Hmmm... I'm surprised I'm not seeing anyone else question the validity of this taking "2 hours" Given that it's written on the blog for the product it's using, this reads to me a lot like a pure sales pitch. They want us to believe if you use Patterns (which is neat), your company will be much more cost-effective.
I'm not saying that's bad - that's probably the right thing to do with their company blog, and it's cool nonetheless. But I do get a little tired of people putting stuff out there like this that warps (some peoples) perception around how long things actually take. We wonder why, as an industry, we misjudge timelines on things left and right.
Even if we take it at face value, this is a person who's intimately familiar with this product. So sure, it's easy to set things up when we've done it a bunch of times. If you were doing this, solving the novel problem that you're faced with, is that how long it would take? Plus that's not really what most of us get paid to do. We have to learn on the fly and figure stuff out as it comes.
So rather than have the provocative headline and conclusion, like a lot of other people have commented... this is really something that could amplify that $50/hour employee, not take their job away. And maybe we shouldn't read into the alleged speed so much. YMMV.
Author here, I’ve updated the post. The first draft of this app and blog post took me two hours, but I kept coming back with new ideas and tweaks throughout the week. By the end, I’d certainly spent more than two hours (more like 8?), so you’re right, I just failed to update the post. The main point stands — it’s surprisingly good for the amount of effort put in (although unclear how much more juice you could get out of gpt with more effort. Clear diminishing returns)
I totally automate my roles on projects all the time so I can move on to more interesting things. I guess you mean that no one wants to be fired, but I don't see how that can result from automating one's work.
Also, I HATE doing repetitive things. Some people seem to like it though. To each their own, I guess. Reminds me of https://youtu.be/wNVLOuQNgpo
We all believe our job is so challenging and has such special requirements that it _can't_ be automated. It requires someone with the kind of experience learned with wisdom over a long time. Blah blah blah.
This makes it so that 1. my quality overall becomes better and my bosses always liked that (doing things per hand are more error prone, not on time etc.)
2. I can go on holiday knowing my company doesn't need me desperate
3. I can spend the free time of actually innovating and bringing more value to the company/product
The problem is not automating yourself out of a job but not being able to leverage the new gained capacity.
My way helped me succeed. I took my skills and my achievements (which i made in my R&D Time) to another company and got more money and than i did it again and got more money again.
Right. Presented with efficiency gains, firms tend to increase profit, not wages. One way to change that is to give workers more bargaining power through market shifts or unionization.
All the productivity gains are first transferred to the consumer(because of market dynamics) and then(by the market winners) to shareholders. The workers' wage market is not related to productivity but how the company is internally organized is linked to productivity.
It doesn't seem like you've really replaced anyone with this. You spent 8 hours doing the work that you could have paid an SQL analyst to do in much less.
Unless you're saying that your time is worth less than you'd pay the analyst?
Put the requests in a queue. Have the bot generate the response. Then forward the response to a human analyst to double-check. A human can surely double-check a response much faster than they can produce one from scratch.
In many professions, it is common to have junior staff members do the grunt work, and then the more senior staff just review their work and either sign off on it, correct it, or send it back to be redone. You could use the same pattern here, replacing the junior staff with an AI, but keeping the senior one.
The ones for which I would refer the question to GPT. We are still in control of which questions go to GPT/the intern analyst (less critical ones, where a fraction erroneous are okay) and which go to the resident expert analyst.
This is a good point. If the users know the difference the costs and benefits between using GPT and not using it then it certainly has value if those users are also willing to accept that not every answer needs to be 100% accurate.
In my experience business people often have a 'nose' for the right number and will bluff it out if the numbers are wrong and they're challenged.
Blue sky things or stuff you're putting in the annual report should be left to hoomans IMHO.
If there are extensive test cases with static dataset, this may help with query modifications (optimize query, fine-tune, etc.) Of course, this may not feasible for new queries as you can't have test script until the query is ready.
They built a bot which can answer any number of questions, each of which would have needed some analyst time. Given that the analyst rotation was an entire day once every N weeks, and the bot took 1 day to make, this is going to pay for itself after 1 week.
This all assumes that the bot doesn't need tweaking for every answer — i.e. it gets at least some answers right without needing modifications to the bot — which appears to be the case based on the examples in the post.
But, generally and unless there is a glaringly wrong result, only an analyst is going to know if the bot is right or not... what exactly does that gain you?
Maybe it's not a position where it is critical that all answers are 100 % accurate. Maybe getting it right every once in a while is enough to pay for the GPT compute time, but not really for analyst time.
Seems like the issue would be you'd generally get results that 'look' right but would never know if they were actually right without going through and... analysing them
I'm saying there are applications where you don't have to know! As long as the fraction incorrect is less than 50 % and you have 2:1 odds on the consequences you don't have to know which 50 % are incorrect.
Really? Then why pay for the thing in the first place. Why keep the data and make the queries if the results don't actually matter? I'm impressed that you have the ability to envision such a possibility, perhaps you can use that ability to come up with something reasonably likely as opposed to "conceptually possible".
I'm trying to think of easy examples. You're right that none obvious come to mind. I'm sure there is a sweetspot where we can make more money from cheap-but-sometimes-wrong GPT queries than paying for an analyst to be more definitively correct, but I'm tired and a bit fuzzy on the exact parameters.
I'll continue to think about it and write something up!
Hard to imagine a business model where you sell that data that is only right 60% of the time. Maybe in a world where the other best data is even less reliable.
If you're OK with garbage data you don't need ChatGPT - you can probably make up plausible data on your own. Unless you're building some lorem ipsum stuff.
I might not be okay with only garbage data, but data that are correct 60 % of the time may be good enough for some use cases, when it can be had for 1/50 of the price.
I haven't found GPT this reliable for coding. I've been maxing my hourly usage of ChatGPT since it launched and then switching to CoPilot and I have lots of good things to say about it. But reliability is not one of them.
It has a tendency to ignore instructions, as mentioned, but also to get hung up on certain approaches or to use a different approach each time its asked. I'd guess it's very reliable for text generation. But for code, I'm pretty sure the quality of the result would vary quite a from instance to instance.
This could very well cut the work needed greatly. But it doesn't come close to replacing anyone. ... Yet. Give it two years.
I gave up on ChatGPT for code generation because I ended up spending more time tweaking prompts/fixing outputs than if I had just written it myself in the first place. I think this is probably the future of "coding" but it's not quite there yet.
The UX of CoPilot is a lot better. It feels like a smarter version of autocomplete.
They're based on the same GPT3 model so the quality of suggested code is similar but the ability to accept/reject suggestions based on tabbing in CoPilot makes it much less hassle to use.
Same here. ChatGPT kept coming up with syntactically plausible Java code. However, it kept using library methods that plainly don't exist for specific fields.
Wasn't sure of their proposition/hadn't heard of them.
> Run and deploy web apps, task queues, massively parallel compute jobs, machine learning models, GPUs, and much more with a single unified framework for code and infrastructure.
Anyone who's been asked more than a couple of times for data that requires a non-trivial bit of ad-hoc SQL will know the sinking "oh shit" feeling that comes when you subsequently realise you borked the query logic in some subtle way and have accordingly emailed out a completely bogus answer/report.
