07:22The Next New Thing · Interview
We ran Fable through 600+ tests. It's amazing
Zapier's Automation Bench ran Claude Fable 5.0 against hundreds of realistic business workflows — here's what the numbers actually mean.
June 9thThe argument in one line.
Jev is a fast, cheap AI model built only to make constrained decisions instead of generating text, and builders are already dropping it into routing, triage, and classification steps across a dozen different agent workflows.
Read if. Skip if.
READ IF YOU ARE…
- You're building AI agents or automations and need a cheap step that routes tasks between multiple models based on cost and difficulty.
- You want to add fast classification (email sorting, ticket triage, lead scoring) to a workflow without paying for a full LLM call on every item.
- You already use Claude Code, Codex, Zapier, or a persistent agent like GrokBot and want a low-cost decision layer inside it.
SKIP IF…
- You want a model that writes emails, code, or long-form copy — Jev can't compose a sentence, it only scores fixed choices.
- You're hoping to use it for open-ended prediction like trading; the video's own closing example shows it fails at that.
The full version, fast.
Jev is a fast, cheap AI model from TypeSafe AI built to make decisions, not write text: it can't compose a sentence, but it returns a probability-scored answer to a fixed multiple-choice or yes/no question in a fraction of a second. The video walks through real demos: categorizing 500 emails, building a voice-controlled browser, routing tasks between cheap and expensive models for a 70% cost cut, triaging support tickets, scoring leads instantly, searching an AI agent's memory with 80% fewer tokens, ranking video clips for shorts, controlling smart-home devices in 300 milliseconds, and filtering content ideas inside a running agent. The one clear failure case: don't trust it with trading decisions.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:36
01 · Jev
Cold open framing Jev as a decision-making model, not a chatbot — it can't write a sentence but makes decisions faster and cheaper than other models.
00:36 – 01:30
02 · Email Categorization
Jev sorts 500 inbox emails into categories and a pie chart in seconds, flagging which ones need a response.
01:30 – 02:42
03 · Setup
Connecting Jev to an agent (Codex, Claude Code, GrokBot) through TypeSafe AI's quick-start prompt and an API key; access has bounced on and off a waitlist.
02:42 – 04:39
04 · Voice-Controlled Browser
A demo where spoken commands (open this site, click this link) drive a real web browser through Jev in near real time.
04:39 – 07:12
05 · How Jev Works
Speech streams to text, a server scans the page for interactive elements, and Jev answers a fixed set of multiple-choice/yes-no questions with probabilities that simple code thresholds act on.
07:12 – 08:15
06 · Zapier + Jev
Zapier ships a native Jev integration that can screen calendar entries for spam invites cheaply and quickly.
08:15 – 09:27
07 · Model Routing
A side-by-side test routing 12 prompts through Jev before deciding which model to use cut costs 70%, since 9 of 12 tasks never needed the top-tier model.
09:27 – 11:15
08 · Customer Support Triage
Jev reads incoming support tickets and instantly decides whether to send an automated reply or escalate to a human, for a fraction of a cent per ticket.
11:15 – 14:06
09 · Lead Qualification
A Typeform-style qualification form uses Jev to score each answer in real time and route highly qualified leads straight to a Calendly link.
14:06 – 18:09
10 · Agent Memory
A folder-based AI memory system (Claudia OS) uses Jev to find the right file to read or update, cutting one recall from 13,000 tokens to 2,756 for about $0.003.
18:09 – 19:30
11 · Video Clip Scoring
Jev scores 17 candidate moments in a long video for short-form potential in about three seconds, though the demo never shows the actual clips it picked.
19:30 – 21:00
12 · Smart Home Control
A Home Assistant integration uses Jev to turn a light on or off with a roughly 300 millisecond response, fast enough to feel instant.
21:00 – 23:33
13 · GrokBot + Jev
Inside an always-on GrokBot agent, Jev filters and shortlists AI-generated content ideas for SEO before a human reviews the drafts.
23:33 – 24:38
14 · Where Jev Fails
Hooked up to a Bitcoin buy/hold/sell signal, Jev performed badly — a reminder that it's a decision router, not a prediction engine for high-stakes calls.
