Average is all you need(rawquery.dev) |
Average is all you need(rawquery.dev) |
A car that starts 50% of the time ?
A plane that stops on 50% of the flights ?
A pacemaker that beats only 50% of the time ?
David Goodenought said that average is enough ..
if anything it makes the world more dangerous
a reckoning is coming
the top decile will be janitors for the rest
Why is that a good thing? Claude didn't ask any obvious follow up questions, like what determined whether a user got an email or not? It is using the ab test terminology in Step 3 without any kind of confirmation that this is, you know, a valid test.
Pass.
"Whereas before, average was expensive in terms of both time and effort, average became cheap."
But nobody bothered to check if it was correct. It might seem correct, but I've been burned by queries exactly like these many, many times. What can often happen is that you end up with multiplied rows, and the answer isn't "let's just add a DISTINCT somewhere".
The answer is to look at the base table and the joins. You're joining customers to two (implied) one-to-many tables, charges and email_events. If there are multiple charges rows per customer, or an email can match multiple email_events rows, it can lead to a Cartesian multiplication of the rows since any combination of matches from the base table to the joined tables will be included.
If that's the case, the transactions and revenue values are likely to be inflated, and therefore the pretty pictures you passed along to your boss are wrong.
Further reading, and a terrific resource:
https://kb.databasedesignbook.com/posts/sql-joins/#understan...
I can write that script faster than I can write the text asking the AI to write the script as SQL is concise and my IDE has auto-complete.
I will never understand Engineers who struggle with SQL lookups. The vast majority of queries are extremely basic set theory
I think this is important, because if his hypothesis is right, then LLMs behave differently here: They really are average in all dimensions. They are the pilots the Air Force thought they had before Daniels made the study.
So if he is right, we'd be changing from a mostly-non-average to a mostly-average society, which would really be a massive change - and probably not a good one IMO.
[1] https://noblestatman.com/uploads/6/6/7/3/66731677/cockpit.fl...
Why didn’t the boss ask the AI for the charts to begin with?
Everyone’s income is going to be below average, because they got fired.
You might say it "still less work" and that's true, perhaps, only for the first few times. After a while you _learn_ how to do it, and understand how to _think_ with the language of your data. With LLMs, you never get this benefit, and also loose your ability to judge the LLM's output properly.
But again, that might be enough on your case, or, you simply don't _know_.
Let's say you start with a report someone else wrote. It seems like you still need to read it and understand what it's telling you. Sometimes plotting all the points helps, or drilling down and looking at the raw data.
When it comes to bs dashboard where "average is all you need", maybe the "better than average" result would be asking yourself if it's even worth doing in the first place?
The Business simply cannot admit that it’s really doing nothing above average. If they did, investment dries up.
There's a market for both, but the furniture slop of Ikea is dominant.
Do you know enough about JOINs and how they work to be able to break those big queries down and figure out whether they are doing exactly what you're asking for in English?
It makes me wonder if Hacker News has a silent majority of people who would actually use AI in this way without wanting to admit it, and a vocal minority of people who wouldn't.
> ninety percent of everything is crud
Yes, thinking about your data and how to check it is so annoying. Much better to do something average, see if the result puts you in a good light, and share that insight into your company's working with ~~everyone on the internet~~ your boss.
Rarely have I seen "we help you create meaningless slop more easily" advertised so explicitly. Or is this also average?
I think many people here work at nice, large places with reasonable and knowledgeable colleagues that are cooperative and mostly rational and try to do the right thing. In my experience that is not a common or widespread thing. Of course I only have small to medium business experience, but that's still a pretty good chunk of the economy. LLMs are an absurd, ridiculous win in those kinds of environments.
And that's when the agent even manages to construct a reasonable naive query. I've seen even Opus 4.6 ignore a `is_demo` column in the schema it was given when asked to construct a query for the number of active users.
Where I've seen text-to-SQL work well enough is when you're pointing it at data that's already been well-modeled for analytics such that the naive query a LLM will construct is correct by default. The data is either structured as a wide table such that no joins are necessary, or all the joins are 1:1 fact <-> dimension joins. All metrics are additive and so can be aggregated without asterisks. Columns follow a consistent naming convention, using the business domain terms a user would use in their prompt to the agent.
