Frontier results, on device - RL Nabors, Arize

AI Engineer · 30 min · 339 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.

00:01Hi there.
00:02I'm Rachel Lee Neighbors, and today I'm here to talk with you about how to use local models to stop paying for frontier models.
00:09Let's dig into it.
00:11So, I've worked on standards that power today's web with Mozilla on Firefox DevTools and the W3C on web standards, and of course on Microsoft's Edge browser.
00:21I've even been on the React team.
00:23Now, I've spent the past 3 years consulting with AI startups and some of our favorite LLM and browser companies on all things web, AI, and UI.
00:33And recently, I've joined Arize.
00:36Have you ever had your CTO ruin your agentic workflow with a slight change of prompt or an LLM migration?
00:42Have you ever been that CTO?
00:44Well, you probably need Arize's observability platform for models and the agents who love them.
00:49Anyway, I'll be actually using one of Arize's open-source projects today, Phoenix.
00:55We'll talk more about that later.
00:57But today, specifically, I'm here to talk with you about how AI is costing you.
01:02Every time you reach for foundation models like GPT-5 or Claude, it's costing you, your users, and the environment.
01:10Let's have a look at the costs of one-size-fits-all inference.
01:17So, first there's security.
01:19Security costs trust.
01:20When you use a large LLM that's in the cloud, you're sending data to remote servers, and it always carries the risk of exposure, interception, and retention by third parties.
01:30We have cases where the use of remote AI chatbots has led to sensitive business data being stored, breached, and leaked to the public.
01:38Latency costs the user experience.
01:40Now, there's been research on mitigating response delays in LLM chats in VR and found that 4 seconds is the limit of believability for users, and many calls that you will make to large models are going to take longer than 4 seconds, as we will see when we check Phoenix.
01:57It cost your business.
01:59Third-party inference costs are uncontrollable compared to API API costs.
02:04Agency compounding levels of inference and this means even if the tokens are cheaper, you may be using more of them or if you're using four of them, they may be
02:13more expensive.
02:14And of course, if you aren't connected, remote models simply aren't going to work, which means that unless your software is connected to the web, nobody can use it.
02:25Big inference going offline costs productivity.
02:28If there's an outage, if you're in a place where you cannot reach Wi-Fi or you're in a very secure environment.
02:35Now, token costs have been falling as of late, but total inference spend has been rising because agent can reasoning workloads consume tokens way faster than prices are dropping.
02:48But we can completely eliminate most of these costs and it starts by asking ourselves exactly how much
02:55[snorts]
02:56is this costing?
02:57Do we really need an LLM to do this job?
03:02All right.
03:03You can use task-specific models, which are small in size and power consumption compared to an AI foundation.
03:10Uh this is my little cheat sheet.
03:12If you're looking for an expert model, is a camera pointing at something?
03:16Division uh you know, is is there a a vision component?
03:20You can use a model like MobileNet, YOLO, MediaPipe.
03:23Is it a microphone recording something?
03:25There are audio models like Whisper and Wave2Vec2.
03:29Chat translation or analysis?
03:30This is where you might use something like a small language model like Gemma or Quinn.
03:36These uh small models are called SLMs uh or smaller language models.
03:42The the definition of small is up to debate here.
03:46Uh these are great for times when we do need the language power of a generative pre-trained trans- transformer, a GPT, but we probably don't need the sum total of human knowledge in a black box at our disposal,
03:59or we don't need multimodal capabilities.
04:01We're not going to be analyzing images and audio at the same time.
04:05Now, SLMs, smaller language models, contain millions to billions of parameters.
04:11An LLM contains billions to, well, trillions.
04:16So, you see the big dot there?
04:17That's one of the smaller LLMs that's out there.
04:20But, the little dot actually represents one of the larger SLMs.
04:23So, you can see that there is a huge difference in parameter size, and this means that there is a vast difference in the size of machine you're going to need to run one of these models.
04:33The good news is you don't need most of what's in the the big green dot there.
