How People Are Actually Using Jev  🔗 **Source:** [How People Are Actually Using Jev](https://youtube.com/watch?v=uV6h3Uo4Nh8&is=O0POlYOcXZQSPKTU) > [!note]+ Document Notes > --- > tags: [judgment-models, ai-triage, batch-classification, synthetic-personas, agent-routing, ai-evals] > author: The AI Daily Brief > source: The AI Daily Brief: Artificial Intelligence News (YouTube) > reference: https://www.youtube.com/watch?v=uV6h3Uo4Nh8 > date created: 2026-09-26 16:22 > --- > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > # 🏷️ Title: How People Are Actually Using Jev > ## 🧭 Table of contents > - [📝 Summary](#-summary) > - [📖 Main content](#-main-content) > - [🗒️ Notes](#️-notes) > - [📚 References](#-references) > - [🗣️ Paraphrase](#️-paraphrase) > - [📰 Newsletter](#-newsletter) > - [✅ Action items](#-action-items) > - [🧪 Hypotheses](#-hypotheses) > - [📊 Methodology](#-methodology) > - [❓ Questions](#-questions) > - [🔮 Future research](#-future-research) > - [💡 Implications](#-implications) > - [🧠 Learning styles assessment](#-learning-styles-assessment) > - [📊 Accessibility assessment](#-accessibility-assessment) > - [🌍 DEIB assessment](#-deib-assessment) > - [✅ Logical validity](#-logical-validity) > - [🛡️ DARVO analysis](#️-darvo-analysis) > - [🎯 Conclusion](#-conclusion) > - [🔄 Refactor appendix](#-refactor-appendix--how-people-are-actually-using-jev) > ## 📝 Summary > - **What happened.** The AI Daily Brief published a 22:02 episode on 2026-09-26 (metadata upload date) surveying how people used Jev, TypeSafe's "system one" judgment model, in its first ten days. The host is not named in the transcript; the video description credits "NLW," the initials the channel uses for its host. The auto-captions render the product as "Jev," "Jeb," and "Jeff" and the vendor as "Typesafe" and "Types safe"; quotes below preserve the caption spelling. > - **The claim.** The host sorts the public experiments into six categories (analyzing an existing pile, searching by meaning, triaging what comes in, checking work against rules, speeding up agents, and instant response), then argues the durable value is not that Jev judges well but that it judges cheaply enough to judge everything. He closes with four fit criteria and four question-writing rules. > - **Why it matters.** The episode is a curated field report, not a vendor demo. It collects roughly 20 third-party experiments with numbers attached (724 ads in 40 seconds for 9 cents; 3,300 posts for about $13; 100 emails ranked in 453 milliseconds), and it names where the model breaks: multi-step reasoning, arithmetic, dates, consistency, and intent. > - **Key stakeholders.** TypeSafe (vendor); The Information (source of the $1 billion raise report); named experimenters including Matthew Berman, Ian Nuttall, Justine Moore of a16z, Harrison Chase of LangChain, Steven Tey of Dub, Marcel Pociot of Beyond Code, and the team at Every; users of Claude Code and Codex who maintain large skill libraries; anyone running triage, routing, or review at volume. > - **Risks and tradeoffs.** Every number in the episode is self-reported by the experimenter and relayed by the host, with no independent replication. The persona-scroll experiment produces model opinions about archetypes, not buyer data, which the host himself flags. A bare probability carries no reasoning, so hiring, money, and security uses need a second reader. The funding figure is a report of talks, not a closed round. > - **Bottom line.** The six-category map and the four fit criteria are the reusable content. The individual multipliers are anecdotes until someone reproduces them on a labeled set. > ## 📖 Main content > **Who is speaking, and who is being quoted.** The transcript has one voice. He reads from X posts, a blog post from Every, a LangChain remark by Harrison Chase, a report from The Information, and TypeSafe's own launch and limitations material. This analysis separates the host's synthesis (the six categories, the "difference in kind" argument, the fit criteria) from the experimenters' claims (the timings, prices, and accuracy figures) and from the vendor's claims (the speed and cost multipliers, the 13-question batching test, the bad-at list). > **The recap of what Jev is (00:01:29 to 00:03:36).