Lines worth screenshotting.
- Jev can't write a single sentence, but it makes constrained decisions faster and cheaper than any general-purpose model.
- Turning a decision into a fixed multiple-choice or yes/no question lets Jev return a probability score in a fraction of a second instead of reasoning through the problem like an LLM would.
- Routing 12 prompts through Jev before an expensive model cut total cost 70%, because 9 of the 12 tasks never needed the top-tier model.
- A voice-controlled browser built on Jev works by having a server re-scan the page for up to 100 interactive elements every time a new speech fragment arrives, then asking Jev a fixed set of yes/no questions about intent.
- Searching an AI agent's memory with Jev used 2,756 tokens instead of 13,000, an 80% cut, and cost about $0.003 for the recall.
- A customer support triage system used Jev to instantly route only the tickets that need a human, at a fraction of a cent per decision.
- Jev scored 17 candidate moments in a long-form video for the strongest short-form clips in about three seconds.
- A smart-home voice command routed through Jev returned in roughly 300 milliseconds, fast enough that there's no perceptible lag between the command and the light responding.
- One creator plugged Jev into a Bitcoin buy/hold/sell signal and it performed badly, showing the model is a decision router, not a forecasting engine for financial markets.
- Zapier's own CTO reportedly pushed the company's native Jev integration live before it was even indexed by Google, treating the release as too useful to wait on.
- The mechanism behind Jev's speed is a set of simple probability thresholds in plain code (for example, ignore a command below 0.5 confidence), not further reasoning by the model itself.
Where a fast decision engine beats a full AI model
SPEED VS REASONING
A model that can't write a sentence but answers fixed questions with a probability score turns out to be the cheapest, fastest piece to drop into routing, triage, and classification steps across a dozen real agent workflows.
02Email Categorization
- Feeding an inbox to a probability-based classifier can sort 500 emails into categories and flag which ones need a reply, before a general-purpose model even finishes reading the first ten.
- Categorization is a fundamentally cheaper task than composition, so routing 'sort this' to a fast classifier and 'write the reply' to a bigger model plays to each tool's strength.
03Setup
- Getting a fast decision model running inside an existing agent is a copy-paste quick-start prompt plus one API key, not a rebuild.
- Access to a new high-demand model can toggle on and off a waitlist; a fallback route through an aggregator keeps a workflow from stalling.
04Voice-Controlled Browser
- A voice interface doesn't need a full LLM turn to feel responsive — a fast classifier reacting to partial speech is enough to control a browser conversationally.
- Real-world demo speed can be inflated by inconsistent conditions, so treat a single demo run skeptically rather than as a guaranteed benchmark.
05How Jev Works
- A fast decision engine works by converting an open-ended judgment into a fixed multiple-choice or yes/no question and returning a probability, not by reasoning through the problem step by step.
- Plain code, not the model, makes the final call: simple probability thresholds decide what happens next.
- The tradeoff for that speed is scope: a model built to answer constrained questions can't compose free text, so it always needs a second, slower model nearby for anything that requires writing.
06Zapier + Jev
- Wiring a fast classifier into an existing automation platform turns a one-off demo into something that runs on every calendar invite or email that flows through, at negligible per-decision cost.
07Model Routing
- Running the same batch of tasks through a cheap router first and only escalating what actually needs it cut costs 70% in one test, because most tasks never needed the expensive model.
- The saving comes from measuring, not assuming — without a routing layer, it's easy to default every task to the priciest model out of caution.
08Customer Support Triage
- Urgency detection is a natural fit for a fast classifier: deciding 'human or automated reply' doesn't require generating the reply itself, so it can run ahead of whatever drafts the response.
- A ticket that reads as escalated should route to a human regardless of how cheap the automated path is — speed doesn't override judgment calls.
09Lead Qualification
- Scoring a lead's fit in real time, field by field, lets a form change its own follow-up before the visitor even submits.
- A qualification model is only as good as the criteria it's given — an under-specified test scored a lead as merely qualified until the context was tightened, then it correctly read as highly qualified.