04:37You don't need history, you don't need philosophy, you don't need all those Reddit chats, you don't need a lot of what the models have learned and been trained on.
04:47Most of us are using our models for things like summarizing a chat thread, or detecting if this person's being a jerk right now.
04:54And it takes a surprisingly smaller amount of parameters to determine those things.
05:01Now, the nice thing about smaller language models is that they consume the same or less energy as large language models to produce correct uh awareness.
05:09We have uh correct responses.
05:11We have good good research that shows this.
05:14Small language models um they come in all sizes and shapes.
05:20Most small language models uh for mobile and web are deployed with quantization, that is to say, 8-bit, 4-bit, and that can have a quarter disk and memory requirements.
05:301 billion parameters fits on about 2 GB in FP16.
05:35Uh this this grid, by the way, gives upper bound estimates.
05:39So, these are just to give you an idea of what will and will not fit on different devices.
05:45These are so lightweight that they can be put on devices.
05:48Speaking of, my Pixel Pro ships with one.
05:51And I I of course bought the the Pixel 10 Pro as soon as I could because I wanted to know what it would be like to work with an on-device model.
05:59An SLM of my very own.
06:02The good news is that small language models are production ready.
06:05Nvidia called SLMs the future of agentic AI.
06:09Once again, great research paper from 2025 that found that SLMs are sufficiently powerful for running agentic task loads.
06:18And they consume less energy than language models.
06:22Let me Let me give you a bar chart here to take a peek.
06:26So, let's say this is the total amount of, you know, energy consumed to perform a task that an LLM would take.
06:34An SLM takes about 25% of that.
06:38And a task-specific model takes about half of that over.
06:41So, as you can see, it's from an energy perspective much cheaper to run the smaller language models and task-specific models.
06:50Um So, what are the benefits?
06:53Once more, more secure, works offline, no fees, more efficient, and lower latency because it's on device, no round trips.
07:03Now, I've been using local AI for some time.
07:07I for instance first started using local models.
07:10You see on the left you've got Claude, but on the right you have Goose, which is an open an open agent harness that has a a really nice interface
07:19for chatting with models.
07:21And this one here, it's running Gemma, and it's able to answer.
07:26It takes a little longer.
07:28Smaller models can be a little longer depending on what device you're running and how big the model is.
07:34But one of my favorite questions to ask a model is how likely is it that ichthyosaurs, marine reptiles from the dinosaur era, had echolocation cuz if you look at their skeletons,
07:46they look a lot like dolphins.
07:48And it takes reasoning about biology to be able to determine whether or not that's possible.
07:54And it's hilarious to me that even I think this was Gemma 3 was able to come up with a good a good response that there is no evidence in the fossil record of their having developed a melon or
08:09the specialized bones required for conduction that dolphins have.
08:14Of course, Claude hedged it by saying, "Well, we just can't know."
08:17Uh about one out of every three times, which was interesting.
08:22Claude has always been a little nervous about about coming forth with an opinion.
08:28So, how do you pick the best model for the job?
08:31This is the hard part.
08:32Uh I recent I So, how do you pick the best model for the job?
08:43I built a framework with Google that I use on my own projects.
08:48You can find more about it at web.dev, but I'll run through it here.
08:52Now, first off, I like to think of this as prototype big, deploy small.
08:56Just repeat this to yourself.
08:57Prototype big.
08:58Think big.
08:59Go big.
08:59Deploy small.
09:01Um you want to convert the parts of your system over to SLMs and specialized models for production, but you can prototype on a foundation model, no problem.
09:12Uh the first step in this process is to determine whether or not what you're trying to do is even possible at all.
09:19You use the largest, most capable model, and you you just see if you can get it to do the thing.
09:25See if you can get the model to recognize people by their handwriting or the way they they write.
09:32If that is possible, then you know that another model can probably handle it, too.
09:37So, you could use like a foundation model like Gemini, or you could use a really really tough uh task-specific model for like handwriting recognition for instance.
09:46Now, this is a feature of the product that I like to build on the side.
09:51It's Mima.
09:52It is a a client for all your social networks.
