Do Diffusion LLMs Scale? Testing Mercury 3 + Decide — Transcript

Full transcript

1. 0:00Okay, so most of the big AI labs are
2. 0:02racing in one direction and that's that
3. 0:05they're fully focused on intelligence.
4. 0:07Every new release is about topping some
5. 0:09other benchmark. And really when it
6. 0:11comes to speed, that's not their main
7. 0:13concern at all. And for a long time
8. 0:16that's basically just been the story
9. 0:18with auto regressive models, right?
10. 0:20You're predicting one token at a time.
11. 0:23Well, diffusion language models have
12. 0:24been the exception here. They've always
13. 0:26been much quicker than normal LLMs, but
14. 0:29for a long time they came with a
15. 0:30trade-off. They just didn't perform as
16. 0:32well. You could kind of think of them as
17. 0:34the fast little model that you could use
18. 0:36for auto complete, but not the model
19. 0:38that you could hand a real agent task
20. 0:40to. Well, I've got to say with Mercury 3
21. 0:43and Mercury decide, that's all changed.
22. 0:46Mercury 3 itself runs at over a thousand
23. 0:48tokens a second and on Inception's
24. 0:51benchmarks, it sits in the same range as
25. 0:53GPT-6 Luna and Claude Haiku 5.5. So,
26. 0:57you're not having to give up all your
27. 0:58intelligence just to get speed anymore.
28. 1:01And Mercury 3 is not the only thing that
29. 1:03they've been working on. Alongside that,
30. 1:05there's Mercury decide, which is their
31. 1:07take on a Jeb style decision model. And
32. 1:10as I'm recording this, it's currently
33. 1:11free on OpenRouter. Now, when Inception
34. 1:14reached out to me to give me access to
35. 1:16this early model and sponsor the video,
36. 1:19I was really curious to see just how far
37. 1:20diffusion models have actually come for
38. 1:23LLMs. So, in this video I'm going to go
39. 1:25through the two models. We'll look at
40. 1:26benchmarks, both some from Inception,
41. 1:29but also some third-party ones and I'm
42. 1:31going to show you some tests that I've
43. 1:32run myself and actually see where
44. 1:35diffusion language models are these
45. 1:36days. Okay, so if you don't know
46. 1:38anything about diffusion LLMs, let me
47. 1:41explain a little bit about them. So, how
48. 1:44are they actually different? Well, most
49. 1:46LLMs are auto regressive. You can kind
50. 1:48of think of that as them writing like a
51. 1:49typewriter, one token, then the next,
52. 1:52then the next and every single token
53. 1:54costs a full pass through the model.
54. 1:57That's why decoding speed is so
55. 1:58important. The diffusion on the other
56. 2:00hand works much more like an editor.
57. 2:02They start with a rough masked out draft
58. 2:05of the whole response, and then they
59. 2:07refine lots of tokens in parallel over a
60. 2:10handful of passes. So, each pass does
61. 2:13more useful work. And because inference
62. 2:15is limited mostly by memory bandwidth,
63. 2:17not compute, getting lots of tokens out
64. 2:20of each pass is where the speed up
65. 2:22actually comes from. And if the idea of
66. 2:25diffusion sounds familiar, it should,
67. 2:26right? It's the same basic idea behind
68. 2:29most of the image and video generators
69. 2:32that are out there now. And I'm going to
70. 2:34point out that's why it's no coincidence
71. 2:36that Inception CEO Stefano Ermon is one
72. 2:40of the co-inventors of some of the key
73. 2:42diffusion methods behind modern image
74. 2:44and video generation. And the founding
75. 2:47team at Inception have worked on
76. 2:48techniques like this for quite a long
77. 2:50time now, and many others that we use
78. 2:53like DPO, flash attention, etc. So, this
79. 2:56team is clearly a bunch of people who
80. 2:58got diffusion early. And they've really
81. 3:00been working hard for a while now to
82. 3:02prove that this can work just as well
83. 3:04for text, not just images. And so, the
84. 3:07big question with diffusion LLMs has
85. 3:10always been, do they scale? Back at ICML
86. 3:13in 2024, I attended Sander Dieleman's
87. 3:16diffusion circle, and specifically asked
88. 3:18him at the time about this. Did it look
89. 3:20like diffusion was going to take over
90. 3:22LLMs just like it had done with image
91. 3:24generation
92. 3:25and video generation, and even things
93. 3:27like music generation. And at the time
94. 3:29then, it really was an open question. No
95. 3:31one had been really able to show that
96. 3:33these things could actually scale up to
97. 3:34sizes where they became very strong
98. 3:37language models. Now, Inception are
99. 3:39really the ones who've shown that this
100. 3:41can scale. And it's pretty cool that a
101. 3:43reasonably small startup is actually
102. 3:45doing this. So, back at the start, they
103. 3:47started with Mercury 1. They showed how
104. 3:49you could build a commercial diffusion
105. 3:51model. And if you look at this chart, it
106. 3:52really just shows what's happened since.