10Agent Memory
- Searching a folder of memory files by keyword-matching filenames is expensive and unreliable; scoring which file is relevant with a probability-based question set cut one recall by about 80% in tokens.
- The same weakness applies to writing memory: without a fast way to decide which file a new note belongs in, an agent tends to just append everything to one daily file, which degrades the system over time.
11Video Clip Scoring
- Scoring candidate clips from a full transcript in seconds is a plausible use case, but a demo that never shows its actual output should be treated as unproven, not confirmed.
12Smart Home Control
- A roughly 300-millisecond response is the difference between a smart-home action feeling instant and feeling broken — the perceived reliability of voice control is mostly a latency problem, not an intelligence problem.
13GrokBot + Jev
- Plugging a cheap filter into a persistent, always-on agent controls the cost of letting that agent run unsupervised 24 hours a day.
- Automatically checking new ideas against what's already been published turns a big raw idea list into a much smaller, already-vetted shortlist.
14Where Jev Fails
- A model built to make cheap constrained decisions is not the same as a model that can forecast an unpredictable market — don't hand it trading signals.
- The honest failure case in a video full of wins is worth more than another success story: it marks the actual boundary of where the tool is safe to trust.
Terms worth knowing.
- Jev
- A fast decision-making AI model from TypeSafe AI that answers fixed multiple-choice or yes/no questions with a probability score instead of generating open-ended text.
- TypeSafe AI
- The company behind Jev; its site provides the quick-start prompt and API key used to connect Jev to an existing agent like Codex, Claude Code, or GrokBot.
- Model routing
- Automatically sending each task to the cheapest AI model capable of handling it, reserving expensive models only for tasks that actually need their reasoning.
- GrokBot
- A persistent, always-on AI agent platform that creators use to mine and filter content ideas from their own material.
- Claudia OS
- A folder-based system, built from daily markdown files, that gives an AI agent long-term memory of past conversations and work.
Things they pointed at.
01:16toolTypeSafe AI (Jev)
02:37toolOpenRouter
00:36channelNate Herk
01:30channelLukas Margerie
02:42channelMoritz
07:12toolZapier
08:15channelJay E
09:27channelMayank Aggarwal
11:15channelDavid Ondrej
14:06toolClaudia OS
18:09channelGreg Isenberg
19:30toolHome Assistant
19:30channelSyntax
21:00toolGrokBot
21:00channelEric Siu
Lines you could clip.
00:15
“It can't even write a sentence. It simply makes smart decisions way faster and way cheaper than any of the other models.”
06:54
“I'm going to overly simplify it and say that Jev turns everything into a multiple choice test where most things like, Opus turns it into an essay.”
08:24
“It actually resulted to 70% savings because nine out of those 12 tasks never needed the top model anyway.”
15:28
“This request took almost eighty percent fewer tokens because we're using Jev, and the cost of this recall was 0.00297.”
23:35
“I would not put this model in front of like your stock portfolio or Bitcoin or anything like that. This is just for like routing or other sort of decisions.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogy
everyone is talking about jev it is so fast watch this that is not sped up that is actually real time but what can you do with it and what does it do we're gonna get into it it's not a chat bot It can't even write a sentence. It simply makes smart decisions way faster and way cheaper than any of the other models.
My friend, I've watched endless YouTube videos and talked to people about how they're using Jev. I've got the best use cases that you and I can actually use. I'm so excited to show them to you.
Presented by Zapier, the AI automation company. Let's get into the first one. And this one is just categorizing email.
Watch how fast Jev is going through 500 of me. And so it's running through all my emails. It's gone through 80, 90, 100, 110, 120, 130.
And you can see here that it's actually categorizing every single email. It's deciding when we need to respond. There's a lot of emails I need to respond to today.
I'm not that great at checking my email. And you can see here, it's going through all of these and it is done. It has gone through.
500 out of 500 emails, and then it creates this pie chart of all of the different types of emails. I can so see this. tagging my email, making, helping me figure out what I should be responding to, how I could respond to it.
