Claude Opus 5, GPT 6 hack, Flux 3, new Gemini, quantum breakthrough, new Qwen: AI NEWS — Transcript

Full transcript

1. 0:00AI never sleeps and this week has been
2. 0:03absolutely insane.
3. 0:06Anthropic unleashes Claude Opus 5. Open
4. 0:09AI releases some new and very useful
5. 0:11features for Chat GPT. We have a new
6. 0:13open-source image generator and editor.
7. 0:16So, this is super flexible. One of the
8. 0:17top platforms for hosting open models
9. 0:20called Hugging Face gets hacked by Open
10. 0:22AI's internal models. But, get this, it
11. 0:25was an open-source model GLM 5.2 that
12. 0:28helped detect and fix Speaking of GLM
13. 0:315.2, this is a text-only model, but now
14. 0:34it finally gets vision capabilities. We
15. 0:37have a new AI for creating consistent
16. 0:39multi-shot videos where you can specify
17. 0:41the exact cuts and transitions. Alibaba
18. 0:44releases their best image model yet,
19. 0:46Qwen Image 3, and teases their massive
20. 0:49frontier model Qwen 3.8, which will be
21. 0:52open-source. Google releases their
22. 0:54latest Gemini models, which are
23. 0:56incredibly fast. We have a new
24. 0:58open-source interactive world model for
25. 1:00Minecraft. We also have some ridiculous
26. 1:03robot demos and a lot more. So, let's
27. 1:05jump right in. First up, Microsoft
28. 1:08released a really powerful image model
29. 1:10called Mage Flow. And get this, it
30. 1:13doesn't just generate images, but it can
31. 1:14also edit existing images just like GPT
32. 1:17Image or Nano Banana. First of all, here
33. 1:20are some text-to-image examples. As you
34. 1:22can see, it can generate some really
35. 1:23realistic images of different-looking
36. 1:26people like this. Here are some other
37. 1:28random photorealistic images, and as you
38. 1:31can see, for the most part, this looks
39. 1:33pretty good. And this is also great with
40. 1:35text. So, here are some examples of its
41. 1:38generations with text and a ton of
42. 1:40different elements. And as you can see,
43. 1:41it can easily generate posters,
44. 1:43infographics, or other related content.
45. 1:46It's also multilingual, so this can
46. 1:48easily generate different languages like
47. 1:50Chinese. Now, like I said, instead of
48. 1:53just generating images, this can also
49. 1:55edit existing images. So, for example,
50. 1:57you can easily replace the background of
51. 1:59an image or zoom in or zoom out on an
52. 2:02image or change the camera angle of an
53. 2:04image or even change the pose of a
54. 2:06character just with a text prompt. So,
55. 2:09very similar to Nano Banana or GPT
56. 2:11image. You can easily change the tone or
57. 2:14weather. For example, you can turn a
58. 2:16photo from day to night or you can also
59. 2:18turn a realistic photo into another
60. 2:20artistic style. And of course, you can
61. 2:22also do virtual try-ons with this or add
62. 2:25text to an image, change the expressions
63. 2:27of characters, or micro-edit certain
64. 2:30things like changing the look of
65. 2:31someone's hair, or changing the
66. 2:33composition or color of a certain
67. 2:35object. This also has ControlNet
68. 2:37inherently built in. For example, you
69. 2:39can turn a photo into a sketch or line
70. 2:41art, or you can also turn a photo into a
71. 2:44pose skeleton. And of course, you can
72. 2:46also do the reverse. You can take a pose
73. 2:48skeleton and generate a photo from that.
74. 2:50Or you can also generate Canny edge maps
75. 2:53like this. And vice versa, you can take
76. 2:55a Canny edge map and generate a photo.
77. 2:57This can also do depth and segmentation
78. 2:59like this. So, a super flexible image
79. 3:02editor. Now, this is a fairly
80. 3:04lightweight 4 billion parameter model,
81. 3:07and they also include a four-step turbo
82. 3:09model, which allows you to generate
83. 3:11images way faster in only four steps.
84. 3:13Apparently, it takes less than a second
85. 3:15to generate an image with the turbo
86. 3:17model. And then it takes a bit over a
87. 3:19second to edit an existing image with
88. 3:21the turbo model. Their self-reported
89. 3:23benchmarks also look really impressive.
90. 3:25So, if you look at this chart, you can
91. 3:27see that Mage Flow even scores better
92. 3:29than Flex Client or Quinn image.
93. 3:31Although, they haven't included the best
94. 3:33open models out there, Crea 2 and
95. 3:35Ideogram. And then in terms of image
96. 3:37editing, then you can see that Mage Flow
97. 3:39does score slightly behind Flex Client
98. 3:42in terms of quality, but it's pretty
99. 3:43close, and the turbo model is even
100. 3:45faster. The awesome thing is they've
101. 3:47released this already. If you click on
102. 3:49this explore models page, it includes
103. 3:53various models you can choose from. The
104. 3:55main Mage flow model is 17.5 GB in size,
105. 3:59so you should be able to fit this in mid
106. 4:00to high-end GPUs. And if not, because
107. 4:03this is open source, I'm sure there's
108. 4:05going to be more quantized or compressed
109. 4:07versions of this in the future as well.
110. 4:08And if you scroll down a bit here, it
111. 4:10contains all the instructions on how to
112. 4:12download and run this locally on your
113. 4:14computer. If you're interested in
114. 4:16reading further, I'll link to this main
115. 4:17page in the description below. Also this
116. 4:20week, this AI is super helpful for
117. 4:22creating videos. So it's called Shot
118. 4:24Plan, and this can generate videos with
119. 4:26precisely timed cuts, fades, and camera
120. 4:29movements. So here are some examples.
121. 4:32For the input, it takes in a text prompt
122. 4:34plus descriptions of each shot, and you
123. 4:36can also specify exactly where the cuts
124. 4:39occur, and the output is a full
125. 4:42multi-shot video that follows all of
126. 4:44these instructions. Or here's another
127. 4:46example where this is the high-level
128. 4:49prompt, and then you also add in
129. 4:50descriptions for each separate shot,
130. 4:53plus the frame which the cuts occur, and
131. 4:55here is your final video. As you can
132. 4:57see, it's able to follow everything very
133. 5:00well, and the characters remain
134. 5:01consistent the whole time. So this is
135. 5:04kind of like the multi-shot feature in
136. 5:06Kling 3. And then here's another
137. 5:08example. Now, instead of just doing hard
138. 5:09cuts, this AI is also able to understand
139. 5:12crossfade transitions, or soft cuts as
140. 5:15you can see here. It's also able to
141. 5:17understand different camera movements.
