Recently, natural language processing by artificial intelligence has evolved explosively, so I have summarized it. [Part 2]

Chapters 1 through 3 mainly introduced general-purpose language models. A general-purpose language model is a versatile language model that supports various language processing tasks such as summarization, translation, document classification, and question answering; BERT, T5, GPT-3, and PaLM are all general-purpose language models. Among the language models introduced so far, only Meta's NLLB-200 is a single-function specialized language model specialized for translation functions.
From Chapter 4 onwards, I will mainly introduce these single-function specialized language models.

4. Image generation from text

First, among the single-function specialized language models, I will explain DALL-E, which was announced by OpenAI in 2021, and the subsequent image generation from text which has been gaining excitement with new models being announced one after another.

① Four types of image generation models

First, not limited to image generation from text, generally speaking, image generation models that automatically generate images using deep learning can be broadly divided into the following four types of models.

(1) Variational Autoencoder (VAE)
VAE is a type of autoencoder with a mechanism that compresses input data into features and then restores those features back to the original data. By introducing a probability distribution to these features, it is a model that can stochastically generate unknown data.

(2) Generative Adversarial Network (GAN)
GAN is composed of two networks: a generator and a discriminator. When generating images, the generator outputs images that look just like the real thing based on training data, and the discriminator determines whether the image is real or fake. The generator learns to deceive the discriminator, and the discriminator learns to judge more accurately. In this way, GAN is a model where two networks compete for conflicting goals, thereby improving the accuracy of the generated images.

(3) Flow-based generative model
A flow-based generative model is a model that utilizes a technique called normalizing flow to explicitly model probability distributions, thereby enabling it to generate new samples based on complex distributions.

(4) Diffusion model
A diffusion model is a model that reverses the process of gradually adding noise to original data until it becomes complete noise (Gaussian noise), and models the process of restoring data by gradually removing noise, using it to generate new data. Diffusion models have recently attracted attention for their
training stability and high quality of generated images, and are adopted in the latest image generation models such as DALL-E 2 and Imagen.

② Image generation by GAN

Before the release of DALL-E, the most commonly used image generation models for creating images from text were those utilizing GANs.
GAN is a deep generative model announced in 2014 by Ian Goodfellow and his colleagues, which developed significantly, especially in the field of image generation, and surprised the world by, for example, automatically generating images of non-existent human faces that are indistinguishable from real photographs.
The following image is an example of a non-existent human face generated by "This person does not exist", which is published on the site StyleGAN. (Refreshing the page will switch to a new image.)

Since many GANs that generate images from text have been developed, I will introduce some of them.

(1) GAN-INT-CLS (2016)
Announced in 2016, GAN-INT-CLS is an image generation model that used GANs for the task of generating images from text for the first time. Although it could generate 64x64 pixel images, the output image data had an overwhelmingly larger amount of information compared to the input text data, which led to the challenge that it could not generate high-resolution images due to the limitations of the generator's learning capacity.

(2) StackGAN (2017)
Announced in 2017, StackGAN is a model that enabled high-resolution image generation by using a two-stage GAN structure. It is designed to generate a low-resolution 64x64 pixel image based on the text in the first stage, and then generate a higher-resolution 256x256 pixel image in the second stage.

(3) AttnGAN (2017)
Microsoft announced AttnGAN in 2017, which uses the Attention mechanism, a system that performs efficient machine learning by focusing on important words, to utilize information down to the word level, thereby enabling the generation of images that follow text instructions more faithfully than ever before.

(3) From OpenAI's DALL-E to DALL-E 2

(1) DALL-E (January 2021)
DALL-E is an image generation model announced by OpenAI in January 2021 that generates images from text. The mechanism of DALL-E's image generation is as follows.
[During training]
256×256 pixel image data for training is compressed into 32×32 feature quantities by VQ-VAE and paired with caption text data describing the image to create training samples. Then, using these samples, machine learning is performed to associate the feature quantities of the image data with the text data using a 12 billion parameter model of GPT-3.
250 million pairs of samples were prepared for DALL-E's training. Also, VQ-VAE is a type of VAE that represents feature quantities with a finite number of vectors.
[During image generation]
The input text data is passed to the trained GPT-3 to generate image data feature quantities that match the content of the text data. Then, using VQ-VAE, these feature quantities are restored into an image.
DALL-E can combine multiple elements within text to render unprecedented images with high precision, as in the following example.