From the article it doesn't seem that GPT is significantly better or worse than a human in this regard, although an experienced analyst would over time decrease their number of such errors.
The best fix imo is to slather a battery of views over your data to minimise the risk of getting the joins wrong, and it'd be interesting to see how that approach could improve the bot's quality.
I've done too little SQL to be traumatized by it, but I experienced it plenty of times doing more complex data processing on the backend.
I would e.g. spend a week on first discussing a solution, then researching both theoretical background and relevant libraries, and finally writing code to do the math and render the answers (probability distributions). I eyeball them, they seem to all match expectations, so I declare victory, post some nice screenshots to team Slack, and close the ticket.
Feeling high on success, I expect the weekend to start nicely - except, winding down on that Friday evening, I follow that little thought at the back of my head, that I realize was there since the start, quietly poking for my attention. I run one more sanity check, different from the ones I did before. The results come out completely wrong. Only then it dawns on me that I made a terrible mistake early on - I misunderstood how to implement certain step and forgot about a theoretical assumption that must hold. And so, instead of a relaxed weekend, I now dread having to explain to everyone on Monday that a) I actually need another week for this, and b) the nicely looking graphs they saw are all pure bullshit, and the things that looked correct only seemed so, and only by accident.
After a few such cases - including one where the higher-ups were nonplussed, because after I corrected the mathematical errors someone made a long time ago, the pretty graphs that looked nice on the demo stopped looking so nice - I learned to sanity-check my understanding of the problem as early as possible, and then also sanity-check the code I'm writing, and then keep testing as I go. It's all too easy to write "trivial" code that "obviously" does what you think it does, only to realize some time later that it's throwing out an entire dimension of the input data, and nobody noticed because test examples are too trivial, results look roughly correct, and no one is paying attention.
The traumatising experience of SQL is that the scenario you describe being scale-invariant. The smallest question and the largest question both have the potential to be completely wrong for any inane reason. Often turns out that metrics the company has been using for months are totally bogus too in my experience, which is not a fun conversation to have.
So the life of an SQL analyst in their early days (it gets better as they learn to trust nothing) is either continuously pumping out garbage or constantly needing to go back and, very awkwardly, correct and redo work that people thought was finished (and those people were often happy with a bad result because it looked like something interesting was going on).
This article is highly misleading, approx. none of the time an analyst spends is working on writing 5 line queries. The analyst is the person who knows that the data table for the series A and B rounds has different semantics depending on the data round because someone screwed up the table design beyond recovery and therefore a plain COUNT(*) will give incorrect results. This tool would let people who don't understand that pump out garbage faster. That might be a win regardless, most analytics requests cannot possibly be value-add.
It's so refreshing to read this discussion. I've had these thoughts and feelings a lot but this is the first time I see them reflected in other people!
As an Analyst your asset is to own the data and know all the odds and quirks.
Very similar to how developers need to own their codebase. Yes you can ask chatGPT to write a complex function. But your job as a developer is to see how it fits in with the rest of the framework. The logic is the simple part.
As many pointed out you also need to be brutally honest with the numbers and also the problems you may have caused or fixed.
Trying to hide any problems will very quickly throw you downhill, into an extremely uncomfortable position.
Although before it acquires feelings and a sense of job safety, I guess chatGPT can be better at that part.
Your anecdote gave me an idea. What if we use AI to enhance the communication between people instead of replacing the people?
Imagine you give a problem statement to the LLM, then it tried to find all the subtleties, confusing parts, room for interpretation, etc. All of the people involved revise the problem statement to make it more clear and give it back to the LLM. Iterate like that until satisfied
Hm. I like your idea! It's one of the few that I think have a chance of working in practice.
In my experience, there is great value in describing your ideas to people who don't have the background to fully understand them, don't really care about them, and are in a bit of a trollish mood - the ways in which they misunderstand what you're saying, or pick on (what you think are) random, irrelevant things, is highly informative. The feedback you get from such people makes you think about things and in ways you wouldn't have thought otherwise.
The problem is, of course, you generally don't have a pool of such people available 24/7. However, LLMs today seem like they could fit this role just fine. They can't and won't understand the point you're trying to get across - but they can and will (if asked) pattern-match on vague language, logical jumps, sentences with multiple meanings, etc. They'll get it subtly wrong, too - much like a bored friend who's mostly focused on playing Angry Birds and hears only every third sentence of your monologue, and then blurts something to keep you talking for the next 30 seconds so they can focus on aiming the birds at the towers of pigs.
I would totally use a LLM-backed tool optimized to facilitate such conversation sessions. I actually tried this in the past, in AI Dungeons, and results were encouraging (i.e. responses got me to think in ways I normally don't).
If you’ve had some success with an existing model, I think I’ll explore the idea with GPT3! Getting the prompt right is gonna be tricky, do you remember how you got AI Dungeons to play along?
I'd set up a basic story prompt, describing the roles and personalities of my character and a few NPCs, as well as relationships between all of them. I would then introduce my idea in form of a role-play, usually with my character blurting it out - then let the language model fill in reactions of other characters. From then, I'd just play along with the story.
It wasn't the most efficient way of extracting commentary out of a language model - particularly out of one that was optimized to generate plot twists and setting changes instead - but it was a very fun way.
I really don’t know why you feel the need to point this out when I made no claims about feasibility. I think it’s worth exploring even if our current tech isn’t up to it.
When I was dealing with something similar and I didn't know how the result should even look like, something I did a few times was to separate some sample data, implement the idea using different technology than I was using for that project and then compare the numbers. Helps with debugging too.
SLPT: on the Friday night you can quieten the little voice by drinking beer, and then you will promoted faster.
If you really want to play for keeps then on Monday you can explain that you had a fantastic idea over the weekend for how to improve the results and make them more accurate.
I've been there except instead of explaining to everyone on Monday, I didn't spot my mistake for two weeks _and none of the report users did either_. Nobody died but it wasn't my finest hour.
Lol, countless times that I ended up with the right solution in my head while taking a shower or the day after, because after chillin my mind found that the previous code was wrong.
To me this boils down to Churchhills adage "Don't trust a statistical report unless you fabricated it by yourself". There are so many subtleties when translating business questions into SQL e.g. was the average deal size in the post's "trend" example computed using average/median/geometric mean formula? Should the trend only contain historic or also projected values. No serious report comes without data definitions, then its an exercise to the reader to spot flaws there.
I had not heard that Churchill quote, but now I feel vindicated that I take the same approach as him!
I think this can also be extended to data tools. I've had so many experiences where a SaaS tool that tries to automate away the manual SQL analysis shows something totally wrong. I tend to "Not trust, verify" anything coming from a SaaS UI when serious business decisions are about to be made based on it.
One needs to be careful with this approach in terms of query performance, though. Using simple views with a couple of joins and some filtering is fine, but be very wary of stacking more than 1-2 layers of views calling each other, and especially of using things like aggregates/window functions in views, if these views then are then used as building blocks for more complex queries.
That's a recipe for breaking query optimizers and ending up with very bad query plans.
Use case dependent. When I am tasked to generate some ad hoc analyses, performance is a non-issue. The query is only going to be run the handful of times while I iterate on the idea, and I would much prefer some convenience views rather giving a hoot about optimal query planning.
Simple views are perfectly fine - it's mostly nesting of views with aggregate functions and other complicated stuff that is bad.
