Just-in-Time Ontologies: What Cheap Intelligence Means for Data — Hrishi Olickel, Southbridge AI

AI Engineer · 21 min · 233 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:00[music]
00:04So, I really love this article from Hannes at Duct Tape B, where he asks if we lost a decade chasing distributed architectures that presumed compute and storage were slow and expensive.
00:15Capability in every era though defines architecture and tooling, and the tooling likes behind consistently.
00:22We first made compute cheaper than storage, and then we gave it to everybody, and then we made it faster.
00:28And every single time architectures undergo this massive change.
00:32All of our data infrastructure today is built on the idea that transformation is expensive.
00:38What if that wasn't true anymore?
00:41I believe we've solved the final bottleneck in transformation, general intelligence or judgment, and we've made it accessible on tap all along the Pareto frontier.
00:52Which puts us right on the verge of a paradigm shift.
00:55So, in this talk, we're going to cover why single ontologies always fail, the three roles models play in pipelines, intersecting model priors for better results, and Pareto serving.
01:06How today's models can tune a workload for cost, speed, and accuracy all at the same time.
01:12Here's a quick example.
01:14When Jeff launched, we already had an entity resolution pipeline running, working out Ohio's campaign finance records and whether they belong to the same person, company, or committee.
01:25It was projected to cost about $10,000.
01:28Our systems took that new model, worked out where it fit, tested over a dozen approaches, and integrated it.
01:3510 days after that first call, we had final ground truth for every donor case in Ohio under $800.
01:42That is the power of Pareto serving and just-in-time transformations.
01:47So, let's get into it.
01:52Ontologies are the enterprise version of mind maps.
01:56It's this carpet that sucks us all in, including me.
02:00I talked about a single universal schema when I founded Southbridge 2 years ago.
02:06The popular conception, and it's not wrong, it's just small part of the picture, is that migrations and ontologies fail because of a lack of buying.
02:14And they're abandoned halfway through, that organizations and people never switch over.
02:18All of that is true.
02:20100%.
02:21But buying isn't the primary reason.
02:24Say, if you could integrate and migrate all of your information into a single super schema in the blink of an eye, you still have to contend with the truth that there is no super schema.
02:35There actually are four schemas.
02:37The first one belongs to the source, wherever the information comes from.
02:42I take temperature as an example.
02:43Say, I ask you to write down the temperature.
02:46The source schema is whatever you have access to.
02:49Your phone gives you your region's temperature, sensor on your desk, which is less accurate, but it's more local, or your watch, say.
02:56But even your patience in following through on my request is part of that source schema.
03:01The second one is the storage schema.
03:04There's only so much you can write down.
03:07You can't write down the firmware of the device you're using.
03:10You can't even write the time down to the second because you couldn't finish writing all of that inside of a second.
03:16This is the schema imposed on you by storage.
03:19The third one belongs to the decision.
03:22Why do I need you to write the temperature down?
03:24Maybe you're figuring out what makes you comfortable, in which case I need temperatures from you spread throughout the day, and I need to make sure that it's you that's doing the measuring.
03:33Maybe I'm building weather migration maps, in which case I don't care who does the measuring, but I care a lot about your exact location, the time, the device you used, and whether that's changed.
03:44Or maybe I'm deciding whether to install sensitive equipment and where, and you yourself become this variable that I need to cancel out.
03:53There's a fourth one though, which we'll get to when we talk about priors.
03:56It's this recent addition.
03:58The true problem is that you will never know all four schemas at the same time except at the moment that the decision is being made.
04:06In almost every enterprise and every home, most of the data being recorded has little idea of what it's going to be used for.
04:13Take a customer, the first box in every ontology.
04:17To building, a customer is whoever pays.
04:20To supported, whoever calls.
04:22To compliance, it's whoever we're on the hook for.
04:25And the only time you know is when the decision is known.
04:27Should we refund this order?
04:29Can we extend this account more credit?
04:31Suddenly, what you need to know, how and to what degree, becomes extremely clear.
04:38This is why I was wrong.
04:40There is no universal schema.
04:42There is no concept of a customer that's divorced from the question that you want to answer.
04:47And pulling these schemas together is the process of transformation, of making information actually useful.
04:53Taking the context that it was recorded in and merging it with the context of the decision.
04:59And that takes discernment.
05:01Doing that merge when the decision space is known instead of ahead of time is what I mean by a just-in-time ontology.