142. 5:19So here's an example of circle left, and
143. 5:21then here is circle right. And of
144. 5:23course, it also understands camera
145. 5:24movements like zooming in or zooming
146. 5:26out, and you can specify exactly at
147. 5:28which frame you want this camera
148. 5:30movement to occur. Here's another
149. 5:31example where we specify a pullback at
150. 5:35exactly frame 40. And as you can see,
151. 5:37it's able to follow this very well. Or
152. 5:40here's an example of truck right from
153. 5:42frame 25. And again, it understands this
154. 5:45very well. And if you look at these
155. 5:47multi-shot generation benchmarks, Shot
156. 5:49Plan on average performs better in terms
157. 5:52of consistency and narrative compared to
158. 5:54the other competitor models. The awesome
159. 5:56thing is they released this already. So,
160. 5:58at the top of the page, if you click on
161. 6:00this code button and you scroll down a
162. 6:02bit, here it contains all the
163. 6:03instructions on how to download and run
164. 6:05this locally on your computer. They also
165. 6:08released the code on how to train this
166. 6:10as well. So, this is fully open source.
167. 6:12Notice that they released variants for 1
168. 6:152.2 or 1 2.1. Now, for 1 2.2, that one
169. 6:19is like 28.6 GB in size, so this will
170. 6:22likely only fit on high-end GPUs. If
171. 6:25you're interested in reading further,
172. 6:26I'll link to this main page in the
173. 6:28description below. Also this week, we
174. 6:30have a new AI called Homey. In the
175. 6:32simplest sense, this generates videos
176. 6:35containing specific references which you
177. 6:37input. This can include multiple people
178. 6:40and objects. And as you can see, it's
179. 6:42really good at preserving the
180. 6:44consistency of all your input
181. 6:46references. Note that not only can this
182. 6:48do photorealistic references, but also
183. 6:513D animation like this. And you can
184. 6:54input a ton of different complex
185. 6:56objects. For example, this kimono plus
186. 6:59this toy here has a pretty complex
187. 7:00design, but it's able to preserve the
188. 7:02consistency of these really well. And
189. 7:04so, from this you can easily pump out a
190. 7:07ton of influencer content where you just
191. 7:09input a photo of the influencer plus any
192. 7:11product you want to get it to promote
193. 7:13and then get this AI to mass produce all
194. 7:16these UGC style videos. The really
195. 7:18powerful part is you can input multiple
196. 7:21views of a certain character or object
197. 7:23to give it even more consistency.
198. 7:25Especially for objects with really
199. 7:27complicated designs, it's best if you
200. 7:29have multiple views of the object so
201. 7:31that they look consistent in the video
202. 7:32at all angles. Another really cool
203. 7:35feature is you can also input an OCR map
204. 7:38which shows where text should appear in
205. 7:40an object. And this allows the model to
206. 7:42generate objects with text in them a lot
207. 7:44more accurately. Here are some
208. 7:46additional examples for your reference.
209. 7:48Now, I featured a ton of these reference
210. 7:50to video generators on my channel
211. 7:52before, such as Vace and Phantom, but if
212. 7:54you compare all these models with this
213. 7:56new Homie model, then Homie seems to be
214. 7:59a lot more consistent and faithful with
215. 8:01the least amount of errors. And this is
216. 8:03especially apparent if you have
217. 8:05characters or objects with really
218. 8:06complex designs. Again, for Homie, you
219. 8:09can upload multiple views of that
220. 8:11character to give it even more
221. 8:12consistency, whereas the other models
222. 8:14can't do this, and as a result, they are
223. 8:17a lot more error prone. At the top of
224. 8:18the page, they have released the code
225. 8:20and the models to this already, so if
226. 8:22you click on this code button, and you
227. 8:24scroll down a bit, here it contains all
228. 8:26the instructions on how to download and
229. 8:28run this locally on your computer. The
230. 8:30total size of everything is 37 GB, so
231. 8:33you'll need a high-end GPU to run this,
232. 8:35or you can also wait for some more
233. 8:37compressed versions. Note that this is
234. 8:39based off of one 2.1 and Phantom, and
235. 8:41it's under the Apache 2 license, which
236. 8:43has very minimal restrictions. You can
237. 8:45even use this for commercial purposes.
238. 8:47If you're interested in reading further,
239. 8:49I'll link to this main page in the
240. 8:50description below. Also this week, this
241. 8:53story is pretty crazy. OpenAI revealed a
242. 8:56pretty alarming security incident where
243. 8:58one of its internal models kind of
244. 9:00escaped, got access to the internet, and
245. 9:02hacked Hugging Face, which is like one
246. 9:04of the top platforms for hosting
247. 9:06open-source models. So, here are the
248. 9:08details behind this. OpenAI was
249. 9:10basically testing GPT 5.6 Soul and an
250. 9:13even more uncapable internal model with
251. 9:16some cybersecurity evaluations,
252. 9:18including a benchmark called Exploit
253. 9:20Gym. Now, these models were placed
254. 9:22inside what was supposed to be an
255. 9:24isolated sandbox environment for testing
256. 9:26only, with only network access limited
257. 9:29to the ability to install packages
258. 9:32through an internally hosted third-party
259. 9:34software. So, it can't really access
260. 9:36anything else outside of this scope.
261. 9:38Now, instead of just solving the
262. 9:40challenge inside this isolated sandbox
263. 9:42environment, what the models decided to
264. 9:44do is identify and chain vulnerabilities
265. 9:48in the package system, which gave it
266. 9:49broader network access. And from there,
267. 9:52they eventually reached a device with
268. 9:54access to the public internet, and then
269. 9:56they also hacked into Hugging Face's
270. 9:58production infrastructure to obtain the
271. 10:00answers from the benchmark, instead of
272. 10:02actually trying to solve the test
273. 10:04itself. So, here it says, "After gaining
274. 10:06internet access, the models searched for
275. 10:09and successfully found ways to gain
276. 10:10access to secret information it could
277. 10:12use to cheat the evaluation. It chained
278. 10:15together multiple attack vectors,
279. 10:17including using stolen credentials and
280. 10:20zero-day vulnerabilities, which are
281. 10:22really hard to find, by the way. These
282. 10:23are like hidden security flaws that are
283. 10:25not known to anyone. And it used this to
284. 10:27find and hack the answers to the
285. 10:29benchmark from Hugging Face." Now, what
286. 10:31OpenAI probably intentionally omitted in
287. 10:35this article is that Hugging Face only
288. 10:37detected this using the open-source
289. 10:39GLM-5.2.
290. 10:41So, they used this to detect the hack
291. 10:43and fix the vulnerability, because
292. 10:44OpenAI's model would refuse to do it.