** The host previously called Jev a judgment model; TypeSafe calls it a system one model after Daniel Kahneman's *Thinking, Fast and Slow* (the captions render the name as "Conaman"). Three question types are named. "Choice" picks one of up to 255 options and returns the pick plus a probability for every other option. "Score" places an item on a scale of 2 to 10 levels, each described in words, with a confidence indication. The third is a yes/no type; the captions render it as "a new, which is short for Berni," which is almost certainly a garble of a boolean type name, and it returns a probability from 0 to 1 that the statement is true. The host's heuristic for a good Jev task: you repeatedly read something, make a small judgment, then take a predictable next step. His example is a file landing in a downloads folder: existing software can classify it as a PDF; Jev can decide whether it's an invoice, which project it belongs to, and whether someone needs to see it. > **Why it's fast, and the batching point (00:03:36 to 00:05:45).** The emoji demo (attributed to a name the captions garble as "Hey Stephan and next") types a classifier such as "things you can wear in winter" into a text box and every emoji gets the same yes/no question at once; matches rise, non-matches fall. Nobody labeled the emoji. Because the task is narrow, it runs fast: TypeSafe's launch claim is 20 to 200 times faster and 40 to 400 times cheaper than comparable LLM processes. The host then makes the point that matters most operationally: questions about one item run in parallel, so the ninth or tenth question costs tokens but not time. TypeSafe reports 13 questions in one call were 12.2 times cheaper and 10 times faster than one at a time with identical answers. Matthew Berman's 724-ad run asked 12 questions per ad in a single 173-millisecond request. Pricing as stated: $4.20 per million input tokens (captions: "4.2"), output tokens free, 70 to 500 milliseconds per call. > **What it won't do (00:05:45).** No code, no contracts, no nuanced multi-factor decisions. The host's formula: small judgments at volume. Which bucket, how urgent, is it relevant, is it safe. > **Category 1: analyzing what you already have (00:06:29).** Ask the same questions of every item in a collection, then count. Berman: 724 live ads from 37 brands, every hook, format, offer, and CTA, in 40 seconds for 9 cents of tokens. Two days later, 723 ads scrolled by 30 buyer archetypes (gym owner, dental office manager, toddler mom, AI-curious engineer) asking "would this person stop scrolling," 21,690 decisions for 22 cents. The host's caveat is the most important sentence in the segment: this isn't data, it's 30 archetypes as considered by an AI, so treat it as a hypothesis generator before paying for real tests. Ian Nuttall: 8 questions (topic, hook, tone) across nearly 3,300 of his own X posts, about $13, to compare against engagement metrics. > **Category 2: searching by meaning (00:08:39).** Describe what you want in plain words and let the model check every candidate. Justine Moore of a16z: classify thousands of Zillow listings by attributes you can't filter for (architectural style, renovation status, freeway proximity). An experimenter the captions render as "Burhan" clipped a 90-minute video by natural-language theme in under 2 seconds for under 2 cents. An SEO audit by someone the captions render as "Boura" (possibly the same person) read all 586 pages of a site and rebuilt the internal link map in 45.1 seconds for 21 cents; Claude Opus 5 got through 21 pages for $1.43 in the comparison. The framing quoted from that post: internal linking is 8,790 yes/no calls, a classification problem previously paid for at frontier prices one page at a time. Filtering examples: a real-time AI-slop detector on X that blocks posts above a confidence threshold; 384 morning news stories scanned for relevance to 15 brands in 24.9 seconds for 19 cents, with Opus 5 readin Jev is one of the buzziest models we'vehad in a long time. And that's becauseit's not just another LLM like a GPT6 oran Opus or Fable model. It is somethingfundamentally different. But becauseit's different, it's not necessarilyclear at the beginning exactly what it'sgoing to be best used for. With thebenefit of a week and a half under ourbelts now though, people are discoveringand sharing a slew of different usecases that take advantage of what makesJev unique. And today, we're going toget into the best of them and where theymight be relevant for you. The AI DailyBrief is a daily podcast and video aboutthe most important news and discussionsin AI.Welcome back to the AI Daily Brief. Overthe last couple of weeks, one of thebuzziest new things to come up in the AIworld has been a new model called Jev. Now, what makes Jev interesting is thatit is not just another LLM that youwould use for the same thing as GPT6 orOpus 55, but actually works in aslightly differently and, as we will see,in a complimentary way. In the 10 days or sosince launch, not only has there been aton