But that's a much thinner niche that what rawquery is proposing. You can't get around the analytics engineering effort involved in constructing a quality analytics dataset; the LLM will be a best a fuzzy fronted to your data warehouse, coextent with your BI tool.
Note: I do see value in value in rawquery's CLI-first approach to accessing data. In the right hands agents are very helpful at rapidly exploring datasets and validating assumptions on source data; but all the cloud data warehouse products I've interacted are all somewhat fiddly to access locally.
It's a post claiming average AI is useful... by a for-profit "data platform with a CLI that LLM agents can use directly". What are they going to do? Criticize the whole industry they are selling to?
And there is a lot of that type of work to do if you're trying to grow a business. But, something in there should be trying to be exceptional or else you have no moat. Claude will probably not be able to breeze through that part with the same amount of ease...
The question is, do we have good enough feedback loops for that, and if not, are we going to find them? I would bet they will be found for a lot of use cases.
/end extreme over optimism.
I think you can have LLMs do that too, and then generate synthetic training data for "high-effort code".
where it gets interesting is when you have a custom system that your LLM surely never saw (custom ERP) that has 50 sometime cryptic tables, unclear look up tables and unexplained flags.
something no text2sql solution solved for us.
we built a second mcp that lets the agent look up business logic (generated from source code) and then does better queries. that i think is something i never read in a blog post about a text2sql solution.
You could use claude code for the "text2sql" kind of part, but this is not why this tool exists. Nor what the article advocates.
This is not only average. This is actual magic.
So let's be real: the SQL is average. The joins are average. The chart is average. And that took us less than 5 minutes and that was amazing, that is the entire point.
You did not need a data engineer to model your HubSpot data, or a meeting to agree on whether it should be last-click or first-click or linear or time-decay or whatever.
You needed a query, written fast, on data you already own. Your LLM wrote it. You confirmed it made sense. Your manager got a link.
Honestly, average is clearly magic; prove me wrong.
It is like nails on a chalkboard.
Not all context is documented, and some context has to even be changed because it doesn't make sense.
I find AI very useful, but I think a lot of this AI SQL products are misleading.
No one has ever differentiated themselves based on how good of a ticket taker they are. Coding especially on the enterprise dev side where most developers work has been being commoditized since 2016 at least and compensation has stagnated since then and hasn’t come near keeping up with inflation.
In 2016, a good solid full stack, mobile or web developer working in the enterprise could make $135K working in a second tier city. That’s $185K inflation adjusted today. Those same companies aren’t paying $185K for the same position.
My one anecdote is that the same company I worked for back then making $125K and some of my coworkers were making $135K just posted a position on LinkedIn with the same requirements (SQL Server + C#) offering $145K fully remote.
I 100% agree here.
AI has been a huge boon for me personally, because I stopped spending most of my writing code years ago. I was reviewing code, writing procedures, handling incidents, and generally just looking for pain points across the entire company and solving them before they became critical.
Those skills have transferred directly to working with AI.
I might not agree with the point, but I can see that idea that many things just need to be "good enough" (which we might define as "average") and we save our real expertise for the things that really matter.
s/average/median
the other being how well the ai can use it, and how ai SEO-ed you are.
vercel and next.js for isntance are absolutely loved by claude
The people who need to be above average and exceptionally are senior management and maybe a few bright sparks in middle management. Most of the value-add happens there that builds social machines that then do the work.
> If average is all we need, then anyone can do it.
Pretty much, yes. That is why the range of salaries on offer is pretty compressed compared to the range of returns capitalists get.
That is the dream. Upper management can get software made without talent.
But is seems to be the greatest ideas in the last 30 years didn’t start in board rooms. They started with a couple coders creating a new idea.
No boardroom could have invented Google. It was so fundamentally different than what other search engines were doing.
We have this myth that upper management is so important. It is as the business grows in size, they are excellent for coordination. But ideas come from people closer to the problems.
How stable that is on the long term, I don't know any more than the next guy, but it is where I'm contributing now.
