basically do all our assignments, right, without, very little prompting. And of course, there is a lot of training data for it because all of our projects are open source, they're all on GitHub. Not just students at Carnegie Mellon University, but also students outside of university, we let them use it. So there is a lot of training data for them to implement things.

[28:55] Andy Pavlo
Um, and so I think agents basically could implement anything, you know, you would want to build a database management system now. You know, obviously you have to prompt it the right way and hold its hand and make sure, you know, you you have, you generate the right design or it produces the the implementation based on the design that you want. Um, at the end of the day, I think these agents are very capable to build, to do this.

[29:16] Matt Turck
So they could build an entire database? Because so my my experience of uh building databases as a venture investor is that it's a 10-year journey of pain where nothing much happened for three years and you have some of the smartest people in the world getting together to solve what seems each time like insurmountable problems. So are we now saying that uh it can build the whole thing?

[29:34] Andy Pavlo
Yeah, the old adage from from database systems uh is that it takes 10 years to build a database system. You can build the first 90% in three years, and then the remaining 10% takes the next 70 years, or seven years. Um, so yeah, no, I I I think that the agents are very capable, you know, with enough tokens, of course, and then with enough guidance, people can, you can vibe code an entire database management system, and there are certainly companies that are doing this now.

[30:00] Andy Pavlo
Uh, and pretty much every single database company is using agents to help them develop things. So one of my, again, I love databases, one of my side projects is the Database of Databases, dbdb.io. Uh, and one of the things we do now is we keep track of every single database system that I know about, and for the open source ones, every single night we pull down all the latest commits on GitHub.

[30:21] Andy Pavlo
And we, let me track to see which ones are actually being, you know, co-signed by Claude or or Codex and things like that. And obviously, if people turn that feature off, you don't know whether it's actually been generated by an agent, but most people don't do that. And at this point I think, like, 60%, over 60% of the open source database systems are being, they have commits from, coming from agents.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment