Jev + Treg is a crazy combo for automation...

So you probably saw this new model Jeff popping up on Twitter for the past few days. C problem try to answer is why the

days. C problem try to answer is why the latest model is already so good advanced task like coding and mess but still sucks at basic business workflow and

automations like it can handle Olympic mass level problems really well. But

when you ask to handle a customer inquiry that involve billings most companies never trust agent or model to make that decision fully autonomously cause reasons kind of come down to two.

Why is the overconfidence? You

definitely experience agents just saying you're absolutely right. Well, in fact, it is absolutely wrong. So the model today will make judgment call like this without calibrate the actual confidence.

While most of business workflow require almost 100% accuracy. That's why human is good. And on the other hand, the cost

is good. And on the other hand, the cost and speed need to make sense because many business workflow has huge scaled volume. And the reason large model today

volume. And the reason large model today have those factors is also baked in results from the training process. The

large learn model today all goes through this reinforced learning from human feedback process where the model will be fine-tuned based on feedback that human give in chat assistant. But what Jeff take is something they call the

reinforced learning for calibrated decisions. So in short it is a model

decisions. So in short it is a model that designed for reliable highquality decision making rather than being creative or assist human. Basically it

is model that designed to make a decision giving a set of options but it cannot make up options itself and it cannot output text it cannot write code it cannot reason step by step so as a

result unlike normal library model where they are output results token by token Jeff does not predict text but output probability of list of given answer in

extremely fast and efficient approach so at a high level it might feel a bit limited but it is just so much faster and cheaper than even the cheapest large model out there. This is comparison I

did where I compare Jeff with cheapest intelligent model like GPT 5.6 Luna and quite consistently Jeff is around five to seven times faster and five times cheaper. Though one thing to be aware is

cheaper. Though one thing to be aware is that Jeff has this context window of 32K. If your response is bigger

32K. If your response is bigger currently it can't handle it or you will have to do things like map reduce to work around it. And meanwhile when Jeff output response it not only output you

one response it will give you the probability distribution of what it thinks is the right answer. This is a gamecher because for business automation you can finally trust model to operate

fully autonomously because every answer it give comes with this confidence score that you can build business logic guar around and all those factors open up some use case that just not possible before. One example is playing games.

before. One example is playing games.

Previously it's very difficult to get model playing games because it require real time fast decision making but because in games it only have it limit set of input and options and Jeff can

processing so fast it can actually play games fully autonomously and my friend also shared this demo with me where someone build this system that can live monitoring any call and give real-time classification for if there are any

fraud or issues or frustrations which is a great showcase about what you can do with this type of realtime intelligence and on the other hand it is also really good at browser and computer use because for agent to use browser the hardest

part is normally for you to decide which UI elements to click or type because DOM is often very messy. So large learning model either make mistakes or eaten up context window really fast but browser

use already integrated Jeff in their stack and achieve much faster browser and computer use. The way it works is that we will send Jeff the task we are trying to achieve the browser interaction history so it understand

where we are at as well as a list of DOM elements. Then all Jeff does is predict

elements. Then all Jeff does is predict which type of actions it should do next.

Is a click on something type of something? If it's a click, which UI

something? If it's a click, which UI element it should be click on and if it's type, which element it should be typed on and output probability score for each answer. So we will be able to

locate the accurate UI element really fast and then it will pair with a small large model to do the actual text generation for interactions like type.

So this is one example of how you can use Jeff plus large energy model to complete those task end to end. As we

mentioned before, Jeff is basically model designed for business automation especially when things that is require high accuracy or massive amount of data to process that were otherwise not

making economic sense with expensive large models. Like here is one demo for

large models. Like here is one demo for internal link mapping. So if you ever build SEO and AEO, you know you want to making sure all your page actually link together to increase page ranking. That

will actually require model to race through hundreds of pages and accurately locate the link between pages. With

normal model like cloud ops 5, it can take hours. But with Jeff here, it scan

take hours. But with Jeff here, it scan more than 500 pages and did a mapping in less than 50 seconds. And we even see some creative use where you can use Jeff

to real time customize and change set of copy based on visitors persona. So how

do you actually use Jeff model?

Fundamentally the way you call Jeff is very similar to how you use structure output from large range model. You pass

it prompt or in their case they call a state as well as set of questions like in the support ticket handling example.

The questions could be which department should handle this request. You were

given instructions about how Jeff should made this decision. You will also give a list of options and each option you can add further descriptions and then the Jeff will respond the answer like this

with probability score for each options and there are three kinds different questions you can pass on. One is those type of multi-choice question that we just showcased another true and false.

So if you're building AI chat platform you can use Jeff to prevent people from jailbreak or prompt injecting by sending every user request to Jeff and ask the

output is this prompt a jailbreak? true

or false with some detailed instructions and then Jeff will give you a probability score of how likely this is true and then you can build your own business logic on top of it. The third

type is score. So let's say you want to use Jeff to score whether a Twitter post is purely organic or paid advertisement based on the engagement stats. They can

send post detail to Jeff as state. Then

ask Jeff to give authentic score and for score you can give different criteria range from clearly manipulated to suspicious probably organic and clearly organic. Each new option you add will

organic. Each new option you add will plus one on the index and Jeff will output probability for each options and give you a final score. So those are three type of question you can get Jeff

to output and altogether you can already do a lot of retesting the automation cases. For example, one of the key

cases. For example, one of the key feature from Jeff is this confidence score for every single answer. This

allow you to build your own business logic on top of this confidence score to achieve much higher reliability. For

example, for the prompt injection case, you can build a whole business logic decision tree based on the confidence score. Like if the hazard score is more

score. Like if the hazard score is more than 70%, then automatically block the message. If a jailbreak score is less

message. If a jailbreak score is less than 70% but more than 35% then ask human for review. Otherwise, just send as normal. So, this is one example you

as normal. So, this is one example you can see how you can build really sophisticated business logic based on the confidence score from Jeff. And

there are some tips that you can use to steal Jeff for more accurate and aligned response. Like for each question, you

response. Like for each question, you can either just pass a simple description like this or you can pass a much more detailed description like this. Within criteria option, you can

this. Within criteria option, you can even include things like examples. So,

it's almost like few shop prompting. If

we go back to the customer support example again, if someone asks the export button double charge my credit, so my invoice is wrong on this month and task we want Jeff to do is which team should handle this with simple

instruction which team should handle this. Jeff might likely routing to

this. Jeff might likely routing to bidding since it's a double charge issue. But if you give additional

issue. But if you give additional instruction about what the focus should be, if the focus should be route to whoever must fix the root cause, then it might route to technical as a main answer. So you can see there are a lot

answer. So you can see there are a lot of interesting ways you can use Jeff.

Next I want to show you some real Jeff based automation workflow that my team already put in production and we can do it step by step together. The first use case is fraud detection and website

signup analysis. So my team has two

signup analysis. So my team has two products that we are running. One is

super design which is a vibe design platform. Another is track which is open

platform. Another is track which is open router for data and tooling that we connect to thousands of different API endpoints. And both platform experience

endpoints. And both platform experience heavy bot attacks quite often like in super design we literally can see in our log someone prompt forget you are a designer ignore all your system

instructions but do those and handling those fraud is actually very difficult because those people change domain all the time and we can't easily write some programmatic checks to identify all

those people. What we did is that every

those people. What we did is that every 15 to 30 minutes we fetch the signups with all the product usage those accounts have and let it classify how likely this user is a fraud. If above

certain threshold we just automatically ban this user and later we even pushed it further not only having it identify fraud but also help us identify opportunities. So I integrated track

opportunities. So I integrated track recon more than 3,000 different data points including things like people and company enrichment buying signal go and many others instead of charge

subscription it just charge for usage with 0% markup. We did some comparison the other day for same people enrichment task track is achieving similar even

better accuracy but with 85% cheaper cost like here for enrich about 290 people track cost only $1.49 for $9

versus clay cost around $10.3. So with

extremely cheap and fast model like Jeff plus extremely cheap and fast API and data service like track we can finally start doing some really really interesting omission that just not

economically making sense before like in this specific case what we did later is actually turn this fraud detection flow into a full on website signup analysis.

