YT YouTube2Text

YouTube2Text → Transcripts → Opus 5.5 모든 effort로 동일한 과제 시켜봤습니다

Opus 5.5 모든 effort로 동일한 과제 시켜봤습니다 — Transcript

by 개발동생 · 3,836 words · 581 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Everyone, in this video, I tested two
0:02tasks by inputting various port levels
0:04into the Office 5.5 model to determine
0:06which port level is the most efficient.
0:09You can view this as me testing up to
0:11five port levels for each task. For the
0:13first task, I was asked to create a
0:15game where you can explore the subway
0:17history of Gangnam Station. I will not
0:20reveal the port level for now. You can
0:22consider this to be a simulation of the
0:24Gangnam Station created using one of
0:26the first port levels, ranging from Low
0:27to Max. This is a simulation created
0:30with a different port level, where the
0:32subway station entrance appears. Also,
0:35when you come down to the waiting room,
0:37you can see scenes like people walking
0:39toward the ticket gates and looking at
0:41old maps. I also made a music video,
0:45and I will show that briefly as well.
0:49Regarding the music video for "When the
0:54Stone Opens in This Life, It Spins and
0:58Spins,"
1:01>> I incorporated AI-generated music and
1:03requested that the AI independently
1:06generate all the essences and graphics.
1:09The Office 5.5 model was released on
1:11September 22nd. Since then, it has
1:13garnered tremendous attention, and we
1:15can see a trend where many results
1:17utilizing Office 5.5 are still emerging
1:18. In fact, according to those actually
1:22using the Office 5.5 model, some even
1:23say that it feels like using tokens
1:25almost without limit, which suggests
1:27that the token efficiency and
1:28performance have improved significantly
1:30. Let me make a quick advertisement
1:34before we go. This video was produced
1:35with the support of Gilbeop Publishing.
1:38Today, I would like to introduce a book
1:40titled "Starting the Claude Code
1:41Properly." The distinguishing feature
1:43of this book is that it covers
1:44everything from the Claude Code
1:46Executive to Simma in a single volume.
1:47The book is divided into two parts: the
1:49"Simple" and "Spicy" sections. The
1:50Simple section covers everything
1:53step-by-step, from Claude Code
1:54installation to service development and
1:55deployment, so that even those with no
1:56prior knowledge of Vibe Coding can
1:57follow along. The Spicy section, on the
1:59other hand, allows those already
2:00familiar with Vibe Coding to take their
2:02skills a step further by covering a
2:03wide range of topics, from Context
2:05Engineering to Spec-Driven Development
2:06and Harness Engineering. I read the
2:09book myself, and I liked that it is
2:10structured in an easy-to-follow manner,
2:11allowing you to immediately follow
2:13along with exercises such as automating
2:14YouTube shows, creating Discord AI
2:15assistants, and building stock reports.
2:17Furthermore, since this book ranked#1
2:19on three platforms—Kyobok, S24, and
2:20Aladin—I highly recommend it to
2:22anyone who wants to learn how to
2:23properly use Claude Code. Also, Deal
2:25has generously provided ten copies of
2:27this book for the subscribers of "
2:28Development Dongsaeng." Please leave a
2:30comment mentioning what you would like
2:31to learn from this book. If you would
2:33like to enter the giveaway to receive a
2:35copy, please click the link in the
2:36pinned comment to enter. We look
2:38forward to your participation. The
2:39motion graphics I showed you earlier
2:41did not use any external SS at all.
2:43However, if you were asked to create
2:45something like this after adding
2:46elements like mid-range animation, it
2:48could produce something like this.
2:50Isn't the quality amazing? Currently,
2:51Claude's port levels are broadly
2:53divided into five categories. You can
2:56view them as features where, as you
2:58move up from Low to Medium, High,
3:00X-High, and Max, the cost increases and
3:02the work time lengthens, but it allows
3:04for much more meticulous attention to
3:07detail. The default port level for the
3:10current Claude Offset 5.5 model is
3:12considered to be Medium. In this video,
3:17I will help you judge which port level
3:19is most efficient by showing you the
3:22results of assigning the same task to
3:24each of the five port levels. One
3:28prompt asked me to create a music video
3:31based on the Seoul Subway Line 2 song.
3:34The other asked me to create a
3:36simulation game called "Gangnam Station
3:37Exploration." Let's take a look at each
3:40prompt. First, let's look at the prompt
3:43asking me to create a music video. Well
3:46, I had already generated the music
3:48file in advance using Labs. I created
3:51music like this. The lyrics are like
3:56this.
