JEV Explained (Real Use Cases) — Transcript

Full transcript

1. 0:00Today we're talking about Jev, and how
2. 0:01you can use this as a developer. So, Jev
3. 0:03is a new model from Type-A AI that is
4. 0:05getting a lot of hype online. Jev is not
5. 0:07a large language model. This is a
6. 0:08completely different type of model.
7. 0:10What's interesting [music] about it is
8. 0:12it's much faster and much cheaper than a
9. 0:13traditional large language model for
10. 0:15very specific use cases. So, in this
11. 0:17video, I'm going to walk you through
12. 0:18some practical [music] real examples of
13. 0:20how you can build software, AI systems,
14. 0:22and automations using this new model.
15. 0:25I'll link all of the resources, the
16. 0:26code, [music] and the sample
17. 0:27applications in the description below.
18. 0:29So, let's get into it. So, this is the
19. 0:31model we're going to be talking about
20. 0:32from Type-A AI. They call it a system
21. 0:34one model. If you're familiar with the
22. 0:35book Thinking Fast and Slow, the idea is
23. 0:37we have two types of thinking, fast
24. 0:39thinking where we make instinctive
25. 0:41decisions, and much slower and more
26. 0:43deliberate thinking. So, Jev is all
27. 0:44about making fast decisions,
28. 0:46essentially.
29. 0:48So, you can see here that you have an
30. 0:50input, which is a question. You pass in
31. 0:52a request to the model, and what's
32. 0:54different about Jev in particular is you
33. 0:56get back a typed answer. So, a specific
34. 0:59schema or structure comes out that you
35. 1:01can depend on. And the schema can have
36. 1:04probabilities and confidence scores,
37. 1:06essentially, that allows you to then
38. 1:08make a decision. So, this is going to be
39. 1:09useful in specific use cases. This is
40. 1:12not a general-purpose AI model. I would
41. 1:13use this for two specific use cases. So,
42. 1:16the first use case I would use this for
43. 1:17would be routing, and the other one
44. 1:19would be classification. So, these are
45. 1:21really common in AI systems, and I think
46. 1:23this is where I would use Jev myself.
47. 1:26So, the model has three different
48. 1:27primitives, and we're going to take a
49. 1:28look at each of these now in Python. So,
50. 1:30you have a choice, a score, and a null.
51. 1:33And we'll break down when to use these
52. 1:35and what they are. All right, so the
53. 1:36first thing we're going to need to do is
54. 1:38get an API key. So, in order to use
55. 1:39this, you need to get an API key. There
56. 1:41are two ways to use this. The first way
57. 1:42is to get an API key from the
58. 1:45playground. So, if you go over here to
59. 1:46the console,
60. 1:48you can see here, click down here, and
61. 1:49you can get an API key. I've already
62. 1:51created some. The other way you could
63. 1:53use this is with Open Router. If you're
64. 1:55not familiar with Open Router, it's a
65. 1:57way of using different models through a
66. 1:59standard API. So, this can be very
67. 2:00useful, especially if you're trying out
68. 2:02the model. You can get this from
69. 2:04OpenRouter. So, either option will work,
70. 2:06but you're going to need an API key
71. 2:07either way.
72. 2:08Okay, so we're going to start by looking
73. 2:09at some code examples cuz I think this
74. 2:11is what probably the easiest way to
75. 2:13understand this new model. So, the first
76. 2:15type we're going to look at is this yes
77. 2:17or no type, which is called a null.
78. 2:19Essentially, you need to import the Type
79. 2:21Chat SDK. You also need to make sure
80. 2:23your API is configured for this to work.
81. 2:25But once you got this set up, you can
82. 2:27use the client like this. It's pretty
83. 2:28straightforward to use this client. You
84. 2:31instantiate it, and then you have
85. 2:32client.system.one. We pass in the state,
86. 2:35and then we can define this null type.
