start	end	text
80	1600	shooting out. I believe that was
1600	5670	something I've seen.
5680	8160	The trees are quite reminiscent of a
8160	10639	pooh. Deepseek V4 Pro has just been
10639	12960	released in general available version,
12960	15120	which also happened with the V4 Flash
15120	17920	model back about 12 or 13 days ago on
17920	21279	0731. Now, that model when tested proved
21279	23600	to be a significant significant increase
23600	25600	in capability over the preview version
25600	27760	that was previously around. And today
27760	29279	we're going to be finding out if the
29279	31279	same can be said for the newly updated
31279	33920	checkpoint general availability for V4
33920	36000	Pro dubbed0813.
36000	37520	So before we get into it, do feel free
37520	39680	to subscribe so we can hit 70K and then
39680	41680	100K with the plaque. And with that,
41680	43520	basically there's not a lot to say about
43520	45200	this right now because it seems to have
45200	47200	been more or less silently released.
47200	48879	There's no official announcement post
48879	50640	yet on X, which I do believe is
50640	52719	generally where we would see this. But
52719	54719	they do mention here just in the website
54719	56879	a couple references to it where if you
56879	58960	use this model identifier right here for
58960	61440	V4 Pro, you are being routed to the new
61440	64479	V4 Pro0813 version. Additionally to
64479	66880	that, it is also live on open router and
66880	68560	this makes me feel comfortable that it
68560	70640	is actually genuinely released and this
70640	72159	is the model we're going to be using.
72159	73920	So, from the bit of information we have
73920	75840	right here, aside from just the model
75840	77600	identifier now being today's date,
77600	79840	depending on what local one is in, we
79840	81600	can see that the model has a million
81600	83600	context window with a maximum output
83600	86240	length of 384K, which really is a lot.
86240	88080	Now, something that can sometimes happen
88080	90240	with DeepSeek, at least historically, is
90240	92240	there may be a lot of thinking. Now,
92240	93920	this does support selectable thinking
93920	95759	modes, as you can see if we do click on
95759	97759	this. So, it's possible to kind of tone
97759	99520	that down, but in the interest of seeing
99520	101680	the maximum potential performance, all
101680	103600	of our coding tasks for today are going
103600	106560	to be run on this maximum effort. Now, I
106560	108240	want to mention the following with a
108240	109759	grain of salt because I cannot
109759	111600	definitively confirm that this is ground
111600	113600	truth, but I would imagine it probably
113600	115600	is. There were some references and
115600	117439	photos floating around which apparently
117439	119920	came from a group related to deepseek. I
119920	121600	believe they said it was a WeChat group
121600	123759	about Deep Seek or something. And if
123759	125520	this is to be believed, we can see that
125520	127200	these would hypothetically be the
127200	129200	benchmark results for Deepseek V4
129200	132080	Pro0813 or the general availability
132080	134239	model. And basically my main interest
134239	135680	here is just seeing the leap in
135680	138160	capability over the V4 Pro preview to
138160	140959	now because I was initially disappointed
140959	143200	with the V4 Pro Preview as well as the
143200	145440	V4 Flash preview. They felt very
145440	147520	undercooked and probably because they
147520	149280	were named preview. That's actually like
149280	151440	fair. It makes sense. But again, I want
151440	153120	to reiterate the leap in capability
153120	155840	between flash and then the 0731 general
155840	157920	availability version of flash was
157920	159840	absolutely massive. This is like the
159840	162319	best local model right now for like a
162319	165519	dual spark setup. So I mean in a niche,
165519	167760	you know, but still if that leaping
167760	169760	capability remains, then what we can see
169760	172160	right here just in this deep sore
172160	175760	increase from 12.8 to 62.7 would then
175760	177440	hypothetically put this above things
177440	180319	like Opus 4.8 8 and make it very very
180319	182159	competitive being especially that this
182159	184400	is an openweight model or will be.
184400	186400	That's also extremely exciting because
186400	189519	this is not like horrifically gigantic I
189519	191360	guess when considering things like Kimmy
191360	194319	K3 or Quen 3.8 Max. It's still big at
194319	196640	1.6 trillion parameters. It is a mixture
196640	199280	of experts model with how many active?
199280	201120	So if we do estimate that this is going
201120	203120	to be the same size as the preview
203120	205519	version of V4 Pro. It's a 1.6 6 trillion
205519	207760	parameter mixture of experts model with
207760	210159	49 billion active. So it is still very
210159	212799	very large but it's not as gigantic as
212799	214720	like the closing in on 3 trillion
214720	217040	parameter model. So additionally to that
217040	218720	pricing and this is something that we
218720	220720	need to basically take just as it is
220720	222640	right now because they mention and have
222640	224159	been for quite a bit of time that their
224159	226480	API pricing is going to be significantly
226480	228959	increased soon. Now significantly is
228959	230640	relative to what it currently is which
230640	233440	is like very very very cheap. So, right
233440	235519	now, pricing per million input tokens
235519	240480	for DeepSeek V4 Pro is 4.35.
240480	243280	So, okay. No, I know what this is, but
243280	245519	like last time I messed up on this one
245519	247599	and I said it was 2 cents and it wasn't
247599	251200	and then I got roasted. So, this is Z 43
251200	253920	cents per million input tokens and 87
253920	256000	cents per million output tokens is the
256000	258079	current price. Now, if this gets like a
258079	260079	significant increase, say it goes up
260079	261840	like five or 10 times, it would still
261840	264479	hypothetically be competitive. So, it's
264479	266160	all relative to what it currently costs.