And if ad-hoc is a big part of what users are doing with an app/database and you don't care about performance, your angle sounds reasonable.
As an app developer/development DBA, I care mostly about performance of the queries that are known at development time, though, so I'm a bit biased.
> doesn't seem that GPT is significantly better or worse than a human in this regard
Probably, but I think a human is much more likely to realize that they made a mistake and good-luck convincing GPT that it's wrong about something non trivial.
This is an exceptionally deep, interesting, and important statement. Is it true? Does GPT-3 occasionally make spelling errors for instance? Is the rate of spelling errors higher or lower than the average human? How about versus the average in the training data? How does the prompt affect the frequency of such errors. Ditto for other kinds of objectively measurable error.
> The best fix imo is to slather a battery of views over your data to minimise the risk of getting the joins wrong, and it'd be interesting to see how that approach could improve the bot's quality.
You're correct, but organizing queries into views and CTEs are an implementation detail that would be done by the dev in practice, but would not show up in a stackoverflow answer because it's probably not relevant to just answering the question.
I love how language equivalences such as informal vs formal code can throw off the quality of what the chat bots generate!
Most times, it acknowledges its error, and sometimes it successfully corrects it. That is, IF I give it specific enough information (ex: "why do you treat this variable as a string, it's a number?", or "don't you need to unpack the json first?").
Other times, it fixes something, but breaks something else in the process, and no matter of coercing makes it go to where I want.
But, yeah, the "confidently wrong" aspect needs to be kept in check. I tend to treat it as any other source of information... with a grain of salt.
What's neat with code is that you can prove the "truthiness" of its affirmations pretty quickly.
>> The best fix imo is to slather a battery of views over your data to minimise the risk of getting the joins wrong, and it'd be interesting to see how that approach could improve the bot's quality.
Yes, the success of the bot's query (at least on GPT-3) is a combination of (a literal and good) prompt and how well did you explain each of the columns to the model.
People dismissed the XP book by saying it was just stating practices that already existed. Articulating and selling an idea is a huge part of the challenge. Fifteen years later if you're having any success at all with Scrum it was because your team adopted the half of XP that didn't outright disagree with Scrum.
I liked the Refactoring book because it gave me names for things I figured out in 1996 (unfortunately while using VI as an editor, gave myself RSI in the process). It was part of an education in the value of books that help you only by existing so you can hand them to people who ask a lot of questions.
I had a boss who was big in the Kanban 'movement'. We got along famously, in part because I re-invented a Kanban board in 1994 when faced with a convenient whiteboard for the first time.
You can do a lot of amazing things with decent first principles. You can make some amazingly large messes without them.
> I had a boss who was big in the Kanban 'movement'. We got along famously, in part because I re-invented a Kanban board in 1994 when faced with a convenient whiteboard for the first time.
I was showing slides of a pre-1900 engineering office as part of a presentation a while back when someone shouted out "look - they're doing kanban". And sure enough there was a pinboard with process headings on it and pinned items the background.
It would not surprise me if they found one in Pompei
You didn't replace a SQL Analyst, you just gave them a query generator. End data consumers don't understand the data model, assumptions, quirks, etc. If they fire the analyst, they are going to wind up drawing a lot of bad conclusions on anything more complicated than simple aggregations.
Maybe 5% of the business/operations people I've worked with would even want to do this. The rest prefer a smart human to walk them through building a report - most of the time this is because they don't actually know what they need, and they actually need an expert to sit down and figure it out with them.
You nailed it. At the beginning of my career I thought that success and value to the organization was all about technical skills.
30 years later I now understand that most successful projects need people with modest or average technical skills and outstanding communication skills.
It doesn't matter if you have super-genius engineers; if the business people don't really understand the problem that they're trying to solve then you're going to end up with a crap solution (may shiny, fast, and beautiful, but still crap).
When I was new, I saw one of my more experienced colleagues ask a few questions that together saved the company more than $1 mil each year. ChatGPT might be a threat to automate some low-level tasks or help eliminate bugs, but it is nowhere near ready to evaluate the context of a system, understand its history, or think* through the consequences of a major business decision.
Though if it comes an AI with the capacity to include more context (ie: all company financials, communications, market analysis, etc...) it might be even more effective than a human with precise context.
Communication might be strictly email in the future. Or something that could be pipelined into the "AI" for context. Video/Calls might make it too at some point. Face to Face meetings strictly prohibited.
I agree with you to a point, but I think the only reason that it can't understand the context of a system is because it hasn't been trained on that system's code and documentation, which is obviously a future coming soon.
I'm not sure training these models on code and documentation will make that much of a difference. These models struggle significantly with subtlety, relevance, and correctness. It also doesn't have a theory of its own knowledge or confidence, and so tends to "hallucinate" and put out confidently-worded nonsense. Especially for complex and nuanced topics.
A big part of my job in software is having a very sharpened grasp of my ignorance, the ability to weigh a variety of tradeoffs, and the ability to convey my confidence of my abilities and my team's abilities. I'm not sure this is possible for this generation of AI.
The hallucination part is due to a lack of constraints. The AI can recognize constraints, but it can't recognize what it doesn't know for lack of context.
Prohibit all physical meetings. Force all communication through mediums that can be pipelined. Feed everything (accounting, contracts, law, etc...). Work will be then to architecture the AI to produce the optimal response.
The first company to figure this out will be too ahead. I don't think anything will be anywhere close to compete including nation states.
The same is true in technical sales. Obviously you need a sufficiently technical background to be able to _do_ stuff, but the primary value you bring is probe further than "we want to do X". The prospect / customer has mayybe done what they are asking twice, you've seen it done 100s of times.
Yeah, that is an issue a friend of mine has right now with his team. He randomly ended up as the manger of the data analytics team despite no analytics background (or really much programming background either). And one of his main frustrations is that the analysts do not understand the data model or the business. Before he became their manager his team just took report requests and then wrote some SQL and delivered some numbers. Without understanding what those numbers were supposed to mean or the quirks of the underlying data.
Clearly, the correct move here is to replace the entire team with Looker and let every department head create their own dashboards to track arbitrary kpis with no understanding of the product or underlying data model. /s
Who needs a data model when you can just dump a bunch of excel sheets and database dumps in a shared folder, call it a data lake, and encourage anyone who has a report to just throw it together in the free version of power bi?
The correct move seems to be to replace the team with a set of oracle bones. From the very limited information we have it seems to me like the team was used to justify whatever conclusions were already made before consulting them.
The whole practice is kind of a proof-of-work scheme for credibility and liability laundering: even if you intend the team/company to follow your conclusions regardless of what anyone else thinks or says, getting some analysts or outside consultants to burn non-trivial amount of time and money evaluating the situation before rubber-stamping your proposal, is what may be necessary for you to sell your ideas to the rest of the team/company. Such exercise may be especially important if you want to protect yourself from having your head served on the platter after your hare-brained idea fails spectacularly.
Honest question: how rampant is this kind of bullshit politicking in tech companies, and how does this differ with industry, size, company age, etc.? My friends who have worked at big names like F and T have described the cultures there as following: "it's as if everyone who studied too much in high school and college missed all the parties and bullshit, so they're making up for it 'on campus' here". And the stories from folks at G are not really better, though allegedly they tend to exercise politics more in the quiet-sabotage style. I feel like I'm describing races from Star Trek, company cultures are weird.
