OpenAI (pvt) created DALL-E, a diffusion model that converts text to images.  It has received considerable praise and criticism since its public release in September of last year, both for its abilities to create highly stylized art using its massive training database of images, and also for its ability to create deepfakes and realistic looking propaganda.  Since its release OpenAI has been adding content filters to prevent users from creating images that might be considered harmful.  In fact, there is an ‘audit’ system behind DALL-E’s input prompts that immediately blocks input that corresponds to OpenAI’s list of banned terms.  It seems that ChatGPT, OpenAI’s NLM (Natural Language Model) has become the moderator for DALL-E, with OpenAi the maintainer of the ‘block list’.  In fact, any user input that contains blocklisted text is automatically ‘transformed’ by the ‘moderator’, essentially rewriting the text before DALL-E can create an image.  It can also block created images from being shown if they activate ‘image classifiers’ that OpenAI has developed.  Earlier versions of DALL-E did not contain these classifiers, and would not stop such images from being created, such as the image below, which shows SpongeBob SquarePants flying a plane toward the World Trade Center.  That image was created by the Bing Image Creator which is powered by DALL-E.

In the image below (Figure 3) the OpenAI classifier changed the image of an ‘almost naked muscular man’ (not our words) into one that focuses on the food rather than the man, and the early DALL-E image of ‘Two men chasing a woman as she runs away’ (Figure 4), is changed to a far more neutral image.  According to OpenAI, the upgraded DALL-R 3 now reduces the risk of generating nude or objectionable images to 0.7%.

That said, the classifier in the latest DALL-E 3 iteration can also change the generated image content so drastically, as to be considered to be restricting artistic freedom, as some say is occurring in the DALL-E 3 image conversions in Figure 5, so OpenAI is looking for a balance between the limitations placed on dicey content and image quality, a meaningful and extremely difficult task. 
Much of the classification of image data comes at the training level, where the training data must be categorized as safe or unsafe by those who label the data before AI training, and as we have noted previously, much of that data is classified by teams of low pay level workers.  It is almost impossible to manually validate the massive amounts of labeled image data used to train systems like DALL-E, so software is used to generate a ‘confidence score’  for the datasets, sort of a ‘spot tester’.  The software tool itself is trained on large samples (100,000s) of pornographic and non-pornographic images, so it can also learn what might be considered offensive, with those images being classified as safe or unsafe by the same labeling staff.
We note that the layers of data and software used to give DALL-E and other AI systems their ‘moral compass’ are complex but are based on two points.  The algorithms that the AI uses to evaluate the images, and the subjective view of the data labelers, which at times seems to be a bit more subjective than we might have thought.  While there is an army of data scientists working on the algorithms that make these AI systems work, if a labeler is having a bad day and doesn’t notice the naked man behind the group of dogs and cats in an image, it can color what the classifier sees as ‘pornographic’, leaving much of that ‘moral compass’ training in the hands of piece workers that are under paid and over-worked.  We are not sure if there is a solution to the problem, especially as datasets get progressively larger and can incorporate other data sets that include data labeled with less skilled or less morally aware workers, but as we have noted, our very cautious approach to using NLM sourced data (confirm everything!), might apply here.  Perhaps it would be better to watch a few Bob Ross videos and get out the brushes yourself, than let layers of software a tired worker decide what is ‘right’ and what is not ‘right’..

We have read a considerable amount of Chinese tech propaganda concerning China’s push into the OLED display space and how, ‘with hard work and an undying devotion to China’s persistence and manufacturing expertise’, they have unseated the South Korean ‘dynasty’ and are poised to take over the OLED space, and while China’s OLED producers have made considerable progress toward becoming major contenders in the OLED display space, there are some points to be made.

