YouTube2Text

Insane Jev Use Cases You Need To Use Right Now — Transcript

by AI LABS · 2,806 words · 385 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Jev is the fastest model out there and
  2. 0:02people have been using it in a lot of
  3. 0:04insane ways ever since it was released.
  4. 0:06But the biggest difference we found was
  5. 0:08when we connected Jev with the coding
  6. 0:10agents that we use every day. So we
  7. 0:12tested Jev in our own workflow and found
  8. 0:15seven places where Jev makes a huge
  9. 0:17impact. And not only does Jev make your
  10. 0:19agent way faster and cheaper to run in
  11. 0:21these places, but it also fixes some
  12. 0:23problems that your agent always had. If
  13. 0:25this is your first time, we're a
  14. 0:26software company and this is our
  15. 0:28channel, AI Labs. And in this video,
  16. 0:30we're going to show you all seven Jev
  17. 0:32use cases, what each one changes in your
  18. 0:34workflow, and how to use them. This
  19. 0:36[snorts] section is for those who don't
  20. 0:38know what Jev is. And if you already
  21. 0:39know it, you can skip ahead to the next
  22. 0:41section using the timestamps in the
  23. 0:43description below. Now, Jev is an AI
  24. 0:45model made by a company called Typesafe.
  25. 0:47Unlike other models like Claude and GPT
  26. 0:50which write their answers out in text,
  27. 0:52Jev doesn't do that and instead it picks
  28. 0:54from a given set of options which is why
  29. 0:56it's also called a decision model. So
  30. 0:58instead of the way you prompt Claude and
  31. 1:00other models, Jev chooses from options
  32. 1:02whose answer can be either a yes or no
  33. 1:04or a list to pick from or a score to
  34. 1:06give to anything and it will respond by
  35. 1:08choosing between them. And along with
  36. 1:10its pick, it also gives you a number
  37. 1:12called confidence value. Confidence
  38. 1:14value is a number that shows how sure
  39. 1:16the model was when it chose between the
  40. 1:18options. And Jev is also way faster than
  41. 1:20other models. This is because models
  42. 1:22like Claude and GPT write their answer
  43. 1:24one token at a time. But Jev doesn't
  44. 1:26work like that. Jev answers all of your
  45. 1:29questions at the same time and gives you
  46. 1:31a number at much higher speed. Now you
  47. 1:33might think that models like Claude can
  48. 1:35also make these decisions themselves.
  49. 1:37But the problem with these models is
  50. 1:38that they have to think and write every
  51. 1:40answer. So each one costs you time and
  52. 1:42usage. But Jev doesn't write huge
  53. 1:44amounts of text like those models
  54. 1:46because it just gives one decision. So
  55. 1:48each decision takes a fraction of a
  56. 1:50second. Jev is also a lot cheaper than
  57. 1:51the usual models because it isn't
  58. 1:53writing tokens like the others are
  59. 1:55doing. You only pay for what you send to
  60. 1:56Jev and the answer it sends back is
  61. 1:58completely free. Now [snorts] Jev isn't
  62. 2:00a model that works inside Claude code or
  63. 2:03codeex as a replacement for all of the
  64. 2:05models that you have on it. Jev is just
  65. 2:07used for decision-making only because it
  66. 2:09doesn't write text like the other
  67. 2:10models. It can't be used to write code
  68. 2:12or let the model use tools. It can be
  69. 2:15used just as a decision maker inside
  70. 2:17tools like claude code and codeex. You
  71. 2:19can get access to jev via types safe's
  72. 2:21own platform. But the problem is that
  73. 2:23new signups are paused as of now because
  74. 2:26of the increasing demand. But it is also
  75. 2:28available on other platforms like open
  76. 2:30router and versal AI gateway. Both of
  77. 2:32them are platforms that give you one key
  78. 2:34to use AI models from lots of different
  79. 2:36companies and Jev is one of those
  80. 2:38models. You can use whichever you want,
  81. 2:40but we went with Versal AI gateway. To
  82. 2:42create the API key, you need to open
  83. 2:44your Versal dashboard and go to the API
  84. 2:47keys page inside AI gateway. When you
  85. 2:49create a key there, versal only shows it
  86. 2:51to you once, so you need to copy it
  87. 2:53right away because this is the exact key
  88. 2:55that you will be using in your workflow.