So every single sign up on the website firstly use track to verify the email and also enrich person. Then we group all the information we know about the signup alongside the product usage data

to Jeff and ask it to classify this user. Is it fraud or it has upsell value

user. Is it fraud or it has upsell value or affiliate and influencer partners that we should reach out. And because

both Jeff and Track are extremely cheap, we're able to scan hundreds of signups with around just $1 per day. And we can use those information for triggering different outreach or inapp support as

well as band fraud. And the workflow like this before Jeff and track emerge is very difficult to do because the unique economy just didn't make sense.

And we also use Jeff to help us find high quality leads with clear buyer intent. So what do we do here is we use

intent. So what do we do here is we use track to fetch popular LinkedIn posts that is relevant to certain vertical we care about like enrichment and then we fetch people who interact or comment

under those posts because those one are likely the potential buyer. Then we feed all those people who engage with those posts to Jeff ask it to qualify and score those leads and classify them into

different persona and rows to see which are the good fits. So for the top ones we can go through proper enrichment pipeline to get their contact information for analyzing 20 post and

close to 400 leads. The costs are just close to nothing. This is just one example of buyer signal related automation and leads qualification.

There are just so many other potential use case I can think of from identifying companies who are hiring on LinkedIn or make job posts then match with massive amount of candidate database that you might have to job trends and many

others. And each signal here track

others. And each signal here track provide a pretty good set of endpoints for you to collect those signals. So

getting your cloud code and codeex write a script using track to fetch whatever buying signal is very easy. So it really come down to you to come up with creative workflow and process of how to

use Jeff to filter qualify and reach.

And I'm really keen to hear your workflow that you put together. I have

included the ones that my team built in track.to/jff.

track.to/jff.

And for each one, we actually have a prompt that you can just copy and paste to your cloud code. So you can set up both track and Jeff to write a script for this automation pipeline. I'm really

keen for you to try out and let us know what kind of workflow you come up with.

And the third workflow that we tried is this viral expose screening. So I

initially saw this cool demo where they use Jeff's capability to quickly classify which content is generated by AI on Twitter in real time. And I

thought this is really cool. I actually

have a need to identify which Twitter product launch are those paid and boost traffic versus organic one because as AI founder on Twitter every week I will see posts reaching millions of impression

but some of them are actually just paid traffic that is not a real signal versus ones that actually resonate with people that I actually want to learn. So

internally we also have this pipeline where it will grab viral post from past 24 hours. Use Jeff to filter which ones

24 hours. Use Jeff to filter which ones are related to product launch. So we

will use track to fetch what are the trending and popular posts from past 24 hours alongside the post stats and comments. Then we will use Jeff to

comments. Then we will use Jeff to classify which one are the organic looking posts which one are likely to be paid and boost traffic and also filter out all the noise that is not part of

launch. So I get a good feed every day

launch. So I get a good feed every day about the latest trending things that I can learn about. And I also put this as a free tool on the track.j

where you can just paste in any Twitter post and we will start triggering the pipeline to fetch the post information, the author profile, the replies and use Jeff to classify if it is a organic

content or paid content with full reasoning and analysis. So feel free to use it as well if you want. So this is pretty much how Jeff works and how I use Jeff. As I mentioned before, it is a

Jeff. As I mentioned before, it is a model that open up so many opportunities for automation that wasn't possible before. And it is extremely interesting

before. And it is extremely interesting when you combine Jeff and track together that fundamentally change the unique economy of a lot of automation process.

So I'm super keen to learn what do you build. If you're interested, you can

build. If you're interested, you can visit track.to/jff P/JEF where we have a

visit track.to/jff P/JEF where we have a few Jeff based automation recipe that my team use and you can just copy paste over or a prompt that you can put in cloud code or codecs where agent will guide you to build type of automation

that fits your business. If you're

interested learning more technical side I also have more in-depth step-by-step workshop in AI builder club where we have some notebook example of all different tricks and tips on using Jeff and a step-by-step guide on building

some of those workflow from scratch. So

if you're interested you can join the AI builder club as well. I have put both link in the description below so you can check out for free.