107. 3:55From Mercury 1 Mini to Mercury 3,
108. 3:57they've scaled total parameters by more
109. 4:00than 40 X, and their intelligence
110. 4:02average has gone up by 3 X. Now, the bit
111. 4:05I find interesting is the jump from
112. 4:07Mercury 2 to 2.5 was roughly the same
113. 4:10size model. So, it seems like that gain
114. 4:12came from training, not scale. But then
115. 4:15from 2.5 to 3, they've really scaled up
116. 4:18to a much bigger model here. So, both
117. 4:21levers seem to be working for diffusion,
118. 4:23which is exactly what we want to see if
119. 4:25we're going to see this take off. Don't
120. 4:27forget we're talking about a startup
121. 4:29that doesn't have anywhere near the
122. 4:30funds of the big players like OpenAI and
123. 4:33Anthropic, et cetera. Okay, so how does
124. 4:35Mercury 3 compare to models in its
125. 4:37class? This is against basically GPT-6
126. 4:41Luna, Claude Haiku 5.5. Where it's
127. 4:44strong is in reasoning, instruction
128. 4:46following, and search. It's doing very
129. 4:48well on GPQA Diamond there. In fact, you
130. 4:51can see that it's doing well on many of
131. 4:52these benchmarks, including the Aether
132. 4:55search benchmarks. While it's not at
133. 4:57Opus 5.5 level for Aether coding, it's
134. 5:00still well-suited for medium-level tasks
135. 5:03and as a default model for coding
136. 5:05sub-agents. And that whole idea of
137. 5:08basically having like swarms of agents
138. 5:11is really suited to this model. When
139. 5:13you've got something so fast like this,
140. 5:15you can deploy it for a whole bunch of
141. 5:16different ideas like recursive
142. 5:18self-improvement loops, understanding
143. 5:20large code bases, and anything we've got
144. 5:23a really large amount of data that you
145. 5:25want multiple sub-agents to go across
146. 5:27and interpret that data. And if you're
147. 5:29building something like a search agent
148. 5:30or a support bot, where following
149. 5:32instructions really matters and speed
150. 5:35matters, Mercury 3 is looking really
151. 5:37strong here. Now, that's just comparing
152. 5:39the intelligence benchmarks. If we start
153. 5:42to look at speed, then Inception's
154. 5:44numbers really take off. So, Mercury 3
155. 5:47can get 1,068 tokens a second on Nvidia
156. 5:51Blackwell. If you can compare that to
157. 5:53those other models, that's about eight
158. 5:55times what Luna gets, and even 4.5 what
159. 5:58Claude Haiku is getting. And as you'll
160. 6:00see in the demo that I showed, this
161. 6:02thing just flies at being able to
162. 6:05generate text out. And you also have the
163. 6:07ability to turn the reasoning effort up
164. 6:09and down. And you'll see just how quick
165. 6:12that I can actually can make answers
166. 6:13coming back. All right. So, where does
167. 6:15this actually matter? So, Inception have
168. 6:17a few early customers sharing numbers.
169. 6:20The one I like was Octane, who do search
170. 6:22on smart glasses. Their search was
171. 6:24taking about 80 milliseconds. So, the
172. 6:27model was the bottleneck there. With
173. 6:29Mercury 3, the generation time is 0.9
174. 6:32seconds. And you can imagine on smart
175. 6:34glasses, anything that's going to take
176. 6:35just a few seconds is suddenly going to
177. 6:37make people think that something is
178. 6:38broken there. Now, the price for this at
179. 6:40launch is going to be 10 cents in and 35
180. 6:43cents out per million tokens. It was
181. 6:45actually 50% off. So, it's going to be
182. 6:47interesting to see how long they run
183. 6:48that. So, longer term, it may be a
184. 6:50little bit more expensive than Luna or
185. 6:53Haiku. But, you've still got the
186. 6:54advantage of things like cached inputs
187. 6:56being 90% off. And if your agent resends
188. 7:00the same context every step, that's
189. 7:02going to [clears throat] really add up.