  89. 2:57Then you need to use the export command
  90. 2:59in the terminal to save the key as an
  91. 3:01environment variable. An environment
  92. 3:03variable is basically a saved value that
  93. 3:05anything you run in that terminal can
  94. 3:07read. So when you open Claude in that
  95. 3:09same terminal, Claude can find the key
  96. 3:11and use it. The first use case of Jev is
  97. 3:14replacing the compaction method that
  98. 3:16coding agents like Claude code use. The
  99. 3:18usual compaction basically sends the
  100. 3:20whole conversation back to the model and
  101. 3:22asks it to summarize it by following a
  102. 3:24set of instructions and that can take a
  103. 3:26long time. Jev is a good choice here
  104. 3:28because it can make a decision on
  105. 3:30whether something is important or not.
  106. 3:32It can actually read your conversation
  107. 3:34so far and decide what is important in
  108. 3:36the conversation and the model's
  109. 3:38responses. So instead of the compaction
  110. 3:39summary these models by default go for
  111. 3:42Jev gives the things that are actually
  112. 3:44important. To do this there is already a
  113. 3:46plug-in called fast Jev compaction and
  114. 3:48what it does is replace the compaction
  115. 3:50summary with Jev decisions. You can
  116. 3:52install it by using their install
  117. 3:54workflow but that repo doesn't support
  118. 3:56the versal AI gateway that we were using
  119. 3:58because it was built for Typesafe's own
  120. 4:00platform. So what we did was copy the
  121. 4:02plug-in into the project that we were
  122. 4:04working in. Then asked Claude Code to
  123. 4:06make some small changes in it so that we
  124. 4:08could use it with the gateway as well.
  125. 4:10This plug-in uses a hook to redirect
  126. 4:12Claude from following its usual compact
  127. 4:14command instructions. Since it's stored
  128. 4:16in the project folder rather than
  129. 4:17installed through the Claude
  130. 4:19marketplace, you'll need to ask Claude
  131. 4:20to update its settings. So it recognizes
  132. 4:23this plugin as well. Once you install
  133. 4:24it, you just run the compact command
  134. 4:26like normal. The plug-in hook takes over
  135. 4:28and replaces the usual compaction. It
  136. 4:30compacts much faster, usually in less
  137. 4:32than a second. There is also a limit for
  138. 4:34when it uses Jev for compaction. If your
  139. 4:36session is below 25%, it will not use
  140. 4:39Jev. Instead, it will use the normal
  141. 4:41compaction with clawed models. This
  142. 4:43helps make sure the Jev API is only used
  143. 4:46when it can make a big difference. But
  144. 4:47before we move on to the next use case,
  145. 4:50it would be great if you subscribe to
  146. 4:51the channel and hit the hype button.
  147. 4:53This small gesture of support goes a
  148. 4:55long way for us.
  149. 4:56>> [snorts]
  150. 4:56>> The second use case basically makes use
  151. 4:58of Jev's speed and low cost. And it is
  152. 5:00about using Jev as a checking layer in
  153. 5:03between that makes sure whatever you
  154. 5:04want is done right. When you're working
  155. 5:06on a project, there are a lot of docs
  156. 5:08and files that describe what features
  157. 5:09you want and what features need to be
  158. 5:11built. The thing is that there are a lot
  159. 5:13of document files which can actually get
  160. 5:14hard for the agent to keep track of.
  161. 5:16Since your default model is the one that
  162. 5:18worked on the feature, we can't let it
  163. 5:20be the judge. In order to make use of
  164. 5:22Jev's speed, we created a hook that runs
  165. 5:24whenever Claude edits the files that
  166. 5:26control who can access what in our app,
  167. 5:28and it asks Jev whether each rule in our
  168. 5:31docs has a test for it. This helps
  169. 5:33because it lets us find which rules were
  170. 5:34listed in the docs but never got a test.
  171. 5:36To use it, you just need to prompt
  172. 5:38Claude to check any particular part of
  173. 5:40your app and it will load the Jev skill,
  174. 5:42show you which rules are missing a test
  175. 5:44and then you can ask Claude to write
  176. 5:46tests based on the report Jev gave. The
  177. 5:49third [snorts] use case is about skills.
  178. 5:50Jev is actually a great model when it
  179. 5:52comes to making decisions. Agents only
  180. 5:54load the skills name and description in
  181. 5:56the context window. But there is a
  182. 5:58problem. When you're using a lot of
  183. 6:00skills on a project, the model has to
  184. 6:02consider all of them in the context
  185. 6:04window. And then from those skills, the
  186. 6:06model has to decide which ones to use
  187. 6:08and which ones not to use. That choice
  188. 6:10is actually something that you can make
  189. 6:12better by using Jev because Jev can make
  190. 6:14this decision much more quickly. So for
  191. 6:16that reason, we created a hook called
  192. 6:18the skill picker. This skill picker
  193. 6:19basically works by reading the names and
  194. 6:21descriptions of every skill just like
  195. 6:24Claude Code does and then it reads your
  196. 6:25prompt and picks the one skill that fits
  197. 6:27it or none if the prompt doesn't need a
  198. 6:29skill. So when you're actually
  199. 6:31submitting anything, it will read the
  200. 6:33prompt first and then tell Claude which
  201. 6:35skill to use for that task. So Claude
  202. 6:37doesn't have to work that out from every
  203. 6:38skill on its own. So when you're working
  204. 6:40on any task, you'll see that after every
  205. 6:42prompt, you give a hook kicks in and all
  206. 6:45your skills get selected way faster than
  207. 6:47it would have been normally without Jev.