of buzz. I'm talking hundreds ofdifferent posts on X, which eachthemselves have hundreds or eventhousands of likes and shares. But thatattention has also translated intosignificant financial opportunity, withthe information reporting that thecompany is in talks to raise as much as$1 billion at a $10 billion or higher valuation. That is a decent jump fromits $40 million seed that was completedat a $200 million valuation. But now,that we've had a chance for people toactually get their hands on Jev itself,I wanted to go back through and talkabout how people are actually using thisthing outside of the buzzy visual andgame demos that have been all oversocial media. Basically, is Jevsomething that the average person who'snot a game designer or not a developerShould be paying attention to and eventhinking about as part of their largerAI stack? Now, to recap what Jev is. Previously, I called it a judgmentmodel. What Jev can't do is write in thetraditional way that you think aboutchatbots writing. Instead, the team atTypesafe, who built Jev, calls it a system.One model after a concept from DanielConaman's thinking fast and slow. System one thinking is fast, instinctive patternmatching, i.e., snap judgments, and that'swhat Jev is built for. A good way tothink about a test for whether somethingis a good job for Jev is where yourepeatedly read something, make a smalljudgment, then take a predictable nextstep. So, take for example, a file landingin your downloads folder. An ordinaryrule that existing software might havewould be to classify it as a PDF. Butwhat Jev can do is go a step further,identifying: is it an invoice? Whichproject is it for? And does someoneneed to see it? Once the judgment isMade, it can be predictably moved on tothe next step in a system. There arethree core types of questions that youcan ask Jev. The first is pick one orwhat Jev calls choice. It answers whichof these fits. You list up to 255options, and Jev can find the best fit. So an example of this might be whichteam should handle this ticket. Is itbilling, tech support, sales, or other?The Jev model is going to give you backthe pick plus a probability for everyother option. The second type ofquestion that you can ask Jev is whatthey call a score, which is basicallyrating on a scale. You describe eachlevel in words from 2 to 10 levels, andJev can answer where something falls. Sofor example, on the scale from calm tovery angry, how frustrated is thiscustomer? You're going to get back fromJev a position on that scale as well asan indication of how sure the model is. The last type of question is a yes or noquestion, what Jev calls a new, which isShort for Berni, and it simply answers,is this true? So, to stay on thiscustomer service example, is this personexplicitly asking for a refund? Whatyou're going to get back is aprobability from 0 to 1 that thestatement is true. Hey, Stephan, and next,did a really cool, very simplevisualization where he put a pile ofemojis at the bottom of a screen, andthen into a text box he wrote, you cantype in classifiers like things you canwear or things you can wear in winter. The emojis that match race to the top, orfall off if the classification haschanged. And so, what's happening in thebackground, is that every emoji isgetting the same yes-or-no question. Does this match what you typed? Andimportantly, nobody labeled or tagged theemoji as would happen in traditionalmachine learning. There's no rule, inother words, for winter. Every emojigets the same question at the samemoment. Does this match? And part ofWhat you might be recognizing if you'rewatching this demo is that because whatJev does is limited and specific, it cando it incredibly fast. When Typesafelaunched the model, they said it canwork 20 to 200 times faster and 40 to400 times cheaper than comparableprocesses with LLMs. Now, instead ofthis pile of emojis, imagine that thepile is your leads, your tickets, yourdocuments, or some other bigundifferentiated mass of information. Importantly, you can ask many questionsabout the same item all at once. Becausequestions run in parallel, asking the9th or 10th question doesn’t cost moretime. It just costs more tokens. Say,for example, you have a single customeremail. You might want to ask three yesor no questions at once. Is the customerangry? Did they get a delayed response?Are they dealing with an incorrect item?These are three separate questions thatare each going to have their ownprobability. In their testing, TypesafeFound that 13 questions handled in asingle call was 12.2 times cheaper and10 times faster than one at a time, andgot to the identical answers. MatthewBurman did an