190. 7:03So, the idea is you're paying for speed
191. 7:05here, not saving on tokens. Another
192. 7:07example that they talked about was Tiny
193. 7:09Fish, who build web agents. And in their
194. 7:11blog post, they talked about other
195. 7:13examples of people basically being able
196. 7:15to do a lot more just because the model
197. 7:17is so much faster. And the cool thing
198. 7:19here is that this is fully OpenAI
199. 7:21compatible, so switching is basically
200. 7:24just changing a base URL change. It
201. 7:27supports tool calling, it supports
202. 7:29structured outputs, and they even
203. 7:31support a context window of a million
204. 7:33tokens, up from 260K in their last
205. 7:36model. And it's great to see also that
206. 7:38they've made this very easy for people
207. 7:40to try out. Not only is it available on
208. 7:42their own API, it's on Open Router, Base
209. 7:46Ten, and Vercel. Okay, so Mercury 3 is
210. 7:49the thinking model. It's the system two
211. 7:51side of things. But, as I've been
212. 7:53talking about on the channel for quite a
213. 7:55while, a lot of things you want a really
214. 7:57quick system one kind of decision.
215. 7:59That's the whole sort of thing like
216. 8:01agents just needing really quick calls.
217. 8:03Is this a billing ticket? Is an agent
218. 8:05stuck? Did a page actually load? And
219. 8:07I've talked about it quite a lot on the
220. 8:09channel. That's what decision models are
221. 8:10for. These models don't generate any
222. 8:13text at all. You give them some state
223. 8:15and a question, and they give you back
224. 8:16an answer and a probability in one call.
225. 8:19So, here's where Mercury Decide fits in.
226. 8:22It's a diffusion decision model. You
227. 8:24pass in a state, and the cool thing
228. 8:25about this is that can both be text, but
229. 8:28it can also be images as well. You pass
230. 8:30in a set of typed questions, and then
231. 8:33the model will process all of that and
232. 8:34return its answer in a fraction of a
233. 8:37second. So, how good is it? So, rather
234. 8:39than just take Inception's word for it,
235. 8:41let's look at some of the independent
236. 8:43evals for this one. Vala's AI ran it
237. 8:45through the same tests they used on Jev.
238. 8:47On claim verifications, it was already
239. 8:50matching Jev. On legal bench, Mercury
240. 8:53Decide is clearly ahead, scoring around
241. 8:5587% versus 74%. And here's the cool
242. 8:59thing, it costs about a quarter of what
243. 9:01Jev does. And there's a free version
244. 9:03that you can try on Open Router as well.
245. 9:06On Artificial Analysis Decision Index,
246. 9:09this model is third on not hallucinating
247. 9:11at 32%. And like I've talked about in
248. 9:13the other videos about decision models,
249. 9:16it really matters when you want a model
250. 9:18that says it's not sure instead of just
251. 9:20guessing. And then on speed, Vala's
252. 9:22measured a median of about 0.16 seconds
253. 9:26for a single judgment and 0.4 seconds
254. 9:29for 64 judgments. So, clearly a super
255. 9:32use case with this is where you're
256. 9:34passing in one context and asking lots
257. 9:36of questions about that at the same
258. 9:38time. Now, the new thing with the latest
259. 9:40release of Mercury Decide is Vision. And
260. 9:43I really think that this is a pattern a
261. 9:44lot more agents are going to use. Cheap
262. 9:46System 1 checks wrapped around a single
263. 9:49System 2 call. You save the expensive
264. 9:51thinking for the parts that need it, and
265. 9:53you get a confidence score at every step
266. 9:55that you can use to escalate to a human
267. 9:57if you need it. Okay, so jumping into
268. 9:59the demo here, I put together a little
269. 10:01app for testing this out, and I'm going
270. 10:03to show you a number of different things
271. 10:05in here that I think are really cool
272. 10:06with this testing suite and with my
273. 10:07normal testing suite. So, first off, if
274. 10:11you've never actually used a diffusion
275. 10:12LLM, you're probably going to get a sort
276. 10:14of shock at one, how quick it is, but
277. 10:17two, also just how it actually does the
278. 10:19tokens. So, let's look at this. I'm
279. 10:21going to ask it to explain how diffusion
280. 10:23language models differ from
281. 10:24autoregressive ones. So, you'll see that
282. 10:27it's going to take a bit to kick off.