  208. 6:48But before we move on, let's have a word
  209. 6:50by our sponsor, Zapier. If you've built
  210. 6:53an app, you know how quickly users start
  211. 6:55asking you to connect it to the other
  212. 6:56tools they use. What surprised us was
  213. 6:59how little it took. The whole
  214. 7:00integration is really just writing
  215. 7:02Node.js, and that one build reaches
  216. 7:04Zapier's 1 million users. We scaffolded
  217. 7:06a new integration from a template, then
  218. 7:08wired up our own app, quicklink, a
  219. 7:10little URL shortener we built. We
  220. 7:12defined its short link object once, and
  221. 7:14that single definition became a new link
  222. 7:16trigger and a shorten URL action with no
  223. 7:19extra wiring. We tested it locally and
  224. 7:21pushed it up, and it showed up right in
  225. 7:23the Zap editor, ready to use in a real
  226. 7:25Zap. And once it's in there, your app
  227. 7:26sits alongside Zapier's 9,000 plus app
  228. 7:29ecosystem, ready to connect to the tools
  229. 7:32people already use. and Zapier handles
  230. 7:34the plumbing for us from storing our API
  231. 7:36key to polling for new links and quietly
  232. 7:38skipping the duplicates. So, we only
  233. 7:40write the logic that's actually ours.
  234. 7:42Start free with the 14-day trial and the
  235. 7:45docs are linked in the description
  236. 7:46below. The fourth [snorts] use case is
  237. 7:48about finding files faster. Now, you
  238. 7:50might already know that whenever the
  239. 7:52model has to find a file in your
  240. 7:53project, it always searches the files by
  241. 7:56actually running an agent called
  242. 7:57explore. It runs in a separate context
  243. 7:59window and just returns the exact file.
  244. 8:01So it doesn't fill up the main agents
  245. 8:03context window with information from the
  246. 8:05files it opened that weren't needed. Up
  247. 8:07until a while ago, the default model
  248. 8:09Claude used for running Explore was
  249. 8:11Haiku, but then they changed it so
  250. 8:13explore now uses the same model as your
  251. 8:15main session. That means if you're
  252. 8:17working with Opus, Explore runs on Opus
  253. 8:192, which makes every search slower and
  254. 8:21more expensive. So to make use of Jev,
  255. 8:23you can create a skill that asks Claude
  256. 8:25to do what Explore does using Jev. You
  257. 8:27can also ask it to create an agent that
  258. 8:29replaces explore for this project only.
  259. 8:32But we chose a skill because a skill can
  260. 8:34also store the scripts and other things
  261. 8:36it needs. The skill works in two steps.
  262. 8:38First, it runs a normal keyword search
  263. 8:40to find the files that might be related
  264. 8:42to your question. Then Jev scores those
  265. 8:44files 20 at a time on how closely each
  266. 8:46one matches your question. And Claude
  267. 8:48only opens the few files at the top. So
  268. 8:51whenever you give it a prompt, it
  269. 8:52basically loads the skill and runs that
  270. 8:54search and ranking. And within a few
  271. 8:56seconds, Claude knows which files to
  272. 8:58read without opening all of them. Like
  273. 9:00it took 1.6 seconds to rank the 35 files
  274. 9:03and it put the right file first. The
  275. 9:05fifth use case is about reviewing code.
  276. 9:07Now you might already know that the
  277. 9:08agent that built a feature shouldn't be
  278. 9:10the one that checks it. That's why when
  279. 9:12we are working on our own projects, we
  280. 9:14always ask a different model to review
  281. 9:16the work. But there is a problem. This
  282. 9:18review that runs whole is actually slow
  283. 9:20because of the way Claude normally does
  284. 9:22step by step. So, for that reason, we
  285. 9:24added Jev before the review starts. Jev
  286. 9:26doesn't replace Claude as a reviewer.