09:56And I built a feature for it that summarizes long conversation threads.
10:01Because I don't know about you, but sometimes I go to bed and I wake up and there's 50 comments on something I posted.
10:06And I just want to know, are people angry at me?
10:09And this is what uh this is a feature that saves me saves me a lot of heart attacks in the morning.
10:16So, I first prototyped it out with Claude to prove that it's good enough.
10:20So, you can see the little you can see the little summaries there, and says a pretty good job of of saying who's talking about what.
10:30So, first step for this was to collect a set of inputs and outputs, that is to say in in my case, I wanted to collect a set of threads,
10:40and then how I would summarize them.
10:42Or in this case, how Claude summarized them seemed adequate to me.
10:45So, I I exported a golden data set.
10:49Now, a golden data set is a curated high-quality collection of preferably human-labeled input-output pairs that you're going to use as the ground truth to evaluate, validate, and benchmark your model.
11:02This is what the the data looked like.
11:04Um this is all public knowledge, but you know, you can see it's got the author handle, the content, created at, whether or not it came from me.
11:14And I created a big JSONL of all of these.
11:17So, there were 14 threads, and I evaluated each one for summaries and annotations.
11:24Because some of the summaries actually link deeper to ex link deeper into the conversation.
11:31So, it was the same threads, but had two different outcomes for them.
11:35One was a short summary, and one was a summary with references.
11:38All right.
11:39So, that means 28 examples all together.
11:43Um yeah.
11:45So, these are the things that I was going to be measuring.
11:49Um you need to before you start doing anything, you need to know what it is that you're measuring.
11:56What is success?
11:59In this case, was the JSON being output by the summarization process correct?
12:05Um an easy way to test is to try JSON parse.
12:09And if it works, yay.
12:10If it doesn't, boo.
12:12Um the reference structural validity, you know, if it's pointing to different parts of the conversation inside the summary, do those parts actually exist?
12:23Factual consistency, that is to say that it's actually able to summarize the content of the threads within reasonableness.
12:32So, if we're talking about cats, it doesn't give a summary saying we're talking about puppies.
12:37Uh this is this is something that we need an LLM to judge or humans, but the LLMs will be cheaper, I assure you.
12:45Um length compliance, making sure it stays within a certain word count, and of course checking the latency.
12:50P50, this is, you know, the median across the eval set, and the P95 is the worst-case scenario.
12:58All right.
13:00So, you collect these, uh and now you're going to test and do different models, and compare how those models rank against the big models.
13:12You test from small to large.
13:15I had to choose a selection of models, but first I need to find a framework, a tool to do the measuring.
13:21And I chose Phoenix, which happens to be created by Arize.
13:25It is open source, it's free and the engineers who work on it are pretty darn awesome if I do say so myself.
13:34We're going to perform first of all a what is known as a capability eval.
13:39A capability eval asks, you know, what can this agent do well?
13:44Where we compare the performance of the large model against the performance of a selection of smaller models.
13:51We say this is what Claude Opus produced and this is what Gemma 4 produced.
13:56How do these two stack up against each other?
13:59Now, I used Claude Sonnet for the baseline.
14:03And you can see actually there in the bottom number one, you can see the the baseline.
14:08It's looking pretty good.
14:09It's got an average latency of 2.9 seconds.
14:13It costs about $0.22 $0.22 to run this to run 14 of these tasks.
14:22So, I actually did some math and it turns out I'm using about a dollar worth of inference every day using Mima.
14:29I don't have the money to pay for that many teenage girls using Mima every day.
14:34So, we're going to have to run this on device.
14:38Good news is the total cost column for all these small local models is absolutely zilch because that inference has been pushed to the consumer.
14:46It runs on their device.
14:47They're the one who has to charge the phone so that it can draw energy from the battery to run the model.
14:52Of course, you want it to not suck suck the battery dry, but that is another conversation for something that we could be testing and evaluating in the future.
15:02Now, I chose Let's look at our Let's look at our contestants here.