266160	267759	And with that, I am going to be using
267759	269919	this from within Codeex. When we tested
269919	272800	V4 Flash, the general availability model
272800	275440	on the 31st of July, I did also use it
275440	276960	through Codeex because I believe they
276960	278800	had suggested to do so and the
278800	281360	performance was very, very good. I was
281360	283040	happy with it. So, I'm going to be using
283040	285199	Codeex as well for this. Now I should
285199	288080	also mention that they did release in
288080	290800	beta a desktop app for codeex or
290800	293280	chatgpt. So I also have it routed and
293280	294960	working through this desktop app as well
294960	296880	and it is doing a skate game for us in
296880	298720	C++. I would imagine that will take
298720	300240	quite a bit of time. So I did just want
300240	301919	to start it and verify this app was
301919	303840	actually working properly on this system
303840	306160	and it does seem to. So keep in mind any
306160	309120	mention of GPT or codecs here is routed
309120	311280	directly to DeepSeek on max reasoning
311280	313759	effort for the 0813 model. And with
313759	316080	that, let's take a look at our Okay, so
316080	320720	apparently the OpenAI beta application
320720	322880	set itself to be the default for all
322880	324720	file types. That's infuriating. All
324720	327759	right. All right. Very Quen 3.8
327759	330080	aesthetic. I'm going to turn sound on
330080	331759	because I did see some mentions of it
331759	334240	just talking about like
334240	335919	sounds. All right. It starts in our
335919	339759	notepad. Can we resize this? All right.
339759	343120	Acceptable. Minimize. Unminimize. Good.
343120	345440	The big one. Is there a right click?
345440	348870	Yeah, there is.
348880	350720	also the correct date and time
350720	352400	in my local in the bottom right. And
352400	354080	there is what I would imagine here to
354080	355759	probably be the special feature. It is
355759	357759	the holographic desktop feature. So, we
357759	359759	will save that for last. Let's start out
359759	362160	by looking at the files app, which I
362160	363840	don't think I've seen a files icon that
363840	365600	looks like this before. All right. We
365600	367199	have some nice transparency to the
367199	369199	application backgrounds and definitely a
369199	371039	nice synth wave theme, which I do
371039	372639	personally like. So, unfortunately, it
372639	374240	seems our fake file system is just
374240	375600	static and we can't browse through it.
375600	377440	Nonetheless, can we create new things in
377440	379680	it? Okay, we can. Can we remove the new
379680	383280	things we created? We also can. Can we
383280	384720	remove things that were there by
384720	387360	default? Yeah, we can. All right, cool.
387360	388960	That makes me a little happier. And then
388960	391360	up would just go there. Oh, okay. Not
391360	394160	bad. Decent bit of immersion, at least
394160	395919	in terms of what directories would exist
395919	398319	were this a real computer. Notepad we
398319	400160	did kind of see because it opened with
400160	401759	this just to give us some information.
401759	403199	The only thing I like to see is will it
403199	405440	save it as a text file on the system. No
405440	408000	such directory. Didn't I delete the
408000	410319	documents directory? Hold on a second.
410319	412720	That's okay. Oh, it remembered that I
412720	414319	did that. I might have messed that up.
414319	416880	So that's going to be that's on me. Uh
416880	420240	okay. This right here is just not right.
420240	422960	It should not have a like visible text
422960	425680	box. That happens sometimes. Help did
425680	429759	give us things. Let's just do Neoetch.
429759	431680	That's actually pretty well done right
431680	434309	there.
434309	434319	there.
434319	439440	Nebula or Nebula. Needula. Still, that's
439440	441039	that's actually all right. I'm just
441039	443759	being kind of mean. Memory. This is
443759	446800	fine. Oh, it actually put a bit of
446800	449840	cleverness in there. 1.44 megs out of
449840	451599	infinity. This is fine. Isn't that how
451599	453280	much a old school floppy disc would
453280	455440	hold? I don't know cuz those are older
455440	458560	than me. Next up, calculator. Okay, so
458560	461520	47 *
461520	464390	8
464390	464400	8
464400	465909	8 376.
465909	465919	376.
465919	467199	I haven't done one of these in a while
467199	472560	divided by six. All right, settings.
472560	475440	Definitely seeing a bit of like Quen 3.8
475440	477280	in here just in terms of some of the
477280	479520	style here. What is that and why do I
479520	482160	like it?
482160	485680	Aurora. Oh, okay. Sunset
485680	489440	City. This No one would understand that
489440	491599	reference. This looks like something one
491599	493199	of my friends in high school was making
493199	495039	as a game project. It actually looks
495039	497440	like a ton like the splash screen for
497440	500319	his project. We have customizable
500319	503280	accents as well. Difficulty easy,
503280	506080	normal, and hard. Oh, for games. Okay, I
506080	508080	probably set it to easy as you may be
508080	510479	familiar with my quick to anger game
510479	513200	playing style. Next up, let's try our
513200	516880	games. The speaker is on. So, City
516880	521110	Heist.
521120	526949	Oh, you absolute.
526959	528640	Okay. So, you have to press enter to
528640	532160	start and not click the button.
532160	539750	Please shift gears. Ah, okay. Okay.
539760	543670	All right. It's
543680	551910	Oh, cool.
551920	553920	Oh, okay.
553920	558470	It's troubled,
558480	560959	but
560959	563040	the sounds for like the smack are kind
563040	565519	of funny. Like, all right. And we
565519	567680	wrecked and we only grabbed a 100red out
567680	570080	of a thousand, but it did work. So, keep
570080	571839	in mind we need to press enter to play
571839	573519	these games and not hit the button that
573519	584389	says launch. Oh. Oh, all right.
584399	587440	Do I go through these? Yeah. I I just
587440	589440	assume based on the sounds. I wonder if
589440	590959	that's how like dogs are since they
590959	592800	don't understand language. But like that
592800	594240	to me seemed like a positive
594240	595839	reinforcement sound. So, it makes me
595839	597760	want to do it more. And I think we did
597760	599440	just discover something about how dog
599440	602240	sprains work.
602240	604959	Okay, the score. I don't know that we
604959	607360	need that many decimal points just to
607360	609040	show a score. I think maybe just
609040	612160	rounding up to 734 or down would have
612160	614800	been acceptable, but I applaud it for
614800	618000	its intricacy. And then, oh, read me.