And then actually at times passing it on to a smarter LLM that can actually suggest what to write. But this is what it's great at categorizing and speed and doing it inexpensively. Let's go into the next example that I want you to see.
This is actually, before we get into an example, this is how to install it, how to get it started. I hate how many people in all these videos just get going without telling me how it turns out. You just have to go to type safe .ai and then here's how to connect it to Codex.
It's very similar to connect it to everything. else let's take it away lucas available to you but in order to get this started all we have to do is go over here to this quick start and you can just copy this agent prompt and then you can open up any agent you want in this case i'm going to be using codex but you can use cloud code you can use rockbot as well and you want to open up a new workspace in my case dev and we just want to paste in that prompt that we got from the website and then hit submit.
And also another thing that you have to do is you have to generate an API key. So you can just click on this over here and create a new key. And then you can copy and paste that to like a simple plain text file where you store your API keys.
One thing that you should be aware of is they went into waitlist and they took off the waitlist and you can sign up immediately. Now they're on the waitlist again. It just goes back and forth.
Keep trying it. And if you need direct access to it and it's not available, it is available through open router without having to do this whole waiting list nonsense, but keep checking the site. I think it's great and it's worthwhile.
All right, next I set an example. Let's get into an example. This is my absolute favorite one and we'll get into more practical ones.
But what, um, more it's did is. He had to create a web browser. Boy, did I catch him in a weird pause here.
He created a web browser that you can speak to. Here's the prompt. I'm going to 2x the speed because he's talking kind of slowly into it.
Okay, for the first example, I want to build a voice controlled web browser. So let's write a prompt here. I want you to build an app using the Jeff API.
It should be a web app where I can record or I can speak into. It automatically transcribes what I say and controls the web browser for me. So I can say, for example, open this website, click on this link, etc.
And the browser just does all of that in real time. Okay, this is done. So let's check it out.
The dashboard control page is here. We can put it. Okay, so on the right is the dashboard control.
Now let's see the browser and how he's in. interacting with Wikipedia in it. So I can click here, start, and then it should be recording my voice.
Okay, go to wikipedia .org. Okay, it's working. Okay, click on the first link.
Okay, click on the Aferi Sinks link.
Scroll down the page a little bit Okay, I'm gonna pause it there more You get the picture. The idea is that it's acting pretty quickly by analyzing what's on the page He did say later on in his video and I'll have a link to his video and all the other videos below but he did say that that at other times it was working faster, that he thought that there was something going on with his internet when he did this demo.
We've seen a few other examples of this where people have gone and breezed through it quickly. Let's understand how it works. And I pulled this clip from Moritz's video.
Can I hit play? So think of it in three parts that passes a note down the line many times a second. So your voice becomes text while you're still talking.
The dashboard page in Chrome listens to your microphone in terms of speech and text. Okay, that makes sense. And then the next step is...
Jeff answers a quiz about that text. Okay. Each time a new fragment arrives, the node server looks at the controlled browser window.
So it looks at what's in the browser, makes a short list of what's on the page. Okay. Like up to a hundred things.
Like there's a link which says new, there's a button which says search, search, and it sends that list. plus your words to Jeff, okay? With a fixed set of multiple choice and yes and no questions.
Like, what does the user want? Like, do I want to navigate? Do I want to search, click, type, and so on?
Which element on the page do they mean? Which website are they talking about? Is the sentence finished yet?
Is the user even talking to the browser or to someone in the room? Okay, interesting. And which is action, buy, delete, or send something.
Okay. Cool. And after that, then plain code decides and acts.
So the server has a handful of if statements with thresholds. If is this a command is below 0 .5, ignore it. So these are the probability that are basically returned.
If it's under 0 .5, ignore it. And if is the sentence finish is below 0 .6, wait for more words and so on. Okay.
So it has like a bunch of rules and based on the probabilities that it. it's it's getting back it's making those decisions and like acting on it basically and all of that is happening in real time and so the thing here is that if an llm were to make this decision, it would take a lot longer to actually arrive at that decision, right?
Because it needs to reason through the steps. It would need to like look at everything that's there and then kind of reason through it. Whereas Jeff is just, it gets all of that data and then it can make very fast decisions on what probability is it that this button should be clicked.