15:07We've got Quen 2.5 Instruct weighing in at 1.5 billion parameters, only a handy 1 GB on disk, and its sister Quen 3, 1.7 billion, just a little bit bigger.
15:22And Llama 3.2, weighing in at 3 billion parameters and a tidy 2 GB on disk.
15:29And then there was Gemma 4 E2B, which is 5 billion parameters, a hefty 3.1 GB, but this was the one that so many engineers I spoke with when I was picking a model were like, "Oh, Gemma 4 is the best.
15:42You got to use Gemma 4." And I think that's important here because if I had just gone with what my buddies told me, I may have given the user What?
15:50Pardon.
15:51Not may have.
15:52I would have given the user an extremely different experience, not a good experience.
15:59You're going to want to select the smallest model that gives acceptable responses for your use case.
16:05Or as I like to call it, the SAGE model, the small and good enough model.
16:10I'm trying to make this a thing.
16:12Bear with me.
16:12I hope we can make SAGE happen.
16:15Now, at first, I thought I wanted the the Qwen 2.5 model because it was the fastest.
16:22You can see that Qwen is all the way here.
16:25Let's see, where'd it go?
16:25Yeah, 2.5 is the gray circle in the lower left-hand corner.
16:29It came in around 1 second total in latency on the P P50.
16:35That that's amazing.
16:36Like, that's really fast for for an AI summary.
16:40The problem was that its accuracy was pretty pretty low compared to everybody else.
16:46The orange square is Gemma 4 E2B and the blue diamond is Claude Sonnet, which is our ceiling.
16:53It's the most accurate, but it's also a little pudgy, a little slow.
16:58It's weighing in around 3 seconds in latency.
17:01So, when we take accuracy into account, the winner was actually Llama 3.2, which is the big green circle that's weighing in right here around the 90% for accuracy.
17:15And you can see that both uh both Claude and uh and Llama 3.2 are much faster than Gemma 4. Gemma 4 was coming in around 8 seconds.
17:31Now, maybe that was because Gemma 4 needed a different kind of prompt, but this was pretty consistent.
17:38I was testing each one of these three times and then averaging the results.
17:43This is what the results look like.
17:44I recommend when you're running evals with something like like like Phoenix, you open up the experiment and you actually take a look.
17:52You can see what the raw responses were, uh what the expected responses were.
17:56You can actually compare them against each other.
17:59Um I found that in many cases, Llama's response was so close to Claude's response as to be pretty much indistinguishable from one another.
18:11So, yeah.
18:12Uh Llama 3.2 was the ultimate winner and I decided to move forward with them.
18:16And this makes sense because Llama 3.2 is created by Meta and Meta has a strong interest in creating models that do a good job with human inputs and summarizing human things on a social network.
18:28Makes sense, right?
18:29I mean, it's a social thing.
18:31Benefits from a a social model.
18:34So, recap.
18:36Here's how you right-size your model in four steps.
18:39Number one, you prove it's possible.
18:41You test whether what you're trying to accomplish is possible at all by using the largest possible model.
18:46This could be a foundation model like Gemini or a task-specific model.
18:50[snorts]
18:50And you set success criteria.
18:52You collect a set of inputs and outputs um that that you want to see coincide with one another.
19:00This will be the bar.
19:02And then you test from small to large.
19:03You compare the outputs of small models against your test criteria and you work your way up from the smallest model until you you within an acceptable range of that small and good enough model, that sage model.
19:15That's when you select your sage model, the smallest model that gives acceptable responses for your inputs.
19:24But you're probably wondering, what are we going to do about that gap?
19:26I mean, 90% accuracy, that sounds like it could be uh you know, that could be a deal breaker.
19:33Well, we can squeeze better performance out of smaller models with prompt engineering.
19:37This is important in cases where you can't control which model you're using.
19:41Some people might, for instance, create a distilled model that's been trained to do this one task really well.
19:48But if you're working, for instance, with a mobile app, you might not want to be using a distilled model because every time you add capabilities, you'll probably have to train the model slightly,
19:59and then you'll be shipping a new 1 or 2 GB model every time to your users.