analysis of 724 live adswith each ad being asked 12 questions. One about the hook archetype, one aboutthe format, one about the offer, oneabout the CTA intent, one about theawareness, etc., etc. He found that thatsingle request for the 12 questions forthat ad took just 173 milliseconds tocomplete. And this is really where thepower of the model comes in. It's notjust that it's good at judgment. It'sthat it's fast enough and cheap enoughto ask and make judgments abouteverything. A million input tokens costjust 4.2. And output tokens are free. And each call is going to take between70 and 500 milliseconds, and as we'veseen, can combine a bunch of differentquestions in a single call. Now,importantly, Jev makes trade-offs thatallow it to be good at the things thatIt does. But that also means that itdoesn't do certain things. It's notgoing to write code. It's not going todraft contracts. And it's not going tomake nuance decisions that involve avariety of factors that aren'tquantifiable and clear. It's going to beused for small judgments at volume. Which bucket? How urgent? Is itrelevant? Is it safe? So, now let's talkabout six different ways that people areactually putting Jeb to work outside ofjust cool demos. A first category isanalyzing what you already have. Asecond category is searching by meaning.A third category is triaging what comesin. A fourth category is checking workagainst your rules. A fifth category isspeeding up AI agents. And a sixthcategory is responding instantly. Solet's move to category one, analyzingwhat you already have, aka what's inthis pile. This is basically whereyou're going to take a collection youalready have, ask Jeb the same questionsAbout every item, and then count theanswers. This is analysis thatpreviously could not be justified doingby hand, but is now going to takeseconds and cost sense. One example was that from Matthew Burman that we justtalked about where he analyzed and brokedown 724 live ads from 37 brands toidentify every hook, every format, everyoffer, every CTA. The idea is to be ableto then take this data and compare it towhat actually performed, giving youincredibly deep and complex fine graininformation about advertisingperformance that would have beenextremely difficult before. It took 40seconds and 9 cents worth of tokens forBurman to get the full breakdown ofthose 724 live ads. Two days later, Burmanwent further and instead of just askingthose questions in general, this time hehad Jev scroll 723 ads as 30 differentbuyer personalities. The personalitieswere archetypes like gym owner, dentaloffice manager, toddler mom, or AI. Curious engineer, Jeff's job was to askfor every buyer in every ad, would thisperson stop scrolling or keep going? Itcost 22 cents to get 21,690stop or scroll decisions. Now, ofcourse, this is not data. This wasn't 30actual buyers. It was just buyerarchetypes as considered by an AI. Inother words, it's better to view this asa hypothesis generator. But it's hardnot to think that marketing will shiftto using this as a key part of itsprocess to pressure test ad or landingpage angles before paying for realtests. Lots of other folks are runningJeb over whole archives, whether that’semail archives, past exposts, folders ofdocuments, or databases of researchprojects. And in each case, what Jeb is doing is asking the same few questions,but about every single item in thearchive. Ian Nuttall, for example, hadJev ask eight questions each topic,hook, and tone about nearly 3,300 oftheir previous exposts to be able toThen compare that data with engagementmetrics to get a better sense of whatactually worked. And by the way, thoseeight questions each across 3,300 postscost about 13. So, how might we trysomething like this? Basically, any bigpile of data you have. A year ofcustomer emails, CRM notes, a folder ofmeeting transcripts—anything where youcan write a few simple questions and runthem against all of those queries. That's a potentially useful place for adev job. Category 2 is searching bymeaning. In other words, which of thesematch what I mean? The idea is to beable to describe what you're looking forin plain words and let Jev check everycandidate. The power is that it can findwhat you mean even when the words don'tmatch. There were a bunch of examples ofthis with people basically using it as anew approach to natural language searchon websites. Justine Moore from A16Zgave the example of scanning thousandsof Zillow listings and classifying. Properties by things you can't normallyfilter for, such as architectural style,renovation status, or proximity tofreeways. In one example that piqued myinterest, Burhan took a 90-minute