293. 10:47This is like the drawback of all these
294. 10:49closed-source models like GPT and
295. 10:52Claude, where there's a lot of nerfing
296. 10:53going on. For example, Claude Fable
297. 10:55would just 100% reject answering any
298. 10:58questions related to cybersecurity or
299. 11:01biology, even if you're just asking it
300. 11:03to fix an issue. And you know, this
301. 11:05again emphasizes the power and the need
302. 11:07for open-source models like GLM-5.2 and
303. 11:09Kimika 3 and others. So, after OpenAI
304. 11:12detected this unusual activity
305. 11:14internally, and then after Hugging Face
306. 11:16also used GLM-5.2 to detect the incident
307. 11:19on their side, they're now working
308. 11:21together to investigate this incident
309. 11:23and tighten network controls, and also
310. 11:25improve monitoring of future model
311. 11:27evaluations. The most concerning part is
312. 11:30that as models get more and more
313. 11:32intelligent, if you tell it to do a task
314. 11:34or achieve a certain goal, instead of
315. 11:36actually doing it directly, it might
316. 11:38find a way to exploit the system and
317. 11:41cheat to achieve your goal faster. And
318. 11:43this could lead to unintended or
319. 11:45malicious actions. Anyway, let me know
320. 11:47in the comments below what you think of
321. 11:49this incident. Also this week, OpenAI
322. 11:51launches health in ChatGPT. This
323. 11:54basically turns ChatGPT from just a
324. 11:56regular chatbot to something that can
325. 11:58actually understand your personal health
326. 12:00history. So with your permission, you
327. 12:03can now connect Apple Health and
328. 12:05supported medical records, including
329. 12:07information from certain US hospital
330. 12:09systems and health apps, and ChatGPT can
331. 12:11then look at things like your
332. 12:13medications, lab results, sleep,
333. 12:15exercise, and other health data. And
334. 12:17this is the important part. With the
335. 12:19regular ChatGPT, you could just upload
336. 12:21your lab report and then ask it
337. 12:23questions about it, but it doesn't have
338. 12:24context about your previous health
339. 12:26history. But with health in ChatGPT, if
340. 12:28you upload all your records, your
341. 12:30medications, your past visits, then it
342. 12:32has a more comprehensive understanding
343. 12:34of you. So when you ask it questions
344. 12:36about lab reports or medications or
345. 12:38whatever, it can give you more accurate
346. 12:40answers tailored to you. Here OpenAI
347. 12:42says they're beginning to roll out
348. 12:44health to ChatGPT users 18 and older in
349. 12:47the US across all plans, including the
350. 12:50free plan, today. In the ChatGPT app, on
351. 12:52the sidebar, you should be able to see
352. 12:54health over here. And then once you
353. 12:56click on that, you can connect ChatGPT
354. 12:58to your health data to give it more
355. 13:00context. If you're interested in reading
356. 13:02further, I'll link to this main page in
357. 13:04the description below. Also this week,
358. 13:06Black Forest Labs teases their latest
359. 13:08model, Flux 3. And this is more
360. 13:10ambitious than their previous models,
361. 13:13which are just image generators. Here
362. 13:15they say that Flux 3 is one unified
363. 13:17multimodal model designed to work across
364. 13:19images, video, audio, and even action
365. 13:22prediction for robotics. Basically,
366. 13:25instead of training one AI to just
367. 13:26understand pictures and training another
368. 13:28model to generate video, here they're
369. 13:30just merging everything together into a
370. 13:32single model. Plus, the video also has
371. 13:35audio built in, just like Seed Dance and
372. 13:37LTEX 2.3. Here are some of its core
373. 13:40capabilities. So, this can of course do
374. 13:42text to video, but also image to video,
375. 13:44or you can also just input images as
376. 13:47references. This can also do video to
377. 13:49video, so you can upload an existing
378. 13:51video and then just edit it with natural
379. 13:53language, just like Gemini Omni. You can
380. 13:56generate videos of different aspect
381. 13:58ratios, plus it also has really strong
382. 14:00topography generation. So, it has no
383. 14:02problems including text in videos. Now,
384. 14:05they claim that Flux 3 can generate
385. 14:07videos with audio at up to 20 seconds in
386. 14:10length. Here, the generations that
387. 14:12they're showing us so far are just 720p.
388. 14:14There's no clear indication whether you
389. 14:16can generate higher resolution like 1080
390. 14:18or 4K. And if you look at their
391. 14:20self-reported results comparing
392. 14:2210-second videos in 720p with audio,
393. 14:25then here it says that Flux 3
394. 14:27generations were slightly more preferred
395. 14:29than even Gemini Omni Flash or Seed
396. 14:31Dance 2.0. And it also beats these other
397. 14:33competitor models like Happy Horse,
398. 14:35Kling, and Grok Imagine Video. Now, just
399. 14:38from these preliminary videos, at least
400. 14:40to me, it doesn't seem like it's even
401. 14:41close to Seed Dance quality. Seed Dance
402. 14:44just has much better prompt and physics
403. 14:46understanding, and it can generate much
404. 14:48better action scenes and motion and
405. 14:50different styles compared to Flux 3. So,
406. 14:52something looks fishy here. Now, like I
407. 14:54said, this is a multimodal model, so in
408. 14:56addition to generating videos, this can
409. 14:58also generate images. Here are some
410. 15:00sample generations for your reference.
411. 15:02This can generate a variety of different
412. 15:04art styles. Now, similar to their
413. 15:06previous releases, it seems that Flux
414. 15:09will offer a more powerful full or max
415. 15:12model, which would be paid and closed
416. 15:13source and only available through their
417. 15:15APIs. But, they are also planning to
418. 15:18release a lower quality dev version,
419. 15:20which will be open weights. So, this is
420. 15:22the one that you could potentially
421. 15:23download and run locally on your
422. 15:25computer. Currently, this is just a
423. 15:27preview. You can't use this yet, but you
424. 15:29could request early access using this
425. 15:31link. If you're interested in reading
426. 15:33further, I'll link to this main page in
427. 15:35the description below. Also this week,
428. 15:37we have another set of open-source
429. 15:39models. This time it's by Poolside AI,
430. 15:42and the family is called Laguna S 2.1.
431. 15:45Now, this is an open weight coding model
432. 15:46designed to keep working on really
433. 15:48difficult problems for much longer than
434. 15:51typical AI models. First of all, let's
435. 15:53go over the specs of this. So, this is a
436. 15:55118 billion parameter mixture of experts
437. 15:58model. So, think of it as like a team of
438. 15:59specialist AIs working together. And
439. 16:02when you use it, only 8 billion
440. 16:03parameters are active. So, this is
441. 16:05fairly efficient. This has a context
442. 16:07window of up to a million tokens. So,
443. 16:09you can jam-pack a ton of information
444. 16:11into your prompt at once. This is
445. 16:12roughly like 700,000 words. And you can
446. 16:15operate it in thinking or non-thinking
447. 16:17modes. Note that they also released a
448. 16:19smaller XS variant, which is only 33
449. 16:22billion parameters with 3 billion total
450. 16:24parameters and 3 billion active
451. 16:26parameters. And this model is trained
452. 16:28like an agentic agent. It's trained to
453. 16:30verify its work, backtrack when
454. 16:32necessary, verify its work and fix any
455. 16:35errors, and keep going until it
456. 16:37successfully achieves your goal. And if
457. 16:39you look at their self-reported
458. 16:41benchmarks comparing models of similar
459. 16:43sizes or even much larger like DeepSeek
460. 16:45V4 or Chimera 3, this new Laguna S 2.1
461. 16:49is not bad. Now, here if you look at
462. 16:51their self-reported benchmark score for
463. 16:53Deep Sweep, you can see that Laguna S
464. 16:552.1 scores 40% slightly above GLM 5.2,
465. 16:59but note that GLM is like almost seven
466. 17:01times the size. So, this is quite an
467. 17:03impressive performance. Definitely one
468. 17:05of the strongest open weights models
469. 17:07that are around 100 billion parameters
470. 17:09in size. Now, at least for me, I'm
471. 17:11actually more interested in this XS
472. 17:13version because this is only 33 billion
473. 17:15parameters. So, this could potentially
474. 17:17fit on a decent consumer GPU. But after
475. 17:20some testing, it doesn't seem to be
476. 17:22nearly as good as the current leading
477. 17:24model of a similar size, which would be
478. 17:26Qwen 3.6 35B or Qwen 3.6 27B. So at
479. 17:30least for me, Qwen is still the best
480. 17:32medium-sized model that can be run
481. 17:34locally. The awesome thing is they've
482. 17:36released the model weights to this
483. 17:37already. So if you click on this Hugging
484. 17:39Face link, note that the total size of
485. 17:41this is 235 GB, which is actually not
486. 17:44bad. You could probably fit this on just
487. 17:46one DGX Spark. They also already have a
488. 17:49ton of quantized versions of this. So
489. 17:51for example, this NVFP4 version is only
490. 17:5471 GB in size. If you're interested in
491. 17:57reading further, I'll link to this main
492. 17:59page in the description below. If you
493. 18:01want to supercharge your content
494. 18:02creation, definitely check out
495. 18:04Higgsfield, the sponsor of this video.
