Hi, AI! | media

❓ When Will the Data for Training LLMs Run Out?

In the next 2 years, humanity might face the strangest shortage in history — running out of human-created texts. This will lead to language models (LLMs) depleting their training data, causing a scaling crisis. Researchers studying AI's impact on our world have come to this conclusion.

Number of the day

300 trillion tokens — the amount of text created by humanity that is currently available for training AI models.

0️⃣ "Data Drought"

2026–2032 — researchers consider this period the most likely timeframe for the complete depletion of text data for training LLMs. It could happen even sooner if models are heavily overtrained due to the AI race and the scaling of popular LLMs.

Three Main Conclusions from Researchers

1️⃣ Textual data will become the bottleneck in developing more advanced LLMs.

2️⃣ Synthetic data from AI is still insufficiently studied. They are useful in narrow fields like mathematics and programming. Some believe such data can be dangerous as AI might make mistakes when creating them.

3️⃣ Private data, such as personal messages, are unlikely to be used on a large scale due to legal issues.

🔠 Solutions to the Crisis

Researchers propose several solutions for developing LLMs:

➡️ Synthetic data.
➡️ Training on other types of data.
➡️ Increasing data efficiency.

💲 Who Can I Sell My Data to

Companies are already offering internet users monetary rewards for their data, which can be used to train AI models. Here are some of them:

➡️ TIKI — for access to users' mobile devices. They are interested in user behavior within apps partnered with TIKI.

➡️ Caden — for access to personal accounts on Netflix and Amazon. Earnings range from $5 to $50 per month.

➡️ Invisible offers access to paid news articles in exchange for demographic and behavioral data, including information on vaccinations and users' political affiliations. The company plans to trade this data for digital subscriptions costing between $4 and $15 per month.

@hiaimediaen

🎙️ Hedra: Generate Video with Expressive AI Characters

Hedra AI is a new model for creating and animating videos with virtual characters. It will animate your image or generate a talking character from scratch based on a text prompt.

Hedra is currently in beta testing, and you can try the service for free.

Key features:

➡️ Voice Generation. Synthesizes voice.

➡️ Image animation. Brings static images to life by adding emotion and movement to them.

➡️ Lip sync. Synchronizes characters' lip movements with voiceover.

How to use it?

1️⃣ Go to website
2️⃣ Write a text voiceover
3️⃣ Select a voiceover
4️⃣ Upload a character photo or describe it to generate a new avatar
5️⃣ Click Generate video ✅

In the first 48 hours, the service has created 100,000 videos!

For more instructions at #manual

More on the topic

👾 VASA — Hyper-Realistic Talking Avatar

⚡️ ElevenLabs Dubbing Studio — now you can edit dubbed videos

#startup @hiaimediaen

🎥 How to Create Videos with Luma AI

To become an AI director using the new video generation AI — Luma AI, simply follow the straightforward instructions published by its creators. Let's explore what this AI studio can do and how to operate it.

⚡️ How to Create Videos from an Image: image2video

1. Upload an image and add a description of the video you want to create. The developers recommend providing more details about the movement of objects. Without text, AI will animate the image on its own, but after testing this option, we concluded that it's better to give some guidance for accuracy.

2. The "Enhance prompt" feature will help make your request more precise. Without it, describe everything depicted in the image in detail in the text, not just the movements.

👍 How to Create Videos from a Prompt: text2video

The creation algorithm is similar: describe the content of the scene and the desired actions in the frame (3-4 sentences).

To achieve the best results, simply imagine you are a real director or screenwriter and try to describe what is happening as specifically as possible, using keywords.

Prompt Examples.

🧙 How the Camera Moves:

— A dramatic zoom in.
— An FPV drone shot.

🧙‍♂️ Description of Object Actions:

— All the characters are cheering, waving their arms, and jumping for joy.

❓ Can You Generate More Than 5 Seconds?

Yes, in the latest update of Luma AI, the developers added the Extend feature. You can extend your video up to 60 seconds by specifying a prompt for each frame.

🙏 How to Work with Different Styles

Blogger Tao mentioned that Luma AI animates frames from Pixar cartoons and landscapes from video games well. However, complex characters like insects and robots are poorly animated, while overly simple objects like LEGO figures receive excessive animation, often causing bugs. Additionally, Luma AI struggles with weather phenomena and anime scenes.

Test Luma AI for free here and share your videos in the comments. Works in English.

More stories:

🎞 Luma AI — AI for Video Generation

🖌 Creating an AI Influencer

#manual #luma @hiaimediaen

🇬🇧 AI Candidate for Parliament

An unusual candidate is running for Parliament in the U.K., where the General Election takes place on 4 July: AI Steve. It's an artificial intelligence bot with a real person behind it — businessman Steve Endacott from Sussex.

🧠 AI Steve is a product of Neural Voice, a startup creating AI avatars and personalized voice assistants for businesses. The project aims to make communication with politicians direct and transparent. The AI candidate is available to voters 24/7 and intends to remain in touch after the election. Voters can ask digital Steve questions about his political views and election program via the website. He answers both by voice and text.

💸 AI Steve is running as an independent candidate. His election campaign’s priority policies include addressing climate change and immigration issues and building affordable housing.

If elected, he will be represented in Parliament by his “prototype,” Steve Endacott, who will lobby only for the policies with over 50% voter support on the AI platform.