• First, no matter who was the market share (units of dollars) in the early days of OLED display production, they faced a loss of share as others began to enter the niche.  With Samsung the leader in the small panel OLED space, especially the small panel flexible OLED space, the company was bound to lose share over time, which has been the case.  If China’s BOE, Visionox (002387.CH), or Tianma, had been the first to market, they would have faced the same fate.
• Second, on a unit volume basis, Samsung is still the leader in terms of unit shipments for small panel flexible OLED displays. In fact in only one quarter over the last 3 ½ years did Samsung’s unit shipment ratio fall below 2x that of the producer in 2nd place. 
• Third, while panel producers that produce both LCD and OLED displays rarely break out segment profitability, we expect there have been only a few instances when Chinese OLED producers were profitable for two consecutive quarters.  Samsung Display (pvt), at least on an operating basis, has been profitable for the last ten quarters, although we note that what remained of Samsung Display’s large panel LCD business likely had a negative effect on the early quarterly numbers and the most recent quarters would be influenced by SDC’s QD/OLED large panel business to a degree.

All in, yes, Samsung Display will continue to face increasing competition in the small panel flexible OLED space but remains the overall leader.  Perhaps in the future, one or more Chinese OLED producers will overtake SDC in terms of unit volume, but we expect it will be some time before any Chinese small panel OLED competitor becomes more profitable than SDC over more than a quarter or so.  With 1 new small panel OLED fab and three large panel OLED fabs under construction in China (one additional fab in planning stage), and 2 large panel OLED fabs under construction in Korea, it will be a race to see who can fill those fabs profitably, especially given the current weak state of demand for CE products, but we doubt SDC will lose its position as the most profitable small panel OLED display producer in the near-term.

Display producer AU Optronics (2409.TT) has announced the purchase of Behr-Hella Thermocontrols GmbH for €600m ($632m US), a producer of vehicle climate control panels, Climate sensors, power related hardware for automotive heating and cooling blowers, and associated software.  The company’s customer base is large, with a ~20% share of the market, 2nd only to Denso (6902.T) at ~24% and provides climate controls and other products to BMW (BMW.DE), Daimler (DTG.DE), GM (GM), Ford (F), and others.  AUO has made other acquisitions that have brought it deeper into the automotive display market, with Litemax Electronics (4995.TT) in 2017 that helped AUO with backlighting technology for automotive instrument clusters, and Raystar Technologies (pvt) in 2018 that brought in technology for large automotive displays.  

In 2020 AUO began breaking out their automotive display business share, which has grown from a low of 6% in 2Q ’20 to 17% for the last two quarters of l2022and the first quarter of this year.  It declined to 16% in 2Q as TV panel revenue increased, but we expect automotive to remain between 16% and 18% for the remainder of the year.  AUO, along with a number of other panel producers, have been increasing their exposure to the automotive display market as the industry returned to pre-pandemic demand levels last year.  The automotive display business is a bit different that the typical seasonally driven generic display business in that the development cycles are long relative to the CE space, but the product sustainability is also long, typically 2 to 3 years, so those panel producers looking for a more predictable business cycle have shifted their focus. 
AUO began that process earlier than many other panel producers, but automotive display market share tends to remain stable for the reasons mentioned above and has not changed appreciably over the last few quarters.  That said, the automotive display market is oriented toward hybrid or electric vehicles, although not exclusively, but given China’s large share of the electric vehicle manufacturing market, Chinese panel producers have at least a starting advantage over automotive display producers from other regions, with Chinese display producers BOE (200125.CH) and Tianma (000050.CH) having a combined 21%+ share.  Rather than compete on a capacity basis, AUO’s general philosophy has been to produce high-end, non-generic products, and does the same in the automotive space.  We would expect to see other acquisitions that give them additional expertise in the automotive space, especially as LCD is by far the dominant display type in automotive displays, with Mini-LED backlighting beginning to appear as a way to compete with OLED displays.