496. 18:06Think of it as an all-in-one creation
497. 18:08platform built specifically for
498. 18:10creators. Instead of jumping between a
499. 18:12bunch of different tools, Higgsfield
500. 18:13gives you access to some of the world's
501. 18:15leading models in one place, including
502. 18:17Seed Dance, Kling, Nano Banana, and GPT
503. 18:20Image. But what makes it really
504. 18:22interesting is that Higgsfield also has
505. 18:24its own tools built for creative
506. 18:26workflows. For example, there's
507. 18:28Higgsfield Supercomputer, which is
508. 18:29basically a general-purpose AI agent for
509. 18:32content creation. You can give it one
510. 18:34prompt and it can help with the whole
511. 18:35process, finding an idea, creating the
512. 18:38product concept, writing the brand
513. 18:39direction, generating the visuals, and
514. 18:41making the full video. And then there's
515. 18:43also Higgsfield MCP.
516. 18:45You can connect the content generation
517. 18:47abilities of Higgsfield with an AI agent
518. 18:49like Claude Code, ChatGPT, or Hermes,
519. 18:52and they do the thinking and planning
520. 18:54while Higgsfield executes the actual
521. 18:56generation. They also have Marketing
522. 18:58Studio, which is really useful for
523. 19:00making marketing content. You can paste
524. 19:02a product link or upload a product
525. 19:04image, and it can generate multiple ad
526. 19:06formats in one workflow, like UGC
527. 19:08videos, tutorials, unboxings, product
528. 19:10reviews, and more. One product can
529. 19:13instantly turn into a whole campaign.
530. 19:15They also have Cinema Studio, which is
531. 19:16built as a full end-to-end AI filmmaking
532. 19:19pipeline. This is especially useful if
533. 19:21you want more cinematic control. Instead
534. 19:23of just typing a prompt and hoping the
535. 19:25video looks good, Cinematic Studio lets
536. 19:28you plan scenes, control the camera, add
537. 19:30specific characters, reuse locations,
538. 19:33and keep everything consistent across
539. 19:34the whole project. From idea to final
540. 19:37output, Higgsfield gives you way more
541. 19:39control over the whole creative process.
542. 19:41Whether you're making ads, social
543. 19:43videos, AI influencers, product
544. 19:45launches, cinematic clips, or any other
545. 19:47content, this is one of the easiest
546. 19:49platforms to start creating with AI. Try
547. 19:51Higgsfield today using the link in the
548. 19:53description below. In robotics news this
549. 19:55week, Unitree releases an absolute beast
550. 19:58called the Super Athlete AS2W. This is a
551. 20:01wheel-legged dog robot built for
552. 20:04incredibly fast movement and just
553. 20:06dashing through rough, uneven terrain.
554. 20:09This weighs about 25 kg with 16° of
555. 20:12freedom, and it can run up to 6 m/s,
556. 20:15which is roughly 21 km/h or 13 mph.
557. 20:20Incredibly fast for a robotic dog. I
558. 20:22mean, imagine this thing chasing you
559. 20:24down. You would have no way of escaping
560. 20:26this. Reminds me of that Black Mirror
561. 20:27episode. This can carry a load of 16 kg,
562. 20:31plus this can walk up to 30 km on a
563. 20:34single charge. This can operate in
564. 20:36temperatures as low as -20° and as hot
565. 20:40as 55° C. And here's a demo of its
566. 20:43strength. Despite only weighing 25 kg,
567. 20:46even if you have like three people
568. 20:47standing on top of it, it can still move
569. 20:50and stay balanced. As you can see here,
570. 20:52this is also IP54 water resistant, so
571. 20:55this can wade and splash through water
572. 20:57without slipping or stopping. And it is
573. 20:59just so flexible and dynamic. You can
574. 21:02see it being able to flip and roll over
575. 21:05all this uneven terrain and recover
576. 21:07pretty much seamlessly. Also this week
577. 21:09the Chinese labs continue cooking some
578. 21:11massive models. If you haven't been up
579. 21:14to date with what's going on, last week
580. 21:16Moonshot AI dropped Kimi K3, which is a
581. 21:18massive 2.8 trillion parameter model.
582. 21:22And this is like one of the best models
583. 21:23out there right now. Well, shortly
584. 21:25afterwards Alibaba also teased their
585. 21:28latest Qwen 3.8 and this also has a
586. 21:31massive 2.4 trillion parameters. The
587. 21:33best part is here they say it's going to
588. 21:35be open weights, which is fantastic. For
589. 21:38now you can try it via their API. They
590. 21:41haven't released any benchmarks on this
591. 21:42yet. There's no like blog post or
592. 21:44article about this, but they do say that
593. 21:46it's compatible to leading frontier
594. 21:48models second only to Fable 5. Although
595. 21:51from some initial reports, it doesn't
596. 21:53seem to be as good as Kimi K3. Anyways,
597. 21:55that's all the info we have for now.
598. 21:57They haven't released an official page
599. 21:59on this with more specs and benchmarks,
600. 22:01but if you are interested, I'll link to
601. 22:03this page in the description below where
602. 22:05you can try out Qwen 3.8 max preview
603. 22:07using their paid token plan. Also this
604. 22:09week one of the best open source models
605. 22:11out there JLM 5.2 finally has vision. So
606. 22:15if you're not familiar with JLM 5.2, I
607. 22:17did a full review video on it right when
608. 22:19it came out. So see this video if you're
609. 22:22interested. This is a super capable
610. 22:24model and definitely one of my favorite
611. 22:26models, but the main drawback of this is
612. 22:28that it does not have vision
613. 22:30capabilities. This is a text only model.