618000	619680	Okay, so our special feature is
619680	621680	something like holographic. And I would
621680	623440	imagine having a couple apps open just
623440	625279	when we trigger that is probably the
625279	628720	move. So, oh yeah. All right. So, it
628720	630160	just puts like an interesting card
630160	631839	effect. This is what it looks like. This
631839	633839	looks like an Apple Vision Pro like app
633839	635680	demo. This is the way they look from
635680	639120	inside the Vision Pro headset. Um, just
639120	642160	if you have used one, you may resonate
642160	645600	with what I'm saying. Okay. Overall, it
645600	647600	was fine. Everything worked. Nothing
647600	649120	really blew me away. But again, that may
649120	651600	be a limitation of the test and less of
651600	654399	a limitation of the model. So alas, huh,
654399	656560	it's kind of neat. We will continue. So
656560	658800	our skateboard game is still going. And
658800	661120	again, I'm using this with the newly
661120	662959	releasleased beta version of the chat
662959	664959	GPT desktop app. I would imagine this
664959	666720	will take quite a bit of time. So let's
666720	668959	check to see just like Oh no, cuz we
668959	671040	already had looked at the cost. And keep
671040	673920	in mind this is partially for the skate
673920	675839	game as well as the browser OS. So 20
675839	678720	cents total across two separate tasks.
678720	680480	One which is still ongoing. All right.
680480	684320	So, after 68 minutes and 24 seconds,
684320	687839	Skate NYC is complete. So, oh, you
687839	689360	idiot. You spoiled it with these
689360	690959	screenshots and you don't even have
690959	693829	vision.
693829	693839	vision.
693839	696079	I'm infuriated. Nonetheless, now that we
696079	698480	know what the game looks like, I had to
698480	700000	stop myself from saying something a
700000	701760	little more colorful. Let's just take a
701760	705040	peek at at the game. All right, so
705040	708000	here's our NYC skate game. Should be in
708000	711440	outputs, right? Yes.
711440	716160	All right. It's not quite as good as the
716160	719760	Quen 3.8 model, but
719760	721600	that model does have 800 billion more
721600	724870	parameters.
724880	727200	Does that say NYC skate on the ground?
727200	729040	It's inverted, but it does. And it says
729040	732399	established 1999. All right. Uhoh.
732399	735440	Sounded like the beginning of uh the Van
735440	738000	Halen song. I don't know what exactly
738000	742000	is. All right. Uh, let's just maybe not
742000	743760	do that same thing. There are sound
743760	745279	effects, though. I don't normally hear
745279	746639	sound effects. Can we actually move
746639	750959	beyond? All right. Not quite. I will say
750959	753519	it's not wonderful. Oh, that's actually
753519	755760	kind of cool, though. The way it like
755760	762310	stylized the game title. Nice. Nice.
762320	763839	Well, we don't need to do the drum kit
763839	765829	sim.
765829	765839	sim.
765839	769509	The trees are
769519	772959	quite reminiscent of a poo if I were to
772959	775200	be honest, but all right, the the legs
775200	782069	are like kicking.
782079	783839	This may be the funniest skate game
783839	785760	result we've ever seen, which is, you
785760	787120	know, something. All right, we have a
787120	790480	street sign there. The text is a little
790480	792800	sideways, as are the building faces.
792800	794720	This is definitely a troubled result.
794720	796959	However, I do like the water fountain
796959	798480	shooting out. I believe that was
798480	802069	something I've seen.
802079	805120	Okay. All right. I've seen that in I
805120	806959	think the Quen result. Now, here's what
806959	809120	I would like to just quickly do for some
809120	810720	reference. This is the result we
810720	812959	received for the C++ skateboard test
812959	815920	from the preview version of Deep Seek
815920	818160	V4. We can see right here. I'm going to
818160	820560	turn the speaker off. This is pretty
820560	823760	bad. It was like so bad that I actually
823760	825600	remember to go to this specific thing
825600	828160	just to showcase it. So, the leap in
828160	830720	capability based off of what we see
830720	832959	right here to what we just saw in the
832959	835200	skate game that it had created for us,
835200	836639	which we can just look at in these
836639	838800	screenshots here is still pretty
838800	840480	significant. Next up, I'm giving it the
840480	842800	Subway FPS test. I've also added on to
842800	844399	the end of the prompt. Don't worry about
844399	846000	checking over the script. I will
846000	847360	manually check it and if there are
847360	849120	issues I will report them back to you
849120	850800	just because this will really like
850800	853279	extend every single result by like two
853279	856399	times in terms of like timewise length
856399	858079	just in the interest of something I
858079	860320	could do visually in like 5 seconds and
860320	861920	then be like hey this doesn't work. So
861920	863440	we're going to do that just to maybe
863440	865279	like keep things a little tighter. All
865279	868079	right so it did finish the Subway FPS
868079	869680	game. The problem is it just keeps
869680	872079	trying to do like troubleshooting which
872079	874079	it's not good at doing on this system
874079	875440	for some reason. All right, let's check
875440	879199	out our subway FPS. M Rand Avenue. All
879199	880949	right,
880949	880959	right,
880959	882959	going to turn the speaker on. So, it
882959	884800	gave us two options. We can either
884800	889040	explore the station. Okay, it's Hey, you
889040	892560	know what? I think that is the best
892560	894880	subway car that I've ever seen drawn in
894880	897360	this test. And you're going to be like,
897360	899120	well, it doesn't look that great. I
899120	900800	know. It's just that they never really
900800	902639	include these. And when they do, they're
902639	906160	not really very good. So this actually
906160	907680	what was that? If that thing starts
907680	910480	moving I would have probably become more
910480	913440	excited than would be reasonable for you
913440	916160	know responding to such thing. This is
916160	918000	very clean. It's simple. It's nothing
918000	919760	that's blowing me away to be completely
919760	921600	honest with you. Queens Boulevard line.
921600	923279	Okay. We have stairs. We have some
923279	924720	lighting elements and things of the
924720	927120	sort. We have ace if you want to go that
927120	931360	side. And then tunnels
931360	933440	like indicators for the train's trash
933440	937360	bin. Floating particles more simple. And
937360	938639	we can't go up the stairs, which is
938639	940480	fine. We don't want to. Now, of course,
940480	943360	the other feature that was included here
943360	946639	is the FPS. So, let's go back to menu
946639	948639	and then enter the survival FPS. Let's
948639	950000	see if it changes the lighting or
950000	952480	anything. Oh my. That's a good weapon
952480	957600	model. Does it? How? Oh, it does.