87. 2:37So, in this example, what we're doing is
88. 2:39basically we're passing in a like a fake
89. 2:41customer support ticket. Does the
90. 2:43customer request contain personally
91. 2:45identifiable information? And then we're
92. 2:47going to return a score. So, let's go
93. 2:49ahead and run this example and see what
94. 2:51comes back. [music] So, we can type in
95. 2:52response.nulls
96. 2:54to get back the and what you can see
97. 2:56here is the contains PII field, and then
98. 2:59the result is 0.95 is the score.
99. 3:02>> [music]
100. 3:02>> So, this gives you back a number between
101. 3:050 and 1. So, 0.5, anything above that is
102. 3:08going to be yes, and then anything below
103. 3:10that is essentially going to be no.
104. 3:12So, the score here is pretty high, so we
105. 3:13have pretty high confidence that this
106. 3:15does contain personally identifiable
107. 3:17information, which makes sense cuz I put
108. 3:19my email address in that, so that's
109. 3:20correct. And then if I was to remove
110. 3:22that, and then if we run this again, you
111. 3:24should see that
112. 3:26we get back a negative answer because it
113. 3:28doesn't contain identifiable
114. 3:30information, so we have a much lower
115. 3:31score.
116. 3:32And if you want to use this as a boolean
117. 3:34in your applications, what you can do is
118. 3:35assign
119. 3:36>> [music]
120. 3:37>> this value, and then anything greater
121. 3:38than a 0.5, you can mark that as a
122. 3:41positive that it contains PII. So, you
123. 3:44can see here the documentation for this
124. 3:45type. So, use this whenever you need a
125. 3:47yes or no answer. For example, does this
126. 3:49resume mention distributed systems, or
127. 3:52does this comment contain personal data?
128. 3:54So, those are the kind of use cases you
129. 3:56would use this type for. So, if you need
130. 3:57a very fast classification for something
131. 3:59like this, this data type would be the
132. 4:01one you'd want to use. So, the next type
133. 4:03we're going to take a look at is the
134. 4:04choice. This is going to be really
135. 4:06useful if you're doing a classification,
136. 4:08which is a very, very common operation
137. 4:09if you're building an AI system.
138. 4:11Essentially, what you need to do is take
139. 4:13an input and then put it into one of a
140. 4:15number of categories. So, in this
141. 4:17example, we have a customer support
142. 4:18ticket and we're trying to classify it
143. 4:20as
144. 4:21billing, technical, product, or other.
145. 4:23So, essentially, we're trying to find
146. 4:24the right support department to route
147. 4:26this customer support ticket to. And you
148. 4:28can see here we have a choice. We have
149. 4:30instructions and then we have a
150. 4:32description of the criteria. So, we have
151. 4:33the billing
152. 4:35category and then we have the main
153. 4:36request contains concerns a refund,
154. 4:38payment, invoice, or subscription
155. 4:39charge. So, essentially, what we're
156. 4:41doing is just defining the category and
157. 4:43then giving a description of what the
158. 4:44category is. And then the AI model is
159. 4:46going to use that to make a decision for
160. 4:49this given input, which of these
161. 4:50categories does it fall into? So, for
162. 4:52the this example, we've got I need to
163. 4:54get an invoice for my purchases. So, if
164. 4:56we run this example,
165. 4:58we should in theory see that this is a
166. 5:00billing question. So, you can see here
167. 5:02it's classified as billing with a
168. 5:04probability of 1.0. So, a very high
169. 5:06probability, very high confidence that
170. 5:08this is definitely a billing question.
171. 5:11And then if we were to change this
172. 5:13question, then obviously we'd get a
173. 5:14different answer. So, I need to reset
174. 5:19reset my password as an example. And
175. 5:21then if we go and run this example one
176. 5:23more time, you can see here now it's
177. 5:25classified it as technical.
178. 5:27So, this is a really common use case as
179. 5:29I said. If you're building an AI system,
180. 5:30this pattern comes up all of the time.