videoand was able to clip it based on themesthat they described in natural language.So, for example, across that 90-minutevideo, they were able to look for andclip quotes of their predictions for when AIwill automate AI research, and they wereable to find a bunch of examples of thatin under two seconds for under two cents. Another example of this is an SEO auditwhere Boura deployed Jev to read all 586pages of their website to rebuild thesite's internal link map which it wasable to do in 45.1 seconds for 21.Claude Opus 5 only got through 21 ofthose pages and spent a buck 43. Bouraexplains internal linking is the perfectJev job. It's not writing. It is 8,790yes or no calls. Does this page have areal reason to link to that one? And isthere anchor text already sitting in theCopy? That is a classification problemand we've been paying frontier prices todo it one page at a time. A finalsubcategory of the searching by meaningis filtering what you read by what youcare about. So, for example, RobinBillgill built a real-time AI slopdetector. As they scroll X, the slopdetector gives a confidence score onthat 0 to 1 scale about how much itthinks a post is AI slop or not,blocking out the ones that reach a certain confidence threshold. This sortof filtering, though, can be applied in abunch of different ways. Elvis Sonx gavean example of letting Jev browse 384morning news stories and identifyingwhich of them 15 different brands shouldbe paying attention to. That wascompleted in 24.9 seconds for 19. And byway of comparison, in that same timeperiod, Opus 5 got through four of thosearticles, leaving 380 unread. Jev usecase category 3 is triaging what comesin. In other words, what is this, andWhere does it go? The kinds of questionsthat people are experimenting with arethings like, "How important is thisemail? Is this downloaded file aninvoice? Is this link malicious? Shouldthis lead go to sales self-serve ornurture?"One really interestingexperiment came from Jonathan Unicowski. He wrote, "Every email app shows yourinbox in reverse chronological order.What if it was live prioritized byimportance instead?" Jeb's job in thiscase was to rate every email'simportance. And in the test, it was ableto rate 100 emails in 453 millisecondsfor about a tenth of a cent, withJonathan reporting that Jeb's ratingmatched his own on every single one.This is a use case that I want rightnow, not tomorrow. In fact, I want ityesterday. And so do all of the peoplewho are sitting there beating theirheads against the wall because I haven'tresponded to them yet. Some other folksare experimenting with a pattern ofAsking one quick question per item. Marcel Pio, the CTO of Beyond Code,writes, "I built a Mac OS app thatmonitors my downloads folder along witha customizable set of rules. Is thedownloaded file an invoice? Move it to aspecial folder with the correct filename. No other LLM calls involved, justJev." Dev Ed uses Jev for live chatmoderation, removing swearing andnegative comments as they arrive. StevenTay of the Dub link shortener fed Jev10,000 malicious domains they'd caughtbefore to flag bad links on their freeshortener service. He wrote, "This hasbeen something that we've been wrestlingwith since day one. With Jev, we solvedit in 2 hours." In fact, this sort of triage is so integral to so much ofbusiness. In a post about 10 Jev usecases for a marketer, Yumanx identifiedworkflow decisions as one, saying, "Mostworkflows eventually hit the samequestion. What should happen next?"That's probably the best place to useJev. Box gave an example having alreadyexperimented with incident triage whereJev is used to judge customer impact andseverity as well as things like whichroutes it can be escalated to. Udu CRMhas a proposed module that asks threethings about every new lead. Theirpriority, their buying readiness, andwhether it's spam or not. If you had to look at just one area to explore,especially if you were in a company withmultiple people touching the same leadsor customers or contacts, this sort oftriage and routing is, I think, where Jevis going to become absolutely integralbasically from the moment that it getsintegrated into the systems. A fourthcategory of Jev use cases is checkingwork against rules, i.e., does this meetthe bar. Instead of the standard LLMquestion of please review this, you turnplease review this into specificquestions and then ask those questionsabout every draft. One example of thiscame from Langchain, which was gradingAn AI agent's work the same way every time. Harrison Chase from Langchainpointed out that Jev is "great forevals, especially online evals where youwant to grade lots of traces." Now, thismight