4:01>> Yes. I roughly generated music like
4:03this. So, I inserted a 30-note song
4:06about Seoul Subway Line 2 and included
4:08the lyrics file as well. Then, the
4:11prompt asked me to create the music
4:12video for this song from start to
4:14finish on my own. It instructed me to
4:15demonstrate how skilled a motion
4:17designer I am, avoid the typical AI
4:19video feel, choose a consistent style,
4:21and stick with it until the end.
4:23Requirements were also included below
4:24in this manner. Specifically, you can
4:26understand the request as asking for
4:27music videos to be generated solely
4:29through code, without using any
4:30external image or video generation
4:31models, stock sets, or APIs.
4:32Additionally, there was a request to
4:34create a Gangnam Station exploration
4:35simulation game. For this, the request
4:41was to create a simulation where the
4:43player acts as an office worker on
4:45their way home at 9 PM, slowly walking
4:48through the journey from the
4:50ground-level exit, passing through the
4:52turnstiles in the waiting room, and
4:54boarding the train at Platform 2. Here
4:57too, the request was to generate all
4:59forms, sign text, and sounds entirely
5:01through code, without using 3D models,
5:04textures, images, or sound files. So,
5:07you can view the requirements as being
5:09set in this way. However, actually,
5:11before these two tasks, a different
5:13assignment was originally given first,
5:15which involved requesting the creation
5:17of things like 2D survival games. Well,
5:19I had also included a task where you
5:22become a train driver and simulate
5:24subway operation. So, I initially
5:26submitted these two tasks at the port
5:28level. However, since the low-level
5:30port level performed both tasks
5:32exceptionally well, you can consider
5:33that I modified the tasks accordingly.
5:36Therefore, what I want to emphasize
5:38here is that the 5.5 model performs
5:40most tasks well even at the low-level
5:42port. While there are some differences
5:45in the details, this racing game was
5:47also made exceptionally well. You can
5:49view this as a game entirely created at
5:51the vertical level. From using boosts
5:55to drifting, everything was implemented
5:58perfectly. So, I will now show you the
6:01actual main task, starting with the
6:02first one. I requested that a music
6:05video be created for the AI-generated
6:07Line 2 music. You can consider it as
6:09having been implemented entirely
6:11through code, without using any
6:12external assets or anything like that.
6:15I requested the Chin Off S 5.5 model to
6:17create a music video by simply changing
6:19the port levels in this manner.
6:22Starting from here, let's take a look
6:24at the results when the same task was
6:27applied to the five port levels: Low,
6:29Medium, High, Xi, and Max. Passing the
6:35line, amidst the waves of people, oh,
6:37earphone volume gathers, today the door
6:39closes once again. Turning and turning,
6:46continuing, my life is a succubus,
6:48instead of Hannah, when the door opens,
6:51turning and turning...I can't help but
6:53feel that
6:57>> overall, the music video itself seems
6:59well-made, but differences emerge when
7:01looking at the details. I will show you
7:04them one by one. First, let's look at
7:07the Low port. Samsung, first, in the
7:11Low setting, it depicted the subway
7:14moving like this, and in "
7:18>> stuck in the waves of people," it seems
7:21to have simply expressed the rippling
7:24of the waves. And if you look here,
7:37there are some awkward parts because
7:38the blank space is quite large.
7:43>> The beat is also slightly off here.
7:45That was the Low Eport level; let's
7:47take a look at Max right away. It looks
7:49a bit more dynamic than before.