181. 5:32It's one of the patterns I use most
182. 5:34often in classifying an input, deciding
183. 5:36where to send it. You could imagine a
184. 5:38rag system or retrieval augmented
185. 5:40generation system where you need to
186. 5:41choose which database to look in for a
187. 5:44particular bit of information. This is
188. 5:46one way you could do that. You could
189. 5:48classify the question and figure out
190. 5:50where you need to get the information
191. 5:51from essentially. So, this is going to
192. 5:52be a really useful pattern. It's very
193. 5:55very fast. It's very very cheap. And if
194. 5:56you're doing this kind of thing, this
195. 5:58model is going to be really useful for
196. 5:59that. So, you can see here that a choice
197. 6:01is a question for to select one option
198. 6:03from a defined set. The answer includes
199. 6:05the selected option and a probability
200. 6:06for each option and a confidence score.
201. 6:09So, you can see here some more examples.
202. 6:10What programming language is this code
203. 6:12written in? And you give it a bunch of
204. 6:13options. So, this kind of thing, this is
205. 6:15going to be one of the more common use
206. 6:17cases for Jiva, I would say. So, the
207. 6:18next type we're going to take a look at
208. 6:19is a score. You can use this when you
209. 6:21need to have something that's on a
210. 6:22spectrum. So, for example, how severe is
211. 6:25this bug? Or how happy is a customer? Or
212. 6:27how much Python experience does a
213. 6:28candidate have? That's when you would
214. 6:30use this.
215. 6:31And so, we've got an example here where
216. 6:33we say, "The PDF export fails. I can
217. 6:35finish today's report using CSV
218. 6:37instead." So, we're trying to figure out
219. 6:39from this customer support question how
220. 6:42stuck the this customer is and how
221. 6:44urgent this request is.
222. 6:46So, we have three levels of
223. 6:46classification here. We have category
224. 6:48one where they can complete their work
225. 6:50without a workaround.
226. 6:52Category two is where they can complete
227. 6:54their work using a workaround. Or then
228. 6:55category three, they have no way to
229. 6:58solve this problem. And you can see here
230. 7:00now we've been moved into this one of
231. 7:01two categories. So, we have these
232. 7:03different levels. So, level zero and
233. 7:06level one. So, we've been put into this
234. 7:07level. So, the customer can complete
235. 7:09their work using a workaround, which is
236. 7:11the correct answer for this use case.
237. 7:14So, you can use this model with a coding
238. 7:16agent if you want to. So, you can
239. 7:17install this with Claude Code or Codex
240. 7:19if you want to. There are probably some
241. 7:21reasons why you might want to do that.
242. 7:22If you're doing things like local email
243. 7:24classification, things like that, you
244. 7:26can use that. If you want to do model
245. 7:28routing is another good use case here.
246. 7:30You can do that. And it's pretty easy to
247. 7:32get this installed. You have an agent
248. 7:34skill which allows you to install this.
249. 7:36And if you're using Claude Code, you can
250. 7:37just run this plugin. So, what we're
251. 7:39going to do is start up Claude Code and
252. 7:41then go and install the plugin and then
253. 7:43what we could do is something like
254. 7:44asking Jev to classify my email. So I
255. 7:47think Claude has access to my emails.
256. 7:49So what we could do is something like
257. 7:51use Jev to pull the latest
258. 7:55five emails from my inbox
259. 7:58and then classify them into
260. 8:01sponsorships,
261. 8:04business, or other.
262. 8:06Just choosing categories somewhat at
263. 8:08random. So we're going to pull the
264. 8:09latest five emails from my inbox. You
265. 8:10can see here we're now loading the Type
266. 8:12Safe AI skill. So the way I would
267. 8:14probably do this is to write a Python
268. 8:16script or something and then use Claude
269. 8:18code to dispatch to that code. So
270. 8:20another example of how you might use
271. 8:21this would be to build something like a
272. 8:23customer support ticketing application.