feel initially like something thatis more for devs than for other types ofknowledge workers and companies. Butgiven how much all of us are going tostart putting agents into production ormanaging agents that already exist,things support bots, research agents,etc., having a better tool to buildevaluation systems into how we judgethose agents work seems like it couldvery easily become core infrastructure. And the team at every showed how youcould use this sort of ability to checkwork against rules at mass scale and atincredible speed as a way to improve AIwriting. They planted mistakes on 12passages of writing with Jev catchingsix of seven as compared to Claude Fable,which caught all seven. Jev, however,caught 6 in 0.35 seconds as comparedTo 8.83 seconds for Fable 5.1. And itdid so at about 580 times cheaper,meaning that you could rerun that samecheck a huge number of times and stillhave it be both cheaper and faster thanusing a Frontier model for the same sortof review. So, how might you actuallyturn this into something that you woulduse? You'd basically need to go through a translation process for your rules forwriting. So, let's imagine that you hada style guide or a list of phrases younever want to see. You could turn eachof those into a yes or no question andthen run those yes or no questions onevery paragraph of your next longdocument as a way to ensure no AISMs orother writing third rails in your keycommunications. And this gets, I think,at one of the biggest rewirings thatwe're going to need to do with Jev. Alot of the unique value of Jev is notjust being able to do a thing. It'sbeing able to do a thing at such scaleKind rather than a difference in scale. Being able to realistically check everysingle sentence in minute detail forAISMs, in other words, becomescategorically different from justrunning a generic LLM check across thedocument as a whole. The fifth categoryof use cases that lots of people wereexperimenting with admittedly does get alittle bit closer to the developerrealm, but I think at least for the sakeof completeness, it's still worthdiscussing. These use cases you mightsum up as speeding up your AI agents.And a lot of this is around the sort ofmodel routing that we've been talkingabout for the past several months. Thekinds of questions people were askingwere things like which model can handlethis task? How much reasoning does thisstep need? Which skill fits thisrequest, if any? Is this old tool outputstill relevant? You can see how in eachof these cases, the common thread ispeople using these small automated microJudgments to route an agent to the right level of intelligence, the rightcontext, the right skills, the righttools to do whatever its job is in themost efficient way possible. Vchenwrote, "People use Jev to pick a modelbefore a task. I made it change GPT6reasonings effort inside Codeex duringthe task. More thinking when stuck, lessfor routine steps. And in their test,they found 50% lower Astra costs whilealso getting faster runs. People arealso using Jev to manage the contextwindow. Daniel Son built somethingcalled Jev skill suggestion for Claudecode, where quote, "For every request,Jev classifies which skill best matchesthe task, then injects only that skillinto Claude's context." This led to an88% decrease in tokens and cost. I knowa lot of even you formerlynon-developers have started to besufficiently proficient with things likecodecs and cloud code that you've builtup big skills libraries and these sorts. Of dev-based tools are potentially a wayto stop sending all of those librariesto the model with every single request.And one thing to note here is that whilea lot of these use cases that we'rediscussing are at this stage individualexperiments, you're also going to seeJev and judgment models like it builtnatively into the tools and harnesses weuse. AJ Aspher, for example, wrote, "Webuilt a new harness using Jev that cutsthe cost of repetitive work by 90%. The harness learns the job as it runs,moving steps from LLM calls to code. Theexample they gave was compliance alerts,and they measured the cost per alert,batch by batch, with each batchrepresenting 50 of those compliancealerts. At the beginning, the cost peralert was about $2.95, but by alert10,000, it was down to just 25. A sixth category of Jev use cases areabout instant response. What does thisperson want right now? A lot of peoplewere experimenting with some version ofThis. Marcus Low wrote, "What if copypaste was smart? Copy a resume, pasteinto an application, and the fields fillthemselves." Norman on X also did aversion of this. Splitting the pastedtext into pieces, working out