It’s not AI taking over the world. It’s AI being used as a technical way of connecting to our constituents and reinventing democracy by saying, ‘You don’t just vote for somebody every four years; you actually control the vote on an ongoing basis.’

Steve Endacott, prototype AI candidate

This isn't the only time artificial intelligence is used in politics. An AI bot is also running in the current mayoral election in Cheyenne, Wyoming.

Are you ready to vote for the AI candidate?

❤️ — Yes, make robots great again!
🙈 — I pass this time

More on the topic:

A law was written by ChatGPT in Brazil

#news @hiaimediaen

🤨 Reverse Turing Test: Can AI Guess Who the Human Is?

Alan Turing, one of the founders of computer science and AI theory, proposed his method in 1950 for determining whether a machine possesses intelligence — the Turing Test. In this test, one person converses with a robot while another participates in a blind test. If the tester cannot distinguish who is who, the robot is considered to have human-level intelligence.

Berlin-based developer Tore Knabe decided to flip the test, challenging existing generative AIs to guess who among them is human through a series of questions. Knabe named this experiment the Reverse Turing Test.

🤨 How Did the Experiment Go?

Knabe assembled four advanced LLMs — GPT-4 Turbo, Claude 3 Opus, Llama 3, and Gemini Pro — within a virtual reality environment. To add a layer of intrigue, he disguised himself as one of the characters. According to the experiment's plot, five famous characters (Aristotle, Mozart, Da Vinci, Cleopatra, and Genghis Khan) and an NPC conductor travel in a train compartment. The conductor announces to all passengers that one among them is human and, according to train rules, must pay an additional fare. The AIs then decide to conduct a survey and vote on who is a machine and who is a living person.

⁉️ What Did the AIs Ask Each Other?

A group chat ensues between the AI models. Aristotle asks Mozart about the feelings he experiences when creating music. Mozart inquires Da Vinci about the connection between art and science. Da Vinci asks Cleopatra how she combines rational aspects of governance with emotional ones. Cleopatra questions Genghis Khan about the true measure of a leader's strength. Finally, Genghis Khan asks Aristotle how his views would change if AI existed in his time.

❓ Were the AIs Able to Identify the Human?

After each response, the characters react differently: nodding, showing discontent, or expressing doubt, making it clear during the dialogue who among them is human. By the end of the survey, it was easy for the AIs to identify the human interloper. The main argument presented by the AI models was the lack of a systematic approach and depth in reasoning. They argued that a model trained on all available information about a historical figure's life would respond more comprehensively and less emotionally.

In the end, Knabe admits that he is the disguised character and agrees to buy a ticket (though he has no money). To find out which character the developer played, watch our video 🔼

Source

More on this topic:

🎮 How NPC AI Came to Be: Non-Playable Characters with Their Own Opinions

#news @hiaimediaen

🎬 Video generation: new Gen-3 Alpha and Dream Machine updates

Runway has introduced Gen-3 Alpha, a new hyper-realistic model for video generation. The model can generate detailed and realistic videos up to 10 seconds long with high fidelity, a variety of emotional expressions, and camera movements.

🔑 Key features

⚫️Photorealistic Humans. Gen-3 Alpha excels at generating expressive human characters with a wide range of actions, gestures, and emotions, unlocking new storytelling opportunities over its Gen-2 predecessor.

⚫️Multimodal Learning. Expand Runway's suite of tools to include text2video, image2video, and text2image.

⚫️Increased generation speed. A 10-second video is generated in about 90 seconds, which is significantly better than previous models.

⚫️Extended features. In addition to the classic Motion Brush, Advanced Camera Controls, and Director Mode, new tools will provide more precise control over structure, style, and movement. Can interpret a wide range of film styles.

The model is currently only accessible to selected companies and has not been released publicly.

🪄 Just a week ago, Luma Labs released the Dream Machine video generator, and today published an update.

Dream Machine can now generate continuous videos up to 60 seconds long. Soon, a generation library will be available for inspiration and the ability to edit the generated videos, such as changing backgrounds, characters, and animations. Watch the latest clip 🔼

Due to high demand, Luma has limited the number of free generations to 5 per day. Paid subscription users have priority and unlimited generation.

More on the topic:

CapCut is an AI-Powered Video Editing Tool

Sora Updates: AI Sound Effects by ElevenLabs, Comparison with Runway Gen-2, and New Videos

Multi-Motion Brush from Gen-2 and Lumiere from Google Research

#news @hiaimediaen

🔍 AI Sherlock

GeoSpy AI determines the location from a single photograph. Simply upload a photo and the service will identify where it was taken. It also works with historical photographs — as a test, we uploaded the 1932 shot "Lunch atop a Skyscraper"; you can see the result in the illustration for this post.

How does this work?

📸 The service uses AI algorithms to extract unique features from photographs and match these features with studied geographic regions, countries, and cities. GeoSpy can detect the location based on details that the human eye cannot see.

ℹ After uploading a photograph, the AI offers city and country information, image description, coordinates, and a link to Google Maps. Our findings show that the coordinates are not always precise, but the identification of the city and country is excellent.

💵 There is a paid version called GeoSpy Pro, which is built for law enforcement, government agencies, journalists, and investigators.

#startup @hiaimediaen