20:03So, this is a great example of like, "Ooh, I don't think I'll be able to control that model.
20:09Once it's on the user's device, I'm not going to ask them to download new ones.
20:12That would eat up their data plan."
20:14Um you might join a team, and they're already committed to Gemma 3, and they're just not going to be implementing Gemma 4 until they've rolled out an an an update that uh you know, addresses all of the evals.
20:28So, let's take a look at closing the gap between Saona and Llama.
20:32Now, the the green and blue dots in that upper left-hand quadrant.
20:37So, the measures I I honed in on, the ones that they really seem to have different results with, were JSON reference uh JSON and reference structural validity, factual consistency,
20:49um P50 latency, and P95 latency.
20:53And I decided that the P50 latency the cutoff is, you know, one 1 and 1/2 seconds for P95 would be 3 and 1/2 because, remember, 4 seconds is that worst-case scenario for people
21:05feeling disconnected from an AI powered experience as per the research.
21:13So, where do we start?
21:16I recommend that you optimize one step at a time.
21:20You want to isolate one variable per prompt variant to test whether what you're trying to accomplish is moving the needle when you're using the different prompts.
21:33So, I created five prompts.
21:36Well, I created four prompts and I had the original prompt as the baseline.
21:40The original prompt was pretty good.
21:41V2 used number input.
21:44The same prompt reformatted the threat the threat as numbered messages instead of using JSON.
21:49The hypothesis being that smaller models could track natural language indexing better than array offsets in heap of JSON.
21:58The second one was a few shot.
22:00It added a couple of examples and outcomes to the prompt.
22:05The hypothesis here being that small models learn format from examples faster than from rules.
22:11Then there was the strict rules version.
22:13This one was a house of no prompt.
22:16It had explicit negative constraints.
22:18No preamble.
22:19Don't count words before responding.
22:21I mean, do count words before responding.
22:23And the hypothesis was that small models respond to literal commands and that they like to be bossed around a bit.
22:29And then lastly, there was chain of thought which forced the model to identify key moments before writing.
22:34The hypothesis just being that thinking out loud would improve grounding.
22:39So,
22:39[snorts]
22:41ran the tests again with these new prompts and just just just to the llama 3.2 and we discovered a couple of things.
22:50We I mean, myself.
22:52I just ran this locally on my machine and was able to compare the results.
22:56You can even put the results into something like Claude and have have conversation about the trades if you like.
23:02So, stop that.
23:09All right, we're back.
23:12So, in this case, the baseline wasn't really good at at the determining how short it should be so it would fit within a certain section of content.
23:23The ref accuracy was 91.2%.
23:27It was factually correct 87.1% of the time and the latency was 1 second.
23:33Reformatted input didn't really make a difference.
23:37Explicit rules actually didn't made things worse.
23:41The model responded very negatively to being told what it couldn't do couldn't couldn't do.
23:47It um it was a a naughty child didn't like to take instructions.
23:51Chain of thought didn't have that big a it it did a little bit better on the length and unfortunately that came at increasing the latency by 600 milliseconds.
24:05The best performing one was the few shot one that provided a couple of threads and a couple of examples.
24:10It was much better at getting the length right.
24:14It was more accurate and in its references it agreed with the Claude model a little bit more about what was said and it only increased the the latency by 200 milliseconds.
24:27So, that sounds like a pretty good deal, right?
24:31So, it was the few shot prompt that one that made the biggest improvement.
24:36Let's have a look at how it stacks up to the original.
24:41Bum bum bum.
24:43So, you can see the bar here Claude Sonnet versus Llama 3.2 3B.
24:47Um I actually did a couple of things at this point because I wanted to close that gap completely.
24:57So, Llama 3.2B with a few shot prompt was actually able to get within a reasonable error area of um a reasonable error margin of error.
25:12The P50 latency went to less than 1 and 1/2 minutes, so it was totally green.
25:19Uh the structural validity 91.7% factual consistency was also at 92.9% Uh the P95 latency was well under 750 milliseconds uh less than Claude.