(2) CLIP (January 2021)
OpenAI announced CLIP at the same time as DALL-E. It is an image classification model that allows users to freely set classification categories and outputs the similarity of images to text. CLIP was trained on 400 million image-text pairs collected from websites on the internet and can classify even unseen datasets well.
CLIP has been applied to various tasks in fields related to computer vision as a highly capable multimodal model. Also, the text-to-image generation model combining CLIP and VQGAN announced by Ryan Moulton in July 2021 became a hot topic for producing many very high-quality artistic images.

(3) GLIDE (December 2021)
Developed by OpenAI in December 2021, GLIDE (Guided Language-to-Image Diffusion for Generation and Editing) is an image generation model that adopts a diffusion model with up to 3.5 billion parameters, capable of generating high-resolution images that follow complex text instructions more faithfully than DALL-E. Additionally, by masking a part of an image and providing text instructions, you can generate an image that modifies that part according to the instructions.

(4) DALL-E 2 (April 2022)
In April 2022, OpenAI announced DALL-E 2, an image generation model that combines CLIP and a diffusion model. DALL-E 2 DALL-E 2 is a two-stage model that uses a pre-trained model to
generate CLIP image features from input text, and then generates high-resolution images from these features using a decoder based on a diffusion model. Note that CLIP image features are image data compressed into features when classifying images with CLIP; in the case of DALL-E 2, it uses the reverse process to generate image data from these image features. This method achieves both diversity and high resolution in the output images. Also, like GLIDE, you can
specify and modify parts of photos or images
.

Since last month (July 2022), the DALL-E 2 beta version has been released, and it has become available to users in the order they registered on the waitlist and received an invitation email. One credit allows for the generation of 4 images; 50 credits are granted for free in the first month, and 15 credits are granted for free every month from the second month onwards. Any usage beyond that is paid. You can register for the waitlist from the following page , so please give it a try.

In my case, I received an invitation email in less than a week and was able to start using DALL-E 2. The image below is one I actually had DALL-E 2 create. It seems very useful as I can easily create images for presentation covers or illustrations. By the way, the input words were "Photo-style image of artificial intelligence at the moment of consciousness generation."

④ Google's Imagen and Parti

(1) Imagen (May 2022)
In May 2022, the Google Brain team announced Imagen, an image generation model that creates images from text. Imagen showed
MS COCO (Microsoft Common Objects in Context) and other image datasets in benchmark tests, demonstrating higher performance than DALL-E 2 and other image generation models.

Imagen generates images through the following three processes.

1. A text encoder using a pre-trained giant language model(in this case, T5) is used to convert the input text into feature vectors.

2. A text-to-image model using a U-Net style diffusion modelis used to generate a 64x64 pixel low-resolution image from the feature vectorsobtained in step 1.

3. Furthermore, by passing the U-Net style diffusion model twice, the 64x64 pixel low-resolution image is upsampled to a 1024x1024 pixel high-resolution image.

The Google Brain team argues that to improve image generation performance, increasing the number of parameters in the language model is more effective than increasing the size of the diffusion model that converts text to images.
Also, for the diffusion model used for the final stage of resolution upscaling, they developed an Efficient U-Net that is more efficient than the conventional U-Net.

Furthermore, Google states that at this time, Imagen is not suitable for public use and will not be released to the general public. This is because the generated images are so natural and high-quality that they are indistinguishable from real photographs, creating a risk of misuse for fake news or revenge porn, and because it uses data collected from uncurated websites, it may contain ethically problematic data such as discriminatory claims or violent expressions, and
the problem of bias has not been solved. Google claims that it will continue to investigate social and cultural biases and address these issues.

(2) Parti (June 2022)
The month after announcing Imagen (June 2022), Google Research announced yet another new image generation model. The new image generation model is called Parti (Pathways Autoregressive Text-to-Image model), and this time it adopts an autoregressive model instead of a diffusion model. An autoregressive model is a mathematical model where the model's output at a certain point in time depends on the model's previous outputs, and
giant language models like GPT-3 also use Transformers to become autoregressive models that predict the next word after a given word. Also, Parti can

generate images that faithfully follow even long and complex text instructions and depict photorealistic images. Parti's 20-billion parameter model showed performance exceeding DALL-E 2 and Imagen
in benchmark tests using MS COCO, and achieved even higher performance in the fine-tuned version.