7:51>> It's like people without waves, people
7:53within waves; previously, it just
7:55looked like rippling waves, but now it
7:57actually creates human waves and
7:59expresses them like this. The earphone
8:02volume and
8:11>> subtitles are also coming through well
8:13without the beat lagging. When
8:21comparing the two, there is indeed a
8:22difference. So, in the Low Eport level,
8:25flat graphics were utilized a lot like
8:27this. However, in the Max Eport level,
8:31we can see that more three-dimensional
8:33graphics have been created. Also,
8:36converting lyrics into images like this
8:38is something that couldn't be done in
8:39Low E, but you can see that Max handled
8:41all of that. There are also incredibly
8:43detailed parts; if you look at the top
8:47right corner in Max, you can see the
8:49time ticking down. So, the workday
8:51starts at 7 AM and goes straight
8:53through to 8 AM. Then, when it is time
8:56to leave, the time changes again, like
8:59this: 8 AM, 9 AM, 11 PM, and so on. You
9:02can see that the departure times also
9:04change in this manner. Regarding the
9:07remaining Medium, High, and X-High
9:08levels, due to time constraints, I
9:10would recommend that you refer to the
9:12five comparison videos I showed you
9:13earlier. Overall, let's compare the
9:16time and cost involved in creating the
9:18same music video assignment for all
9:20five Port levels. Low Port took 12
9:22minutes, with an converted cost of
9:23approximately $ 1.45. Medium took 75
9:29minutes and cost $ 5.02. High took 46
9:32minutes, which was slightly less than
9:34Medium, and cost $ 7.55, resulting in a
9:35slightly higher cost. And X-High took
9:4054 minutes and cost $ 11.96. And at Max
9:43, the time suddenly jumped almost
9:46twofold, going from $ 30.35 to $ 35. So
9:50, as expected, you can view it as a
9:51structure where both time and cost
9:53increase as you move from Low to Medium
9:56, High, Xi, and Max. Actually, you can
9:59see that there was a significant
10:01difference in the details when moving
10:02toward Max. In this music video, I
10:05think we can say that the density of
10:07direction and imagination has increased
10:09. Next, we will try exploring Gangnam
10:12Station. First is the Port Level at Low
10:15. If you look, people are walking
10:17around, and the quality of the people's
10:20faces seems to be somewhat cylindrical.
10:23And the escalator seems to be working
10:27well. We came down to the main hall,
10:30and the lights above seem to be
10:32spreading out a bit white. The ticket
10:35gate is well-designed to prevent entry
10:37without swiping a card; usually, it
10:39opens when you tap your card, but there
10:41is something strange remaining in the
10:44middle. I passed through successfully,
10:45and now I have actually come down to
10:47the station. The screen doors look like
10:49this, and passengers are waiting. This
10:52is what it looks like when viewed from
10:53the end of the station. This was the
10:55Low Port Level, and now let's take a
10:57look at the Medium Port Level. In
10:59Medium, it seems to be implemented a
11:01bit darker, and unlike before, the
11:03buildings appear to have more detail.
11:07Also, the people aren't just
11:08cylindrical figures; this time they are
11:10wearing clothes and carrying bags, and
11:12there are bags like this one. However,
11:15it looks like the sparkling effect here
11:17wasn't handled very well, but the
11:19escalator is going down smoothly.
11:21Alright, let's hurry down. When I
11:23arrived at the main lobby, it was
11:24well-implemented with this kind of
11:26atmosphere, and there is a display like
11:28this here as well. Let's go check out
11:29the convenience store. Okay, the
11:31convenience store door is open like
11:33this. When you enter, there is a staff
11:35member,
11:36>> and it seems the items are sorted by
11:40color. I’ve come to the ticket gate,
11:43and it seems to be implemented
11:44similarly to the Low Port level earlier
11:46; you cannot pass through without
11:48swiping your card. If you swipe your
11:50card, you pass through smoothly, and
11:52the strange indentation in the middle
11:54is gone in Medium. Now, let's go down
11:57further to the platform. When we come
12:00down to the platform, it is implemented
12:02like this, and it does seem like the
12:04glare is slightly less than before.
12:07This is what it looks like when viewed
12:09from the end of the platform in Medium.
12:11Next, let's look at the High Port level
12:14. When we arrive at High, the night is
12:17implemented like this; it has become a
12:19bit softer, with shadows appearing, and
12:22you can see that the quality has
12:24improved. The clothing worn by the
12:26person also seems to have been
12:28well-executed this time. They are also
12:30wearing a bag like this. I’ll take
12:31the escalator down. The escalator works
12:33well, of course. When I got down to the
12:36main lobby, it looked like this. They
12:37are selling things like Deli Manju and
12:39walnut cookies, and the rain smudging
12:41is a bit severe here as well. The
12:43smudging is quite bad, but when I
12:45arrived at the convenience store,
12:47compared to before, there are some text
12:49written on it and the boxes are a bit
12:51more diverse. If you go to the ticket
12:53gate, tapping the transportation card
12:55to enter works perfectly. After taking
12:59the escalator down and arriving at the
13:01platform, it feels like there are a few
13:03more passengers now, doesn't it? Shall
13:06we try riding the train? We boarded the
13:08train like this, and you can see that
13:11things like the handrails are
13:13well-implemented. This was the High
13:16Eport level, and now for the fourth one
13:18, let's look at the X High Eport. In X
13:20High, the night is implemented like
13:22this. As for the outside environment, I
13:25personally felt it was a bit more
13:26natural in High. And this time, I’m
13:29going to try going down the stairs. It
13:36seems the passageway in the middle
13:37isn't implemented very well here,
13:38unlike in Hi X. However, walking around
13:39looking at your phone seems to be
13:40implemented this time. You can see
13:42someone looking at their phone here,
13:43too. You can also enter and view the
13:45convenience store; this time, it seems
13:47to be distinguished only by color,
13:49without any text. There is a flower
13:52shop, a walnut shop, and tapping the
13:54transportation card to enter works fine
13:56. Now, there is something floating in
13:59the middle of the escalator like this;
14:00that seems like an error, doesn't it?