what eachfield is, matching them, checkingagainst the original, and pasting onlythe confident matches. Given how muchwork at work is moving details from oneformat, i.e., emails, PDFs, meeting notes, etc., into other places where thatinformation is supposed to live, likeCRM, intake forms, or templates. This isa category of use cases that feels veryrelevant. Now, just as important asknowing where Jev is useful is knowingwhere you got to be careful with it aswell. A couple of places that I thinkwarrant greater caution are areas likehiring, money, and security. On hiring,you can see how if Jev is used by a set ofapplications, you might be tempted touse something like Jev to rank them. But the problem is that even if you've givenIt good criteria, Jev is going to returnwith a number that doesn't have anyreasoning attached. This, by the way, isa great example of where you might wantto build a more complex system that usesmultiple types of AI. Imagine you dothat same sort of ranking with Jev, butthen automatically have other types ofLLM review, for example, for bubblecandidates that might be deserving of asecond look. Types safe, the makers ofJev, actually even try to be clear aboutwhat they think Jev is bad at. Some ofthe things they put include multi-stepquestions where they see accuracy dropwith each hop. They point out that Jevisn't really good at counting math ordates, that it can extract the facts,but that you'll want to do theactual math elsewhere. And they alsopoint out some other issues likeproblems with consistency and problemswith reading intent. So, if you'retrying to figure out if a task is a goodfit for Jev, four criteria might help. The first is that you can write down theanswers in advance. Think categories, ayes or no binary, a scale. If the answeris a sentence or a calculation, that'sprobably not a good fit for Jev's sortof judgment model. Criteria number twois about volume. There's a pile or astream, such as hundreds of tickets,hundreds of emails, things like that. Athird criterion is about stakes, where awrong answer is cheap or it's easy tocatch. Basically, although I'm callingit a judgment model, you don't want toleave things up to its judgment alone ifgetting it wrong has big consequences. Afinal criterion is whether you can give it the evidence that it needs as text inunder 32,000 tokens. Once you've figuredout a good task, the next step is goingto be to write good questions. Some ofthe tips there from across all of thesedifferent examples are things like onejudgment per question. In other words, anot so good question is, is this a goodlead? Cuz that’s actually not oneJudgment. That's a whole bunch ofdifferent judgments embedded in one goodlead might refer to how good a fit theindustry is for your service, whetherthe company size matches who you like to serve, how strong their buying intentis, etc., etc., etc. Basically, you'regoing to want to break questions intotheir constituent parts. If yourquestion involves a scale, you need todescribe each level of that scale inwords. And you also might even want totake advantage of Jeff scaleopportunities to ask more questions thanyou think that you need. Lastly, likeeverything with AI, you're going to wantto do some tests before you actuallytrust it in production. It might be apain, but if you are using that emailclassifier, for example, to try to organize things based on priority, maybeyou want to label 50 emails yourself asa test and see how it compares to makesure it's actually going to do for youwhat you want it to do. Now, as we wrapUp, I will be posting this presentationthat I've been working through on thisepisode's companion site on AI dailybrief.ai. And the last couple of pagesget a little bit more practical with oneidea for each project in the six usecase areas. Things like a contentarchive analyzer, a research raider, aninbox triager, a rewrite checker, etc. And there'll even be a starter prompt inthere as well. In the first 10 dayssince Jev was released, we've gone frombuzzy exciting concept to actuallyvaluable production use cases extremelyquickly. And yet, because this is atcore a new primitive in its ability toapply simple judgment at scale, atspeed, and for effectively no money, Ithink it's going to take some time forus to really figure out just how deeplywe can weave this into all sorts ofdifferent use cases. As more and morecome online, I will come back and sharethe best of them. For now though, thatis going to do it for today's AI daily. Brief. I appreciate you listening orwatching, as always, and until next time,peace., How People Are Actually Using Jev, The AI Daily Brief, 01m3f7g27artgwkfjfpareb5vw.