Parti consists of a two-stage model of an autoregressive model and an image tokenizer.
For the autoregressive model, both the encoder and decoder of a Transformer model with up to 20 billion parameters are used. By scaling up this autoregressive model, image quality is improved.
For the image tokenizer, ViT-VQGAN is used to perform pre-training to convert sample images for learning into image tokens and restore the original images from the image tokens. Note that ViT-VQGAN is a version of the image generation model that combines CLIP and VQGAN introduced in (2) of ③, where the CLIP part is replaced by the Vision Transformer (ViT), which is an image recognition model.

To generate images from text, first, the input text is converted into a sequence (array) of text tokens using the pre-trained text encoder of the autoregressive model.
Next, the sequence of text tokens is converted into a sequence of image tokens using this autoregressive model.
Finally, the output image is generated from the sequence of image tokens using the image tokenizer.
With the emergence of DALL-E 2 and Imagen, it was thought that diffusion models would become the mainstream for future image generation models, but the feature of Parti is that, despite being a simple autoregressive model, it achieves the highest quality image generation that surpasses these diffusion models by increasing parameters and scaling up the model.

5. Multimodal AI

Multimodal AI refers to AI models that can process multiple types of data combined, such as images, videos, and audio, in addition to text information. Among the general-purpose language models introduced so far, Google's Pathways and China's WuDao 2.0 and M6 are multimodal AIs that can process image and audio data in addition to text. Also, the image generation models that generate images from text introduced in Chapter 4 are also multimodal AIs that handle text and images.

As multimodal AI develops, learning from images and videos, which contain more information than text, will become possible, further advancing artificial intelligence learning and potentially bringing us closer to the realization of artificial general intelligence.
From this perspective, Google's DeepMind team, whose ultimate goal is the development of artificial general intelligence, is particularly actively working on the development of multimodal AI.

① Flamingo (April 2022)

In April 2022, the DeepMind team announced Flamingo, a multimodal AI that can understand text, images, and video simultaneously by combining them with just a single machine learning model.Flamingo was announced.
Flamingo was developed by DeepMind in April 2022 using Chinchilla, a language model with 70 billion parameters, and pre-training it by adding visual learning elements such as images and videos, as a new visual language model with 80 billion parameters.

Flamingo can perform tasks that combine images and videos with text in addition to general natural language processing tasks such as question answering and text generation. Also, when you want Flamingo to perform a new task, it can do so after being additionally trained with just a few examples of the task.
For example, if you train Flamingo with a few combinations of animal images and text explaining the animal's name and habitat, and then give it an image of a flamingo, it will output the text, "This is a flamingo. It can be found in the Caribbean region and South America."
Flamingo has recorded results in many benchmark tests for image and video understanding, such as visual question answering and captioning for images and videos, where even with few-shot learning, which only shows a few task examples, it outperforms the performance of models fine-tuned with thousands of times more task-specific data.

② Gato (May 2022)

The DeepMind team's rapid progress continues. Gato, which the team announced in May 2022, is a multifunctional multimodal AI that can perform not only text and image output but also various actions.
Specifically, Gato has the ability to perform a total of 604 types of tasks, such as playing video games, captioning images, chatting, and controlling a robotic arm to stack blocks, all with a single machine learning model.

Gato adopts the Transformer model, similar to machine learning models such as PaLM and Flamingo.
As for the specific mechanism, first, data for different tasks, such as video game image data, chat strings, and robotic arm movements, are each divided into tokens (the smallest units such as words or motion vectors) and split.
Then, these tokens are learned by a single Transformer neural network without being distinguished.
Finally, this learning result is used to predict target text, actions, etc..

Although Gato's processing ability for individual tasks is often inferior to dedicated programs, the developers of Gato aim to create a highly versatile artificial intelligence program, and they believe that if computing power is increased in the future, the shortcomings can be compensated for.