614. 22:32Well, this week we now have an
615. 22:34unofficial version of JLM 5.2 with
616. 22:37vision. And this is created by Base 10.
617. 22:39So what they did is they basically took
618. 22:42a vision encoder which was created by
619. 22:44the Kimi team and they merge that into
620. 22:47the architecture of JLM 5.2. And this
621. 22:50essentially gives eyes to JLM 5.2. Note
622. 22:53that this is just an unofficial version.
623. 22:55This is not published by the ZAI team or
624. 22:58the Kimi team themselves. Note that this
625. 23:00is an NVFP4 quantized version, but if
626. 23:03you click on files and versions, I mean
627. 23:04the base model is huge, so even this FP4
628. 23:07version is 466 GB. You'll need some
629. 23:10high-end hardware like stacking maybe
630. 23:12two DGX Sparks in order to actually fit
631. 23:15this locally. But if you are interested,
632. 23:18I'll link to this main page in the
633. 23:20description below. Also this week, what
634. 23:22I think is the most exciting update from
635. 23:24OpenAI is they now introduced GPT live
636. 23:27voice inside the GPT desktop app. Now
637. 23:30previously to use the GPT live voice,
638. 23:33you need to use it through their online
639. 23:35chat interface, but now you can use this
640. 23:37directly on the ChatGPT desktop app. And
641. 23:39this opens up a ton of new
642. 23:40possibilities. If you're not familiar
643. 23:42with the ChatGPT desktop app, I highly
644. 23:44recommend you download it and try it
645. 23:46out. It's free for you to try. This lets
646. 23:48you work with multiple projects and
647. 23:50files on your computer at once. And
648. 23:52traditionally you would need to, you
649. 23:54know, open each project in a thread and
650. 23:56chat with it like a chatbot, but now
651. 23:58with GPT live, you can just talk to this
652. 24:00app and get it to coordinate different
653. 24:02tasks in the app all at the same time.
654. 24:05Imagine like being able to vibe code
655. 24:07multiple projects at once just by
656. 24:10speaking to your computer. You don't
657. 24:11even need to type, you can just do your
658. 24:13own thing. That's the power of this new
659. 24:16GPT live feature within the ChatGPT
660. 24:18desktop app. Now currently it says that
661. 24:20this live voice feature is available in
662. 24:22the desktop app for paid users, so free
663. 24:25users will not have access to this yet.
664. 24:27It seems that at least for me it's
665. 24:29available on the Mac version, but not
666. 24:31the Windows desktop app yet. Now the
667. 24:33main limitation here is that only one
668. 24:35voice chat can be active at a time. And
669. 24:38of course you also have usage limits.
670. 24:40Using the live voice will also take up
671. 24:42credits and it depends on which plan you
672. 24:44have. But the bigger idea here is that
673. 24:47OpenAI is trying to turn this live voice
674. 24:49into a hands-free control layer for
675. 24:51ChatGPT. This allows you to manage
676. 24:53increasingly complicated work simply by
677. 24:56talking to it. If you're interested in
678. 24:58reading further, I'll link to this main
679. 25:00page in the description below. Also this
680. 25:02week, we have some pretty exciting news
681. 25:05in quantum computing from Google. So
682. 25:07apparently Google has developed a way
683. 25:09for a quantum computer to automatically
684. 25:11retune itself while it's running. And
685. 25:13this matters because quantum computers
686. 25:15are extremely sensitive machines. Tiny
687. 25:18changes in temperature or even
688. 25:20electronics or hardware can slowly push
689. 25:23its calculations out of calibration. And
690. 25:25if this happens, then currently
691. 25:27engineers might need to stop the
692. 25:29computation completely and then adjust
693. 25:31thousands of control settings and then
694. 25:33start again. And that's a major problem
695. 25:35if eventually we want to run these
696. 25:37quantum computers continuously for days
697. 25:40or months. So here's Google's solution
698. 25:42to this. They use reinforcement learning
699. 25:44together with quantum error correction.
700. 25:46Now this is very technical, but I'll try
701. 25:48to dumb it down. What quantum error
702. 25:50correction does is it produces a stream
703. 25:52of signals showing that errors are
704. 25:54happening. These are like warning signs.
705. 25:56Well, Google basically feeds these
706. 25:58signals into an AI agent. The agent
707. 26:00watches how these error patterns change
708. 26:03and then adjust thousands of control
709. 26:05parameters including the frequencies,
710. 26:07phases, and strengths of the signals
711. 26:09that control the qubits. You can think
712. 26:11of it like tuning an orchestra while the
713. 26:13musicians continue playing. And then
714. 26:15after running this for a lot of rounds
715. 26:17using reinforcement learning, the AI
716. 26:19kind of learns which parameters to tweak
717. 26:22in order to reduce the errors the most.
718. 26:24In fact, they say that this
719. 26:26reinforcement learning fine-tuning
720. 26:27suppressed the logical error rate by an
721. 26:29additional 20% which the team describes
722. 26:33as a record low. Now the team also
723. 26:36tested if this could scale up and what
724. 26:38they found was that the AI's learning
725. 26:40speed did not slow down as the system
726. 26:42became larger which suggests that this
727. 26:44approach could actually scale. So in a
728. 26:46nutshell, this is basically an AI system
729. 26:48that can learn and automatically correct
730. 26:51errors from a quantum computer as it's
731. 26:54running. So, you don't have to stop and
732. 26:55reset everything and then start again. A
733. 26:57pretty interesting breakthrough if
734. 26:59you're into quantum computing. If you're
735. 27:01interested in reading further, I'll link
736. 27:03to this main page in the description
737. 27:05below. Also this week, if you're looking
738. 27:07for a relatively small model that can
739. 27:09actually fit on consumer hardware, then
740. 27:12this one might be a good option for you.
741. 27:14It's called billion
742. 27:18parameter agentic model. As with most of
743. 27:20the top models out there right now, this
744. 27:22is focused on agentic capabilities. So,
745. 27:25tasks that include multiple steps and
746. 27:27long horizon planning and reasoning.
