Translation by Wang Xin
About the author
Kane Wan, former editor of The Novoland Fantasy and an editor of Eastern Wood. Currently a writer on AI trends at Geekpark, he is a sci-fi enthusiast of the Golden Age period.
About the translator
Wang Xin is a translator and writer with a Master’s degree in Translation. He currently teaches at a university in Guangzhou. Wang has translated many sci-fi works from English to Chinese including Some of Your Blood, Medusa Uploaded, and Seasons of Glass and Iron. He has also contributed to the English localization of mobile games such as Perfect World, One Punch Man, and King of Warships. His approach to translation combines creative interpretation with a deep respect for the original text.
Word count: ~5600 | Est. read time: 30 mins
Main text:
An Introduction to AI and Large Language Models
When the famous sci-fi writer Isaac Asimov proposed the “Three Laws of Robotics” in 1942, he likely didn’t foresee the paradox of AI’s development— that they would progress slowly yet suddenly.
We consider such progress slow simply because, even after 80 years, we remain far from creating the robots he envisioned. As of 2024, the most commercially successful household robot is but a vacuum cleaner that can be easily troubled by pet messes.
Yet, the sudden advancements are nothing short of astonishing. November 2022 witnessed the launch of ChatGPT, an AI platform that has triggered major changes in societies worldwide in less than two years. From Silicon Valley tech leaders to Beijing schoolchildren, everyone is talking about AI, which serves as a technological catalyst for solving global challenges as vast as interstellar exploration and as near as building the metaverse. Everything seems to be possible with AI.
So how did AI evolve from excelling merely at Go and DOTA to becoming today’s “universal assistant”?
Will AI awaken one day to dominate humanity like a real-life Terminator?
Many might also wonder why it might seem as though our lives remain unchanged, even after 2 years of AI hype?
Today, let’s delve into whether the AI boom from 2022 will destroy humanity—or reshape our future entirely.
01 The “Sudden Emergence” of AI: Still a Black Box
To understand the difference between today’s AI and the AI we’ve known in the past—or the omnipotent AI often seen in movies, we need to grasp three key concepts: emergence, the Transformer architecture, and the black box.
When discussing AI today, another term frequently surfaces in the media: Large Language Models (LLMs). These models are so named because of their massive scale, typically starting at billions of parameters. ChatGPT, for instance, falls under the category of text-generating AI products powered by large models, which suggests that it specialises in producing text but cannot generate images or videos.
Meanwhile, there are also other large models designed specifically for generating images and videos. While their technical frameworks share similarities with text-generating models, they also have distinct differences. Image-generating tools like Midjourney have already entered the commercial market, whereas video-generation models like Sora or Colossal remain in experimental stages due to limitations like short content duration and uncontrollable character outputs.
Let’s begin with the first term: Transformer architecture, one of the key technological foundations behind today’s large text-generating models. It originates from a 2017 paper by Google titled Attention Is All You Need, which, originally aimed at optimising machine translation tasks, proposed a new neural network structure—the Transformer architecture. This structure, based on an attention mechanism, has proven particularly effective in the field of natural language processing (NLP). OpenAI’s GPT models are built upon this very architecture.
To put it simply, the GPT architecture works by predicting the next token. For example, if you input the phrase “Tomorrow looks ___” into a GPT model, it will try to guess what token should be filled in the blank and generate various possibilities. And after receiving human feedback during training, it will tend to choose reasonable options like “sunny” or “rainy”.
It should be noted that when GPT outputs the word “rainy”, it doesn’t mean that the model truly understands the sentence. It might have simply selected the most probable answer based on human preferences from a vast pool of possibilities.
In other words, the “understanding” exhibited by AI is a highly complex process of pattern recognition and probabilistic prediction based on statistical analysis of vast amounts of data. By identifying patterns in the input, AI generates corresponding outputs without truly experiencing perception, emotions, or consciousness in its processing.
In the early versions of OpenAI’s GPT models such as GPT-1 and GPT-2, the large models performed reasonably well after being trained on certain datasets and computational power, but they didn’t quite surprise anyone. In many cases, the system still required fine-tuning for specific tasks to achieve optimal results. However, with the release of GPT-3, the model’s parameters increased by a staggering 116 times compared to GPT-2, reaching a massive 175 billion parameters. It seemed as though the model’s capabilities suddenly surpassed the threshold of a singularity.