And then it just goes and clicks that button. Okay. I'm going to overly simplify it and say that Jeff turns everything into a multiple choice test where most things like, Opus turns it into an essay where you're almost scored based on how much information you know and how much information you share.
And the fact that it's just making decisions like that means that it's going quicker. Obviously oversimplify. Let me know in the comments if you've got a better analogy here.
Let's tell you about Zapier. Listen to this. I know a lot of you are skipping through some of my Zapier ads.
I want you to know why this is the one to never skip because Zapier now has Jev integration in it. And here's why this is important. Jev makes decisions like it could look at your calendar and say, Is this another nudnik who's now posting that you have to, like, do you ever get those people who are spamming your calendar with you've got a voicemail or you've got to show up to this webinar or whatever?
Zapier can send each one of your calendar entries over to Jev, which will then decide is this spam or not quickly. jeff is inexpensive and zapier's pulling of this is inexpensive you've got a very inexpensive decision maker and that's just one way to use it this is so live that the cto of zapier actually rushed it out and said to the whole team we got to get this up they did get it up but it's not even in google yet so when i googled for the link i couldn't find it i got it directly from zapier my sponsor and i'll put it for you in the description so you can go directly in and try it with zapier zapier makes you happier let's move on to the next here's the next example You know, everyone tells you send some things over to the expensive model, some things over to the inexpensive, faster model.
How do you do it? It's always been a pain in the, you know, let's get into what Jason said. There wasn't really a quick and cost -effective way for us to automate model routing up until Jeff.
And so to set this up, you can just use this prompt for you to get started. And just to give you a visual demo of the test that I set up, essentially what I asked Cloud to do is to do a comparison of around 12 prompts with Jev and another one where it's running with 5 .1 every time. And you can see here that because of Jev and the fact that it's actually routing to the right model, depending on the task, it actually resulted to 70 % savings because nine out of those 12 tasks never needed the top model anyway.
So that is quite useful. Okay, he's got the prompt in his video. Of course, I'll link to it.
The other thing that he said in his video is create a skill for yourself that will allow you to turn on this Jev integration and turn it off. So it could be like a slash Jev. When you turn it on, it will decide what to send where.
And sometimes you just turn it off and you can focus on sticking with Fable because that's what you want. All right, really good video. And by the way, if you like this kind of analysis where I find you the best use cases out of, what, hundreds of videos?
I've got more coming up, so subscribe. I'm working so hard on these. Next, this is customer support triage, where you know how when email comes into your customer support, somebody has to decide, is this something that we can just ignore?
Is this something we can send one of our automated responses to, which you can see here on the bottom, or do we need a human being to respond to it? It's been really tough to do that in the past and expensive. What Jeff does is handles it quickly.
This video from Mayank, I thought is really good, but because of his accent, too many people are missing it. Let me show you a little bit of the video. Subject, third time asking.
This is the third time asking about my broken export. Yeah, you know, you feel that, right? As a human being, you see that and you go, third time asking about my broken export.
You understand this cannot have an automated message. This is unacceptable if it isn't fixed this week. So let's try the ticket.
Jeff will quickly send it to human coup. Q basically it should be sent to human. Okay.
And that is what happened. Fraction of a fraction of a penny is what it costs to do that. And let's take a look at a race between Jeff and Kimmy.
Kimmy, of course, being not equally inexpensive, but also inexpensive. Let's take a look at that. And if I try a ticket, pay attention.
So Jeff got it very nicely. Kimmy is still taking time. So it is still taking this much time in real part to get you the answer.
Yeah, and one of the things that he said was, use Jev for the routing part because he can't write messages. And then in his case, I think he used Kimmy to draft the replies when they were necessary. So it's not a do -all as you heard at the beginning of the video.
Boy, I'm talking fast, right? Like if you see a model talk fast, you almost want to talk faster. Okay.
All right, this one is actually going to take a little bit of time for you to understand, but I think it's such a good use case. What he said is that he wanted to create a type form competitor, but it's more than that. What David is doing is he's saying, I have people who are coming in and filling out a qualification form to see whether they should be working directly with me.