6. Automatic Programming

Before the rise of large-scale language models like GPT-3, most automatic programming technology development used proprietary programming languages with simple syntax to generate code, and it was considered difficult to target general programming languages such as Python or C.
Also, in recent years, app development methods called low-code development or no-code development, which allow even those without much programming knowledge to develop simple apps by combining prepared parts using mouse operations, have begun to appear, but it has been said that there are issues with the freedom and extensibility of app development.

① Automatic code generation by GPT-3

The GPT-3 large language model announced by OpenAI in May 2020 can perform natural language processing tasks such as text generation, translation, and dialogue, but beyond that, it can also automatically generate code from everyday language (natural language).
Specifically, it can automatically generate code in common programming languages like Python or C in response to instructions from users in natural language. It is also possible to have GPT-3 write the rest of the code that a user has started writing, or have GPT-3 explain the content of the code.
The emergence of GPT-3, which can automatically generate code from natural language, has become something that greatly changes the history of automatic programming to date.

② GitHub Copilot (June 2021)

In June 2021, GitHub, a software development platform under Microsoft, cooperated with OpenAI to release a programming support tool that automatically completes the rest of the code that the user has started entering called GitHub Copilot.
Initially, it was provided for free as a technical preview version, but from June 2022, it has been provided as an official service for a fee (with exceptions such as students being free).

GitHub Copilot is compatible with Microsoft's code editor VSCode, Visual Studio, JetBrains products, and more.
When a user enters code halfway through in a development tool, GitHub Copilot suggests code candidates that follow the context of that code and coding conventions. Also, when a user enters a comment, it similarly suggests code candidates that match the comment. And the user can choose whether or not to accept the suggestion.
The supported programming languages number in the dozens, and it seems that highly accurate inference results can be obtained, especially with languages like Python, JavaScript, and Ruby.
According to GitHub, on average, more than 27% of the content of code files is generated by GitHub Copilot, and it reached about 40% for Python and others.

GitHub Copilot uses OpenAI Codex, an automatic code generation system developed by OpenAI based on GPT-3.

③ OpenAI Codex (August 2021)

In August 2021, OpenAI began providing a beta API for OpenAI Codex, an AI system that automatically generates code from text such as English.
OpenAI Codex serves as the foundational system for GitHub Copilot to automatically complete code, and it supports more than 10 programming languages, including Python.
With OpenAI Codex, if you enter the command you want to execute in plain text, it will automatically generate the corresponding code. OpenAI has released demo videos of creating simple games, analyzing datasets, converting programming languages from Python to Ruby, and creating answers to math problems.

OpenAI Codex was based on the pre-trained GPT-3 and was trained on billions of lines of Python code collected from public GitHub repositories (storage locations for files, etc.). Note that OpenAI Codex has 12 billion parameters.
The OpenAI research team developed a new benchmark test called HumanEval, consisting of 164 programming problems, to evaluate OpenAI Codex. The evaluation results show that 28.7% were solved correctly on the first try, and 77.5% of problems had a correct solution within 100 attempts.

To try the beta version of the OpenAI Codex API, please register for the WAITLIST on the following OpenAI sitefrom here. Once you receive an invitation email, you can use it for free.

④ AlphaCode (February 2022)

In February 2022, Google's DeepMind team developed an AI system that automatically generates code called AlphaCode, and announced that it achieved results equivalent to the top 54.3% in competitive programming contests.
AlphaCode is a model that improves upon OpenAI Codex developed by OpenAI, enhancing its accuracy until it became capable of creating competitive programming code.
AlphaCode has up to 41.4 billion parameters, which is about four times the size of OpenAI Codex.

The specific mechanism of AlphaCode is as follows.
First, it uses a large amount of code published on Github to have the AlphaCode model perform pre-training, and then it is fine-tuned on datasets such as problems actually presented in competitive programming contests.
The fine-tuned AlphaCode model creates a large number of candidate codes that serve as answers to a given problem, and by filtering those candidates to select promising code, it achieves high-precision programming.
Since competitive programming problems are highly difficult, an accurate comparison cannot be made, but it is considered that the accuracy is significantly improved over OpenAI Codex.