That is to say, the “sudden emergence” of AI occurs when its parameters and computational power scale to astonishing levels.
- The model can rapidly learn and execute new tasks with minimal examples—a capability not directly trained but naturally developed through processing vast amounts of textual data.
- Task generalisation. GPT models excel at performing numerous tasks for which they were never explicitly trained, such as translation, summarisation, and answering questions. This versatility emerges progressively during large-scale text processing.
- GPT models also demonstrate a certain level of reasoning abilities. They are capable of logical and commonsensical reasoning, far surpassing the scope of simple text prediction.
- With a foundational “creativity,” GPT models can generate original content including stories, poetry and even computer code. This is not merely a rearrangement of existing material but rather a result of deep language pattern understanding and flexible application.
In essence, this “emergence” is what we call “scale unleashes miracles”.
In other words, once the data volume, computational power and model scale reaches sufficient levels, AI’s capabilities will indeed improve, and this is the direction in which most AI research and development companies are currently focusing their efforts on. Therefore, the number of parameters in large models continue to increase. For example, Meta’s (formerly Facebook) Llama 3 model has reached an impressive 405 billion parameters. With such vast scale, the number of GPUs required also continues to rise. The well-known tech maverick Elon Musk, recently built a 100,000-GPU cluster to provide much greater computational power and parameter capacity to train and enhance the large models developed by his company, xAI.
It’s important to note that while to some extent, the “emergence” can be seen as AI reaching a certain singularity. But this is not the “technological singularity”1 proposed by science fiction author Vernor Vinge, nor does it imply that AI truly understands the content. After all, the internal mechanisms of these models remain a “black box” to us.
In the words of Li Mu, a prominent AI startup entrepreneur2, today’s large language models are pretty much like Chinese alchemy. You put certain materials into the furnace, and then refine the elixir according to a formula. The data for large models is the raw material of the elixir, the computational power is the heat of the furnace, and the algorithms are the formula. However, just as the alchemist has no understanding of the technical details of the alchemy, we too do not fully understand the internal workings of large models. Currently, we only know the inputs and outputs, but the internal technical mechanisms are still under investigation.
It was precisely due to a complete lack of understanding in the mechanisms that Elon Musk, Geoffrey Hinton, and over a hundred other tech leaders jointly published an open letter3 in March 2023, calling for a pause of at least 6 months in training AI systems more powerful than GPT-4, as they were concerned about the emergence of super-intelligence.
Although the final call for action went unanswered, their concerns do carry some weight—today’s AI is still a black box, inside which we don’t know what might happen.
However, there may be some safer methods to control AI than simply halting its development. After all, it’s not something that can be stopped with just a few open letters once a technological breakthrough occurs.
Looking back through history, we can see that in the history of science and technology, the study of principles often lags behind the invention itself. For instance, it wasn’t until around 1930—30 years after the Wright brothers invented the airplane—that humanity fully understood how it worked.
It’s normal that technological advancements usually outpace research. So, the “black box” isn’t so scary after all. Don’t panic!
02 AI Awakening: a Long-Held Fear yet a Distant Reality
The open letter reflects just one of the many manifestations of contemporary fears surrounding AI. Themes like AI awakening and AI threats—popular in sci-fi movies and novels—have once again taken stage centre.
But will today’s AI suddenly awaken, turn against humanity, and eventually dominate or even destroy us?
A basic understanding of today’s AI mechanisms makes it clear that scenarios like Skynet or The Matrix, often depicted in science fiction, are unlikely to happen anytime soon.
Today’s AI lacks the capacity for sudden awakening, nor does it harbour aspirations to dominate humanity.
As we’ve discussed earlier, modern AI mainly functions by “guessing the next word or token.” It draws upon patterns from billions of past datasets and then statistically predicts the response humans are most likely to favour. While its outputs often seem intelligent, they do not stem from genuine “understanding.” Take the sentence “Tomorrow looks rainy” for example, AI does not grasp what “tomorrow” or “rainy” signifies—it simply predicts the word “rainy” as the most probable completion based on prior data.