14:02It shouldn't be set up like this, but
14:04it is trying to come in backwards. The
14:06man is a woman. And when you enter the
14:10platform, it is implemented like this,
14:11and when viewed from the end, it looks
14:13like this. Let's try boarding the train
14:16as well. When you board the train, the
14:20interior looks like this. You are
14:22supposed to be holding the handrail,
14:24but it is just holding onto a handrail
14:26in thin air. Yes. That was X Hi, and
14:28finally, let's take a look at the Port
14:30Level in Max. Max is definitely
14:33different when it comes to port-level
14:35rendering. The buildings and steel
14:38plates look very natural, and the
14:40movement of vehicles is also well
14:42implemented. People are all going down
14:45the escalator, and when I went down the
14:47stairs, the area that was open earlier
14:48is now well-implemented so it doesn't
14:50look like a hole, and even the light
14:52reflections are well-represented. When
14:55I came down to the waiting room and
14:57looked at the map, it seems to be
14:59rendered quite detailed and with high
15:01quality. Shall we try going into the
15:04park and the convenience store to track
15:06our current location? Oh, it looks like
15:09this convenience store has an automatic
15:11door; you open it like this to enter,
15:13and the products this time look like
15:15pictures. The addition of pictures has
15:18added a bit more detail. And if we go
15:20into the ticket gate, hmm, isn't this
15:22the kind of gate actually at Gangnam
15:24Station? It seems like they've even
15:26implemented the part where you tap your
15:28transportation card and pass through,
15:30and then rotate like this. When I came
15:32down to the platform, it had this kind
15:35of feel, and the route map was made
15:37very well. It seems like there are some
15:39disconnections, but anyway, it was
15:40implemented quite well and is fairly
15:41similar. Oh, the train has left, so I
15:43think I need to take the next one. Yes,
15:45I will try taking the train on the
15:48opposite side. Someone has arrived, so
15:50please get off for a moment and then
15:52get on. I will get on then. Yes. It
15:55feels a bit more pleasant now that I'm
15:57inside. There are handrails, people are
16:00standing, looking at their phones, and
16:02holding bags—this was the Ports level
16:04in Max. Overall, we have now examined
16:07both tasks. We looked at the music
16:09video and the Gangnam Station
16:11simulation, and now let's take a look
16:13at the cost when running five Ports
16:14levels for both. First, for Low Port,
16:17it cost $ 1.45, and for the Music Video
16:20, it was $ 1.45, $ 5.02, $ 7.55, and
16:23for X High, $ 11.96. You can see that
16:26the cost suddenly increases
16:27dramatically when moving to Max.
16:30Similarly, in the Gangnam Station
16:31Exploration, you can confirm that the
16:33cost suddenly increases rapidly from Xi
16:35to Max. Overall, looking solely at
16:39quality, I naturally thought Max was
16:41number one. This is the comparison
16:44table from the second assignment, the
16:46Gangnam Station assignment. As you can
16:48see, in Low, it took 25 minutes and
16:50cost $ 9 for $ 3.95; in Medium, it took
16:5261 minutes and cost $ 13; in High, it
16:54took 88 minutes and cost $ 16; and in
16:56XI, it took 171 minutes and cost $ 29.
