Modern Creator
Why Modern Creator?
vogel · YouTube

Jev Classifies 1,500 Real Inbox Emails in Real Time

A live test of TypeSafe's new classifier model Jev, run against the creator's own 1,500-email inbox export to see if it can actually replace manual triage.

Posted
4 days ago
Duration
Format
Demo
hype
Views
66.2K
572 likes
Part of the collectionJev, explainedEvery Jev breakdown, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

A model built to do nothing but classify, instead of generate text, can sort a real inbox at 38 requests per second and a fraction of a cent per email, which is fast and cheap enough to run as a background triage layer rather than an occasional AI query.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You build tools that need to triage or route high volumes of text (email, tickets, leads) and are paying too much in latency or tokens to do it with a general chat model.
  • You're evaluating whether a narrow, rules-based classifier model is a better fit than GPT-style models for a production pipeline.
  • You want a real, unscripted before/after: raw claims from a launch post tested against a real 1,500-email inbox, not a synthetic benchmark.
SKIP IF…
  • You're looking for a general review of chat-style AI assistants — Jev doesn't generate text and isn't a chatbot.
  • You want deep technical documentation on how Jev's rule format works — the video only shows a UI dashboard, not the underlying API or JSON schema.
TL;DR

The full version, fast.

TypeSafe launched Jev, a model that only classifies text against rules instead of generating it, claiming 20-200x the speed and 40-400x lower cost than standard AI models. The creator tests that claim on a real 1,500-email export from his own inbox, scoring each message for category, priority, spam, and whether it needs a reply. A 100-email batch runs at roughly 200ms average latency and 240ms P95, correctly flagging real security and account-violation emails as high priority and reply-needed. Scaling to a 1,000-email batch holds the same per-email speed across 1.1 million tokens, and the entire run costs about 5 cents. The takeaway is that narrow classification is a distinct AI workload from text generation, and treating it that way unlocks speed and cost that general models can't match for high-volume sorting tasks.

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 →
Chapters

Where the time goes.

00:00 – 00:11

01 · Cold open

TypeSafe is introduced as a new company built around a classification-only model.

00:11 – 01:01

02 · What Jev actually is

Jev doesn't generate text, it only classifies input against rules, with headline claims of 20-200x speed and 40-400x cost savings over standard models.

01:01 – 02:14

03 · First live test: 100 real emails

The creator loads 1,500 of his own exported emails and runs a 100-email batch across 8 workers, defining four classification fields: category, priority, spam, and reply-needed.

02:14 – 03:06

04 · Proof in the results, then scaling up

The high-priority and reply-needed flags correctly surface real security and account-violation emails, then the creator scales the test to a 1,000-email batch that holds the same per-email speed across 1.1 million tokens.

03:06 – 04:06

05 · The cost reveal

The full run costs about 5 cents, and the creator closes on what a model built for classification rather than generation unlocks.

Atomic Insights

Lines worth screenshotting.

  • A model built only to classify text against rules, instead of generating text, can run 20-200x faster and 40-400x cheaper than a standard general-purpose AI model.
  • A 100-email classification batch averaged 200ms per email with a 240ms P95, processing 38 emails per second on 8 parallel workers.
  • Four classification fields (category, priority, spam, reply-needed) are enough to cover the two questions that actually matter for email: is this spam, and does it need a reply.
  • Scaling from 100 to 1,000 emails held the same ~200ms-per-email latency across 1.1 million tokens processed, showing the speed doesn't degrade at higher volume.
  • A full run of roughly 1,100 real emails cost about 5 cents in API spend, cheap enough to run as a standing background process rather than an occasional query.
  • Splitting 'does this need a reply' out as its own classification field caught real emails a keyword filter would likely miss, including a 91%-confidence account-violation notice.
  • Treating classification as a distinct workload from generation is the core bet: a model that never writes a sentence can still be the more useful AI for a sorting problem.
Takeaway

Classification is a different AI workload than generation, and treating it that way is cheap.

WHAT TO LEARN

A model that only sorts text against rules instead of writing it can run fast and cheap enough to sit in the background of a real product, not just answer occasional queries.