A similar view is shared by Ted Chiang4, a renowned science fiction writer, who argues that when AI says “sorry,” it doesn’t actually understand the apology or emotions the word represents. Instead, it simply selects this word as the most statistically likely option for the given context.
Furthermore, today’s AI serves only as a responsive system, which has no autonomy or desires at all. It won’t take any action without human commands.
Imagine that AI developed a will for self-action and the goal of dominating humanity—could it destroy humanity with its current powers?
First, AI would need to grow rapidly and enhance its capabilities significantly to realise its supposed ambitions of world domination.
But the first challenge it faces is a lack of “food”. Training large language models like GPT relies on data corpora from the internet, including published books, news, and other high-quality textual datasets. However, nearly all the textual data accumulated throughout human history has already been fed to it to achieve the current “emergence” of AI capabilities.
Some have proposed using AI to generate its own data. Indeed, certain companies have started experimenting with synthetic or AI-generated data for training large models. However, a Nature article5 published in July revealed that training AI on AI-generated data could degrade its performance, even leading to model collapse and regression. While such generated data is usable, its current effectiveness appears insignificant with many risks.
Then, there’s the issue of computational power. Enhancing AI capabilities demands a massive number of specialized GPUs for AI computation, primarily produced by NVIDIA or other tech companies. While consumer-grade GPUs can be used, they are significantly less efficient and prone to power loss more quickly. Even with professional GPUs, the challenge of combining tens of thousands of GPUs into a single system while minimising performance degradation remains unsolved. Currently, one of the most efficient methods to scale computational power is through GPU clusters. Recently, Elon Musk has built a supercluster equipped with 100,000 H100 GPUs, which is likely the most powerful AI training cluster in the world today.
However, the large language model xAI trained by Musk’s company has demonstrated only average capabilities, proving that computational power alone is not yet sufficient.
For many companies and even nations, assembling a cluster of 100,000 GPUs presents challenges beyond data exchange between the GPUs—it also requires a stable power supply, whether from a high-infrastructure-capacity single data centre or a distributed system. Take Musk’s 100,000-GPU cluster as an example: the total power required for its key IT components is approximately 150MW. In comparison, the key IT power of El Capitan, the largest supercomputer at the U.S. national laboratory, is only 30MW, about one-fifth of that, making it seem relatively insignificant. If we make a rough calculation, the servers in this cluster would consume about 3 million kW/h of electricity daily—equivalent to the daily residential electricity consumption of Dongcheng District in Beijing.
This is merely the power demand of today’s AI. In other words, the kind of sudden AI awakening often portrayed in movies, where an AI consumes an entire city’s electricity all at once, is unlikely to happen. The energy requirements of an awakened AI would likely far exceed the power supply of a single city.
But even if AI somehow overcomes challenges related to data, computational power and electricity, the gap between current AI and the real world remains a significant barrier for it to dominate the world.
Today’s AI still lacks an understanding of the real world. Its greatest strength lies in content creation such as generating text, images or videos.
Why is creating text the easiest task? Because text is a one-dimensional form of communication and, in some ways, a relatively inefficient one. In contrast, the real world is far more complex. Following the emergence of GPT models, many experts have introduced the concept of “world models”, precisely because understanding text is comparatively easy for AI, while comprehending the real world remains exceptionally challenging.
Take autonomous driving as an example. After years of development, it has only recently achieved limited success in a few cities. Robots, however, face challenges far more complex than those encountered in autonomous driving. This is why robots have yet to replace humans on a large scale, and the most commonly used robot today is still the humble robotic vacuum cleaner. This is also why AI has not replaced factory workers as some feared, but has instead come close to displacing content creators.
For today’s AI, interacting with the three-dimensional world remains an insurmountable task.
Therefore, today’s AI is far from capable of creating a robot army to exterminate humanity.
If AI truly wanted to destroy humanity, the fastest way would be to gain access to the launch codes of various nations and deploy nuclear weapons. This scenario, however, would require someone to issue such an order and the AI to comply with the command and bypass all laboratory and network constraints to secure launch access. Yet, it is widely recognised that current AI systems are so heavily regulated that even generating profanity or explicit content is prohibited.
That being said, AI still has the potential to lead to humanity’s destruction, albeit in a manner that leads humans to destroy themselves.