I don't want to have them qualified later on and then say to the people who are the right fit, send them a Calendly link. To the people who are not, add them to my email newsletter drip or whatever it is. He needs to make the decision instantly.
And that was expensive and slow before Jev. And so he's going into his, he's going in and asking, what was he using? I think Claude or Opus to build this.
Let's listen in.
reason over it and take 10, 15, 20 seconds for a candidate until you decide whether to show them a Calendly or not. But 100 milliseconds, that's nothing, right? Let's see it.
Help me build basically a competitor to Typeform where we ask a couple of questions and the app is kind of split in like a 60 -40 view where on the left is the 60%. You have the form and we can ask like three or four questions. And then on the right, we actually see like the admin view of what's happening with this new AI model from TypeSafe AI.
Okay, it's also misspelled here. Boom, so obviously I'm doing it. You see that what he's doing is dictating his prompt, right?
Dictation to be faster. We can nicely visually see the probabilities it's assigning for each user, whether he's like highly qualified, qualified, or like mediocre or disqualified, right? Okay.
All right. Now let's see it in action. What you're going to see is on the left is the form and on the right is the analysis.
Me as a potential lead. So obviously I didn't know it was for a business automation service. So let's do a founder executive.
Continue. Let's see how that changes things. Okay.
Slightly more qualified. Let's do 20 ,000 plus. Okay.
Qualified now. A lot more qualified.
Would you like to make it happen as soon as possible? So now we should be even. Yeah.
So we went from qualified to highly qualified. Nice. So obviously my first response wasn't the best because I didn't give Astra the criteria.
It just kind of figured that out. But yeah, this is how Jeff works. All right.
It's really good at making decisions like that quickly. Let me ask you something. And this is not one of these BS ask you a question.
So you just comment. But I really want to know. You know, I had to adjust this table so that you get the view of behind me and the whole thing looks good.
Meanwhile, David over here has got nothing but a white background. He's got his Breathe Right nasal strip still on his nose. Am I putting too much effort into my background?
Should I just be getting right into it? Let me know in the comments what you think. On to the next.
Here is... Moritz again, this time what he's showing you is how he organizes all the files that his AI keeps on what he's doing so that he has a good internal memory system and how using Jev he can get at his data faster and cheaper. Here's how his files are organized.
A bit more context. The way my Claudia OS works is, you know, it's essentially a folder which has a couple of subfolders. And one of those subfolders is this memory folder, which has a bunch of daily memory files where whenever I chat with my AI from inside of this folder, it kind of saves what I've been chatting about.
And so this kind of builds this like knowledge base. And all of these other folders here are things that I've built over time. So it contains a lot of memory.
And this is the kind of stuff that we were all told just, what was it, two months ago when everyone was talking about second brain, add more data, markdown file this, markdown file that. The problem with that is now your LLM, your agent, in order to get you an answer to something has to search through all of it, which was really expensive.
Let's take a look at what Jev does with that. what's the current offer positioning? So now it's doing the recall.
It's kind of waiting and asking Jeff. Okay, so now it did the recall and it returned what you can see here. So it says 2 ,756 tokens versus...
thirteen thousand so this request took almost eighty percent fewer tokens because we're using jeff and it says here the cost of this recall was 0 .00297 okay and um it answers the question so it was able to answer the question okay what i really like about morris's video was he He didn't fully understand this, but he did give us an explanation of how everything worked by basically going back into the AI that he was using to build this.
I forget which one he was using. And then saying, now tell me, why did this work? How did it work?
And here's the answer that he got. I asked for a quick explanation here. So the way it was before, Claudia's memory is a folder of Markdown files.
And the problem is that it's very inefficient the way the LLM is doing it. So it's basically when I ask a question, it's like trying to guess which file the answer. is in based on the title of the file.
So it's, you know, doing a search for like keywords and file names basically. And then it's reading the entire file to see if the answer is actually inside of that file. And then, yeah, it's basically hoping that it's right, which is, you know, pretty bad.