957600	959600	I think.
959600	962880	Yeah, it does. It's like smooth now. I
962880	965040	don't see any enemies. Look at when we
965040	968000	run. It's like a waddle.
968000	971279	All right. So, uh the big thing I'm
971279	973040	noticing here is there are no zombie
973040	976160	enemies. So,
976160	978399	good reload. The waddle is just kind of
978399	983030	funny.
983040	985519	Weird. Hey, look. There are shell
985519	987040	casings that actually drop. That's kind
987040	989279	of cool. All right, this is like Yeah,
989279	991440	it's kind of mid. So, I'm going to give
991440	994399	it a feedback piece. Now, I stopped it
994399	996079	because it was just like it had finished
996079	998000	and it was like I can't use this in the
998000	999680	headless environment to check. And at
999680	1001040	that point, it's easier to just check it
1001040	1003440	and tell it what's wrong. So, I'll give
1003440	1005279	it that honest feedback. So, supposedly
1005279	1007360	our fixed and polished version of the
1007360	1010079	Subway FPS is completed. Let's just take
1010079	1012800	a peek at it. The sound is on. All
1012800	1015279	right. Interesting. I think it swapped
1015279	1018560	the positioning of the buttons. So,
1018560	1020399	all right.
1020399	1022160	Seven hostiles. It's showing me in the
1022160	1026480	top right. I unfortunately just can't
1026480	1028720	locate them. I would have expected it to
1028720	1030959	have done better. And again, this is one
1030959	1032559	where we could go back and reference the
1032559	1035360	previous model of this. So, basically
1035360	1037760	the preview version, it was
1037760	1039679	significantly less detailed than this
1039679	1042710	was.
1042720	1044720	though unfortunately we don't have
1044720	1046640	enemies and given the amount of time it
1046640	1048559	took as well as the kind of multiple
1048559	1050480	times I told it there were no enemies
1050480	1052320	I'm regrettably going to call this one
1052320	1054400	next up I'm giving this a difficult task
1054400	1058320	to create a 3D model of an RB26 inline 6
1058320	1060400	engine it must be something that can be
1060400	1062799	3D printed with little to no support and
1062799	1065280	fit inside of it a tiny DC motor called
1065280	1067600	an N20 motor because this is something
1067600	1069440	that I've told it should like work with
1069440	1071600	my RC car project actually I don't think
1071600	1073200	that's specifically in this prompt, but
1073200	1074960	that's the whole point of this where
1074960	1076960	this is it should understand and it will
1076960	1079280	scale it properly to that motor. So far,
1079280	1080880	I've run this test with a few
1080880	1082640	intelligent models and I've never really
1082640	1084000	received anything that was actually
1084000	1086480	worth printing. So, hopefully sooner
1086480	1088400	than later that will change. So, our 3D
1088400	1090400	model task is pretty interesting because
1090400	1092160	it's just doing a lot of research on
1092160	1094320	like the actual proper orientation of
1094320	1096640	specific items on this engine. Like
1096640	1098160	where is the downpipe, which is an
1098160	1100080	exhaust component? What side is it? What
1100080	1101840	is the intake side? What is the exhaust
1101840	1104880	side? Intake plenum left or right? So, I
1104880	1106960	do like to see the amount of like
1106960	1108880	intricate research that's going into the
1108880	1111200	culmination of this model creation. And
1111200	1113760	it would be really awesome for like us
1113760	1115840	to get a model for this task that's
1115840	1117120	actually printable. It would be the
1117120	1119840	first one. So, in 1 hour and 10 minutes,
1119840	1122640	we received our 3D model response. This
1122640	1125120	is a very, very intricate just bit of
1125120	1127520	text about the thing it made. Now, of
1127520	1128799	course, the other part of that is
1128799	1130480	actually just taking a look at what it
1130480	1132880	looks like. Though, once I remember what
1132880	1134559	I named this folder, we will be able to
1134559	1136720	take a peek. Assembled display STL. I
1136720	1138559	would imagine that to be the case.
1138559	1140799	Please be good. I've really like I need
1140799	1144080	a good one for these.
1144080	1147039	Actually, it's sideways. Son of a gun.
1147039	1150480	This did a really good job.
1150480	1153840	This is actually like
1153840	1157360	so far the best one. It's not perfect.
1157360	1159520	We have some sideways turbocharger
1159520	1162160	issues, though. I mean, this is All
1162160	1163760	right, let's let's peek at this in
1163760	1165360	OpenCAD. And there are a bunch of
1165360	1167200	individual pieces for this. So, we also
1167200	1169039	want to see if the engine that we told
1169039	1170640	it needs to fit inside the little remote
1170640	1172799	control car engine. Let's see if it has
1172799	1176400	any hope of properly fitting research.
1176400	1178240	It has a research folder right here just
1178240	1181679	showing like a bunch of images and stats
1181679	1184559	for this motor. Very cool though. And
1184559	1186160	here we have some better preview
1186160	1188230	renders.
1188230	1188240	renders.
1188240	1191440	This is definitely well done.
1191440	1195280	Turbo side. Okay. Underside. All right.
1195280	1196880	It's not something that's going to
1196880	1199440	really easily be 3D printable. Although,
1199440	1201200	I guess I can't make that determination
1201200	1203360	because this comes in multiple pieces.
1203360	1205440	So, it's not designed to be printed as
1205440	1207520	we first saw it just in one lump like
1207520	1209120	this. It's designed to be printed
1209120	1211600	separately. I would like to see. Let's
1211600	1214720	look now at just the single first piece,
1214720	1216480	which is ideally where it would be
1216480	1219440	placing the 3D printed motor that would
1219440	1221760	go in it.