The proliferation of fake news and deepfakes on the internet has already created significant divisions and ideological conflicts among people.
This, ironically, is currently the most likely way AI could bring about humanity’s downfall—by pushing humanity into civil wars. However, even without AI, humanity has always been capable of self-destruction. So, take it easy… ≧ω≦
03 Can Today’s AI Replace Human Jobs?
Since AI won’t be achieving sentience or destroying humanity anytime soon, let’s move on to the next question: Can today’s AI replace humans?
Or more specifically, can today’s AI replace human jobs?
The simple answer: AI can already replace certain types of jobs, but its primary role isn’t about replacement; it’s about enhancing human capabilities. The specifics, of course, depend on the type of work and the context in which AI is applied.
To evaluate the current technical capabilities and maturity of AI, we can look at the main types of AI models in use today. Broadly speaking, today’s AI and large language models can be divided into three main categories: text generation models, image generation models and video/audio generation models.
Text Generation Models: Mature and Widely Adopted
Text-based AI products include the iconic OpenAI’s GPT series, and other prominent models like Claude, Gemini, Llama, ERNIE Bot, Qwen and Kimi. These models have achieved a high level of technical maturity and have already been deployed in a variety of real-world commercialised scenarios.
Popular products, including ChatGPT and Chinese platforms like Kimi and Doubao, primarily focus on text-based interactions. Paired with well-designed prompts, text generation AI can assist with tasks like translation, document analysis, sentence polishing and email drafting—tasks common in work or study settings. Currently, these text-based AIs can be compared to a well-rounded college graduate: multilingual, possesses strong language abilities, and proficient at explaining concepts, but less dept in fields like advanced mathematics.
Some scenarios in which my friends use text-based AI include generating social media posts, crafting email replies, translating articles from foreign languages, brainstorming ideas, and even co-creating portions of scripts or novels.
Image Generation Models: Advancing To Commercial Use
Image generation models, such as Midjourney, Stable Diffusion, and DALL-E 3, focus on creating visual content. These so-called “AI artists” have increasingly found more commercial applications.
Some game scenes, book covers, and display images of e-commerce products have begun tapping on AI’s generative technology. A popular Chinese app in 2023, StyleArt, also uses AI to generate images—though it generates new images from existing ones rather than using text prompts.
The main issue with these models is their low success rate—akin to a “gacha game” where users must engage with the AI multiple times to generate satisfactory results. Moreover, current image generation AIs are less capable of understanding human language compared to text-based AIs, requiring users to master some “spells” to generate satisfactory results.
Video/Audio Generation Models: Still in Experimental Stages
One of the hottest AI products during the Spring Festival was OpenAI’s Sora, a text-to-video generator capable of producing one-minute-long, highly realistic, high-quality videos. While Sora caused a sensation, it remains unavailable to the public.
Meanwhile, other contenders such as Pika, Runway, and China’s other generative models like Kling, Pixverse and Doubao are still catching up. Initially starting with 3-second videos, these products have gradually progressed to generate videos lasting about 10 seconds, but they’re still far from ready for commercial applications.
Audio products are currently mainly divided into two categories: TTS products, which can automatically convert text into speech, and music generation products mainly represented by Suno and Udio.
Text-to-speech products have already been implemented in many commercial scenarios, such as AI reading novels, adding voiceovers to videos, and even creating AI-generated podcasts.
On the other hand, music generation products are still in the experimental stage, but they’re facing more complicated copyright issues compared to text generation products.
A Toolbox for Everyone: Explore What AI Can Do
By September 2024, competition in large language models has intensified among tech giants like Google, Microsoft, Meta, and Amazon. In China, the ecosystem still includes a mix of established players and emerging startups, though shifts in this landscape are likely to happen anytime soon as AI large model developments are extremely costly and sustainable business applications remain unclear.
This trend aligns perfectly with the predictions of sci-fi writers: tech giants will monopolise everything, much against the will of capitalists.
Here’s a breakdown of popular AI products based on functionality and type. Testing them all might offer you a clearer understanding of what AI can and can’t do today.
Of course, I’ll leave it up to you to decide for yourself, if there’s a need to try them all!