  • Splitting 'what type of task is this' out from 'generate a response to it' lets a narrow classifier model run 20-200x faster than a general-purpose one, per this creator's test.
  • Real-world latency held at roughly 200ms average and 240ms P95 across 100 parallel-processed emails, a workable speed for near-real-time inbox sorting.
  • Throughput didn't degrade going from 100 to 1,000 emails: 1.1 million tokens processed at the same ~200ms-per-email pace, which is the number that actually matters for a production workload.
  • Reducing a messy problem like 'triage my inbox' to four scored fields (category, priority, spam, reply-needed) made a fuzzy task testable and fast to verify by eye.
  • The all-in cost for classifying roughly 1,100 real emails came out to about 5 cents, cheap enough to run continuously rather than ration as an occasional AI call.
Glossary

Terms worth knowing.

Classifier model
An AI model trained to sort input into predefined categories or scores, rather than generate new text. It answers 'which bucket does this belong in,' not 'what should I say next.'
P95 latency
The response time that 95% of requests finish faster than. A common way to measure worst-case speed instead of just the average.
Worker
A parallel process handling one request at a time. Running 8 workers means 8 emails can be classified simultaneously instead of one after another.
Resources

Things they pointed at.

00:00toolTypeSafe
04:06toolJev
Quotables

Lines you could clip.

04:06
“Jev gives all AI properties of code.”
tight closing tagline that reframes the whole demo in one line→ TikTok hook↗ Tweet quote
03:17
“Look at this. I've used 5 cents. That's it. This is insane.”
concrete, surprising cost number lands as proof→ IG reel cold open↗ Tweet quote
00:11
“It's not like a standard AI model where it generates text or anything like that. It just does essentially classification.”
clean one-line definition of the product category→ newsletter pull-quote↗ Tweet quote
The Script

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.

So the co -founder of ChatGPT has created a new company and it is called TypeSafe. What it is, is a classifier AI model. And...
It basically, it's not like a standard AI model where it generates text or anything like that. It just does essentially classification. So you can feed in these sort of like Jason rules and everything like that to be like, okay, based on this input, evaluate this set of rules type of thing, like production capability down, uh, credible churn risk.
I don't know all of that stuff, but apparently according to them, it's 20 to 200 times faster and 40 to 400 times. Cheaper with output tokens free, which is absolutely insane. There's a lot of cool stuff in the post, but I wanted to play around with it with what I think might be one of the coolest implementations of this, which is email classification.
So right here, I've got 1500 of my own emails that I've exported from my email. And we're going to run this classification model because I got access to it and see how well it runs the classification on it. We're going to do a batch of a hundred emails just to start out with eight workers and we'll see how long it takes to do it.
So let's start. Boom, a hundred emails done. It doesn't tell me how long it took.
Average was 200 milliseconds, 200 milliseconds per. Email P 95 was 240 milliseconds and it did 38 per second, which is amazing. Okay.
So what I had it do is I gave it four different things to like evaluate for category, priority, spam, or reply. These are really important because these are the core things that you actually care about in email. Do you need to reply to it?
Or is it spam? Those are basically like the two that it could kind of be gone down to. So priority high, uh, safety identifier in your open AI account.
I would count that as high. Okay. Domain revocation account.
That's high. SES custom mail. And then let's look at the ones where it ranked a high reply.
So we had a 51%. Yep. Looks like someone replying to other ones.
Okay. We got a 16%, 11%. Yeah.
So it would have properly, oh, 91 % account violation. Yep, there we go. So that's another email that like needs reply.
This did an amazing job because it ran super fast. Let's get a little dirty, shall we? Let's do a thousand emails.
We have eight workers still. Let's start and let's just watch it cook. Look at that.
It is flying. It is absolutely flying. We've got 1 .1 million tokens in so far and it is just, it is sticking around that 200 milliseconds.
per email that's being classified. And if we go down, we can just see it all. Man, this is incredible.
This is so fast. I'm not used to this type of speed of AI at all because every other like AI, it's like, okay, well, we've got reasoning and we've got this and that, the other. I think this is more than enough for standard email classification.
This is amazing. I love this. I'm gonna try to think up some more ideas.
to to do this with i'm blown away and i'm gonna see if i can implement this into my own like super local project that i was doing because it's hella cheap too it's so incredibly cheap Look at this. I've used 5 cents.
That's it. This is insane. Oh, wait, this isn't even me.
Wait, this is a shared account. What the heck? Because I see Luke Experiments on Playground.
That's the other people on the team. We only have a balance of $5 down here, which just shows how cheap this model is. Why don't we just go ahead and finish out the run?
Just why not, you know, and just look at that average latency. That is insane. That is full trip request, if I recall correctly.
That is full email submitted, email returned. And if you parallelize this even more, oh my gosh, this is insane. This is super cool.
There's so many different use cases here that I can't even think of, but this is going to be a super cool model called Jev. Jev gives all AI properties of code. This is cool.
The Hook

The bait, then the rug-pull.