05:08So, we'll come back to discernment with a story from my last job where like most of my job was to make sure that certain numbers just did not exist ever.
05:19Before that, a little bit about me and us.
05:21We are best described as obsessives.
05:25Obsessives who embed intelligence into pipelines.
05:28Pipelines that can scale, that can run over and over again to take work truly off our hands, and pipelines that can mature in cost, performance, and speed all on their own.
05:38And most of the work we do in data breaks into three regimes.
05:42The first is write once, run once.
05:46If you do any kind of development, think of these as monster scale immediately invoked functions.
05:51They're built for something single use, like joining a data set, mining information, or answering one specific question, and then they get torn down.
06:00The second is write many, run once, where lots of smaller and larger workloads are built for one large common goal.
06:08The closest thing you might know is like large migrations.
06:11Years of records moved out of systems nobody fully remembers.
06:15Lots of workloads built for one rule and verified individually at source, checked against the others because all of it only happens once.
06:23The third, though, and this is bread and butter, is the most common, is write many, run many.
06:28These are repeatable pipelines that are built and maintained over time as we add more information in.
06:34And at Southridge, we've moved so much work off the main thread and into these repeatable systems, from benchmark design, tracking, coding, to making subtitles for movies that we want to watch.
06:45Almost everything we do in the modern world is a transformation of data, or it's bottlenecked by how hard it is to transform data.
06:52And in all of these cases, our job is simple.
06:55Make sure people can rely on the results.
06:58In just 2 years, we've gone from figuring out how to make early 3.5 turbo level models reliable on things like health care, aging, mining research, to building Hank Wave, our open-source harness for data work, and then using that to get models to work together at terabyte scale.
07:15Today, we have self-optimizing systems.
07:18You give them the problem, a way to fail safely, a place to work, and they find the shape of the solution.
07:24They pick the right models, they run the pilots, they scale up, and they Pareto serve their way to the best option they can find.
07:34So, story at my last job, the question we heard most often was, "Can we make this decision data led?"
07:42And sometimes that was the right question when we had thousands of vessels on the platform.
07:47The problem is that we also asked it when we only had 50 users.
07:51And if you know statistics, you can explain really well why AB testing 50 users is a bad idea.
07:57Not because you're not going to get a result, but because you will.
08:01So the only real option was to do not produce a number at all because the moment a number exists, it argues for a result.
08:10Today though, it is worse.
08:12Interactive agents will always give you an answer.
08:14They'll give it to you quickly.
08:16You ask for a transformation, whatever, and you get a contra and yes, sir.
08:20What this means is mistakes add up quickly.
08:23Moment in chat can be a broad nightmare forever just waiting to be discovered.
08:28The primary reason that this happens is that intelligence, as it relates to data work today, occupies three separate roles that coding agents kind of mash into one context window.
08:39Let's pull those apart because it'll help a lot when we get to decision models because they fit into exactly one of these three.
08:46The first one is building and designing the pipeline, which is breaking down the problem, writing the early specs, authorizing spend.
08:54The second one, and this is usually confused with the first and given to the same models, is working inside the pipeline, actually doing the work, reading the PDF, writing to databases,
09:04checking results.
09:05Before we get to the third one, I want to make it clear that these two roles pull in very different directions.
09:11The first one wants slower models with heavier, thicker tokens, models that can think more, walk around a problem and see the angles, digest documentation, understand how these pipelines get built, and predict how something will do before it even starts, right?
09:26Opus or up.
09:28The models you want inside pipelines though are the opposite.
09:32You want them cheaper, faster, specialized, good at following instructions, repeatable workers with noble defects.
09:39You see the same thing in nature.
09:40A leafcutter ant colony, for example, runs on four different sizes of worker, right?
09:45Small, medium, high, extra high.
09:47As you watch any system that you design, review, or hear about, listen for what slots in where in this layer cake.
09:54Makes a big difference.
09:56Because we had this distinction, bringing in Jev in this newly exploding class of decision models was very easy cuz they're workers.
10:04They're built for quick contact with information and they're deployed at the bottom end of large workloads.
10:10So, third category, though.
10:12This is something we call the hill climbers or servers.
10:15Models that can take a first one and self-optimize.
10:19Given the right knowledge, instructions, and an incredibly stable surface to stand on, they can try new things, vary budgets to limit cost exposure, debug quickly, take notes, surf their way not just to the right solution, but the optimal one.