273. 8:25So again, these kind of applications are
274. 8:27quite common in the business world if
275. 8:28you're building any kind of software
276. 8:30with AI. This is a common use case. So
277. 8:32we're going to create a new support
278. 8:33ticket and then we're going to choose
279. 8:35one of these categories. So these are
280. 8:36like some sample requests and tickets we
281. 8:38might get. So let's take a look at this
282. 8:40one. There's a production outage. So our
283. 8:42product is down for every customer. Our
284. 8:44product API is returning 500 errors. So
285. 8:47we're going to use Jev to classify the
286. 8:49ticket when we create it.
287. 8:51And then we've automatically now
288. 8:52classified this customer support ticket
289. 8:54into critical and engineering. So we've
290. 8:57given it both a department
291. 8:58classification and also a priority as
292. 9:01well. So this would be really useful if
293. 9:03you were building this kind of
294. 9:04customer-facing triaging application.
295. 9:07You could imagine triaging things like
296. 9:09GitHub issues or support tickets, things
297. 9:11like that. This is a really great use
298. 9:12case. You can see the probability
299. 9:14distribution here. So this is a pretty
300. 9:16straightforward use case. So 100%
301. 9:18certain it's engineering and 100%
302. 9:20certain that it's critical, which is
303. 9:22actually kind of correct.
304. 9:24And you can kind of go through more
305. 9:25examples here.
306. 9:26But this works really well. It's pretty
307. 9:28fast. For these kind of really basic
308. 9:30classifications, this is going to be
309. 9:31much, much faster than using a
310. 9:33traditional LLM.
311. 9:34Okay, so for this next example, what
312. 9:36we're going to do is use a skill to do
313. 9:38model routing. So if you want to save
314. 9:39money with AI, you want to spend less on
315. 9:41tokens, you're using up your limits a
316. 9:42lot, you can use an intelligent router
317. 9:44to choose which model to use for a
318. 9:46particular task. So what we can do is we
319. 9:48can say use the Codex router skill
320. 9:52to
321. 9:54review the code base.
322. 9:57So what this is going to do, the Codex
323. 9:59router skill is going to use Jev with a
324. 10:01Python script, and then it's going to
325. 10:04based on this prompt, it's going to
326. 10:05figure out which model to use for this
327. 10:07particular task. Okay, so the model is
328. 10:09now spawned a new thread, so it's using
329. 10:11the GPT-6 Astra high reasoning
330. 10:14for 99% probability. I guess because
331. 10:16this is a code review essentially. So if
332. 10:18we go and open this chat, you can see
333. 10:20here now we've selected the GPT-6 Astra
334. 10:23high model. So in this example, we're
335. 10:25using the GPT-5.6 Luna model with low
336. 10:27reasoning because it's a really routine
337. 10:29simple task. And then if we open this
338. 10:31chat, you can see here we're using the
339. 10:32Luna light model. So this is a good
340. 10:34example of how you can use Jev for these
341. 10:36kind of routing tasks. It's actually
342. 10:38quite useful if you wanted to build this
343. 10:40into your Codex setup. So this is the
344. 10:42code for the Codex router if you wanted
345. 10:44to use it. So essentially what we're
346. 10:45using is the Codex facilities to create
347. 10:48new threads. You can obviously adapt
348. 10:50this for Claude code as well if you
349. 10:51want. But we're actually using some
350. 10:52code, so as well as the markdown file,
351. 10:55we've got a script here which is going
352. 10:56to use the Jev
353. 10:58SDK to do the routing. So we are using a
354. 11:01combination of both the skill markdown
355. 11:03file and deterministic code to make this
356. 11:06skill more powerful. So if you want to
357. 11:07take a look at this skill and use it in
358. 11:09your own workflow, I'll link it in the
359. 11:11description.
360. 11:12So just to give you one more example of
361. 11:13how you could use this. So this is an
362. 11:15example of a stock trading application.
363. 11:17So in this example, what we're doing is
364. 11:19we're using Jev to classify trades.
365. 11:21We've got live Bitcoin
366. 11:23pricing data, and then we're basically
367. 11:25giving Jev a bunch of money, and then
368. 11:27letting it figure out how to trade
369. 11:30the uh the Bitcoin.