ChatGPT
OpenAI’s AI chatbot product, built on its proprietary large language model series GPT, is currently one of the most widely used AI products in the world. It is this product that has sparked global attention and discussion about AI, propelling the current wave of excitement over AI technology.
The core of ChatGPT is a series of GPT models developed by OpenAI, with the most powerful ones being GPT-4 and GPT-4o. Currently, ChatGPT primarily focuses on text-based conversations and does not generate videos or images (though some premium users can generate images), but it is capable of recognising images.
OpenAI’s large model for text-to-image generation is DALL-E 3, released in 2023. Users can simply input a text description, and the model will generate corresponding images.
The large model for text-to-video generation by OpenAI is Sora, which released demo videos in February 2024 but has not yet been made available to the general public.
Microsoft has heavily invested in OpenAI, and the two companies currently maintain a strong, intertwined partnership.
Claude
Anthropic launched an AI chatbot product Claude, based on its own large language model series of the same name. The underlying model, Claude 3.5, is currently a text generation model with capabilities comparable to GPT-4o.
The company has consistently promoted the concept of “Constitutional AI”, positioning itself as one of the leaders in aligning AI with human values. Many people believe Claude excels in tasks such as text and code generation.
Amazon has invested in Anthropic and holds a significant stake in the company.
Gemini
Google has launched an AI chat product based on its own large language model, the Gemini series.
However, due to the position of large corporations and their business models, Google initially adopted a cautious approach to AI development. After all, a demonstration of their AI product at a press conference, which contained factual inaccuracies, led to a sharp drop in Google’s stock, in turn assisting OpenAI in taking the lead. Now, Google is closely chasing after OpenAI in this race.
Gemini’s capabilities are now integrated into many of the Google’s products, such as Google Docs and Gmail, offering users a seamless AI experience.
These are the most widely used AI chatbot products globally. In China, several other products are also very popular, including the following:
Doubao
ByteDance, the parent company of Douyin and TikTok, has launched an AI chat product Doubao based on its own large language model. It is currently the most widely used AI product in China. The product offers a comprehensive range of features, including chat, translation, image generation, and AI search.
Kimi
The AI chat product launched by China’s startup Moonshot AI has gained widespread admiration on platforms like BiliBili and Xiaohongshu. It excels at reading long texts — for instance, reading an entire Dream of the Red Chamber at once — and offers many interesting micro features. It is one of the most widely used AI products developed by a startup.
ERNIE Bot
Baidu has launched its own AI chatbot product based on its own ERNIE Bot large language model, which is also the first AI product in China to be widely available to the public. It is said to be particularly adept at drafting official documents, though its most capable version, ERNIE Bot 4.0, requires a paid subscription.
Qwen
Alibaba has launched its AI chatbot product based on its self-developed Qwen large language model. The product supports features such as speech-to-text conversion, PowerPoint creation, and rapid document reading.
Other chatbot products include: Hailuo AI, Ying AI, Wanzhi AI, and more.
In addition to chatbot products, there are also many innovative products based on large language models primarily focused on AI search, AI companionship, and AI programming.
Perplexity
The webpages on the internet are actually designed for search engines rather than for human readers. Therefore, if AI could assist us in searching for and organising content, it might present search results in a different light.
Take the popular Perplexity AI search for example, it specialises in “knowledge search.” By inputting a sentence, this AI organises the results and presents them to you, with each piece of information clearly marked with its source, ensuring the accuracy of the information.
The goal of Perplexity is not to overthrow Google, but rather to offer an alternative approach to content search and organisation. A similar product in China is Metaso AI search (https://metaso.cn/), which is particularly useful for knowledge retrieval and learning purposes.
Character.ai
What is humanity’s dream for AI? A perfect AI assistant, a friend, or even a lover. And this is precisely why, after ChatGPT went viral, many AI-based companion chat products began to emerge. You can have AI play various roles in conversations with you. After all, loneliness is one of the greatest fears of people in this era.
A similar product in China is Xingye, launched by the AI startup MiniMax.
AI programming products
In addition to the well-known GitHub Copilot, a number of AI programming products have emerged this year, targeting ordinary users with little programming knowledge. This reflects an important aspect of AI’s current capabilities: enhancing human potential rather than replacing it.
Cursor：https://www.cursor.com/
It’s said that a 7-year-old girl used this product to independently develop a game.