As someone who's seen queries that are hundreds of lines long, involve a bunch of CTEs, nested SELECTs as well, upwards of a dozen joined tables with OTLT and EAV patterns all over the place (especially the kind of polymorphic links where you get "type" not "table_name" so you also need to look at the app code to understand it), I'd say that SQL can be too hard for people to reason about well.
Bonus points for having to manually keep like 5 Oracle package contents in your working memory cause that's where the other devs on the 10 year old project stored some of the logic, while the remainder is sort-of-dynamic codegen in the app.
Same as with most app code, it shouldn't be like that, but you sometimes get stuff that is really badly developed and the cognitive load (both to inherent and accidental complexity) will increase until people will just miss things and not have the full picture.
You might want to try a different example, that one rather undermines the point you're trying to make. PageRank [0] was developed by Page & Brin as original research/based on the work of other people who weren't employees.
Part of the problem is that better code is almost always less code. Where a skilled programmer will introduce a surgical 1-3 LOC diff, an incompetent programmer will introduce 100 LOC. So you'll almost always have a case where the bad code outnumbers the good.
Though usually "average" implies arithmetic mean.
But it is useful to question whether that is true in all cases. The cases that aren't normal-distributed might be exactly the cases where it pays off to be neither average or median
I do not sell a wrapper on top of some LLM; you can absolutely write your SQL directly. There is an engine, there are iceberg tables. You can just live your best life doing your own SQL by hand.
Now if you couldnt do it before and you have a sensible understanding, you can likely do a bit more with the CLI tooling. And if you know a lot more, you can still do that. The queries are not hidden, or abstracted, If you need them they will be saved - transparently in SQL.
So I dont know what is the answer to the question "how do people do things they don't know how to do" ?
The statue quo had been to learn SQL or ask a human you trust to check their own work, which hopefully you can reuse.
Now it's ask AIs that are intentionally a bit random, and less likely to (or incapable of) check(ing) their work. Perhaps without seeing the SQL at all, requiring to trust it for every interaction. And in a culture that moves so fast that there is no checking by any(one|thing).
Modern models are quite capable at surfacing and validating their assumptions and checking correctness of solutions.
Oversight helps you build confidence in the solutions. Is it perfect, no.. but way better then most engineers I also ask to check things.
That's like saying 'cars were better made in the 1950's because they used tons of steel'. Like they were 'heavier and more robust' - but that doesn't mean better.
Foundations are way better, more robust, especially weatherized. Windows today are like magic compared to windows 100 years ago.
What we do more poorly now is we don't use wood everywhere, aka doors, and certain kinds of workmanship are not there - like winding staircases, mouldings - but you can easily have that if you want to pay for it. That's a choice.
AI is power and leverage, it will make better things as long as it's directed by skilled operators.
The precision of how the wood or material meets is worse (when cut at the site). There is a huge amount of sloppy work in modern construction.
It seems to me that in the past there probably was lots of shoddy workmanship and just no-one paid attention to it.
But I have no proof of that.
And we can accommodate for 'selection bias'.
We have all of the historical evidence we could ever want for 'how things were built', basically 'infinity examples'.
I think some things were more robust, particularly some of the old framing, like in Europe, with non load-bearing walls etc. Those will stand for 1K years, but arguably unnecessary.
You have to get a representative sample, that's the tricky part.
So there's that!
If you think an LLMs can check their work, then you are doing a terrible job at writing software. Plain and simple.
They even go as far as "cheating", so tests fail, writing incorrect tests, or straight out leaking code (lol) like the latest Claude Code blunder. Is this the tool the original comment "is using wrong, plain and simple"? Or do you have access to some other model that works in a wildly different way than generating text predictions?
in this case to make a local copy of the db, fill it with a set of records with an expected output of the query, then check to see of the query produces what you want.
you could then have it make queries that check the various assumptions that went into that artificial set of data. if it can find the assumptions broken, add records like that to the test set.
same old agentic programming techniques as ever. use your engineering skill to set up feedback loops. stuff that was painful to do as an engineer for checking your work is now straightforward
Otherwise the LLM can just write tests against whatever it wrote and not what is expected. This happens often with the top models too.
Someone needs to check the tests work, review they cover edge cases etc.