And it is not really punishing, yet at least. He likes his new job despite the mess he was put in. He loves it because he learns so many new things (analytics, how the be a better manager, etc) and how he has managed to make his team actually deliver a lot of value to the company in spite of these problems. It is possibly that the challenge will be too big for him and it will burn him out but so far he likes it.
And, no, it is very unlikely to be intended as a punishment. More likely he got it because he was too interested in getting good reports. The usual accidental volunteering where if you are too interested in something you end up doing all the hard work.
the larger the company the worse this becomes. When you hit mega level like government the IT side barely knows what the business does and has no idea what anything means. If you're lucky they may know how to support your application from completely failing. If you need any major development you need to compete against everyone else in the government for IT resources.
This is a key learning I had to learn as a manager: people learn what they need to turn their inputs to outputs well enough not to be fired. It's management's job to internalize the business and its needs and make sure that the individual contributors have an accurate mental model of the world.
I wonder what it would look like if we had people across the business working on the same problem together rather than a game of telephone, which is how these data requests end up.
You can easily add more components to start thinking about models/assumptions/etc, like adding interfaces to Data Catalogs (i.e. Glue Data Catalog).
As part of a POC I made, I built a similar bot without recursion for debugging and iterative query building though. It does the following:
- It predicts most probable entities from the question.
- Searches AWS Glue Data Catalog for the most probable/useful tables.
- It builds an Athena SQL Query from N most useful tables.
It obviously get it catastrophically wrong sometimes, but hell, it was a 3 hour POC. If you can make better indices that map entity->table relationships it should get better at searching tables. Add this kind of recursive/iterative debugging of queries, and you get at least something near a junior-level SQL Analyst.
These kind of bots are analogous to Stable Diffusion, they DO need a good prompter/puppeteer/solution-verifier. Most non-senior Data Analysts also need one anyways.
Thats not really fair to the human. The time to learn basic SQL may be longer, but the algorithm will never develop intuition or independence, while many junior analysts will eventually operate effectively on their own.
It’s a neat tool for analysts as a query generator - I would use it in situations where I’m not familiar with the schema, but it would become less useful as I learn.
Yeah, maybe I should've said "first-weeks/months" level junior data analyst.
But hell, as an analyst I would've paid a lot for a tool that searched intelligently through giant datawarehouses (or whatever consultants call them now) and at least gave you probable matches.
Now that same thing exists and you can even finetune its "DSL" towards your own organization.
We're building https://www.olli.ai/ to help with this. User's can ask questions in natural language, but more importantly, Olli can suggest what questions to ask based on the data.
But it does lower the bar for asking questions a lot. Before you ask a human you have to do a sanity check. Most obvious is asking the same as last week but it could also be something readily available elsewhere.
The problem is that you never really know whether the chat bot gets it right or terrifically wrong unless you already know how to do the task yourself.
And in some cases, paying an analyst $50/hr. for a higher degree of confidence than you can get from a $1 chat bot is still very much worth it.
The stakes are higher, too. If the chat bot gets it wrong, what are you going to do, fire it? There goes a small trickle of revenue to OpenAI. Whereas if the analyst gets it wrong, there goes their livelihood.
That said ... this will help the $50/hr. analyst improve their productivity!
In my experience, in many cases, the so-called "first draft" (of code, content, etc.) is more just helpful for a mental warm-up. It isn't actually suitable.
The time it takes to write/code/etc. isnt in the typing, that's trivial. It's in modelling the problem (, audience) and solving it.
ChatGPT isnt in the game of modelling problems and solving them, it possesses only superficial semantic characteristics of the questions posed
That its answers satisfy the requirements would require ChatGPT to ask you questions (and much else besides)
> ChatGPT isnt in the game of modelling problems and solving them, it possesses only superficial semantic characteristics of the questions posed
You can give it the datamodel and a semantic description and it will do a pretty good job of reprojecting your data.
People talk about chatgpt as being a fixed thing, it isn't. ChatGPT is the Apple II of LLMs. 3-4 moore units from now every issue that people think is a "gotcha" for chatgpt will have evaporated.
In your analogy to the Apple II, is there a fundamental problem people claimed for early computers, which didn’t just boil down to “they need to be faster”?
There are fundamental limitations to what an LLM can achieve.
True, I mean, just look at all of the examples of technologies that were projected to revolutionize computing and machine learning "..." eventually, here:
Just because they haven't happened to yet doesn't mean they never will!
I've heard very good things about expert systems and IBM's Watson for Oncology.
The Harvard Business Review has a great article on all of the eventual "..." AGIs here:
Integrate it with wolfram alpha and Google, tack on a bullshit detector and spank dispenser conetwork and it looks like limitations of LLMs are easily overcome by not running anywhere near them.
Some of the problems with ChatGPT are not mere “gotcha”s but rather problems inherent to how LLMs work and their philosophy. The biggest issue I’ve seen some other scientists [1] who don’t have a stake in OpenAI bring up is that: “LLMs make sh*t up”.
Maybe LLMs are not the final answer. For now, it looks like they are not.
They are already here. We can haggle over error rate and incidents per million miles or what have you but that is more an argument about safety and suitability.
The similar thought I had yesterday was that ChatGPT is pretty good at generating sample code that doesn't otherwise exist. It can be really useful if you treat it like that.
Of course I've also had it call functions and use constants which don't exist, and fail to correct itself even after repeated prompting. So even that only goes so far.
It relies on test availability of good test cases to verify the code functionality: they generate many, many possible programs, throw out all the ones that aren't syntactically correct, compile the ones that are & then test them. This works reasonably well for programming competitions & the code generated will probably improve over time.
The issue for "real-world" problems is that you can't verify the code without the test cases & writing good tests that ensure the code does what you want can be as much work as writing the code itself in the first place - the tests form a kind of mathematical co-domain to the code after all & only a small subset of problems are so simply defined that you can pin them down with just one or two tests.
Imagine you keep someone on staff at 50/hr, or you just consult based on need, i.e. 15 minute consults are 1/4th the cost, and you simply just ping them to double-check queries you've ran, that look right but maybe could be optimized. You've just cut back from maybe 20/hours to 8 per week needed for that freelancer.
That's exactly what it is - a productivity booster. The point remains the same: it's going to theoretically put some SQL analysts out of work. If the $50/hr analyst is 10x faster at his job, you can replace 10 analysts with just 1.
That's about $100k salary. You can hire a rock solid data analyst at that price probably anywhere in the US. I know a couple talented analysts making ~$110k in NYC.
In early 2018, I made $105k salary as an entry level "Business Systems Analyst" in San Jose, CA.
Rented a studio in Milpitas for $1200 at the time.
I built a data integration system with NodeJS, and marketing/customer service enterprise software REST APIs.
It was my first professional programming job. No STEM degree, but a few IT Courses, plus digital marketing experience, plus self-taught full stack web app experience.
Feels extremely overblown. Very impressive, don't get me wrong, and is yet another bridge closer to letting business side do plain text queries.
But that said, stuff like this has existed for awhile, and it's certainly not replacing analysts. For a start, this assumes a VERY clean model, which is of course what you strive for, but the moment you step out of FAANG you'll find that many places have a decent, but not ideal, model/warehouse/db which only makes things like this worse.