1221760	1223840	Unfortunately, printing this without
1223840	1226559	supports is just not an option in any
1226559	1228400	potential orientation that we see right
1228400	1230320	here, which is a bit frustrating. It
1230320	1232000	almost seems like it has these posts and
1232000	1233760	holes in it to be able to snap fin on
1233760	1236799	other parts, which is cool. Okay. Yep.
1236799	1238559	And I think those are definitely like
1238559	1240480	attachment posts, which is pretty cool.
1240480	1242400	This one would be able to be printed
1242400	1244240	more or less without any supports if we
1244240	1247280	were to put it flat. Then this should
1247280	1249760	have the Nissan logo, or at least the
1249760	1253440	text Nissan written on it.
1253440	1256080	Yep. And this one's quite well done. No
1256080	1258799	issues here. The oil pan. Oh, no. Okay.
1258799	1261280	This is where we This is where we decide
1261280	1262400	we're not going to be trying to print
1262400	1265840	this. I'm not disappointed, but I'm not
1265840	1268799	super impressed either. So, I suppose I
1268799	1270799	am lukewarm when it comes to judging
1270799	1272880	this. Definitely, if we just look at
1272880	1275440	this as a render, even forget about
1275440	1276960	looking at the individual parts there,
1276960	1279120	cuz it's not really going to be feasible
1279120	1281600	to print this, it did a really decent
1281600	1283600	job just assembling it. The research
1283600	1285280	that went into this, I think, is almost
1285280	1287679	probably like I hate to sound like an AI
1287679	1289600	here, but like the story isn't the
1289600	1291120	model, it's the research that went into
1291120	1293520	it. That's like generic like AI slob. It
1293520	1296240	really did a lot of decent research
1296240	1298720	here. I mean, did it find this and
1298720	1302000	download it just to see? No. Okay. So,
1302000	1304240	this is something it maybe made or maybe
1304240	1306320	probably found online just to be able to
1306320	1308080	better assess what the size of this
1308080	1310159	asset is, but it took a lot of care to
1310159	1311840	ensure that what it tried to generate
1311840	1313840	would have hypothetically fit the
1313840	1315760	specific specifications that were
1315760	1318960	denoted in the original prompt. So, not
1318960	1321200	half bad. I'm also going to give this a
1321200	1322880	simple front-end design test where it
1322880	1324720	needs to create the watch website. So,
1324720	1326240	it has to have a pretty cinematic
1326240	1328320	panning effect over a 3D watch asset
1328320	1330799	that it creates. I didn't specifically
1330799	1332720	Okay, no, I did specifically denote that
1332720	1334559	the watch needs to be a 3D model. So,
1334559	1336240	we'll see how competent it is at putting
1336240	1338720	the entirety of this together. So, in 1
1338720	1341280	hour and 30 minutes, we got our watch
1341280	1343520	front end result. It again seems to have
1343520	1345760	a pretty substantial like just list of
1345760	1347679	what was built here. So, let's take a
1347679	1349600	look and see how it is. beautiful watch
1349600	1352880	front end website.
1352880	1355280	That's actually it's so close to
1355280	1359039	excellent. The big problem is this, I
1359039	1360559	would assume, is the watch strap
1360559	1363039	sticking out. This is 3D, so don't be
1363039	1365760	fooled by its 2Dishness look. It is
1365760	1368320	actually 3D. It's slowly panning, too. I
1368320	1370480	really Okay, so a couple things I do
1370480	1372880	like the wooden background and then this
1372880	1375360	like leather card or pad that this watch
1375360	1377520	is placed on. Additionally to that, the
1377520	1379280	face looks really good. The materials
1379280	1381520	look very good. It's just like an
1381520	1384320	overall it's a high-end aesthetic, which
1384320	1386080	is exactly what we go for when
1386080	1389039	performing this test. Now, that's
1389039	1391200	actually that right there looks good.
1391200	1394149	The
1394149	1394159	The
1394159	1396320	Let's just sit and appreciate for a
1396320	1398640	minute. Okay. Yeah, that maybe I could
1398640	1401200	have done without that, but it's
1401200	1403600	definitely quite something. And then
1403600	1405039	look at the way that goes away as soon
1405039	1407760	as we start to scroll down. Quiet luxury
1407760	1410480	measured in seconds.
1410480	1412799	Sunburst dial. Okay, it has like a
1412799	1415280	material card here. The sapphire crystal
1415280	1417200	is perhaps something I would have maybe
1417200	1419120	omitted from this section, but the dial
1419120	1421200	looks good. Two watches, one point of
1421200	1423200	view. Please render them. Oh, look at
1423200	1425360	that. That is sick. That really looks
1425360	1426799	pretty good. I love the way they're
1426799	1428480	placed on the table. Really the big
1428480	1429840	thing letting it down is just this
1429840	1432320	leather slab is inhibiting our view of
1432320	1434559	these beautiful straps. This one is the
1434559	1436880	stainless steel option. This one is the
1436880	1439440	gold option and priced accordingly I
1439440	1441679	would imagine. Yeah. Although I would
1441679	1445520	say the jump in price of like it's a
1445520	1447039	negligible amount to get the gold one.
1447039	1448320	I'd probably go with this one
1448320	1450159	personally. Oh, cool. And these ones we
1450159	1452720	can freely rotate so that helps a bit.
1452720	1455440	There is no movement to the watch hands
1455440	1456799	but that's okay. I'll forgive it for
1456799	1459760	that. But these are pretty darn cool.
1459760	1462880	These are well drawn, I would say.
1462880	1466400	Especially the numeral indicators. Look
1466400	1468320	at the way that they actually have multi
1468320	1470320	material on them. So, they're not just
1470320	1472960	like one static material. It's gold and
1472960	1475360	then black infill or silver and then
1475360	1476880	black infill as well. So, that's some
1476880	1479039	attention to detail I do like seeing. As
1479039	1481520	well as the kind of brushed material
1481520	1483120	look on the stainless steel and the
1483120	1484480	gold. It's probably weird we're like
1484480	1487440	assessing like a 3D watch, but it just
1487440	1489840	shows like some capability. Okay, that
1489840	1491120	looks pretty good right there. And then
1491120	1492960	we have a clean footer as well. I like
1492960	1494640	this. I'm happy with this. Even though
1494640	1498720	it's like not perfect, it's darn close.