370. 11:32And so we have three options. We can
371. 11:33either buy, hold, or sell. And then you
372. 11:36can see here the simulated profit from
373. 11:38any action it takes. I would say this is
374. 11:40actually a bad use case for Jev because
375. 11:42you could probably do this much more
376. 11:44efficiently using deterministic code.
377. 11:46But it's just another nice example of
378. 11:48how to use this kind of classification
379. 11:50idea.
380. 11:51So if you take a look at the pricing,
381. 11:53you'll see that there's a massive
382. 11:54difference between a model like Jev and
383. 11:55a large language model for this kind of
384. 11:57classification task. So I asked Claude
385. 11:59to compare the price of processing 5,000
386. 12:02emails and classifying them using a
387. 12:04model like Jev, using Claude's Haiku
388. 12:06which is a cheaper model, or using Opus
389. 12:085.5. So you can see the price per
390. 12:10million tokens is much, much lower for
391. 12:12Jev. And so you can see the cost of
392. 12:155,000 emails with Jev is $0.09.
393. 12:18Whereas with Haiku, this is going to be
394. 12:19about $2. So a very significant
395. 12:22difference in cost here even using
396. 12:23Claude Haiku. But then if you go up to
397. 12:25something like Opus, you can see here
398. 12:27we're put paying like $14 to $29 to
399. 12:30process these emails. So massive
400. 12:32difference here. This is like a 30,000%
401. 12:36difference in cost. So really, really
402. 12:37massive difference. And then in the
403. 12:39processing time, you can see here for
404. 12:41Jev to process 5,000 emails
405. 12:43sequentially, it would be about 49
406. 12:45minutes. And then if you're using Claude
407. 12:47Haiku, this would be kind of similar
408. 12:50timeline. But if you're using something
409. 12:52like Opus, this is a massive, massive
410. 12:53difference in time. So if you are
411. 12:55building an AI system in production and
412. 12:57you're processing, you know, thousands
413. 12:59of emails or you're doing thousands of
414. 13:00classification operations every single
415. 13:02day, the cost and the speed is going to
416. 13:04be a massive uh factor in choosing
417. 13:07something like Jev. It's going to be
418. 13:08dramatically faster and dramatically
419. 13:10cheaper than using a traditional large
420. 13:12language model. But if you're just doing
421. 13:13things on your local computer, you're
422. 13:15not using this at high volume, it's
423. 13:17probably not worth the complexity of
424. 13:19adding something like Jev to your setup.
425. 13:20So I wouldn't use this unless I was
426. 13:23doing a large volume of classifications.
427. 13:25For simple things on my local machine,
428. 13:27I'd probably just use a large language
429. 13:28model cuz honestly, it's not going to
430. 13:30make that much of a difference. But if
431. 13:31you are building AI systems in
432. 13:32production, this difference is going to
433. 13:34be very dramatic. So overall, I think
434. 13:36Jev is a really really good model. And
435. 13:38especially if you're building some kind
436. 13:39of AI system in production, I would use
437. 13:41this model because it's going to be
438. 13:42dramatically faster and dramatically
439. 13:44cheaper. But just for day-to-day,
440. 13:45[music] I'm not going to be using this
441. 13:47model that much personally. So I think
442. 13:49there's a lot of hype around this model.
443. 13:51>> [music]
444. 13:51>> I think it's genuinely useful for any
445. 13:53kind of AI system where you're doing
446. 13:55high volume classification or high
447. 13:56volume routing.
448. 13:58But for day-to-day stuff, I'm probably
449. 13:59not going to use this model myself. It's
450. 14:00not that useful for [music] me
451. 14:02day-to-day. But as I said, if you're
452. 14:03building an AI system in production, I
453. 14:05could see this model being very very
454. 14:06useful.
455. 14:08Thanks again for watching the video. If
456. 14:09you do have any questions or thoughts
457. 14:10about this model, let me know in the
458. 14:11comments below. Thanks again for
459. 14:13watching. I really appreciate it and
460. 14:14I'll see you in the next one. Take care.