10:34We call that burrito surfing and we'll cover it at the end.
10:38But, first, let's talk about our experience with Jev.
10:41Jev launched on a Tuesday when we were deep into an entity resolution pipeline.
10:47You were working on donors in Ohio's campaign finance.
10:50And at this point, a lot of people have seen this chart.
10:53We talked about the results before and adapted some of the ongoing work to the data set.
10:57So, you can download it for yourself, prompts and everything.
11:00What we didn't talk about is everything behind that one arrow and what happened when you ran the same idea on the whole state.
11:08Here all of the architectures we tried.
11:10We started with three designs on a small development set.
11:13A cheap reasoning model on its own, Jev with a reasoning model for anything it wasn't sure about, and a reasoning model proposing answers with Jev checking them.
11:23On about 200 health out families, the cheap reasoning model on its own did as well as Fable for about 2% of the cost.
11:31Jev first got the core answers right, but it missed nine household links because it had confidently said no to every one of them.
11:39So, we rewrote the questions.
11:41Without touching the models of the thresholds, the household question now said what counts as enough evidence, and Jeff went from no on all 13 development questions to yes.
11:50And as we scaled it statewide, we ran over a dozen rounds of such qualification.
11:55Each round fixed a definition, what a committee is, what a filer means, when two people live at one address, what does that mean?
12:02Every one of Ohio's 203,317 cases now has a label that says what the answer rests on and how far it was checked.
12:12This is the hierarchy we discovered for this specific problem.
12:16And here's everything we learned about putting decision models into pipelines.
12:19I'll leave this up so you can screenshot it.
12:22But here's where the money went.
12:24The whole statewide job came to 795 on an $800 budget that was set ahead of time.
12:31Checking cost about four times as much as doing, 545 on second opinions, settling disagreements, and the audit against 128 for first answers.
12:41Before Jeff, we expected this work to cost about $10,000.
12:44And now that the pipeline is built and verified, we expect the audit budget for the next state to be close to zero.
12:50The most amazing thing is that our agent 4A, running on Astra and then later on Opus, found most of these approaches incrementally on its own.
13:01By building the system, running it, debugging it, and trying new things.
13:06It ran for 10 days, tried over a dozen architectures.
13:09Its hangs ran for more than 4,000 times alongside 90,000 calls to Jeff, tried practically every model on the Pareto frontier, and the models inside those runs went through roughly 6 billion tokens.
13:22And once all of that was done, we wrote what we learned back into our library of agentic parts.
13:27This work leaned on these existing ones and we're adding new tested ones like the checking letter.
13:34This is to me the value of taking work off the main thread for humans and for agents because every single one of these iterations was a version controlled Hank.
13:44They were easy to revert, understand, debug, and critique.
13:48So, can models do this natively now?
13:52Not quite.
13:53They're missing four things which we'll we'll go over and there's something that doesn't go away though as models get bigger and more capable, the prior.
14:02Just like humans, models bring their own priors, their beliefs about what should be in the world to every task.
14:08This is the fourth schema that I mentioned before, the one the model brings with it.
14:14Bigger models in a strange way are worse at this sometimes.
14:18They hold onto their beliefs and argue for them even more vehemently.
14:22Here's a few that we ran into.
14:24One pipeline looked for in kind spelled with an underscore.
14:29The file spelled it with a capital I and a hyphen.
14:32So, the model counted 15,658 in kind contributions as ordinary money, didn't double-check.
14:39The loan payment repayments also disappeared the same way.
14:42Another one tried four date formats when the file used the fifth.
14:4799.4% of transaction dates then came out empty and every row count and dollar so matched so the agent moved on.
14:55Several runs noticed that loan amounts were negative, decided that the sign was redundant, and stored the absolute value.
15:03A net of 784,000 became 5.3 million dollars of debt in the end, validated by multiple models.
15:12And this is quickly becoming the largest class of error I see across companies at data work as models take over.
15:17Sometimes you can solve this just by not vibing, by checking results and applying human intuition.
15:23And what you're actually doing there when you do this is interesting.
15:26You're intersecting your priors with the models.
15:30In that first example, any human would catch missing in-kind contributions if they knew campaign finance or thing to check.
15:37So, we found it.
15:38But humans, speaking as a human on this particular task, are expensive.