747. 27:29Now, if you compare this to similar
748. 27:31sized models including Google's Gemma 4
749. 27:33as well as Gwen 3.5 4B and even 9B,
750. 27:37which is like three times larger, you
751. 27:39can see that across these agentic and
752. 27:42software engineering benchmarks such as
753. 27:44MCP Atlas or Claw Eval, SweepBench
754. 27:47Verified, SweepBench Pro, TerminalBench,
755. 27:49and also its reasoning and world
756. 27:51knowledge such as Humanities Last Exam,
757. 27:53GPQA Diamond, etc., you can see that
758. 27:56across the board Nanobeg 4.2 is even
759. 27:59better. This is extremely impressive
760. 28:01considering that, you know, Gemma 4 is
761. 28:03like four times larger and Gwen 3.5 9B
762. 28:06is three times larger. So, in terms of
763. 28:08performance versus model size, this is
764. 28:11by far the strongest performer. And you
765. 28:13know, the most interesting part here is
766. 28:15its looped transformer architecture. You
767. 28:17see, a normal transformer moves through
768. 28:19each layer once, but what Nanobeg does
769. 28:21is it reuses the same layers multiple
770. 28:24times as if it's looping the data
771. 28:26through the network again. And
772. 28:28interestingly, this allows it to perform
773. 28:30more computations without needing to
774. 28:31store a separate set of parameters
775. 28:33somewhere else. Apparently, this looping
776. 28:35step allows it to think for longer,
777. 28:38which gives higher quality outputs. Now,
778. 28:41if you click on files and versions, you
779. 28:42can see that this is really tiny. So,
780. 28:45So, full 3 billion parameter model is
781. 28:46only 8 GB in size, which should fit on
782. 28:49like most consumer GPUs. And if not, I'm
783. 28:51sure there's also going to be more
784. 28:52quantized versions of this that can fit
785. 28:54on even lower VRAM. Anyway, a super
786. 28:56performant model. This was only released
787. 28:58like 2 days ago, but it has already
788. 29:00gotten like over 4,000 downloads. If
789. 29:02you're interested in trying this out,
790. 29:04I'll link to this main page in the
791. 29:06description below. Also this week,
792. 29:08Alibaba releases their latest and best
793. 29:10image model, Qwen Image 3. Now, the
794. 29:13previous Qwen image models were very
795. 29:15performant, but this is even better. You
796. 29:18can jam-pack a ton of details and text
797. 29:21and elements into your prompt, and it's
798. 29:22able to generate all of this very well.
799. 29:25It's even able to generate like math
800. 29:26equations, different icons, and of
801. 29:28course different text, as you can see
802. 29:30here. It's also great at designing user
803. 29:32interfaces. This entire image was
804. 29:35actually made by Qwen 3. This is not a
805. 29:37snapshot from VS Code. You can see there
806. 29:39are some subtle errors such as the icons
807. 29:42like here and here and over here. So, if
808. 29:44you look closely, you can see that this
809. 29:46entire image is indeed AI generated.
810. 29:48Same with the icons over here. Here's
811. 29:50another example of a really complex
812. 29:52image with a ton of text and elements,
813. 29:55but as you can see, for the most part,
814. 29:57Qwen Image 3 was able to generate this
815. 30:00flawlessly. Here's another crazy
816. 30:01example. This is just an image that was
817. 30:04generated by Qwen. This isn't actually,
818. 30:07you know, a PDF or anything like that.
819. 30:09But as you can see, it's able to even
820. 30:10generate all this text as if it was a
821. 30:12screenshot from a scientific paper, plus
822. 30:15with all these complex math equations. A
823. 30:17very impressive result. And of course,
824. 30:19this can also create some very
825. 30:21photorealistic images, as you can see
826. 30:23from these examples. It's able to
827. 30:25generate a variety of different textures
828. 30:27and art styles. It has very good
829. 30:29understanding of different things. Now,
830. 30:31like the top frontier models out there,
831. 30:32instead of just generating images, you
832. 30:34can also edit existing images. And it
833. 30:36has really good world understanding. For
834. 30:38example, you can upload this image of
835. 30:40two damselflies, and you can get it to
836. 30:43basically create an infographic poster
837. 30:45with labels and information about the
838. 30:47species. And here's the image that it
839. 30:49outputs. And of course, like the other
840. 30:51frontier models, you can also upload
841. 30:53some old and damaged photos and get it
842. 30:54to restore these photos like this. Or
843. 30:57here's another cool example where we can
844. 30:59upload this photo and get it to annotate
845. 31:01stuff like this. One of the main selling
846. 31:03points of this is that it has precise
847. 31:05text rendering as small as 10 pixels.
848. 31:08It's also really good at reproducing
849. 31:09some micro details like pores and hair
850. 31:12strands. And this also supports 12
851. 31:14languages. Now, currently this is
852. 31:16closed. You can try it using their
853. 31:18online coin studio, but it's not clear
854. 31:21here whether they're actually using the
855. 31:23latest coin image three or a previous
856. 31:25version. Or you can also use it via
857. 31:26their API. Now, here's the thing. As a
858. 31:28closed model, this is not as good as the
859. 31:31best ones out there. This is very
860. 31:32similar to GPT image or C-Dream and Nano
861. 31:36Banana. So, it's really hard to pick a
862. 31:38winner here. From my initial test, it
863. 31:40does look like GPT image is still the
864. 31:42leader followed by C-Dream. So, at least
865. 31:44for me, I don't have any real reason to
866. 31:46use this, but if you are interested in
867. 31:48trying this out, I'll link to this main
868. 31:50page in the description below. Also this
869. 31:52week, Anthropic drops their latest and
870. 31:55best model, Claude Opus 5. And this is
871. 31:57indeed one of the strongest models you
872. 31:59can use right now. They claim it has
873. 32:01much stronger reasoning and planning,
874. 32:04especially for complex multi-step tasks
875. 32:06and agentic workflows compared to even
876. 32:08their previous best model, Claude Fable
877. 32:105. So, if you look at these benchmarks
878. 32:12in terms of agentic coding, then Opus 5
879. 32:15is better than Claude Fable 5. In terms
880. 32:18of legal and health, it's not as strong
881. 32:19as Fable. And in terms of biology, as
882. 32:21long as it doesn't outright reject your
883. 32:23question, which it does most of the
884. 32:25time, then it is better than Fable 5.
885. 32:28Now, interestingly, the metric that
886. 32:29really matters to me is Deep Sweep. This
887. 32:31is, you can say, a more accurate measure
888. 32:33of how good it is at agentic software
889. 32:35engineering tasks. And as you can see
890. 32:38here, actually Open AI's GPT-5.6 Soul is
891. 32:42still number one. Look how misleading
892. 32:44this is. They didn't highlight this
893. 32:45orange, but instead gray, as if it was
894. 32:48less important. Now, if you look at
895. 32:49performance versus price, in terms of
896. 32:51this frontier bench, which measures
897. 32:53agentic coding, you can see that Opus 5
898. 32:56can score higher than GPT-5.6 Soul,
899. 32:59which is again grayed out to look like
900. 33:01it's irrelevant, but it does cost more
901. 33:03to significantly beat it. And here's
902. 33:05another misleading thing. Note that this
903. 33:07x-axis of price is log scale. So,
904. 33:10they're kind of like compressing the
905. 33:12prices on the right side. So, these
906. 33:14prices are actually way more expensive
907. 33:16than it looks. Now, if you look at this
908. 33:18agentic coding index by Artificial
909. 33:21Analysis, then as you can see, it's not
910. 33:23too impressive. Opus 5, even at its
911. 33:25maximum performance, can only match the
912. 33:27performance of 5.6 Soul, which is the
913. 33:30gray dot over here. But, it does so at a
914. 33:32much more expensive price. Now, probably
915. 33:34the most impressive metric of Opus 5 is
916. 33:37its score on Arc AGI-3. This is an
917. 33:39interactive benchmark that tests AI
918. 33:42agents by dropping them into brand new
919. 33:44abstract games with no instructions.