1498720	1500480	So, next up, just here from within
1500480	1502000	codeex. And I can't show this full
1502000	1504080	prompt because it has an API key. I'm
1504080	1506159	giving this the full-on cinematic Steve
1506159	1508080	the PC repair game test where it needs
1508080	1510320	to create a cinematic game. It needs to
1510320	1512320	create voice assets for these characters
1512320	1514559	just by using an open AI API key that
1514559	1516720	I've given it and it needs to make a fun
1516720	1519200	game that is enjoyable and playable. I
1519200	1521120	will assess this is probably going to
1521120	1522720	take north of an hour. So, we're just
1522720	1524480	going to basically leave Codeex here in
1524480	1526640	the background, the beta desktop version
1526640	1528960	of it to complete this task and we'll
1528960	1530400	maybe work on some other things in the
1530400	1532799	meantime. So, our Steve the PC repairman
1532799	1535679	cinematic game completed in basically 65
1535679	1537840	minutes. So, a little over an hour. All
1537840	1540159	right, here's how to play. I'll probably
1540159	1542400	just do what it says here, not assume
1542400	1544159	that it's already running this in some
1544159	1547360	server for us. So, okay, episode one,
1547360	1550720	the Azure palm job. Okay, click to
1550720	1552880	begin. Headphones recommended. We have a
1552880	1555110	speaker.
1555110	1555120	speaker.
1555120	1558159	>> All right.
1558159	1559600	>> My name is Steve.
1559600	1560240	>> This is good.
1560240	1562159	>> I fix computers.
1562159	1564240	>> Most of my clients are old enough to
1564240	1566480	remember when a hard drive was a luxury
1566480	1569440	item. They pay what they can afford and
1569440	1574400	I make it work. Mostly honest business.
1574400	1576880	>> Oh, thank you, dear. The screen kept
1576880	1579520	telling me the date was 1980 and I
1579520	1580960	thought all fixed.
1580960	1582640	>> I have been traveling long enough.
1582640	1584799	>> Fresh battery. Your computer now
1584799	1585110	believes
1585110	1585120	believes
1585120	1586640	>> I will bring you oatmeal cookies on
1586640	1587110	Tuesday.
1587110	1587120	Tuesday.
1587120	1587919	>> Same as the rest.
1587919	1589440	>> Give Mr. Buttons a scratch behind me
1589440	1590240	ears for me.
1590240	1594080	>> Let me walk you out, ma'am.
1594080	1596720	>> I'll be right with you, Mr. Thomas,
1596720	1598720	>> the speech overlays itself.
1598720	1600640	>> A good boy, Steven.
1600640	1603360	>> If that quiet man needs fixed, you be
1603360	1603909	patient.
1603909	1603919	patient.
1603919	1605039	patient. >> Punctual.
1605039	1608159	>> I like punctual. Door is unlocked.
1608159	1609039	Coffee is warm.
1609039	1610880	>> A man. What is the job?
1610880	1615120	>> Kane for 2 years. My life.
1615120	1617120	>> See the the head tilt movement back and
1617120	1619200	forth is actually there's three people
1619200	1620960	talking now. And he forgets the entire
1620960	1624320	name Charles buy him penthouse private
1624320	1626310	security
1626310	1626320	security
1626320	1627919	wants his legend some people would like
1627919	1630000	to ask him question him slowly the
1630000	1633120	client pays 100% upront I would like is
1633120	1635840	also up in the case all expenses first
1635840	1637840	class two tickets this is why we work
1637840	1640070	together
1640070	1640080	together
1640080	1642480	the top tray is a normal file on a
1642480	1648149	laptop
1648159	1651120	it Quiet is my second favorite thing.
1651120	1652880	Professional. I will not ask what the
1652880	1655200	first is. The phone rings when the
1655200	1657120	helicopter is ready.
1657120	1660080	>> 40.
1660080	1662880	>> Cut scene over. Thank goodness. Click to
1662880	1666400	take control. All right. Aside from the
1666400	1668559	sounds being just totally messed up,
1668559	1671520	this is actually
1671520	1676880	pretty decent. Okay. The cat is
1676880	1679840	There's definitely some
1679840	1681919	geometric issues. What are we supposed
1681919	1684080	to do? Take the briefcase when you're
1684080	1687039	ready. Okay. Where is it?
1687039	1690000	Well, pet buttons are like,
1690000	1692240	>> "Don't look at me like that. Buttons."
1692240	1692799	>> Easter egg.
1692799	1694240	>> A man's got to eat.
1694240	1697919	>> Okay. Where's the briefcase?
1697919	1699520	All right.
1699520	1702960	We have inventory. Zip ties. Okay. E to
1702960	1704399	continue. All right. Is it going to put
1704399	1704630	us
1704630	1704640	us
1704640	1705279	us >> Miami?
1705279	1705679	>> Nice.
1705679	1708960	>> Palm trees, humidity, and one very bad
1708960	1711840	man, Mr. Colton.
1711840	1712559	>> Champagne perhaps.
1712559	1713919	>> I'm just going to click through cuz it's
1713919	1714789	like
1714789	1714799	like
1714799	1717919	>> black like my well coffee
1717919	1720559	>> asset at the resort.
1720559	1721919	>> Poolside at noon.
1721919	1723120	>> Oh, poolside.
1723120	1726720	>> 170 enters. Camera coverage.
1726720	1730080	Elevator. There is no dog. Blind is the
1730080	1733440	west service stairwell.
1733440	1734720	Well, I think we just saw a glimpse of
1734720	1736559	the pool.
1736559	1741279	All right. Oh, great. Azure Palm Resort.
1741279	1743600	Click to take control.
1743600	1746480	It's even got like the like low rent
1746480	1748240	wallpaper that you'd expect at one of
1748240	1751200	these joints. Oh, wow. This is actually
1751200	1752880	like kind of intricate. Oh, good. Okay.