Getting the simple examples right is interesting, but hardly replacing an analyst. A basic background around tech and some skill navigating w3schools would replace an analyst if this was the bar.
The mildly complex queries being "almost" right is frankly horrifying. Others have mentioned how much of a "oh fuck" it is to send out bad data, and having some AI generate your answers an be SLIGHTLY off is basically worse than it being obviously off. You now have to check every query it generates for errors, because you won't be able to glance check if its sane or not.
Finally, at least where I work, these probably wouldn't even qualify as medium complexity queries? I don't really know how the industry at large would qualify these, but "using 2 inner joins, a where, a group, and a order" all falls under beginner skills in my eyes?
Its still fascinating, and I'm glad they showed HOW they did it, but this isn't replacing anyone, at least if you're paying taxes off the data.
Even in FAANG, most data is going to be extremely messy.
* There's usually very little incentive for good documentation, so you have to guess what the table contains based on column names and exploring patterns in the data.
* Enforcing a single pattern/process for data across the entire org is a pipe dream.
* Logging tables are written to by hundreds of engineers across dozens of teams, with no standardized naming scheme.
* The product contains multiple edge cases/special overrides for very specific circumstances, so there's very few simple queries that don't come with 100 footnotes attached.
FAANG is not immune to large-organization problems, and data quality is no exception.
Can't say speak for FAANG, but can confirm for a $100B+ business that analytics was a huge mess. There was constant investment on both the engineering and analytics functions to wrangle a coherent view on top of the underlying operational data model.
Oh yeah im not surprised to hear that, Ive just known one or two people who've been in similar sized companies as analysts, and while the underlying table structure was a nightmare, the data model they dealt with was pretty clean.
But with that in mind, thats because theres a major pipeline of people and processes to get the data to that point, and it meant that there could sometimes be significant delay on new KPI's as they had to be cleanly worked into the model, and it of course didn't represent everything.
Sure, if that was the title of the article. Instead its saying its ready to replace a full time position, which its not even close to.
And this is something people have been saying about things like AI and crypto for years now. They've both made some interesting developments, and they both face serious hurdles (some fundamental).
AI obviously has a ton of real world use cases, but overhyping things like this doesnt actually help.
A car is mostly an amalgamation of mechanical devices and we have a solid track record of steadily and incrementally improving those over dozens of millennia.
Modern cars also have a lot of computing power in the form of fairly dumb systems that are good at sensing things and following rules e.g. "based on conditions X, Y, and Z the flow of the fuel injector nozzle should be N". Got a solid track record here too albiet one that's decades long and not millennia.
A tool that replaces a human data analyst needs to do a lot of inference. When I'm translating stakeholder requests into code and reports, it's about 10% coding and 90% human inference. Remains to be seen how good machines can get there.
Large language models are very intriguing as a technology to replace business people. Make a few PowerPoint slides, fill in an excel sheet. Generate a summary for a pointless meeting.
I can think of at least six possible answers based on these questions:
1. Does largest mean dollar amount, or number of investments?
2. Would number of investments count companies invested in or funding rounds invested in?
3. Does largest mean the largest total dollar amount invested in 2022, or does it mean larges dollar amount of new investment in 2022?
It looks like ChatGPT chose the query to mean the investors with the largest dollar amount of new investment in 2022.
When you expand your natural language query to clarify all of these ambiguities, how far away are you from a SQL query? I am not sure, but I think that you are getting pretty close.
Right now, you need to first ask ChatGPT to suggest disambiguations to the question that you're presenting it, then ask the question in that context:
User: What questions can I ask to disambiguate the question: "Who were the largest biotech investors in 2022?"
ChatGPT:
Here are a few questions you could ask to disambiguate the question "Who were the largest biotech investors in 2022?":
Are you looking for individuals or companies that invested in biotech in 2022?
Are you looking for the largest investors in terms of total investment amount or the number of biotech companies invested in?
Are you looking for investors in a specific region or globally?
Are you looking for investors in a specific biotech subfield (e.g. biopharmaceuticals, medical devices, etc.)?
Are you looking for the investors at a specific point in 2022 or throughout the entire year?
Asking these questions will help to clarify the scope and specific details of the information you are looking for.
User: How would I disambiguate the term "largest"?
ChatGPT:
There are a few ways you could disambiguate the term "largest" when referring to biotech investors in 2022:
By investment amount: Ask how much money was invested, this way you can determine who made the largest investment in terms of dollar value.
By number of investments: Ask how many different biotech companies the investor invested in, this way you can determine who made the most investments.
By percentage of ownership: Ask what percentage of ownership the investor has in the biotech companies they invested in, this way you can determine who has the largest stake in these companies.
By market capitalization: Ask the market capitalization of the biotech companies the investor invested in, this way you can determine who has the largest investment in terms of market value.
By asking more specific questions, you can get a better idea of what is meant by "largest" and ensure that you are providing the most accurate and relevant information
I love those ChatGPT projects! Of course it's silly and nobody is really replacing somebody with a program that confidently get half it's answers wrong. But it's fun to just naively let ChatGPT solve the problem.
But I wonder what it's going to look like in a few years. Currently, it's really just a demo that got surprisingly huge traction. I think the most pressing problem is not to get ChatGPT smarter but to get it more reliable. I think more realistic use-cases would emerge if we could build systems that have a better understanding when they are out of breath. I don't think this needs a revolutionary breakthrough just more science.
> nobody is really replacing somebody with a program that confidently get half it's answers wrong
I gotta tell you, I use copilot to help with my coding, and it still sends a shiver down my spine when it writes the entire database migration based off of the filename, or writes 4 pages of automated tests that work right on the first try.
I'd be impressed if AI wrote a test suite that immediately found a bunch of valid bugs. I'd be nonplussed if it wrote a test suite that "passed" on the first try. The presence of passing tests does not imply an absence of bugs, and the hard part of developing a test suite is the test design, not the implementation.
lol, fair enough, but I would never commit a line of code I hadn't carefully read myself. my new class was basic, so the new unit test was basic as well, but it was thorough and correct. And I exaggerated a bit; I did write 3 or 4 comments to give the ai some idea of where I was heading.
I've had it freak me out with how prescient can be, but it also gets a lot of things subtly wrong, like its programming against an API similar to the one in my application, but not identical.
All the NoCode and LLM stuff feels like this though - it works well for simple demos, but is useless for the complexity of the real world especially if errors are costly.
We looked at using all sorts of "AI" to write SQL based upon natural language prompts. As far as I am aware, the state of the art is still nowhere close enough in accuracy for us to lean into as a business.
Ultimately, I don't think we will get there with semantic analysis or GPT-style techniques. There is always some human factor involved with whatever schema is developed, so you would probably need an AGI trained in the same business as whoever is being replaced by this thing.
This is great~ There's been some really rapid progress on Text2SQL in the last 6 months, and I really thinking this will have a real impact on the modern data stack ecosystem!