And if it's writing memory, it's actually even worse. So when the agent learns something new, it should append something, but it will never really know which file it should actually append it to. And usually it will just, I guess, append it to the daily memory file.
Yeah. So these are. problems okay so how does it work with the new system that uh now includes jeff so it doesn't write text you give it some the files a list of sections a new bullet and a batch of small typed questions yes or no pick one score okay so these are like kind of the criteria and it answers all of them in one request in a few hundred milliseconds with probabilities okay so it's a judgment and the code does the counting, chunking, and definitely makes the small semantic calls.
So the way I understand it is with this new system that Using Jeff, it's just able to much more easily find the right place where the memory is. And it's kind of like doing this search like in parallel and getting like all of these data points.
And then it's able to pinpoint the right file where the memory is located much more efficiently and much. One of the things that I love about him is that he keeps calling it Jeff. I know he's trying to say Jeff.
He's seen it. And I said earlier, I didn't know what he was using. He's using Cursor and he's using Fable 5 .1 in there to build it and to analyze.
Let's go on to the next video. Here is Greg Eisenberg doing an interview with someone who's using it for video. Everyone who makes content is aware of this issue where you make content and you make a longer form YouTube video or something like that, but you want clips.
And the cool part about this is, So this right here, I'm dragging and dropping in a video file. And what this process is going to do, and I'll explain it really quick, is it's going to transcribe the video and get a word -level transcript of it.
And then it's going to pass that entire thing into Jev with some different classifier decisions to find the best clips. And we'll get to see how quickly it works. I wonder if I should be using it to find my clips.
Paste it in, it prepares audio and scores 17 moments in around like three seconds. And each of these moments are like one of the interesting parts of the video. The frustration that I have with this video is that he didn't show the clips.
I've seen people for years talk about how great Opus is at using AI to find the right clips in your long form content. Meanwhile, I've tried it so many times and it's always junk. The people who created it, I believe, do not sit and watch it.
I don't know if this is any good. I would have loved for him to say, this is what it does. Now let me hit play and show you.
Okay. But it's worth a shot for me to try it too. Oh, you know what?
Let's take a look at this quick hit here. That's pretty slick. Now, the last example I have here is with my smart home.
So I hooked this up to the home assistant MCP and I actually have a light back there that is controlled by home assistant. So I can ask it to turn it on and off. So I'll say, turn off the top bulb.
instant. And so there's no movie magic here. This is the other cool thing about how fast this is.
Like if I were to ask an LLM this, it might respond with a tool call and then have to call the tool call. This took 300 milliseconds. Now, most people will not think this is amazing.
A few people who have Alexa run devices like lights in their homes will know exactly why this is amazing. How many times do I say, Alexa, turn on the outdoor lights? And it just waits.
And I'm like, did I get it right? Meanwhile, I have my guests over and do that. Will the light go on?
Will it go off? Do I look like a fool? And for a few seconds, I do look like a fool.
And then even when it works, all they remember is Andrew does this foolish thing instead of just hitting the switch, which works instantly. This is fast. And I should say, I don't think this is the right use for Jeb though.
I actually interviewed the founder of Needle, which is similar to how Jeb works, but it's so small that you can put it on device that I would love to see elect. Love to see the A word, put it into its own device and make it work. And if they won't over at Amazon, you could into the software and the hardware that you build.
But essentially what we're looking at here is a much better way to interact with devices like this because it's a yes or no. Light on, yes or no. Let's go on.
Final one from Eric Su here. No, no, this is the penultimate, the one before. What I like about Eric here is he is not running it in Claude.
He is not running it in a regular tool. What you're seeing up on the screen is. Grok bot on the left is the chat on the right is the setup for Grok bot, where you give it a name, where you give it a description of what it does and so on.
Um, and what he's using it for is an understanding of content ideas that will help them with search engine optimization and answer a agent answer a engine optimization. Okay, here goes. so i've added a massive upgrade to my aeo seo bot inside of grokbot so you can see what this does is it looks for ideas to mine from basically all the content that i've been putting out or maybe internal calls that we've been having or customer calls sales calls things like that it looks for content spikes right and it gives me a set of ideas to look okay so it can do that and so it gave 40 ideas here and what jeff did was it evaluated said hey Do we already have similar content?