1752880	1754159	I was worried we wouldn't be able to
1754159	1756720	just walk through the walls, but okay.
1756720	1758799	We can toss a coin in the fountain. Your
1758799	1762080	wish. No more CEOs errors. 50/50. It
1762080	1764559	seems like Steve's shop and the airplane
1764559	1767360	scene is like oddly placed inside of the
1767360	1769760	location we're supposed to be at. Let's
1769760	1771760	just ignore that.
1771760	1774320	So, check in at the front desk and then
1774320	1777840	reach the 17th floor.
1777840	1780960	>> Welcome to the Azure Palm. Checking in,
1780960	1782880	Mr. Me.
1782880	1785120	>> Coulter. Steven Coulter.
1785120	1785679	>> Mr.
1785679	1788720	>> I believe I have a reservation. Room
1788720	1792080	1408, partial ocean view. I do
1792080	1794080	apologize. The ice machine on your floor
1794080	1795600	is out of order.
1795600	1800149	>> Story of my life.
1800159	1801760	>> All right. We now have a room key in our
1801760	1809590	inventory. So, is this the elevator? Oh,
1809600	1812000	is that
1812000	1815440	reach? Okay. West service stairwell.
1815440	1820789	Let's just assume this is north.
1820799	1842549	Good
1842559	1844080	like
1844080	1848399	what? Well, that is
1848399	1850640	frustrating because it gave us this
1850640	1853279	really like anticipatory music. I'm sure
1853279	1854799	there's a far better terminology for
1854799	1856000	what I was about to say.
1856000	1857200	>> My name is Steve.
1857200	1859840	>> Oh, no. Don't.
1859840	1862000	>> Most of my clients are old enough to
1862000	1864720	remember when a hard drive was a luxur.
1864720	1867600	>> All right. We we got most of the way. I
1867600	1870720	think this is a test that more often
1870720	1873360	than not, we don't complete it. I will
1873360	1875120	say the elements of this judged
1875120	1877120	individually. Definitely the pieces were
1877120	1878559	all there. I don't know that the puzzle
1878559	1880399	made the correct picture that was shown
1880399	1882000	on like the puzzle box, but it was
1882000	1884159	pretty darn close and I saw some
1884159	1885919	truthfully this did a better job than I
1885919	1887760	expected, especially considering like
1887760	1889679	the subway scene game and things like
1889679	1892799	that. So, I'm actually
1892799	1895679	I'm okay with this. So, the final thing
1895679	1898000	was of course the street yeet game,
1898000	1900320	which is weird because this piece of
1900320	1902080	text moving up and down like that while
1902080	1903840	this one remains completely static makes
1903840	1906480	me uncomfortable in some level. So, this
1906480	1909760	is a Bejian issue. All right, good. So,
1909760	1912480	I neglected to um just run it in a
1912480	1917279	server. One fist, zero mercy.
1917279	1924310	I already like this. Oh, it seems like
1924320	1926399	Good music. All right. All right. All
1926399	1930470	right.
1930480	1936789	Let's just explore the city real quick.
1936799	1940000	I don't even know. Okay, so I saw a car
1940000	1945509	get like
1945519	1954789	This one's actually kind of fun.
1954799	1958640	Excuse me. Excuse me. Pardon me. It does
1958640	1960960	like a like a
1960960	1962880	Okay. I don't know that we hit. Is that
1962880	1967669	a Can we hit the bike?
1967679	1974549	Bonk.
1974559	1980950	Uh-oh. Jay-walking pedestrian.
1980960	1982960	Can we punch this parking meter? Yeah,
1982960	1984720	we can just everything except the
1984720	1986159	buildings. All right. There's
1986159	1988240	unfortunately the orientation of some of
1988240	1990240	these things is not quite right. We can
1990240	1992000	see that we're moving sideways when we
1992000	1995750	move forward.
1995760	1998880	And the cars are also
1998880	2001440	Sorry, pal.
2001440	2003110	Oh,
2003110	2003120	Oh,
2003120	2005200	and it's interesting how they uh
2005200	2006960	interact with the scenery as well. Can
2006960	2008720	we eat this statue? I don't think we
2008720	2014149	can, but Oh, yeah, we can. All right.
2014159	2016640	I will say the cartooness of this game
2016640	2018960	is actually kind of well done. I don't
2018960	2020799	know how easy that is to assess just
2020799	2022720	watching someone play it, but like the
2022720	2024880	pacing is decent. I think maybe that's
2024880	2027600	attributed to the music as well, but the
2027600	2029039	way that when you actually All right,
2029039	2030399	the the loop is getting a little
2030399	2032720	repetitive when you actually yeet the
2032720	2034799	things and they also go and interact
2034799	2036799	with other items in the environment. So
2036799	2039360	like how that will bounce off the
2039360	2040880	building and then maybe hit other things
2040880	2043039	in the path. That's pretty cool. And
2043039	2045039	this was
2045039	2047760	actually quite arcade style and
2047760	2049430	playable.
2049430	2049440	playable.
2049440	2051280	Our score is going up. We have combo
2051280	2054480	multipliers overall. 38 here is a
2054480	2056560	success. All right, so that is probably
2056560	2058079	going to conclude our first look and
2058079	2060560	test of the general availability DeepS
2060560	2063760	V4 Pro, the 0813 model designation. Now,
2063760	2065520	this is definitely an improvement over
2065520	2067359	the preview version that felt pretty
2067359	2068960	undercooked and I was not at all
2068960	2070879	impressed with it. But if we look back
2070879	2072560	at some of the tests that were run in
2072560	2074159	that video, which is on the channel, we
2074159	2076079	ran some of the same ones today. This is
2076079	2078879	a significant improvement. I would say
2078879	2080399	though, I think the flash model
2080399	2082159	impressed me more than this, especially
2082159	2084320	considering the size of it. This has a
2084320	2086720	bit more competition, but again, it is a
2086720	2088720	big leap on its predecessor, which I
2088720	2090320	guess is what we like to see. I was
2090320	2092159	impressed a few times with some things
2092159	2094240	like basically results within results.