283. 10:29I've got some settings over here, but
284. 10:30you can see very quickly look at it
285. 10:32going, and you'll notice also as it goes
286. 10:35through, it's already finished. Okay,
287. 10:37what we can do here is I've put in a
288. 10:39feature to replay it and slow it back
289. 10:42down.
290. 10:43If I come in here and replay this, you
291. 10:45can see I've got it step back, so we can
292. 10:47see what actually happened at each frame
293. 10:50as we go along here. And you can see
294. 10:52that
295. 10:53it gradually sort of fixes things up as
296. 10:56it goes through. Works out different
297. 10:58tokens, like, okay, how the different
298. 11:01tokens should come together.
299. 11:03And it's very different than just sort
300. 11:05of like a standard LLM, right? In fact,
301. 11:07I can scroll through this manually, so
302. 11:08we can just slow it down and look at it,
303. 11:11right? So, with a normal autoregressive
304. 11:13LLM, once it's committed to a token, it
305. 11:17can't really go back and change that.
306. 11:19Now, there've been some papers that have
307. 11:21tried different ways of backtracking and
308. 11:22stuff like that, but fundamentally, it's
309. 11:25doing it step by step. Here, you can see
310. 11:28it starts to sort of plot out like a few
311. 11:31different paragraphs, and then it goes
312. 11:33through and actually sort of refine
313. 11:36this. So, we can see that as it's going
314. 11:39along, it's working out what should be
315. 11:41at that start there. And so, it had when
316. 11:44we go back here, it had at there, but
317. 11:47it's still and it's got auto something
318. 11:49SU, but as we go through the frames, it
319. 11:53starts to basically put it all together.
320. 11:55And sure enough, it then assembles whole
321. 11:58sorts of paragraphs and eventually gets
322. 12:02through and gets right to the end where
323. 12:05it's assembled the whole response here.
324. 12:08And you'll see that the response reads
325. 12:10perfectly, right? There's nothing
326. 12:12unusual about the text that the
327. 12:15diffusion LLM makes. It's just how it
328. 12:18makes it is fundamentally different. So,
329. 12:21this is one of the first things that is
330. 12:22really cool to actually see here is just
331. 12:26the speed of how this happens.
332. 12:29And you can see there, that was so
333. 12:31quick. Let me just run that again so you
334. 12:33can see it. I'm not speeding anything up
335. 12:34here. I'm just showing you the actual
336. 12:36speed, right? It takes a little bit of
337. 12:38time for the first tokens to come in,
338. 12:40although it's so quick, I don't even get
339. 12:42a chance to say that. But fundamentally
340. 12:44then, it assembles it all and it's done
341. 12:47very quickly in there. So, this is the
342. 12:49advantage of diffusion, right? And
343. 12:52you'll see if I give it one of my
344. 12:53standard prompts about writing a
345. 12:545,000-word essay,
346. 12:57this is still going to take it a little
347. 12:58bit of time. It's 5,000 words and you
348. 13:00can see the time to first token there is
349. 13:02taking a little bit to get going. But as
350. 13:04it goes through, it basically puts all
351. 13:06of this together
352. 13:08and gives us out. And if I speed up down
353. 13:10here, we can see it's still assembling
354. 13:12the end of the document. So, it's going
355. 13:14through in chunks, putting it together.
356. 13:18But this model is actually very good at
357. 13:20being able to put the whole thing
358. 13:22together. So, in my test so far, I don't
359. 13:25find it doing the things that a lot of
360. 13:26the open models do of where they only
361. 13:29write half [snorts] of the essay or they
362. 13:32prompt it to oh, if I was going to
363. 13:34continue, I would get to this and this
364. 13:36and this kind of thing. So, that's one
365. 13:37of the interesting things here. And you
366. 13:39can see if I come in and actually paste
367. 13:41that output in here, it's 4,200 words.