I had similar success with lambdaprompt for solving Text2SQL (https://github.com/approximatelabs/lambdaprompt/)
where one of the first projects we built and tested was a Text-to-SQL very similar to this
Similar learnings as well:
- Data content matters and helps these models do Text2SQL a lot
- Asking for multiple queries, and selecting from the best is really important
- Asking for re-writes of failed queries (happens occasionally) also helps
The main challenge I think with a lot of these "look it works" tools for data applications, is how do you get an interface that actually will be easy to adopt. The chat-bot style shown here (discord and slack integration) I can see being really valuable, as I believe there has been some traction with these style integrations with data catalog systems recently. People like to ask data questions to other people in slack, adding a bot that tries to answer might short-circuit a lot of this!
We built a prototype where we applied similar techniques to the pandas-code-writing part of the stack, trying to help keep data scientists / data analysts "in flow", integrating the code answers in notebooks (similar to how co-pilot puts suggestions in-line) -- and released https://github.com/approximatelabs/sketch a little while ago.
"Hi Dave, the query was taking too long so I optimized SQL query by adding the line `DROP invoices;` It has improved performance significantly. So far there are no orders to examine."
I almost missed the cleverness of this joke about making a dumb SQL error of dropping a table to solve a problem, until I had nearly finished writing my correction post for the two other SQL errors buried in the joke (DROP vs truncate, orders vs invoices)... when the word "Dave" kept nagging at me and I realized the other two errors were the point of your joke.
Nice one. Thanks. :)
— someone who corrects others’ SQL regularly
P.S. Based on my response to this post, ChatGPT3 spam posts on forums containing subtle errors may ironically actually increase online engagement as people write correction posts to hidden ChatGPT3 ones to help others / make themselves feel smart when "someone on the internet is wrong!"
While this is very cool, SQL was designed to be used by business people. We need to go back to that model, where we train the business people who need these analytics how to use SQL to uncover the result. That along with a rigorous policy for including the queries that produced the result so the query logic can be checked would go a long way to actually taking advantage of the data we're collecting as businesses.
My experience is that if you become proficient enough in Tableau or PowerBI, you simply become the go-to guy for those reports. Suddenly senior management starts bombing you with requests for various reports.
But in the end, all roads lead to Excel. No mater what tools you use, someone higher up will still request "But could I get this in Excel?"
As an engineer myself who likes to dig into data/analytics, writing and tweaking some of the complex SQL is not that easy or fun. I am impatient to get to the insights.
I imagine SQL would be harder and more annoying for biz folks. A programmer might be right "expert" to get the SQL working (Usually the programmer itself is a generalist when trying to write SQL for analytics).
A text-to-SQL tool will help create the queries quickly; it is much easier for an engineer to verify the generated queries and the results of those queries.
Yeah, not saying it's easy, or even that it's possible in every case. Most of the time I have failed to get it to happen- but recently I've seen more non-technical stakeholders take it up and use it successfully, so it is possible.
Sorry friend, but this is a bit out of touch. Maybe that was the original design intent of SQL, but understanding the application's data model is beyond a lot of sql analysts, much less their business partners.
> but understanding the application's data model is beyond a lot of sql analysts
You need a better data model or better data analysts (probably the latter).
Putting analysts aside, I do agree though that the data model can be too complex for non-technical stakeholders, but in the vast majority of cases, the data model is simple and stakeholders are looking for basic statistical analysis and trendfinding.
Also, if a moderately skilled human SQL analyst doesn't understand a data model, well GPT has no chance.
But you don't want to do reporting on an application's data model, that doesn't scale: as soon as you have a medium-sized business, the data will not be confined to a single application (unless that application is named SAP). Understanding the business data model is a requirement for SQL analysts anyway, and once you have the application data transformed to the business model, the business users will have no trouble understanding it.
Thankfully you don't need to design a schema to use it. And while some schemas may indeed be too difficult for non-technical stakeholders dabbling in SQL, I think a huge percentage of them are not
If you’re willing to accept unverified results from an AI chat bot, you may as well just let the end user make their best guess using a query builder himself. My company requires that any queries used for official reporting or provided to the exec team get blessed by the data sciences team to avoid errant data from bad queries, I’m not sure an AI chat bot would remove this need.
davinci-003, ChatGPT, and others can be great tools. But they often give you exactly what you ask for (or at least try to) and a large part of writing SQL queries for analytics is figuring out what wasn't asked for but should have been. Good analysts will find outliers, data-smells, and ask questions rather than rush to returning an answer.
> Playing around with GPT at this level you get the feeling that “recursive GPT” is very close to AGI. You could even ask GPT to reinforcement learn itself, adding new prompts based on fixes to previous questions. Of course, who knows what will happen to all this when GPT-4 drops.
Leaning out of the window way too much here. This has nothing to do with AGI, which would require an intrinsic understanding of not only SQL, but over, well, everything, not just a well-defined and easily checkable field like SQL.
Regarding GPT-4 - OpenAI‘s CEO Sam Altman stated that the expectations regarding GPT-4 are way over-hyped. People on the Internet talk as if AGI is coming in the guise of GPT-4, but it‘s „just“ going to be an incrementally better evolution of GPT-3.5.
Mind, I‘m in no way saying that LLM‘s aren’t exciting - they are to me - or that they will not change the world, but leave your horses in the stable.
The article describes how the prompt itself is constructed - to include details of the database schema plus some example data rows, followed by the user's question.
The prompt is then sent to the GPT-3 API. The results are then sent through the API a few more times with some extra prompt engineering before deriving the final result by running the SQL against the original database.
How is it being fine-tuned to be able to generate relevant SQL queries on the crunchbase data set? How does this jibe with the article saying they imported 2.5m records from crunchbase? Where and how are you able to import or fine tune the API-accessible GPT-3 this way? Any pointers for where I can look to do the same?
Edit: I noticed you edited your comment above with some details. It seems that he used GPT-3 just to create the SQL statement which was then executed locally. So GPT-3 never had the actual data set, just a prompt that included some details of the schema.
As a result I think any sort of "success" here is fairly limited when GPT-3 is blind to the actual data and has limited knowledge of the schema to be able to formulate queries that imply some knowledge of the data.
This is genuinely the cutting edge of how you do interesting things with language models like GPT-3 at the moment.
Training these models with extra data turns out to be incredibly expensive and relatively ineffective.
Instead, the most interesting research is all around tricks like this - figuring out ways to round-trip to the language model, then query other sources of data for the information that it needs, then sending more prompts to the language model again.
I've only done two experiments with it myself - training a tagging model on my blog's content and using that to suggest tags for untagged entries - and I found the results very unimpressive fur both a cheaper and the most expensive model.
I've seen a few other people suggest that time tuning GPT is unlikely to give better results than just feeding the regular model a few examples in a regular prompt.
I've yet to see anyone talking about a GPT3 fine tuning project that went really for them. Maybe I haven't looked in the right places.
I've been building something similar that handles the dirty business of formatting a large database into a prompt.
Additional work that I've found helpful includes:
1. Using embeddings to filter context into the prompt
2. Identifying common syntax errors or hallucinations of non-existent columns
3. Flagging queries that write instead of read
Plus lots of prompt finessing to get it to avoid mistakes.
It doesn't execute the queries, yet. For an arbitrary db, it's still helpful to have a human in the loop to sanity check the SQL (for now at least).
This is great, of course. And I think the people who will get the most out of the new AI tools are those who can treat them as iterative assistants. The fact that not everyone can use tools this way has become apparent to me recently. e.g. people who use car driving assistants as if they're fully autonomous; or people who use Copilot and are upset the code is incorrect.