2094240	2096560	The cinematic Steve the PC repair game,
2096560	2098640	it actually had some elements to it that
2098640	2099920	were like, hey, you know, it kind of
2099920	2101599	thought this through. Just the way when
2101599	2103040	they were speaking, their heads were
2103040	2104640	like moving back and forth. Not the
2104640	2106160	weird chin thing that was happening. I'm
2106160	2108240	not sure what happened there. The music
2108240	2110079	it included in both that game and then
2110079	2112240	the Street Eat game were actually like
2112240	2114480	they definitely fit the scene that it
2114480	2116160	was drawing here and it did a good job
2116160	2117599	making the Street Eek game very
2117599	2120000	cartoonish and the playability of it
2120000	2121680	felt good, I guess, for lack of a better
2121680	2124000	term. The model that we got of the
2124000	2127040	engine was actually it wasn't printable.
2127040	2129119	So, unfortunately, maybe it failed to
2129119	2131040	like check that box in the specific
2131040	2132880	task. However, if we just look at one of
2132880	2134960	the previews right here, it did a pretty
2134960	2137119	decent job. Definitely better than the
2137119	2138880	other models I've tested this with.
2138880	2140640	Though, I have not tested this specific
2140640	2142560	prompt with any big beefy Frontier
2142560	2144640	models yet. So, keep that in mind as
2144640	2146079	well. But, it definitely had some chops
2146079	2147839	here. And the amount of research it did
2147839	2150560	to actually get to that point was quite
2150560	2152480	a lot. We have all the things in here
2152480	2154000	downloading these images when it doesn't
2154000	2155760	have vision. So, I'm not quite sure what
2155760	2157760	to make of that. But nonetheless, I
2157760	2159359	think probably the biggest point of
2159359	2162560	frustration for me is it just goes all
2162560	2164880	out when trying to actually ensure that
2164880	2167040	every aspect of every line of code it
2167040	2169520	made is correct. And with a model like
2169520	2171920	this where it does not have vision,
2171920	2174320	sometimes it takes longer and there is a
2174320	2176000	cost thing to it as well. Assuming
2176000	2177440	you're going to be paying to use these
2177440	2180640	models, it just sometimes is simpler to
2180640	2182480	have it make the result, not try to do
2182480	2184000	like overviews and checks and
2184000	2185839	follow-ups. you look at it yourself and
2185839	2187359	then tell it what's wrong and tell it to
2187359	2190079	fix it. Having to just go all out and it
2190079	2192079	trying to fix it adds a lot of time and
2192079	2194960	expense and tokens to the task, which
2194960	2196960	maybe is more of a consideration when
2196960	2199599	doing zeroot tests like this, but I find
2199599	2201280	that's something I've consistently had
2201280	2203119	frustrations with as well as with the
2203119	2205119	flash model where if I'm having them
2205119	2207040	make like a say street heat, it's easier
2207040	2208720	if they just spit the script out, I test
2208720	2210160	it and say, "Hey, the buildings are
2210160	2211920	disappeared." and then it fixes it
2211920	2213599	rather than it trying to go through and
2213599	2214960	like well let me try to spin up a
2214960	2216640	headless browser and then I'll take a
2216640	2218640	screenshot and like do a heat map
2218640	2220320	because it can't actually see it or some
2220320	2222720	other function. So it's a little
2222720	2224230	tokenhappy.
2224230	2224240	tokenhappy.
2224240	2226320	I did run it on the highest potential
2226320	2227760	reasoning mode. So that could be
2227760	2230560	attributed to it as well, but I still
2230560	2232800	would like to be able to do that without
2232800	2234640	it like going all out to try to make
2234640	2236480	sure it's polishing every aspect of the
2236480	2238480	script because I want max performance. I
2238480	2240400	just don't need that level of depth. So,
2240400	2242720	that may be a personal complaint, but
2242720	2244640	you know. So, for the final thing, let's
2244640	2246320	take a look at everything we spent for
2246320	2248960	the entirety of this test. We can see
2248960	2251920	right here that all cost $1.91.
2251920	2253760	So, that was pretty darn cheap. And you
2253760	2255839	do have to remember this is going to be
2255839	2258320	increased their API prices. They say
2258320	2260640	significant increase expected, but I
2260640	2262400	mean a significant increase might be
2262400	2264320	like, okay, the entirety of this test
2264320	2266640	was $8 or something, which is still in
2266640	2269359	perspective not horrible. The Quen 3.8 8
2269359	2271760	max test video that we did a few weeks
2271760	2274000	ago or like a week or two ago that was
2274000	2276640	about $32. So keep that in mind as well.
2276640	2278560	That may be a bit more performant than
2278560	2281040	this, but there is going to be a pricing
2281040	2283359	consideration as well. And this was very
2283359	2285760	very very cheap. It won't remain that
2285760	2288160	cheap, but even if it gets five times as
2288160	2289920	expensive, it's still not horrible. I
2289920	2291359	suppose that's going to conclude our
2291359	2294000	first look and test of the 0813 Deepseek
2294000	2296560	V4 Pro model, the general availability
2296560	2298560	version. And this is now out of preview.
2298560	2300560	And that's just pretty exciting.
2300560	2302720	Basically, now I think I'm going to test
2302720	2304800	Grock 4.6 because that also came out
2304800	2306960	today. And then I believe Friday Quen
2306960	2309520	3.827B is coming out. So that'll be
2309520	2311200	exciting as well. Seems like there's
2311200	2312960	been this has been like a pretty busy
2312960	2314720	week in terms of model releases. So
2314720	2317599	that's cool also. And when Musepark 1.2
2317599	2319359	comes out, that's going to be awesome. I
2319359	2320720	don't know how big it is, but I'm very
2320720	2322720	excited to try to run it locally. So
2322720	2324400	with that, if you have any questions,
2324400	2325760	please feel free to leave them in the
2325760	2329200	comments. And thanks for watching.