So here's 40 ideas, shortlisted 10. So it eliminated 30 off the bat and said, hey, you know, is this type of content already live on your website already? Okay.
Right. And so I can kind of fix the evals here, but this is its first pass at it. And I didn't really give it any guidance.
So it's saying, hey, this is already live. It gives the justification for it. And then from here, I'm just like, okay, I just quickly look at it and just say, okay, well.
Let's just go for these over here, but here's a compare table and we can just see that Now it's it's basically drafting it and it's putting it into here the the drafts. I can just review it afterwards. I'm so so Having it be part of your agent is fantastic because these agents are made to just run on their own.
You know, Grokbot is not meant to be turned off. It's just on 24 hours a day. And if you have it do things like this, it just goes nuts and it starts to cost you money.
I know it happened to me last week. And so tools like this that you get to add in are very helpful. I didn't know that I could add in Jev until I saw his video.
I didn't know that I could add in Codex into my Grokbot or Opus. and all the anthropic models until I complained that my bill was crazy last week and then a bunch of people said, you can do it, and none of them told me how. They just said, go ask your agent, and they were right too.
Any explanation they would have given me would have been a waste of time. I just asked my agent, you do the same thing with this. It'll take it on and it'll add it.
All you need to know from here is it is possible and it can actually help you with your content ideas for search engine optimization or other uses. Finally, let's look at what it doesn't do well. I like this fail.
I love all fails. They teach a lot. Here we go.
I could hook Jeff up to a Bitcoin signal. So basically every minute it would run and it would have this decision mix right here where it would tell me to buy, hold, or sell. And it does not seem to be doing well, which shows that this model is great, but it does have some regressions.
I would not put this model in front of like your stock portfolio or Bitcoin or anything like that. This is just for like routing or other sort of decisions. Okay, fair enough.
I think that he could have done a little bit better with this, but still, I like the... I mean, I don't think that him tweaking it, what was it, a day after the thing launched was enough for him to understand whether it could do it or not, but I do think that the big lesson here is don't just give it your stock portfolio and expect it to do magic on it.
It's a decision -making machine based on your criteria that... Let's face it, not that you give it, but that your agent gives it. All right, if you like this, I've got another video for you right on the screen.
I'm looking forward to you subscribing and seeing more of these reaction videos that I do. And I'll see you in the next video that's on the screen.
The Hook
The bait, then the rug-pull.
Every AI YouTuber suddenly has a demo of the same tiny model, and the pitch is strange: it can't write a sentence, yet it might be the fastest, cheapest decision-maker most builders will ever plug into an agent. This breakdown pulls the strongest real demos into one place, then asks the question none of the original videos answered on their own: where does the thing actually fail?
Named ideas worth stealing.
05:00model
The Three-Part Decision Loop
- Speech streams into text in real time while the user is still talking
- A server re-scans the page for its interactive elements and sends that list plus the words to Jev as a fixed set of multiple-choice/yes-no questions
- Plain code acts on Jev's probability thresholds (e.g. ignore a command below 0.5 confidence, wait for more words if sentence-finished is below 0.6)
The mechanism behind the voice-controlled browser demo: Jev never reasons about the page, it just scores a short list of pre-defined questions, and ordinary code makes the final call.
Steal forany voice or chat interface that needs to act on partial, still-arriving input without waiting for a full LLM turn
08:27concept
Cheap-First Model Routing
Send every task to a fast, cheap classifier first; only escalate to an expensive model when the classifier's confidence says the task actually needs deeper reasoning.
Steal forany agent pipeline running the same prompt through one model by default, regardless of task difficulty
How they asked for the click.
VERBAL ASK
07:12product
“Zapier now has Jev integration in it... Zapier can send each one of your calendar entries over to Jev, which will then decide is this spam or not quickly.”
Sponsor read woven directly into the Zapier chapter and tied to a concrete Jev use case (calendar spam filtering) rather than a generic ad break; closes by noting the feature isn't indexed on Google yet, so the link only lives in the description.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