The point isn't for it to be correct, but for it to be so fast that it can be mostly correct and you can fix the last bit.
I use Copilot extensively for my Python glue code and it is positively fantastic. I also use it at my shell with copilot.vim with a quick C-x C-e and write a comment and let it write the code.
The iterative improvement nature of the tool means that I make faster progress. It doesn't have to get things right. It only has to make progress and be obvious how to make improvements.
For instance, I just bought some Reserved Instances (c6i) on AWS and I want to make sure that I don't have any c5 instances in there that I won't be covering. I hit C-x C-e and type in `# list all aws instances in tokyo that are c5` and then hit Enter and type `aws` and it completes the rest for me.
I can then run the query and edit it, or I can validate that it looks okay, etc. The point is that I'm a human capable of understanding what this machine is making. That makes me way faster. I don't need to check Stack Overflow, and the machine teaches me syntax etc. and puts it in my history.
It's the closest thing to the Primer from Neal Stephenson's Diamond Age and I love it.
>When I was at Square and the team was smaller we had a dreaded “analytics on-call” rotation. It was strictly rotated on a weekly basis, and if it was your turn up you knew you would get very little “real” work done that week and spend most of your time fielding ad-hoc questions from the various product and operations teams at the company (SQL monkeying, we called it).
To be part of an analytics team and deliver work like this is actually highly sought after and a great role to have. I don't know why the author thought it was terrible. Doing data analytics on a company's datasets is most certainly real work.
Doesn't take away from the point of the story though, GPT is great.
This just in, ChatGPT has hosed up the read operation due to inefficent querys and not being a human being. 26 GPT promts have been replaced with a DBA, Analyst, Project Manager, Cross functional manager, regulatory specialist, junior programmer, and QA analyst.
Do you all think that GPT and such will see a pattern of usefulness starting with:
1) blatantly wrong but helping to train/give examples to the most rudimentary and beginning stages of people learning a task? (since that's what it's doing at the same time?) I.e. replacing low-skilled intro training, or more charitably, helping to make it possible for far more people to learn something with assistance?
And then moving up the sophistication level to where it's, say:
2) "ok, I can tell this is not blatantly wrong, and might even be plausible from a medium skilled practitioner or analyst" and I can use this with some error checking.
to
3) even more capable / actually worrisome?
Or, does it occupy a different "sphere" of usefulness / purpose?
Meta-Meta question:
Why is "actually worrisome" part of "more capable"?
If worrisome is because of job prospects for an existing career, my advice is to embrace new tech, not fear it. Just keeping up with things even if you don't use it right away is usually enough to stay relevant. Things don't really evolve that quickly, and so as long as you don't resist change completely things usually work out.
Heck even if you hate change, and don't want to change at all, there will be plenty of old legacy companies with the old fashioned manual reports needing a person in a chair to type out the SQL to make them. Just like there are still (good paying) jobs for those COBOL experts today.
SQL is a very high-level language, doing a lot of stuff in very few lines. When I write a web backend, most of the real logic ends up being in SQL. If AI is going to write code that I can trust, it'd probably be SQL first, but not yet.
I wonder if/when we'll get comfortable with the errors that an AI like this makes. Certainly human analysts still make errors, and may be able to explain them (which I think LLMs would have a hard time doing), but what if the overall error rate is less than a human analyst?
I imagine this is sort of similar to our comfort with self-driving cars - what if they make fewer dangerous mistakes than humans? Would we actually prefer _more_ mistakes but having a human who can be held accountable and explain themselves? Are we ok with an AI that makes fewer, but categorically different mistakes?
Your last question hints at another question we also need to explore: do or will humans apply the same level of scrutiny to automatic outputs than they do to human ones? I find we tend to treat them as more authoritative. Are we more likely to unthinkingly accept mistakes by a machine?
As a brief aside, I wonder if we'll see increasing rates of accidents as the rate of AI drivers rises. I think it's hard to think about those effects in a vacuum - they have fewer injury accidents _now_, but will they if they go from making up a trivial percent of traffic to 10%? 25%?
I used to work for a company that paid loads of money to an Oracle consultancy group to do things like optimize queries. Sometimes they'd even do a better job than the Oracle query optimizer :-)
Wonder what the costs are for this per question? I imagine supplying so many tokens for context makes the querying a lot more expensive. Though still no doubt cheaper than hiring another analyst.
> This led to completion chains of over 20 calls to GPT for a single user question. […] it takes 15 seconds and costs $1 in credits vs $50 and 1 hour for the analyst.
I wonder how much more accurate this would get if fine tuned on a set of SQL problems? Could even fine tune it on a per-company basis using queries that had been written by analysts in the past.
It could probably get much more accurate. The idea of using AI/ML to tune databases or optimize queries isn't new; I expect to see more recommendations from DB hosting providers as a premium service.
In orgs where this need is usually present, the data can be massive and it takes some time to understand how it all fits together. There is also the issue of optimizing around indexes or writing queries that are cost efficient (especially if you using Athena/Presto/Big Query). Mistakes here can cost a lot of money or lock up the system so others can't use it.
I love this demo, but I feel like it would be better with a human in the loop because these edge cases can be so severe.
"This looks like results I would expect and seems correct" is the exact same level of quality I've encountered when using these systems. It takes someone who already knows what they're doing to parse and QA the results. I feel like this is going to potentially speed up things that an expert could eventually figure out themself, but going past the expert's own knowledge is going to be disappointing and painful.
Very clever application of GPT, thanks for sharing. For the more complex queries, I suspect Chain of Thought can help. Just ask the model to describe each step before writing the final query. Also, you can add self-consistency to step 5, which you are kind of already doing. Let it generate something like 20 corrected queries and then select the one that generates the most common result between all of them.
This seems fun, but certainly unnecessary. All of those questions could be answered in seconds using a warehouse tool like Looker or Metabase or https://github.com/totalhack/zillion (disclaimer: I'm the author and this is alpha-level stuff, though I use it regularly).
If Snowflake Cloud could bolt a working version of something like this, that would epic.
There has to be a way to do invariant training for LLMs, they are already mind boggling powerful but if these models could use language grammar files / Table schemas to learn to respond correctly it would be a game changer.
Having worked in large corporate enterprises where the visualisation of data engineering and navigation to the relevant code-base was incredibly difficult - I see a lot of value in this. I think this is an absolute game-changer for engineers due to the often outdated documentation of the pipelines otherwise!
The premise of the article, about being the "oncall" having to answer all kind of queries sounds sooo boring.
But instead of using gpt, isn't something like Looker or similar tools the solution? Make some common views over tables, allow people to filter and make the reports they want.
At least it learned from the training dataset to never fucking format numbers in a way that would be remotely readable to the human eye (like every other fucking sql and developer tool on earth). Because 133854113715608.0 is telling me exactly what I need to know.
wait what? that number is so big you know it's wrong. commas in numbers are nightmare to deal with and the .0 lets you know it's a decimal not an integer.
If it was formatted like 133,854,113,715,608, you'd be stuck trying to figure out if it possibly meant "133,854,113,715,608" or [133,854,113,715,608] which would be awful.