25:36In general, its latency was really good.
25:39But, let's see about those uh those couple of those 10% here between structural validity and factual consistency.
25:48This is why it's important to actually open up your e-vows and take a look at what's inside.
25:53Uh when it came to factual consistency, it turned out that Claude was just being a very strict judge.
25:59I was using Claude to judge the responses and comparing, you know, Claude Opus was comparing Claude Sonnet's response to uh Llama 3.2's response, and of course Claude was favoring its little sister and being like, "Yeah, um I think I think that uh
26:16you know, I don't think your interpretation of what Jenna said is accurate because you said she was being angsty and uh she was actually being cross."
26:27It was that sort of thing.
26:28And this is why it's important to crack them open.
26:31Now, as for rough consistency and length, those actually could be handled inside the harness and post-processing.
26:39So, making sure that it's, you know, got the right number of references in the thread, that's something very simple to look at.
26:45You can just see how long the thread is.
26:47And if there are more refs than there are members of the thread, that's incorrect uh when it comes to, you know, how long the actual uh summary is.
26:57If it's too long, you can just truncate it.
27:01And so, when we added the post-processing, we're actually able to close that gap pretty solidly.
27:07Now, we've got uh 100% JSON validity, structural validity was 100%, factual consistency, there's only a little bit of a disagreement there, and it turns out that it was the judge being too darn strict.
27:20[sighs and gasps]
27:21We got the P5 uh P50 latency down to around a 1 second, and P95 uh it was under 300 and uh three three three three and a half seconds.
27:34So, that's pretty good.
27:35It actually ended up meeting and beating Claude's on it after doing this little bit of extra effort, and I'm saving about a dollar a day in inference costs.
27:47So, the important thing is, after you've done something like this, you don't want to lose one ground.
27:53You'll want to make sure that uh you can update the prompt in the future or upgrade the model.
28:01You don't want to lose one ground.
28:03You want to ensure that you can upgrade the the model or change the prompt in the future, and it won't cause summaries to expand to be a paragraph long
28:14or to start hallucinating things that are untrue.
28:18You want to keep your evals running for that, and that would be a regression eval, and you run these sort of like you run um CICD tests, you know,
28:29uh it's how you keep your CTO from blowing away your agentic experience by accident one morning.
28:34True story, happened to a founder friend of mine.
28:38So, where do you get started?
28:40Uh how many Claude calls could be Llama calls?
28:42I I challenge you to go home today and take a look at what you're sending to LLM's and ask yourself, is this something that a smaller model could handle and how much money would I save if I did that?
28:55When working on AI projects, keep an eye out for SLM's and specialized models that might be on device already.
29:02For instance, Chrome and the prompt API, they access Gemini Nano which ships with Chrome um natively and that can be really useful cuz that means you don't need to be shipping a model to anyone using the browser.
29:15You can take advantage of what's right there.
29:20Keep in mind that these models are more efficient and they meet your users where they are.
29:24Their information stays on device.
29:26You don't have to worry about PII and you know, the energy consumption is well, a blessing for the environment so to speak.
29:38Remember to prototype big and deploy small.
29:41You may want to prototype your system with a foundation model and then convert it parts of it to small language models and specialized models for production.
29:50Keep in mind you want to prove it, define it, test it and then select your sage model that passes the test.
29:59You can use prompt engineering to help get better results from SLM's and close that gap with remote models.
30:08So, I challenge you to consider your current implementation of prototype and convert one feature to use a smaller local model.
30:17Go home, try it out for yourself and you might be surprised.
30:21Anyway, I look forward to see you out there on the web.
30:24If you enjoyed talking about small local models or the attentive web, you can follow me.
30:28I'm at nearestneighbors.com.
30:30If you want to try running some evals with your current setup, you can start testing with Phoenix at phoenix.arise.com.
30:36And lastly, you can sign up for uh the beta for Mima uh at mima.social.
30:43I've been Rachel Neighbors, and it's been awesome chatting with you today.
30:46Go forth and build your own inference stack.