15:42They are slow, they pause timelines, they're even more non-deterministic than models, and they hate interruptions.
15:50And what it turns out you can also get really far with cross-model prior intersection.
15:54On our 200 held out families, Jev and Luna agreed on 194.
16:00And none of those shared answers was wrong.
16:03Every error sat in the six families where they disagree.
16:06Models from different families, different pedigrees, tend to disagree more and have different priors than models in the same group, for example.
16:13The exception, I'm not going to say who, are models that are distilled outside the same company.
16:20So, with that kind of work, you can shrink the surface quite a bit, and the errors along with it down to a much smaller but persistent class of problems that apply to LLMs as a species.
16:31So, in one benchmark, four different model families produced the same total down to this end because they all shared a sign convention that the answer key didn't.
16:39And once you get that far, the only way forward is to list them, look out for them, and apply specific solutions.
16:47So, going back, what are the four things we're missing to be able to self-optimize?
16:53Number one is a stable deployment surface.
16:56We use Hank Weave for this, a harness that takes the fast-moving surface of agent runtimes like cloud code and pie, and then exposes a consistent, well-tested log and API interface with checkpointing, rollback, budget enforcement, everything built in.
17:10Something you can build on.
17:12But other things fit this description.
17:14To be fair, for headless data work, Hank Weave, I would say, still the oldest, most stable, and the most open.
17:20I'll leave the skill up if you'd like to try it.
17:23Number two is some way to bound your experiments.
17:26You can't test unless you can control cost, time, and access as resources.
17:31The third one is know-how and methodology.
17:35We found that this new class of models can take everything we've learned from two years of deployment, agentic patterns, debugging techniques, prompting, and apply to complex problems.
17:44With the right abstraction layer, you can express it very token efficiently, and the models can implement it.
17:50The fourth and the most important one is a failure surface.
17:54Archimedes said, "Give me a place to stand, and I will move the Earth."
17:58And in our case, we haven't met a problem or a constraint that we couldn't iterate ourselves to solving given a stable stack to build on and a failure surface to iterate against.
18:08That last one has been the hardest to put into words and into tooling and instructions that helps models get there.
18:15The current best models are still only barely good at this, sparks of intelligence, so to speak.
18:21The whole process though involves going through six stages.
18:24First, we inspect the job and we split it.
18:26For Jeff, that means splitting some 800,000 rows into 200,000 smaller, specialized cases, and each of those cases into typed questions.
18:35What kind of donor, one donor, you know, different different questions that needed to get answered.
18:40Then we decide the failure surface.
18:42In this case, we ran good models on a small slice of the data with some of it being manually checked to set a baseline that we can fail against.
18:50In other cases, we've set automated failure systems, we've set schema checks, we've also covered behavior as a thing that we can measure against, but it's very important to have to define a failure surface before you ever start.
19:04But now comes the fun part.
19:06We want minimal bounded pilots to test architecture.
19:10And once something starts working, we move those tags into rigs.
19:13Code that runs between bounded work segments that we call codons.
19:17These codons are then reused, looped independently, validated, and placed in sequence.
19:24Finally, working pipelines get deployed to production, where they can slowly mature.
19:29And that middle part is where the surfing happens.
19:33We find an architecture that gives a solid result, we use it as a baseline, and we swap models and some other parts for faster and cheaper options, compare against known intermediates,
19:42tighten the requirement for success, each time moving closer and closer to the ideal, and you just watch the frontier move.
19:49Being able to break a complex problem into pieces means each piece can move its own to its own best position, as Adam Smith would say.
19:58In a system like this, ontologies exist as intermediates between intelligence and result, which makes them easier to reuse.
20:06That's the other half of what I mean by just-in-time ontologies.
20:10It's built when the job needs it, kept with its receipts, and reused when the next job needs the same thing.
20:17And once you have those, you can make these things cheaper.
20:20You can you can run them in parallel across data sets to scale up.
20:24I close with an apology, right?
20:26That the answer is there is no ontology.
20:29Not a universal one, at least.
20:31There never was one.
20:32The good news is that this makes every ontology and schema valid, as long as you can label and preserve the context around it.
20:39And if we can harness this cheap intelligence is starting to percolate through the world around us, we're rapidly approaching a future where any information, in any format, is made as useful as it can be.
20:51And at Outpace, we're working as hard as we can to make that future a reality.
20:55Thank you.