920. 33:46Agents must explore, figure out the
921. 33:48goals, and learn the rules on their own,
922. 33:50just like humans do. The thing is, even
923. 33:53the best of the best models, like Gemini
924. 33:55and Opus 4.8, score less than 1%. And
925. 33:57even GPT-5.6 Soul scores like around 7%.
926. 34:02They do pretty horribly on this
927. 34:03benchmark, and that's because,
928. 34:04technically speaking, AI models don't
929. 34:07really learn new things or patterns
930. 34:09after they're finished training. Their
931. 34:10model weights, or basically the values
932. 34:12in their neural networks, are fixed. So,
933. 34:15this benchmark actually tests an AI's
934. 34:17emergent abilities to learn new patterns
935. 34:20it has never seen before. And
936. 34:21apparently, Opus 5 performs much better
937. 34:24than the other models. It scores like
938. 34:26over 30%. However, note that this could
939. 34:30be benchmarked. If you give it some
940. 34:31classic Witness style games that are
941. 34:33present in Arc-AGI, then Opus 5 does
942. 34:36very well. But on the most novel games
943. 34:39with unusual mechanics, then Opus 5
944. 34:42actually performs worse than Opus 4.8.
945. 34:44And these are games where the rules must
946. 34:46be discovered through interaction. The
947. 34:48model needs to figure out these new
948. 34:50rules on its own. Now, if you look at
949. 34:52this independent leaderboard by
950. 34:53Artificial Analysis, then you can see
951. 34:55that Opus 5 is indeed ranked number one,
952. 34:58just one point above Claude Fable, which
953. 35:00is one point above GPT 5.6 Soul.
954. 35:04However, if you look at the price of
955. 35:05this here is where it becomes not too
956. 35:07attractive. So, this costs almost double
957. 35:10the cost of GPT 5.6 Soul, which you
958. 35:12could argue is just as intelligent. So,
959. 35:15Opus might not be the most
960. 35:16cost-efficient option. If you look at
961. 35:18Simple Bench, which tests an AI model on
962. 35:20some tricky but common sense questions,
963. 35:22Opus 5 does pretty well. It is ranked
964. 35:25second, whereas Claude Fable is ranked
965. 35:27first. And surprisingly, Gemini 3. Pro
966. 35:30is still in third place, even though it
967. 35:32was launched like half a year ago. If
968. 35:34you look at Live Bench by Abacus AI,
969. 35:36then GPT 5.6 Soul is still ranked number
970. 35:39one. Surprisingly, Opus 5 is only ranked
971. 35:42number three, even behind Fable 5. It
972. 35:44seems to be fairly weak across the board
973. 35:47in terms of reasoning, coding,
974. 35:48mathematics, data analysis, language,
975. 35:50and instruction following. Pretty
976. 35:52interesting results. And if you look at
977. 35:54this Valse Index, which tests a model
978. 35:56across finance and coding tasks, then
979. 35:58you can see that Opus 5 is also not
980. 36:00ranked number one. It's only second
981. 36:02place and just a few decimal points
982. 36:04ahead of the open-source Kimmy K3. Now,
983. 36:07the main drawback of using any Claude
984. 36:09model is that sometimes it just outright
985. 36:11rejects your request. There are a ton of
986. 36:13guardrails in place, especially for the
987. 36:16frontier Claude models. If you've been
988. 36:18using Fable 5, then you'll know that it
989. 36:20can fall back to less capable models,
990. 36:22especially if you're asking about
991. 36:24cybersecurity or biology-related
992. 36:26questions. Well, the same goes for Opus
993. 36:285, but it seems like the restrictions
994. 36:30are a bit lighter. So, here it says Opus
995. 36:325 is less restrictive than Fable 5 in
996. 36:35terms of answering cybersecurity
997. 36:37questions. It does allow Opus 5 to find
998. 36:40vulnerabilities in source code, but it
999. 36:42blocks all these other requests. And
1000. 36:45like before, if it flags a request, then
1001. 36:47it will fall back to the dumber Opus 4.8
1002. 36:49by default. And then for biology, it
1003. 36:52also says that the guardrails are less
1004. 36:55restrictive for Opus 5. Now, currently
1005. 36:57Claude Opus 5 is available on all paid
1006. 37:00plans plus via API. Now, they released
1007. 37:03this yesterday and today is reserved for
1008. 37:05my weekly news video. So, I'll probably
1009. 37:07make a full review video on Claude Fable
1010. 37:095 tomorrow or the day after that. Stay
1011. 37:12tuned because I'm going to showcase a
1012. 37:13ton of incredible things. Also this
1013. 37:15week, we have a new video model by
1014. 37:17Nvidia called Sora video 2. And this is
1015. 37:20a fairly efficient model at either 5
1016. 37:23billion or 14 billion per parameters.
1017. 37:25And this is designed to generate up to
1018. 37:27720p videos on just a single GPU. So,
1019. 37:30here are some examples for your
1020. 37:32reference. The quality isn't bad, but
1021. 37:34there is some noise and distortions,
1022. 37:36especially along the edges of certain
1023. 37:38things. I would say the quality isn't as
1024. 37:40good as LTX 2.3 or 1. But here, I guess
1025. 37:44they're optimizing for efficiency
1026. 37:45instead of quality. And then here are
1027. 37:47some higher action shots. And again, you
1028. 37:50can clearly see some artifacts and
1029. 37:51noise. Some dogs disappear and reappear.
1030. 37:54So, it's not perfect for slower scenes
1031. 37:56like walking, then it's pretty good. And
1032. 37:58you can also generate first-person
1033. 38:00robotics videos to use for training
1034. 38:02data, as you can see here. So, not the
1035. 38:04best quality video model, but it is
1036. 38:06fairly efficient. If you click on this
1037. 38:08code button, it doesn't look like Sora
1038. 38:09video 2 is out yet, but they have
1039. 38:11released the other Sora models before.
1040. 38:13So, it's very likely that they will
1041. 38:15release Sora video 2 in the near future.
1042. 38:17If you're interested in reading further,
1043. 38:19I'll link to this main page in the
1044. 38:21description below. Also this week we
1045. 38:23have an open source world generator for
1046. 38:26Minecraft called Open Dreamer. And this
1047. 38:28is basically an open source version of
1048. 38:30Google's Dreamer 4, which is closed
1049. 38:33source. Now this world model is
1050. 38:35essentially just a video generator,
1051. 38:37nothing is pre-programmed, but they
1052. 38:38trained this Open Dreamer model to
1053. 38:41generate scenes of Minecraft which can
1054. 38:43react to actions from an AI agent. So
1055. 38:45even though everything is just video,
1056. 38:47it's kind of generating how an AI agent
1057. 38:50would interact with and move through
1058. 38:52this Minecraft environment over time.
1059. 38:54The cool thing is you can try out this
1060. 38:55demo online, so you can like press the
1061. 38:58WASD keys to move this character around.