368. 13:45So, that's certainly a lot better than a
369. 13:48lot of the models are producing out. All
370. 13:50right, if we come across to my normal
371. 13:52sort of testing ground here, we can try
372. 13:54it out also on things like the standard
373. 13:57agentic tool calling stuff. And you'll
374. 14:00see that look, it's just going through
375. 14:01it all so quickly that there are 10
376. 14:04tests in here and often it would take
377. 14:07quite a long time for an autoregressive
378. 14:09model to do this and I actually sort of
379. 14:10pause it usually. I'm not getting a
380. 14:13chance to pause it here. It's gone
381. 14:14through, got 10 out of 10, done really
382. 14:16nicely on them, no sort of overthinking
383. 14:19that we see on some of them in there. If
384. 14:21I come down and give it the sort of pro
385. 14:23ones, again, the speed of the responses
386. 14:26is really good here. So, this is doing
387. 14:30really well and this is one of the
388. 14:31things that we saw even some of the meta
389. 14:34models having issues with some of the
390. 14:36things in here. But, I'm going to let
391. 14:37that go through. It's gone through and
392. 14:39it's basically passed all eight of them
393. 14:42perfectly, right? So, this is one of the
394. 14:44cool things about this model, right?
395. 14:45That it can do all those kinds of
396. 14:48things. And for things like rag, a
397. 14:50really nice job. You can see it's gone
398. 14:52through very quickly doing 4,000 tokens,
399. 14:56going through there to basically get the
400. 14:58right answers. It was doing input to be
401. 15:00able to extract things out. So, that
402. 15:02one's Mercury 3, which is their system
403. 15:05two model, right? That's basically where
404. 15:07all the intelligence is, where it's
405. 15:10doing a lot of thinking, reasoning, that
406. 15:12kind of thing. If we change over to
407. 15:15Mercury Decide, we've now got the system
408. 15:18one model. And this is basically where
409. 15:20you've got all your gem style decision
410. 15:23model things going on in here. So, we
411. 15:25can run things like a support ticket
412. 15:27through this.
413. 15:29And you'll see the crazy thing here is
414. 15:30literally I'm almost just pressing it
415. 15:33and it's giving me the answer back so
416. 15:35quickly. So, I will basically let it go.
417. 15:39And you can see okay, there it's gone
418. 15:41through. It's basically done the
419. 15:43classifications on this test. We've got
420. 15:46the normals, we've got the choices,
421. 15:48we've got a score in here. And the
422. 15:51model's done really well there. We can
423. 15:52also see the confidence scores for
424. 15:54things like choice and for things like
425. 15:56the score and stuff in there. If we want
426. 15:58to see actually what got sent in there,
427. 16:00we can see that's just a standard API
428. 16:03for a decision model.
429. 16:05And then our raw response back is just
430. 16:07giving us the probabilities. Again, the
431. 16:09standard for a decision model in here.
432. 16:13Now, the cool thing about Mercury Decide
433. 16:14is not only we can we do all of that
434. 16:17with text, right? We can also do it with
435. 16:20images. If I take a screenshot of this
436. 16:23UI and basically I'm setting it to
437. 16:26actually do a UI screenshot, this is
438. 16:28going to be the screenshot of a page
439. 16:31after a user has clicked pay. And then
440. 16:34we're going to see does it have an
441. 16:35error? Does it look like the payment
442. 16:37went through, etc. And you can see the
443. 16:39response back there was super quick. And
444. 16:41in many ways, this is basically showing
445. 16:44that's our image there, which is
446. 16:45actually just a screenshot of this UI.
447. 16:48So, really there's not a lot related to
448. 16:50payments or anything like that there.
449. 16:52All right, if we try and do something
450. 16:53like an image check. This is a really
451. 16:55good example, I think, of how people
452. 16:57could use this. So, what I've done here
453. 17:00is I've used an image generator to make
454. 17:02up some catalog pictures that would go
455. 17:05on like an e-commerce site or something
456. 17:07like that. Each of them is the same kind
457. 17:09of outfit. One is with a male, one is
458. 17:11with a female, one is with no human at
459. 17:14all. And you can see the questions I'm
460. 17:16asking in here. I've got a whole bunch
461. 17:18of different questions in there. I'm
462. 17:19asking it about like is there a human in
463. 17:22there? Is it safe for work? A photo, a
464. 17:24screenshot? What's the quality and stuff
465. 17:27like that? Let's just run it and see how
466. 17:29it goes.
467. 17:31And you can see already it's done. I
468. 17:32don't even need to pause or anything
469. 17:34with this. So that was the photo, right?
470. 17:35It's a guy in an outfit there. Has a
471. 17:38person, 100% yes. Safe for work, 99.8%
472. 17:41yes. Kind equals photo. Person equals
473. 17:45man. And then quality it's giving us 3.7
474. 17:48there. Let me swap out this image.
475. 17:52So now I've got basically the same
476. 17:54outfit but a woman in there. Let's see
477. 17:56how this one goes.