You can see how AlphaCode actually writes programs on the DeepMind site below. To start creating a program, select a problem from the drop-down list and press the "play" button.
○ AlphaCode Attention Visualization

Microsoft's OpenAI Codex is offered as a paid practical programming support tool, GitHub Copilot, and has successfully achieved monetization.
In contrast, DeepMind's AlphaCode, while showing higher performance than that, appears to be developed as part of research and development toward artificial general intelligence with little thought given to monetization, which seems to reflect the characteristics of each company and organization.

7. Summary of the text

　There are broadly two approaches to text summarization: extractive summarization and abstractive summarization.
　Extractive summarization is a method that generates a summary by extracting the most important parts that represent the content of the text.
　Abstractive summarization is a method that generates an appropriate summary after grasping the meaning of the entire text (abstracting it), just as a person would when creating a summary.

① Extractive summarization

　The algorithm for encoder extractive summarization first solves a classification problem for each part of the text to determine whether it is important or not, and then generates a summary by arranging only the important parts.
　The advantage of extractive summarization is that, because it extracts directly from the original text, the content and expression of the sentences are less likely to break down.
　On the other hand, the disadvantages include the fact that information from parts that were not extracted is not included at all, so important information may be completely left out, and because important conclusions are often written at the end of a text, there is a position bias problem where it is easy to judge which parts to extract based solely on their position in the text.
　Models for extractive summarization include BERT-based BERTSum and DiscoBERT, which use pre-trained BERT to process classification tasksBERTSum, DiscoBERT, and graph-based TextRank.

② Abstractive summarization

　Since abstractive summarization captures the meaning of the entire text to generate a summary, it cannot be solved as a classification problem like extractive summarization.
　To take the abstractive summarization approach, it is necessary to hold information about words not in the input text and appropriately reflect them in the output, and what made this possible are neural language models such as encoder-decoder models.
　Encoder-decoder models convert the input text into feature values representing its meaning, and generate a summary based on those feature values. Therefore, even information not included in the input text can be reflected in the summary if it is in the training data.

　If successful, abstractive summarization can generate a natural summary that skillfully summarizes the entire text, but there was a problem that the quality of the summary tended to be unstable because unnecessary information unrelated to the input text is often mixed in, resulting in inaccurate summaries.
　However, in recent years, many methods have been proposed to improve the quality of abstractive summarization by training large-scale language models with excellent expressive power, and the number of models capable of generating natural summaries is increasing.
　Models for abstractive summarization include T5, BART, PEGASUS, and they are basically all encoder-decoder models.

③ BERTSum (March 2019)

BERTSum was announced in March 2019 as the first extractive summarization model to use BERT as a pre-trained model.
In BERTSum, pre-trained BERT is used to divide the input text into words and phrases, and then score the importance of each word or phrase within the text. Then, the information regarding the importance of words and phrases output by BERT is input into a summarization layer to generate a summary. The summarization layer uses a stack of Transformer layers.
Note that the extractive summarization model BERTSum is also called BERTSumEXT, and there is also an abstractive summarization model called BERTSumABS.

④ BART (October 2019)

BART (Bidirectional Auto-Regressive Transformer) is a Meta (formerly Facebook) announcement from October 2019, a bidirectional encoder BERT and a Seq2Seq (Sequence-to-Sequence) type general-purpose language model combining an autoregressive decoder like GPT-3, which can handle various natural language processing tasks such as machine translation, text generation, and document classification in addition to abstractive summarization. In the case of text summarization,
the input text is converted into features by BERT, and the autoregressive decoder generates a summary from these features. Since BART demonstrated high performance, especially in text summarization, after the advent of BART,
abstractive summarization models using encoder-decoder models became mainstream.

⑤ PEGASUS (December 2019)

Google Brain team announced the abstractive summarization model PEGASUS (Pre-training with Extracted Gap-sentences for Abstractive Summarization) in December 2019. Like T5 and BART, PEGASUS is also a Transformer-based encoder-decoder model.
In language models like BERT, a task called MLM (Masked Language Model) to predict masked words in a sentence is usually used during pre-training, but in the case of PEGASUS, for text summarization, in order to be able to extract features at the sentence level, in addition to MLM, it is characterized by the addition of a task called GSG (Gap Sentence Generation) to predict masked sentences at the sentence level.
Also, efforts have been made to reduce the amount of data required for fine-tuning, and it is possible to generate summaries with accuracy comparable to other methods with relatively little training data.
Note that PEGASUS technology has already been adopted for the automatic summarization feature in Google Docs.