A launch post claims a new AI model is 20 to 200 times faster and 40 to 400 times cheaper than standard models. Instead of taking that on faith, the creator points it at 1,500 of his own real emails and watches what happens.

Frameworks

Named ideas worth stealing.

01:28list

Four-field email classification schema

  1. category
  2. priority
  3. spam
  4. reply-needed

The creator's own reduction of email triage down to four fields Jev scores per message, arguing the two that matter most are spam and reply-needed.

Steal forany inbox, support-ticket, or lead-routing tool that needs fast automated triage
Storyboard

Visual structure at a glance.

open
hookopen00:00
speed/cost claim
promisespeed/cost claim00:13
dashboard armed for first run
valuedashboard armed for first run00:59
100-email batch complete
value100-email batch complete02:01
1,000-email batch mid-run
value1,000-email batch mid-run03:03
Frame Gallery

Visual moments.

open00:01Frame at 00:04 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:04Frame at 00:07 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:07Frame at 00:11 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:11speed/cost claim00:13Frame at 00:16 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:16
open00:01Frame at 00:04 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:04Frame at 00:07 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:07Frame at 00:11 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:11speed/cost claim00:13Frame at 00:16 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:16Frame at 00:20 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:20Frame at 00:23 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:23Frame at 00:26 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:26Frame at 00:29 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:29Frame at 00:32 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:32Frame at 00:34 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:34Frame at 00:38 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:38Frame at 00:41 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:41Frame at 00:44 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:44Frame at 00:47 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:47Frame at 00:50 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:50Frame at 00:54 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:54Frame at 00:57 from Jev Classifies 1,500 Real Inbox Emails in Real Time00:57dashboard armed for first run01:00Frame at 01:03 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:03Frame at 01:06 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:06Frame at 01:09 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:09Frame at 01:12 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:12Frame at 01:15 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:15Frame at 01:18 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:18Frame at 01:21 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:21Frame at 01:24 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:24Frame at 01:27 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:27Frame at 01:31 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:31Frame at 01:34 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:34Frame at 01:37 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:37Frame at 01:40 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:40Frame at 01:43 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:43Frame at 01:46 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:46Frame at 01:49 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:49Frame at 01:52 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:52Frame at 01:55 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:55Frame at 01:58 from Jev Classifies 1,500 Real Inbox Emails in Real Time01:58100-email batch complete02:01Frame at 02:05 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:05Frame at 02:08 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:08Frame at 02:11 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:11Frame at 02:13 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:13Frame at 02:17 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:17Frame at 02:20 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:20Frame at 02:23 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:23Frame at 02:26 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:26Frame at 02:29 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:29Frame at 02:32 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:32Frame at 02:35 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:35Frame at 02:39 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:39Frame at 02:42 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:42Frame at 02:45 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:45Frame at 02:48 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:48Frame at 02:51 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:51Frame at 02:54 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:54Frame at 02:57 from Jev Classifies 1,500 Real Inbox Emails in Real Time02:57Frame at 03:00 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:001,000-email batch mid-run03:03Frame at 03:06 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:06Frame at 03:09 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:09Frame at 03:12 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:12Frame at 03:16 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:16Frame at 03:19 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:19Frame at 03:22 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:22Frame at 03:25 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:25Frame at 03:28 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:28Frame at 03:31 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:31Frame at 03:34 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:34Frame at 03:37 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:37Frame at 03:40 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:40Frame at 03:43 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:43Frame at 03:46 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:46Frame at 03:50 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:50Frame at 03:53 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:53Frame at 03:56 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:56Frame at 03:59 from Jev Classifies 1,500 Real Inbox Emails in Real Time03:59Frame at 04:02 from Jev Classifies 1,500 Real Inbox Emails in Real Time04:02Frame at 04:05 from Jev Classifies 1,500 Real Inbox Emails in Real Time04:05
One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.