  287. 9:28It's just an extra checking step that
  288. 9:30helps Claude review things better. Jev
  289. 9:32basically reads the change and answers
  290. 9:34seven yes or no questions about it, like
  291. 9:36whether this change is actually allowed
  292. 9:38under the rules and whether it's
  293. 9:39something that should be changed in the
  294. 9:41first place. If the answer to all of
  295. 9:43them is no, the reviewer only does one
  296. 9:45quick round. But if any answer is yes,
  297. 9:47or Jev isn't sure, the change gets the
  298. 9:49full review just like before. This helps
  299. 9:51because the risky changes still get
  300. 9:53checked properly and the small ones
  301. 9:55finish much faster. The [snorts] sixth
  302. 9:57use case is about testing your app in
  303. 9:59the browser. When you're working on any
  304. 10:00app, there are times when the agent
  305. 10:02opens the browser and runs the app like
  306. 10:04a real human would and sees if what it
  307. 10:06built works well or not. This helps find
  308. 10:08the kinds of issues that only appear
  309. 10:10when someone is actually using the app.
  310. 10:12Claude code can do this when you give it
  311. 10:14a browser tool and it has its model
  312. 10:16check whether the thing that was asked
  313. 10:18to be built was built properly or not.
  314. 10:20The model opens the browser and
  315. 10:22interacts with the app by clicking
  316. 10:23different buttons and it decides by
  317. 10:25itself where to click next. We can give
  318. 10:27that decision-making to Jev as well.
  319. 10:29That makes testing faster because the
  320. 10:31final call on whether the app works
  321. 10:33still comes from the main model, but Jev
  322. 10:35decides where to click next much faster.
  323. 10:37So to help with that, we asked Claude to
  324. 10:39create a skill for it. The skill opens
  325. 10:41the app and makes a list of every button
  326. 10:43and link on the page. And then Jev picks
  327. 10:45the one that gets it closer to the goal.
  328. 10:47It clicks that, makes a new list from
  329. 10:49the next page, and repeats this until
  330. 10:51Jev decides the goal is done or that the
  331. 10:53page is blocking it. In our app, we used
  332. 10:55it to sign in as an admin, a manager,
  333. 10:58and an employee and check that each of
  334. 11:00them can only reach the pages they're
  335. 11:01allowed to. And then Claude compares
  336. 11:03what happened against the rules in our
  337. 11:05docs. So, whenever you ask Claude to
  338. 11:07check the app in the browser, it loads
  339. 11:09this skill and uses Jev to test it, and
  340. 11:11it saves a lot of time compared to doing
  341. 11:13it normally.
  342. 11:14>> [snorts]
  343. 11:14>> The seventh and last use case is about
  344. 11:16keeping claude to your rules. When
  345. 11:17you've been working with agents, you
  346. 11:19would have noticed that there are times
  347. 11:20when they stop following what you
  348. 11:22already planned in the specs. The rules
  349. 11:24in your claw.md are just instructions.
  350. 11:26So the model follows them most of the
  351. 11:28time but not every time. For that, we
  352. 11:30created a hook because a hook runs every
  353. 11:32single time and the model can't skip it.
  354. 11:34So in our project, we have a lot of
  355. 11:35rules in other files that are specific
  356. 11:37to different parts of the app. Then
  357. 11:39whenever Claude is about to edit a file,
  358. 11:41the hook sends Jev the change along with
  359. 11:43the rules for that file. And Jev answers
  360. 11:45with a yes or no whether the change
  361. 11:47breaks each rule. If Jev is at least 80%
  362. 11:50sure a rule is broken, the edit gets
  363. 11:52blocked and Claude is told exactly which
  364. 11:54rule it broke so it can fix it. When we
  365. 11:56tested it, we asked Claude to create a
  366. 11:58file that reads the employee data
  367. 12:00straight from the browser, which breaks
  368. 12:02one of our rules. Jev decided that the
  369. 12:04change broke that rule, so the edit was
  370. 12:06blocked before the file was even
  371. 12:07created. and Claude told us exactly
  372. 12:09which rule it had broken. Now, all the
  373. 12:11skills and setup you saw in this video
  374. 12:14is available for you to use in our
  375. 12:15community, AIAS Pro. So, if you found
  376. 12:18value in what we do and want to support
  377. 12:20the channel, this is the best way to do
  378. 12:22it. The links in the description. That
  379. 12:24brings us to the end of this video. If
  380. 12:26you'd like to support the channel and
  381. 12:27help us keep making videos like this,
  382. 12:29you can do so by using the super thanks
  383. 12:31button below. As always, thank you for
  384. 12:33watching and I'll see you in the next
  385. 12:35one.

About this transcript

This page contains the full transcript of Insane Jev Use Cases You Need To Use Right Now by AI LABS, generated from the public captions YouTube serves with the video. The transcript has 2,806 words across 385 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.