Wordware：https://www.wordware.ai/
This fully visual and drag-and-drop AI product helps users build websites. Simply put, it’s similar to WordPress.
04 How Can We Avoid Being Replaced by AI?
The simplest way is to use it. Try out the various capabilities of AI according to your own needs.
The more you understand your enemy, the less likely you are going to be defeated by them. And there’s another truth: you won’t be replaced by AI; you’ll only be replaced by people who know how to use AI.
Learn to Communicate with AI First
Communication between people is difficult enough, let alone communication between humans and AI. The first step in using AI effectively is learning how to talk to it.
If I had written this article six months ago, I would have included a long guide on how to craft a good prompt (the instructions you give AI). But now, with AI’s capabilities improving, that’s no longer necessary.
Recently, a Claude engineer shared some insights on how to communicate with AI and write effective prompts during a podcast. I’ve distilled some key points:
- Most of the time, all you need is a very clear task description, rather than trying to create something abstract.
- You need to be able to explain things clearly enough for the model to understand what your task is, and to think and describe concepts well.
- You need to organise and write down everything in your mind that the model doesn’t know. After all, if you don’t say it, how would AI know? (≧ω≦)
- Many people misunderstand what a prompt is. When they see an input box, they treat it like a Google search bar and just type a few keywords. But then again, you wouldn’t communicate with a colleague that way, right?
- Models can understand complex information, so there’s no need to oversimplify it.
Think about how you’d communicate with your coworkers across departments, with your interns, or even with a repairman who’s coming to fix your air conditioner. Just communicate with AI the same way!
Where There’s Strain, Let AI Reign!
The first confusion many people may face when using AI products is: What can I use it for?
Don’t worry, just follow this simple principle: where there is strain, let AI reign! If it is a strenuous task for you to use to write an email in a foreign language, use AI to ease the pain! If it is a strenuous task for you to read in-depth English reports, let AI help you with the translation. If it strenuous for you to understand some technical terms, write a project outline, or grasp an unfamiliar concept, AI can help with all of that.
Even when you run into confusion while using your computer, you can immediately turn to AI for consultation. For example, I recently helped a friend in a planning role solve a very unusual request:
Is there a faster way to generate an image from a few words I type? There are too many lines, and I don’t want to take screenshots manually.This is how I communicated with Kimi:
User: I want to create a cmd batch file that will take each line of text from my txt file and save it as a separate image.
Kimi: To create a batch file that fulfils your request, you’ll need to use some external tools, such as the convert command, which is part of the ImageMagick software package used for image processing. Here’s a simple example of a batch script that will read a text file and save each line as a separate image.
1. First, make sure that ImageMagick is installed on your computer and that the convert command is available in the command line.
2. Create a batch file, such as naming it text_to_images.bat, and paste the following code into it:
--------------------------------------------------------------
Plain Text
@echo off
setlocal enabledelayedexpansion
:: Set the text file and output directory
set "input_file=your_text_file.txt"
set "output_dir=images"
:: Create the output directory
if not exist "%output_dir%" mkdir "%output_dir%"
:: Read each line of the text file
for /f "tokens=* delims=" %%a in (%input_file%) do (
set "line=%%a"
:: Convert each line of text into an image
convert -background white -fill black -font Arial -pointsize 24 label:"!line!" "%output_dir%/line_!line!.png"
)
echo Done.
endlocal
--------------------------------------------------------------
3. Replace your_text_file.txt with the path to your text file, and images is the directory where you want to save the images.
4. Double-click to run text_to_images.bat. It will read each line of the text file and save each line as an image named line_<line_content>.png.
Please note that this script uses the label: option to create text labels, with a white background, black text colour, and a font size of 24 in Arial. You can adjust these settings as needed.
If your text contains special characters or non-English characters, you may need to adjust the font and encoding settings to display the text correctly.
User: Are there any other ways without installing third-party tools?