⑥ Japanese text summarization services

there are few public releases of text summarization models, although there are several websites where you can try out Japanese text summarization for free. Below, I have listed the main web services where you can try out Japanese text summarization.

• Sanbun Tanteki(BuzzGraph) You can input text and URLs, and it summarizes them into three sentences.

• Automatic Summarization Tool(User Local) You can choose summaries of 3, 5, or 10 lines, and it can also highlight important parts.

• ELYZA DIGEST(ELYZA) You can input text and URLs, and it summarizes them into three lines. ELYZA is an AI startup originating from the Matsuo Lab at the University of Tokyo.

Each company does not disclose what language models or technologies they use, but ELYZA states that they have developed the large-scale language AI specialized for Japanese, "ELYZA Brain".

8. Other single-function specialized language models

1. MUM (Search Algorithm)

Google introduced BERT technology into its Google Search in 2019, but at the developer conference "Google I/O 2021" in May 2021, they announced MUM (Multitask Unified Model), a new search algorithm that is the next step after BERT. MUM also uses the Transformer model like BERT, but it is said to be
1,000 times more powerful than BERT. Also, MUM has been trained in 75 languages, including Japanese. The features of MUM are
(1) Enhanced natural language search, (2) Multilingual support, (3) Multimodal support.

(1) Enhanced natural language search
MUM is leveraging the results of recent natural language processing technology to aim to be a helpful search engine that understands the user's intent from questions like everyday conversation and returns the desired answer, without the user having to craft search words. For example, for a complex and ambiguous question like "I've climbed Mount Adams before, but I want to climb Mount Fuji next autumn. How should I prepare differently than before?", MUM understands the intent of the question, grasps things like "comparing Mount Adams and Mount Fuji" and "climbing preparation requires training and purchasing appropriate equipment," and can accurately present the information the questioner needs, such as "Mount Adams and Mount Fuji are about the same altitude," "Mount Fuji is prone to rain in the autumn," and "a waterproof jacket is necessary." MUM also has a feature to present related information that the user might need. For example,


"Things to consider" is a feature that presents related information that is good to know all at once. Also, a "Refine this search" feature that narrows down search results to show more in-depth results, and a "Broaden this search" feature that expands search words to lead to a wider range of search results, will also be implemented.

(2) Multilingual support
The answers or information to a search query are not necessarily on websites written in the language used for the search. In such cases, MUM can gather information from sites in other languages and present answers close to what the user is looking for.

(3) Support for multimodal
MUM is capable of multimodal search that combines text with images and videos, not just text.
For example, by linking with Google Lens, you can take a picture of a broken bicycle part and search for "how to repair it" to find blogs, articles, and explanatory videos about repairs. In addition, they plan to introduce MUM to enable searching for content that appears within video content such as YouTube.

If we can search using everyday language without having to think about keywords, I think the way we use Google Search will change. In that case, since the number of characters entered will increase, it would be better to be able to search by voice. In any case, Google Search is likely to become increasingly convenient and something we cannot live without.

② LaMDA (Conversational AI)

Most general-purpose language models like T5, GPT-3, and PaLM can also handle dialogue tasks. Also, the development of artificial intelligence models specialized for conversational functions such as chatbots has been carried out for some time.
However, looking at the published dialogue records of Google's conversational artificial intelligence LaMDA, which caused a stir in June 2022 when people asked "has artificial intelligence gained consciousness or emotions?", it feels like it is on a different level from previous dialogue systems. As if the artificial intelligence has a personality, one can perceive a deep understanding of the conversation content and its background circumstances, high-level judgment and insight, and logical consistency.

LaMDA (Language Model for Dialogue Applications) is a language model for natural language processing specialized for dialogue announced by Google in May 2021.
LaMDA adopts the Transformer model and first performs pre-training on a massive amount of text such as 1.56 trillion words of public dialogue data and public web documents using up to 137 billion parameters.
Next, it performs fine-tuning to train a "generator" that creates response candidates and a "classifier" that evaluates the safety and conversation quality of the response candidates using response data annotated by humans.
The flow of LaMDA's processing when interacting with a user is as follows.

