Hacker News Reader

Comment by TrackerFF | Hacker News Reader

I work for a gov. agency, and we're in the midst of a project (with your typical MBB consulting firms) to map the use of generative AI/ML tools, and what possible benefits there are.

One immediate internal benefit would be the information retrieval bit - you kind of get rid of the "data" barrier, which involves knowledge in databases/SQL etc.

Gov. agencies typically have lots of bureaucrats with deep domain knowledge in laws, regulations, and whatever the field they're working on - but limited data knowledge. And instead of relying on analysts etc. to retrieve the needed information, these LLMs could bypass that step.

Of course, it's not entirely that straight forward - as you'd need to validate the things the LLM serves you, but that's one of the ideas. Leadership have been discussing AI/ML non-stop for the past 6-7 months, and it seems like these kinds of FOMO projects are popping up everywhere...good times for the consulting firms.

View on HN

I'm not clear on the immediate benefit there. You are asserting it is easier to get answers out, as opposed to having to learn a query language?

If so, I think that makes sense. With the /major/ caveat that you are also making it even easier to get incorrect data out. No?

That caveat is how you end up LEFT OUTER JOINED to a terrorist watchlist.

I guess little Bobby Tables grew up to be a malicious bureaucrat.

A future 'get out of jail free' card for anyone with a good defense attorney.

You'd be shocked and amazed how poorly designed and managed many gov. databases are. Many of these databases follow this pattern:

1) Some worker is fed up with having to sift through hundreds of excel spreadsheets in order to find the information need, and the nightmare of keeping such spreadsheets updated

2) Said worker deploys a small database - just for their own use. After some time the DB gets more users, who make their own views and what not. The database and tables within may or may not follow some rules. Heck, maybe the worker that created it was learning as they went on.

3) Said worker quits or gets a new job, and the database is essentially unmanaged. Some other worker might not now about that DB, and you go back to step 1)

Now multiply a database like that with 100, and span it over 20 year. You get this unimaginable spaghetti monster of multiple DBs, some written in one dialect, some in others. Some are completely unnormalized, others a high degree of normalization.

And to build a report, you may need to access tables from numerous such databases. Even seasoned analysts dread starting on the reports, because they'll spend a good day just to find the right dbs, tables, and all the errors.

So of course when the directors hear about this new fangled AI magic that just spits out the results when you ask it it plain English, they immediately order a use/benefit analysis on it.

Ok, so that may be a bit harsh - but that's the reality of many agencies. Dogshit DB management, and being 20 year behind the digitalization revolution.

But yes, a problem would of course be: How do you know that the data the LLM returns is true? Is it conjuring up fake data? Does it process/calculate data as you want it to?

This summarizes the actual hard parts of my job as a government statistician.

But I'm hoping LLM-generated code (R in my case) can let me stay "in the flow" when exploring. I can spot-check generated code pretty quickly, but finding data, reading their docs, then finding packages to do what I want takes time. I'll often forget why I asked the question in the first place. For example, "How many children lived within 50 miles downstream of X in 2015 and were diagnosed with Y?" I can imagine how the finished code would look, but writing it myself means brushing up on the diagnosis records, two or three GIS datasets, and a GIS package. If we had a nice database or warehouse, this wouldn't be terrible.

I don't fear for my job. The coffee would be generated by including snippets of previous analyses in prompts, and guess who wrote those snippets? My role will become less coding, and more reaching out to policymakers and nonprofits to help them answer questions. I'll tell them what data we do have, what's reliable, and what kind of statistics can answer their questions. At least, that's my dream.

So I work for one of the consulting firms :)

I can confirm the giant FOMO happening, and LLMs projects are popping up everywhere.

Good news is, the "talk-to-your data" use case you're describing is not the only one that can deliver amazing new tools! There's the "explain-me this", "summarize that", and "next-best-action" use cases that shows great promise!

Radio Shack solved the "next-best-action" problem nearly 50 years ago.

http://switchandlever.com/edm/

It might be ok if that data gets validated by experts in the same field, but there have been plenty of stories about lawyers and professors who failed to even review the output they use. Color me skeptical.

>> One immediate internal benefit would be the information retrieval bit - you kind of get rid of the "data" barrier, which involves knowledge in databases/SQL etc.

In that case, I would recommend a search engine (i.e. Lucene, Elastic, Solr, or other) that holds the agency's 'knowledge' before I'd try feeding all of that into an AI such as GPT-4. Granted there are a lot of ML tools that can be statistically rigorous but GPT-4 is generative and therefore not providing access to original source material, which is troubling to reconcile with its potential use in any organization that needs to establish and maintain a ground truth.

How can someone validate the data if they can't query the original source? I understand the argument "an expert is made more efficient", but you are saying that the model enables someone that doesn't know what they are doing to suddenly be able to supervise the model doing it for them.

> - as you'd need to validate the things the LLM serves

And that is not as hard as retrieving that originally, how so then?

Also no "here we have a prototype never intended to be the real thing, now ship it"-mentality there? Oh, certainly not, it is a non overloaded government agency :D

> > - as you'd need to validate the things the LLM serves

> And that is not as hard as retrieving that originally, how so then?

Not necessarily. With good prompts GPT-4 is decently good at citation of exact sentences in the source. And there could be a separate system that verifies that citation text matches the real text for being safe.

> And there could be a separate system that verifies that citation text matches the real text for being safe.

This is what we’re doing with LLMs at my job and it works quite well. Essentially, the LLM generates a space of possible answers that are then validated with various procedural logic checks. The two strategies work well together, whereas either alone wouldn’t produce as good of results.

Any tips for someone currently working to create one of these systems?

I completely agree that greatly increasing data accessibility is a huge unlock and value add.

A package I open sourced recently might be useful for use cases like this, https://github.com/approximatelabs/datadm It's essentially a chatGPT code interpreter, specifically designed to work with data, that can be run entirely on open models (eg. StarChat). True local mode operation.

Anyone that has ever used tools like Grammarly or Copilot can attest to the added benefit of AI augmentation.

Reply on news.ycombinator.com