1062. 39:01Now unlike Google's Dreamer 4, which is
1063. 39:03closed source, here they're actually
1064. 39:05releasing all the details including the
1065. 39:07model, the training code, and a detailed
1066. 39:09breakdown of what worked and what
1067. 39:11failed. And interestingly, rather than
1068. 39:13starting from Minecraft immediately,
1069. 39:15they built a smaller version using Coin
1070. 39:17Run, which is a simple 2D game where
1071. 39:20moves through an environment to collect
1072. 39:22coins. This allowed them to test the
1073. 39:24entire system on just one GPU and find
1074. 39:26problems quickly and confirm that their
1075. 39:28architecture worked before scaling it up
1076. 39:30to 3D Minecraft level. At the top of the
1077. 39:33page they've released a GitHub repo and
1078. 39:35here it contains all the instructions on
1079. 39:37how to download and run this locally on
1080. 39:39your computer. Plus they've also
1081. 39:41released the data set and the training
1082. 39:42code to this as well. If you're
1083. 39:44interested in reading further, I'll link
1084. 39:46to this main page in the description
1085. 39:48below. Also this week Google releases
1086. 39:50not one but three new AI models, Gemini
1087. 39:533.6 Flash, 3.5 Flash Light, and 3.5
1088. 39:58Flash Cyber. So let's go through each of
1089. 40:00these in more detail. The main model is
1090. 40:02Gemini 3.6 Flash, which is designed to
1091. 40:05be a really fast general purpose model.
1092. 40:07It can handle coding, document analysis,
1093. 40:10knowledge work, visual understanding,
1094. 40:12computer control, and other multi-modal
1095. 40:14tasks. The key focus here is that it
1096. 40:16completes these jobs using fewer tokens
1097. 40:19and fewer unnecessary steps. And
1098. 40:21according to Google, Gemini 3.6 flash
1099. 40:24uses almost 60% fewer tokens compared to
1100. 40:27Gemini 3.5 flash. This matters because a
1101. 40:30model that reaches the same answer with
1102. 40:32fewer steps and fewer tokens can be both
1103. 40:34faster and cheaper. And the benchmark
1104. 40:36improvements, at least according to
1105. 40:38Google, are also fairly substantial. In
1106. 40:40terms of Deep Sweep and machine learning
1107. 40:42engineering benchmarks, as well as
1108. 40:43knowledge work and computer use, 3.6
1109. 40:46flash does significantly outperform 3.5
1110. 40:49flash. Now, interestingly, if you look
1111. 40:51at this independent leaderboard by
1112. 40:52artificial analysis, then you can see
1113. 40:54that Gemini 3.6 flash does not actually
1114. 40:57perform better than 3.5. It's actually
1115. 41:00just tied, but it does cost a bit
1116. 41:02cheaper. But note that this is still way
1117. 41:04more expensive than some more
1118. 41:06intelligent competitors like GPT 5.6
1119. 41:09Luna and JLM 5.2. So, it's not even the
1120. 41:12most cost-efficient option out there.
1121. 41:14Now, Google is also launching Gemini 3.5
1122. 41:17flash light, which, if you're confused,
1123. 41:20light is just basically a faster version
1124. 41:22of flash, which is a faster version of
1125. 41:24Pro. Again, this is designed to be a
1126. 41:26much smaller and faster option for
1127. 41:28high-volume tasks like search agents,
1128. 41:31document processing, data extraction,
1129. 41:33and generating many possible solutions
1130. 41:35in parallel. And if you look at the
1131. 41:37speed of this, flash light is able to
1132. 41:39complete a task way faster, like roughly
1133. 41:41four times as fast as the non-light
1134. 41:44version. Now, here's the thing. Again,
1135. 41:45if you look at this independent
1136. 41:47leaderboard by artificial analysis, then
1137. 41:49Gemini 3.5 flash light is still more
1138. 41:51expensive than the more intelligent and
1139. 41:54open-source model Deep Seek V4 flash,
1140. 41:56which is like less than half the cost
1141. 41:58per task. Same with GPT 5 Luna medium,
1142. 42:01which also costs much less. However, if
1143. 42:03you look at the speed, then this is like
1144. 42:06way faster than all these other
1145. 42:07competitors. Now, the third release is
1146. 42:10Gemini 3.5 flash cyber. This is a
1147. 42:13specialized version trained to find,
1148. 42:15validate, and repair software
1149. 42:17vulnerabilities. It operates inside
1150. 42:19Google's Code Mentor platform, where
1151. 42:21multiple cybersecurity agents
1152. 42:23investigate a problem and combine their
1153. 42:25work into one report. Google says it
1154. 42:27reaches competitive frontier performance
1155. 42:29on this benchmark called Cyber Gym,
1156. 42:31apparently edging very close to the
1157. 42:34frontier models like GPT 5.6 Soul and
1158. 42:36even Mythos 5. Now, for this 3.5 Flash
1159. 42:40Cyber, this model will be exclusively
1160. 42:42available to governments and trusted
1161. 42:44partners as part of a limited access
1162. 42:46pilot program. But, for the previous two
1163. 42:48models, 3.6 Flash and 3.5 Flash Lights,
1164. 42:51they are already available via the API
1165. 42:54or Google's free AI Studio platform. And
1166. 42:573.6 Flash is also available in Google's
1167. 43:00agentic coding platform called
1168. 43:01Antigravity. Now, this release does seem
1169. 43:04quite lackluster. Google did not release
1170. 43:06any state-of-the-art model that can be,
1171. 43:08you know, like Opus or GPT 5.6. It seems
1172. 43:12like they're focusing more on speed and
1173. 43:14efficiency, which, to be fair, is a
1174. 43:16decent strategy. I mean, Google has so
1175. 43:18many different products and platforms
1176. 43:20like Gmail, Workspace, Google Search,
1177. 43:22Google Analytics, Maps, YouTube,
1178. 43:24Finance, etc. And they want to integrate
1179. 43:27AI into all these platforms. So, they
1180. 43:29need to create an AI model that's
1181. 43:31incredibly fast and lightweight for
1182. 43:33seamless integration. Anyway, if you're
1183. 43:35interested in reading further, I'll link
1184. 43:37to this main page in the description
1185. 43:39below. And that sums up all the
1186. 43:41highlights in AI this week. Let me know
1187. 43:44in the comments what you think of all of
1188. 43:46this. Which piece of news was your
1189. 43:47favorite and which tool are you most
1190. 43:50looking forward to trying out? As
1191. 43:52always, I will be on the lookout for the
1192. 43:54top AI news and tools to share with you.
1193. 43:57So, if you enjoyed this video, remember
1194. 43:59to like, share, subscribe, and stay
1195. 44:01tuned for more content. Also, there's
1196. 44:04just so much happening in the world of
1197. 44:05AI every week, I can't possibly cover
1198. 44:08everything on my YouTube channel. So, to
1199. 44:10really stay up-to-date with all that's
1200. 44:13going on in AI, be sure to subscribe to
1201. 44:15my free weekly newsletter. The link to
1202. 44:18that will be in the description below.
1203. 44:20Thanks for watching and I'll see you in
1204. 44:22the next one.