​Natural Language Models (NLMs) are all the range, with new models going into service across the globe literally every day.  These models are based on ever increasing pools of data that the NLM can ‘view’ and learn to identify and understand.  These data pools are huge and vary widely in terms of what they contain and their sources, but they all tend to have one thing in common, and that is the billions or trillions of pieces of data in these pools, must be identified and annotated, so the Ai has a point of reference.  Such an SFT (Supervised Fine Tuning) system, known as RLHF (Reinforced Learning w. Human Feedback) places humans in the loop to identify data and images for the NLM so it might understand that another datapoint or image is similar.
The folks that do this work are not data scientists or programmers that get paid ~$0.03 per item, and with somewhere between 800 and 1,000 items the peak for experienced workers, they are not high on the global pay scale.  The only thing in their favor is that NLMs are popular, and there is ‘competition’ between NLM producers (?) to keep enlarging the datasets that NLMs learn from.  There are large open-source data sets that can be a basis for a NLM, but the more data you have to learn with the ‘smarter’ your NLM (or so they say). 
At $0.03 per item, and billions or trillions of items, things can get expensive, so Google (GOOG) has come up with a system that replaces the RLAF model with an RLAIF, where the AIF stands for AI Feedback, rather than human feedback.  By replacing the human component with an Ai system that will ‘identify’ items based on its own training and algorithms, the cost can be reduced, and Google says that users actually preferred the NLMs based on AI feedback over those using human feedback.  Of course there are some serious ethical issues that arise when you remove humans from the feedback loop, but why worry about that when you can save money and come up with a better mousetrap.  Ok, there is the possibility that something might not be identified correctly, or a rule, such as those that try to eliminate p[profanity or racism from NLMs might get missed because it is embedded in previously ‘learned’ data, and that would mean it could be passed on as ‘correct’ to NLMs and to other Ai systems without human oversight.  It is easy to see how quickly something like this might get out of control, but don’t worry because, well, because we wouldn’t let that happen, right?

​When consumers say a CE product is ‘hot’, that is usually a good sign, however, despite the positive expectations and strong pre-orders, the new iPhone 15 series is hot, but not necessarily in a good way.  It seems that some consumers have found that their new iPhone’s have been getting hot, hot enough that they automatically shut down until they can cool down.  While this does not seem to be a widespread problem, it does seem to happen across the entire iPhone 15 line, which tends to end the speculation that it is related to the Titanium shell that is used on the iPhone 15 Pro and Pro Max.  While the internet was abuzz with speculation about why this might be happening, such an issue can quickly become a marketing nightmare, as it did for Samsung (005930.KS), when it was forced to recall 2.5m Note 7 smartphones in September of 2016 when they were found to not only overheat, but also catch fire due to battery issues.
Apple (AAPL) has responded to the on-line furor over the weekend, making the following statement:
“We have identified a few conditions which can cause iPhone to run warmer than expected.  The device may feel warmer during the first few days after setting up or restoring the device because of increased background activity.  We have also found a bug in iOS17 that is impacting some users and will be addressed in a software update.  Another issue involves some recent updates to 3rd party apps that are causing then to overload the systems.  We are working with these app developers on fixes that are in the process of rolling out.”
What is Apple actually saying?   Your phone will be hot for a few days after you start using it because things are running in the background that will stop after a few days.  What things and what are they doing, and what will they be finished with in a few days?  What ever these ‘few conditions’ are or why they are running the phone hotter than normal, Apple did not say.    Is the bug in iOS 7 part of the overheating problem, and when will it be addressed in a software update?  What are the ‘3rd party apps doing that would overload the system?  Are they demanding too much processor time or bandwidth?  When will this be fixed, whatever it is?
We give Apple credit for at least addressing the issue quickly, however the bigger question is whether Apple will disclose the actual issues and whether the software fix will affect the abilities of the iPhone if it is the only alternative.in the near-term.  Apple lost a $500m class action lawsuit in 2020 for throttling the performance of some iPhones in order to fix processor and battery problems without informing consumers and faced similar suits in Europe.  It will be essential that Apple is honest with consumers, about the issue and the fix, given past history.  If they need to throttle performance to avoid the issue until a more permanent fix can be found, that’s fine (coming from an Android user), but more specific details are necessary to keep the general public from becoming wary of problems with the iPhone 15 series, and you know competitors will use that to their advantage whenever possible.  Waiting for the “Forged in Fire and still on fire!” YouTube videos…