16:59Starting with XI, you can consider that
17:01I ran the test twice, as I ran out of
17:03usage once, performed a reset, and then
17:04proceeded again. In Max, it took 173
17:08minutes, which is almost the same
17:10amount of time as Xi, but the cost was
17:12significantly higher at $ 101. However,
17:15if you look here, Max is a bit
17:16different. Regarding tool calls, Max
17:19made a very large number of calls,
17:21whereas the other subs did not make any
17:23calls at all. Since Max utilized two
17:26subs to run the process, it seems there
17:29was a difference in the approach taken
17:31to the task. Therefore, in my previous
17:34evaluation, I ranked Max as number one
17:36in all categories. However, that was my
17:39personal assessment, so to conduct a
17:41more blind evaluation, I will now show
17:43you the results of a blind evaluation
17:46performed by the GPT 6 Astra model. To
17:49start with the results, for music
17:51videos, Max ranked first, followed by
17:54High, Medium, Xi, and Low. Regarding
17:57Gangnam Station, Max took first place
17:59here as well. Max ranked first in both
18:02categories, even in this blind
18:02evaluation. You can view this as Astra
18:06evaluating it as X High, High, Medium,
18:07and Low. The overall chart for the
18:09music video is structured like this. I
18:11suggest pausing the video to take a
18:13look at this. Also, this is a note
18:14Astra left during the blind evaluation
18:16of the music and video: In "Low," the
18:17background screen continues between
18:19earphone lyric lines, and the door
18:20scene is long, resulting in relatively
18:22little direction based on meaning or
18:23variation in movement. In "Max," the
18:25strengths are the richest visual
18:27changes, consistent color schemes, and
18:29key lyrics that are easy to read. This
18:31is the result of the Gangnam Station
18:33exploration, and I suggest pausing to
18:34take a look at this as well. When
18:36looking at Astra's performance, "Low"
18:37had excessive indoor light flare; the
18:39indoor light flare was very strong. The
18:41simplistic portrayal of character
18:43distance and the repetitive audio
18:44filter correction warnings are
18:45disappointing. In Max, the density and
18:47material of the street, shop, and ad
18:49guide succession were the most
18:50persuasive, and the pedestrian tag
18:52boarding completion indicator was
18:53verified; however, the gap where the
18:58ground structure at the stair ceiling
18:58boundary was visible was regrettable.
18:59That is roughly how the evaluation was
19:00given. Here as well, Max took the top
19:03spot. So, looking at the graph overall,
19:07for the music video, Max took first
19:09place with $ 30, followed by High in
19:12second with 22 points, Medium in third,
19:15X High in fourth, and Low in third. For
19:19the Gangnam Station exploration, Max
19:21took first place, Xi in second, High in
19:24third, Medium in fourth, and Low in
19:26fifth. The rankings were taken
19:28sequentially in this manner. Overall,
19:31it seems we can say that as the port
19:33level increased, the quality of the
19:36final product also improved. Of course,
19:38since these results were obtained after
19:40running the task only once, a different
19:42outcome might be obtained if the task
19:44were run multiple times. In terms of
19:46output quality alone, it is true that
19:48Max ranked first in both cases. However
19:50, when considering costs as well, the
19:52difference in costs from Low to X High
19:55does not rise that drastically. If you
19:58look here, it rose to $ 11, and at
20:00Gangnam Station, it increased from $
20:033.95 to $ 29. But as it suddenly jumped
20:06to Max, the cost tripled, and here it
20:08jumped almost three or four times. You
20:11can view the point where costs
20:13increased rapidly between X High and
20:15Max as the point where costs surged in
20:17this manner. If you look at the total
20:20cost of the experiment here, two tasks
20:22were run ten times across five port
20:23levels. So, you can consider the
20:26converted cost to be around $ 220 after
20:28running it ten times. Finally, I would
20:31like to discuss which port level you
20:33should use. Overall, the functional
20:36implementations from Low to X were all
20:38done well. Therefore, if your priority
20:41is simply good functional
20:43implementation, in my opinion, Low or
20:45Medium levels are sufficient to cover
20:48all functions well. However, for those
20:51looking for cost-effectiveness, I would
20:53recommend the High port level. Since
20:56the scores for both assignments
20:58increased most significantly when
21:01moving to High, I would say that High
21:03offers the best value for money. As for
21:06the most ambiguous port level, X High
21:09seems to be the most difficult. There
21:11were quite a few cases where the
21:12results were inconsistent depending on
21:14the assignment. So, XI is a bit
21:15ambiguous. I would say that it is
21:18better to either invest a bit more in
21:20Max to drastically improve the quality
21:22of your output, or use High if you want
21:25to prioritize cost-effectiveness.
21:28However, since this measurement was
21:29performed as an internal benchmark for
21:31a specific assignment and I did not
21:33conduct multiple iterative tests,
21:35please consider this only as a
21:36reference. Because different results
21:39may occur when measuring different
21:40assignments in your own environment, I
21:42recommend that you try measuring it in
21:44your own work as well. If you have a
21:47different opinion regarding Ports Level
21:49, please share it in the comments; it
21:50will be helpful for many others to
21:52refer to. I will return in the next
21:55video with more useful information.

About this transcript

This page contains the full transcript of Opus 5.5 모든 effort로 동일한 과제 시켜봤습니다 by 개발동생, generated from the public captions YouTube serves with the video. The transcript has 3,836 words across 581 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.