1. The generator creates multiple response candidates corresponding to the user's immediately preceding statement.

2. All response candidates are evaluated by the classifier, and a safety score and a conversation quality score are calculated. Safety means eliminating harmful results such as violent expressions or discriminatory expressions, and conversation quality is evaluated based on three points: Sensibleness (being thoughtful and having common sense), Specificity (being specific), and Interestingness (being interesting).

3. First,exclude response candidates with low safety scores, and rank the remaining response candidates by conversation quality score.

4. Output the highest-ranked response candidate as the final response.

Furthermore, by using LaMDA's classifier to filter out training data with low safety or low conversation quality, the accuracy of creating high-quality and safe responses is improved.

In May 2022, Google also announced a new version of LaMDA, LaMDA 2, and announced that it would soon provide AI Test Kitchen, an app for smartphones equipped with LaMDA 2. AI Test Kitchen includes features such as answering what might happen in a fictional place entered by the user, automatically creating to-do lists, and answering questions about specific topics.

Google is quite cautious about the issue of artificial intelligence bias, and currently does not release LaMDA for free interaction, but I would like to verify as soon as possible whether the dialogue records published by Google engineers in June 2022 are genuine. If they are genuine, it will be an important technology that will significantly change the history of artificial intelligence development.

③ Minerva (AI that solves math and science problems)

In June 2022, Google announced that PaLM, the latest giant language model developed by the company, was used as a base to develop Minerva, which achieved the highest performance ever by artificial intelligence in the correct answer rate for solving math and science problems. Minerva is an
artificial intelligence that can solve quantitative reasoning problems, such as math and science problems. Quantitative reasoning is one of the fields where artificial intelligence is far from human-level performance. This is because solving math and science problems requires combining various skills, such as correctly parsing problems using natural language and mathematical notation, recalling relevant formulas and constants, and solving step-by-step using numerical calculations and symbolic operations.

Until now, the highest performance (SOTA) for the MATH (a dataset of US high school level math problems) correct answer rate by artificial intelligence was only 6.9%, and even PaLM with 540 billion parameters was 8.8%. In contrast, Minerva's MATH correct answer rate reached 50.3%, significantly exceeding the past SOTA. Also, Minerva's
GSM8k (a dataset of elementary school math problems) correct answer rate was 78.5% (SOTA was 74.4%), and the MMLU-STEM (a dataset of high school/university level engineering, chemistry, math, physics, etc. problems) correct answer rate was 75.0% (SOTA was 54.9%), both of which exceed the previous SOTA.

The structure of Minerva's machine learning model itself is the same as PaLM, but in addition to that, by learning from 118 gigabytes of scientific papers obtained from arXiv and web pages containing mathematical formulas written in LaTeX format, it improves its ability to understand mathematical formulas and perform quantitative reasoning.

In large language models like PaLM, prompts (the initial input string) can be used to enable the language model to handle new tasks simply by inputting a few examples and sample answers. Furthermore, when inputting these sample answers, if you also input the steps to reach the final answer, the language model will output the answer including the steps taken. This method of strengthening the reasoning capabilities of language models is called "chain of thought" or "scratchpad prompting".
When having Minerva, which is based on PaLM, solve quantitative reasoning problems, this "chain of thought" method is also used. By inputting detailed steps to reach the answer along with examples and sample answers into Minerva's prompt, it has become possible to solve complex problems.

Minerva solves the same problem multiple times to create several answers, and when different answers emerge, it determines the final answer using a majority vote method, where the most frequent answer is chosen as the final one; this is also one of the techniques Minerva uses to increase its accuracy rate.

Regarding Minerva, there is criticism that it is merely generating text mixed with mathematical formulas probabilistically according to learned patterns, and is not truly understanding and solving math problems, but it is certainly true that it is attempting to overcome yet another hurdle that was considered impossible for artificial intelligence.
If artificial intelligence continues to overcome hurdles one after another in this way, will it eventually, in the near future, completely surpass human capabilities and reach the singularity?

[Reference Books]

All three books were published this year and are authored by Daisuke Okanohara, co-founder of Preferred Networks. I believe all of them are books where you can comprehensively learn everything from the basics to the latest trends in AI technology, which has been evolving explosively recently.