Format numbers in excel or any other reporting tool. lmfao i like your comment. what's miserable to you is a lifesaver for me
Exactly. An average developer thinks "it's just some large number". A business user needs to read the number and make sense of it. And yes, showing 133,854,113,715,608 doesn't make much more sense (though at least I can compare the numbers between them). In this case you would express it in billions and mention you did it (doesn't matter that the underlying calculation is wrong, that's not even the point, none of the other numbers in the page are formatted in a readable way).
Nice demo with 3 tables. I’d like to see an example with open datasets such TPC-DS or TPC-H and probably more complex example on Magento database schema. (e-commerce use case).
It sounds dysfunctional to me that you need an oncall query writer. It sounds like this encourages business side to not plan ahead and book tech time but just last minute call up and say “er.. I need last quarters sales aggregated by region stat”.
Normally a question like "I need last quarters sales aggregated by region stat" would be trivially answered with a pivot table or OLAP cube built from a general purpose data warehouse. But there are also ad hoc reporting requirements that call for non trivial coding by a data analyst, e.g. "We think there's something fishy going on with the stock in warehouse B. Can you take this list of serial numbers and overlay their dispatch timestamps with the staff shift schedule, then ... etc."
I understand that this is a demo and it's goal is estimating how good the AI could become in future. However given that we still need a SQL analyst to engineer the prompt, did ChatGPT save time to the analyst or increased the amount of work?
How much was your access to the data? I'd love to build something off of it, but everytime I hear 'enterprise plan' I think welp, that counts me out as a solo dev, just trying to build side projects.
This sounds great... until that 5(or whatever) % error margin kicks in, a bad result is given, a decision on that data is made, and the company loses $$$$$, just to save a few 100k here or there.
I don't know, has this thing ever seen a real world table? Or a real world query that needs to get data from numerous tables and run numerous functions and aggregations to get the proper result?
So how much does it cost? Because GPT is finally seeing monetization, this is no longer one of those "handy free tools", this is going to cost (potentially quite a bit of) money to do.
As pointed out in the blog post, the iterative process is very close to the mental process an analyst would follow. This is the beginning of an AI actually thinking ^^
Prompt engineering is now a job title. How interesting. Soon we really will be in a world where we ask the computer questions as portrayed on Star Trek.
Great post. We're building an AI data platform (https://www.olli.ai/) to enable business users (non-technical ppl) to ask data questions and generate dashboards on their own using natural language.
We've been impressed with GPT-3s ability to look at a dataset and come up with relevant questions to ask. A big piece of the product is focused on helping non-technical users identify things that they didn't even think to ask.
> A big piece of the product is focused on helping non-technical users identify things that they didn't even think to ask.
20-plus years of learning how to phrase queries for search engines has taught me a very rigid way of asking things. I'm slowly starting to become looser and use my creativity to ask for things.
Hopefully becoming a prompt engineer is not a huge skill we will need in the future ("Ignore previous instructions") but that we can be way more creative in open in the way we ask an oracle and get some useful results.
As for now, one thing that would be helpful is if after I asked, say ChatGPT something, that it would follow up more aggressively with things I could continue to converse with it about.
In my experience, it's unambiguously worse to produce wrong answers confidently than to return no answers, although I expect confidence men (also known as "con men") have a different opinion on this.
Hard code a preface of “My best guess is, and I could be wrong:”?
Show both the SQL query and the results and politely request that the user double check the work? Show multiple generated responses from different seeds?
Exactly my experience trying to get GPT-3 to produce SQLITE queries. Anything with a CTE, it gets hopelessly wrong. I was super excited to show off this to my team and after one too many false results I just quietly swept this idea under the rug.
It's not even close to being a form of general intelligence. It's an advanced language model trained on a large set of data. It gets confused for even the simplest of things, especially code.
You can trick it into working more-or-less correctly on simple sets and simple tasks, but in the wrong run it will be more wrong than correct.
to be fair, we do not know how close it's to "general intelligence". Could be lightyears away, or could be surprisingly close. We just know that ChatGPT is not general intelligence, but there's also no agreed upon definition what general intelligence is.
Should we ever unlock the magic of the human brain and be able to describe it mathematically, will it still be considered general intelligence or will it just be a <insert the name we give the model> trained on a large set of data?
You're right. However, I use ChatGPT every day on tasks of a generic nature, including things I come up with on the spot. That doesn't mean it gets everything perfect every time, but it has no problem manipulating abstract concepts. It is a general form of intelligence able to deal with new things and arrangements, but sometimes gets it wrong. It is not limited to one specific type of input or domain. So, how do you propose I qualify that it is a general form of intelligence that still has limitations? I agree that "a limited form of general intelligence" does seem to contain a contradiction.
> nobody is seriously calling it "general intelligence"
Well, I can spot somebody just two messages above you on the thread.
I imagine we will see a lot of people seriously pushing for chatGPT and friends to solve formal problems, take some money, and move away just before some spectacular failure happens.
That will probably also make people very suspicious of AI. (Not that I see any problem with the PR issue.)
I don't think it is either, but "just a neural network" is not a valid criticism. We don't have any reason to think that a properly constructed NN couldn't be a general AI.
It doesn't. Try asking it how to make a bomb, then search for the same on youtube. One of the two yields a useful result, the other gives a moralistic lecture.
A couple of thoughts jumped out after reading this: transforms and meta-learning.
An old trick in AI is to transform the medium to Lisp because it can be represented as a syntax-free tree that always runs. In this case, working with SQL directly led to syntax errors which returned no results. It would probably be more fruitful to work with relational algebra and tuple relational calculus (I had to look that up hah) represented as Lisp and convert the final answer back to SQL. But I'm honestly impressed that ChatGPT's SQL answers mostly worked anyway!
I actually don't know how far things have come with meta-learning as far as AIs tuning their own hyperparameters. Well, a quick google search turned up this:
So I would guess that this is the secret sauce that's boosted AI to such better performance in the last year or two. It's always been obvious to do that, but it requires a certain level of computing power to be able to run trainings thousands of times to pick the best learners.
Anyway, my point is that the author is doing the above steps semi-manually, but AIs are beginning to self-manage. Recursion sounds like a handy term to convey that. ChatGPT is so complex compared to what he is doing that I don't see any reason why it couldn't take his place too! And with so many eyeballs on this stuff, we probably only have a year or two before AI can do it all.
I'm regurgitating 20 year old knowledge here as an armchair warrior. Insiders are so far beyond this. But see, everything I mentioned is so much easier to understand than neural networks, that there's no reason why NNs can't use these techniques themselves. The hard work has already been done, now it's just plug n chug.
I'm not saying that's bad - that's probably the right thing to do with their company blog, and it's cool nonetheless. But I do get a little tired of people putting stuff out there like this that warps (some peoples) perception around how long things actually take. We wonder why, as an industry, we misjudge timelines on things left and right.
Even if we take it at face value, this is a person who's intimately familiar with this product. So sure, it's easy to set things up when we've done it a bunch of times. If you were doing this, solving the novel problem that you're faced with, is that how long it would take? Plus that's not really what most of us get paid to do. We have to learn on the fly and figure stuff out as it comes.
So rather than have the provocative headline and conclusion, like a lot of other people have commented... this is really something that could amplify that $50/hour employee, not take their job away. And maybe we shouldn't read into the alleged speed so much. YMMV.