Kimi: If you don’t want to install third-party tools, you can use Windows’ built-in PowerShell to achieve this function. Below is an example of how to save each line of a text file as an image using PowerShell:
1. Open Notepad or any text editor and copy the following code into it:
--------------------------------------------------------------
PowerShell
$inputFile = "your_text_file.txt" # Replace with the path to your text file
$outputDir = "images" # Output directory
if (-not (Test-Path $outputDir)) {
New-Item -ItemType Directory -Force -Path $outputDir
}
$text = Get-Content $inputFile
$text | ForEach-Object {
$line = $_
$image = New-Object System.Drawing.Bitmap 800,200
$graphics = [System.Drawing.Graphics]::FromImage($image)
$graphics.FillRectangle([System.Drawing.Brushes]::White,0,0,800,200)
$font = New-Object System.Drawing.Font "Arial",12
$solidBrush = [System.Drawing.SolidBrush][System.Drawing.Color]::Black
$graphics.DrawString($line, $font, $solidBrush, 10,10)
$graphics.Dispose()
$image.Save((Join-Path $outputDir "line_$($text.IndexOf($line)).png"))
$image.Dispose()
}
--------------------------------------------------------------
2. Replace your_text_file.txt with the path to your text file, and images is the directory where you want to save the images.
3. Save the file with a .ps1 extension, such as text_to_images.ps1。
4. Right-click the saved .ps1 file and select “Run with PowerShell”.
This script uses the System.Drawing namespace from the .NET Framework to create and save images. It reads each line of the text file and saves it as a PNG image.
Please note that this script may need adjustments in certain cases, such as resizing the image, adjusting the font size or changing the text position. Additionally, if your text file is very large, this script might consume a significant amount of memory.You see, all I need to do is clearly state my requirements, and I can even “negotiate”.
Still Not Sure How to Use AI? Check Out These Four Examples!
If you don’t have any specific needs, try these scenarios, and you might enter a whole new world.
Translation
This is the feature I use most frequently with AI. With the help of immersive webpage plugin translations, I read a large amount of tech news every day, and even dive straight into English novels.
For example, I read the article written by Ted Chiang published in The New Yorker6 with the help of Claude, which was able to generate both literal and free translations for me, and I finished it within half an hour.
Learning Assistant
AI is ideal for explaining unfamiliar terms or tech concepts. You can ask it to explain a topic as if to a high school or even middle school student. When you want to understand a new technology or field, try to describe it as clearly as possible to AI, and it can offer significant assistance.
It can also be used for paper interpretation, especially beneficial for breaking down lengthy academic papers into understandable bites. However, to truly understand a paper, try having AI ask you questions instead of just summarising the content. After all, the best gauge of mastery is the knowledge you can generate by answering questions.
AI can also assist with language learning. Many people use ChatGPT as their language tutor, and you can try that too.
Creative Writing Assistant
AI can help with article brainstorming, script outline developments, image concept generations, or even detailed descriptions of scenes.
However, don’t entrust AI with the entire creative process. Think of it as your creative assistant or brainstorming partner, not your ghostwriter.
Work Assistant
Do you need to adjust a specific format in Word, have trouble recalling a formula in Excel, experience a writer’s block and can’t find the right words to use, or want to make your email reply sound more polite or professional?
AI can help solve these problems, or at least provide instructive ideas.
In an article7, DeepMind machine learning expert Nicholas Carlini detailed how he uses ChatGPT in his daily tasks such as writing codes, optimising codes and fixing bugs. While it’s more focused on an engineer’s mindset and scenarios, the problem-solving approaches are broadly applicable.
For repetitive tasks that have been bothering you or that you’ve been trying to optimise, try handing them over to AI—you might be pleasantly surprised.
No matter what, just start using it.
References:
- In 1982, Vernor Vinge first introduced the concept of the “Technological Singularity” at the annual meeting of the American Association for Artificial Intelligence at Carnegie Mellon University, suggesting that after the Singularity is reached, superintelligence will emerge. ↩︎
- Full transcript of Li Mu’s lecture at Shanghai Jiao Tong University ↩︎
- Full text of the open letter co-signed by Elon Musk and Geoffrey Hinton ↩︎
- Ted Chiang: Why A.I. Isn’t Going to Make Art ↩︎
- Nature paper; AI models collapse when trained on recusively generated data ↩︎
- Ted Chiang: Why A.I. Isn’t Going to Make Art ↩︎
- A DeepMind scientist: How I Use “AI”— A List of 50 Conversations with AI. ↩︎
Translation Editor: Xuan