478. 18:00And you can see pretty much everything
479. 18:02else stayed the same. The only thing
480. 18:04that changed was a woman, right? And
481. 18:08that was quite deliberate. Like we want
482. 18:10to have the same sort of look in here.
483. 18:12We're just changing one little thing in
484. 18:14the actual image, but the model's able
485. 18:16to detect that and detect it really
486. 18:18well.
487. 18:19Let me try a last one where we've got
488. 18:21the outfit, right? The same outfit that
489. 18:23those two people were wearing, but now
490. 18:25there's no human actually wearing this.
491. 18:27Let me run it.
492. 18:32And you can see sure enough it's
493. 18:33basically has person is 0%. Still safe
494. 18:36for work, in fact slightly more safe.
495. 18:39And then person no human in there. So
496. 18:41this ability to do these decisions
497. 18:44really quickly really benefits from
498. 18:47something like diffusion, right? You've
499. 18:49got that really fast response time.
500. 18:53You've got the ability to ask lots of
501. 18:55different questions and stuff in here.
502. 18:57Even in their benchmarks, when you ask
503. 18:59sort of 64 questions, it's not really
504. 19:01slowing the whole process down too much.
505. 19:04It's still extremely quick with its
506. 19:06responses. So just to wrap up the demo,
507. 19:08the thing I would say is if you've never
508. 19:10played with diffusion language models
509. 19:13before, it's a lot of fun. I got to say
510. 19:15it really is just fun to see how quick
511. 19:18these things work. And I find it
512. 19:20interesting how they're putting their
513. 19:22whole sort of ideas together in a very
514. 19:25different way than auto-regressive
515. 19:27models. All right, so one cool thing
516. 19:29that you can do is put these together.
517. 19:31So this is an example of a pipeline
518. 19:33that's using both Mercury Decide and
519. 19:36Mercury 3 in here. So you can see that
520. 19:39we've got some kind of ticket going in.
521. 19:42And what it's going to do is basically
522. 19:44do a bunch of classifications, go
523. 19:45through, ask a bunch of questions. Then
524. 19:47it's going to use the system two model,
525. 19:50Mercury 3, to actually draft a response.
526. 19:54So let's run this, and you'll see that
527. 19:56it is actually spending the triage time
528. 19:59of doing the decider stuff, then it's
529. 20:01got the Mercury 3 running, and then it's
530. 20:04got more checking at the end. So for
531. 20:07this, it worked out that it was billing
532. 20:09related. It worked out that it needed
533. 20:11immediate attention. It needs a reply.
534. 20:14It's a refund request. And then the
535. 20:16Mercury 3 came in with that information,
536. 20:19was able to basically draft a reply very
537. 20:23quickly, and then we can then check it
538. 20:25again with the decision model just to
539. 20:27make sure did we do anything like
540. 20:30over-promise? Is there anything that's
541. 20:32sort of not in there? What is the tone
542. 20:34like? That kind of thing. So you can see
543. 20:37that in this case, we've basically got
544. 20:39it where we got a flat-back acceptable
545. 20:42to a bit of a warm response back. And in
546. 20:45literally a couple of seconds here,
547. 20:46we've had the whole thing be processed,
548. 20:49checked, classified, all these sorts of
549. 20:52things in a really fast time. All right,
550. 20:55so where do these actually fit? Well,
551. 20:57Mercury 3 makes most sense anywhere
552. 20:59latency compounds. So agents, voice,
553. 21:02search, tutoring, all those kinds of
554. 21:05things, this is going to be a really
555. 21:06useful model. Mercury decide on the
556. 21:08other hand makes sense for high volume
557. 21:10checks. So, anything like moderation,
558. 21:13routing, and then especially anything
559. 21:16where you can use the confidence score
560. 21:18to escalate the uncertain cases. Both
561. 21:21models are currently on launch pricing
562. 21:23right now, so definitely worth checking
563. 21:25out. And I think in closing, really what
564. 21:27Inception has done here is that they
565. 21:29show that diffusion does seem to scale
566. 21:32for LLMs. Anyway, let me know in the
567. 21:35comments if you've actually tried out
568. 21:37these diffusion language models. If you
569. 21:38haven't, I really suggest trying Mercury
570. 21:403 out and just seeing how well, and more
571. 21:43importantly how quickly it can actually
572. 21:45work for your use cases. As always, if
573. 21:48you found the video useful, please click
574. 21:50like and subscribe, and I will talk to
575. 21:51you in the next video. Bye for now.