Chinese spy satellite, Stability AI’s language model and other AI news

We aim to inform readers not only about developments in the Bitcoin industry but also about what’s happening in related technologies — cybersecurity and the world of artificial intelligence (AI).

ForkLog AI has gathered the week’s most important AI news.

Reddit begins charging API fees for dataset collectors

Reddit will begin charging API access fees from companies that use it to collect data for training large language models.

Pricing information is not yet available, but it is understood to be tiered. The company says the ‘new premium-class access point for third parties’ will include additional capabilities, higher usage limits and expanded rights.

Reddit founder Steve Huffman said the platform’s data are valuable and cannot be given away for free to some of the world’s biggest companies.

Other platforms are expected to follow suit. Stack Overflow said it would begin charging API access by mid-2023.

Media: Google employees call Bard a “pathological liar”

Employees allegedly urged Google not to launch Bard, describing the chatbot as ‘worse than useless’ and a ‘pathological liar’.

Bloomberg reporters surveyed 18 current and former employees and reviewed internal emails. One said Bard often gave users dangerous advice on questions such as boarding a plane or diving underwater. Another recommended never to interact with the bot.

According to the publication, Google even “canceled the risk assessment” presented by its internal security team and marked the system as not ready for general use. However, in March 2023 the company nevertheless opened early access to the ‘experimental’ AI.

Adobe unveils AI features for photo and video editors

Adobe announced a suite of new AI features for its photo and video editing software.

The Firefly AI editor has been expanded with new tools to reduce routine work. They allow, via text prompts:

• adjust brightness and saturation levels in video, change time of day or season;
• animate titles and text;
• analyse the script content to automatically create cutaways in clips;
• select sound effects and background music;
• automatically create storyboards and preliminary 2D or 3D visualisations to simplify shooting video.

The Firefly editor can also generate personalised practical guides on using the tools.

The AI editor features are already available in the Adobe ecosystem, including Premiere Pro, Illustrator, After Effects and Photoshop, as part of closed beta testing.

The company also expanded the Select People masking tool in Lightroom with options for automatically selecting hair on the face and clothing.

Additionally, adaptive presets include Darken Beard, Polished Portrait and Enhance Clothing, designed for portrait work.

The first filter darkens the beard, the second smooths skin, makes facial features more pronounced and adjusts lighting. The third preset enhances clothing appearance by increasing contrast, saturation and texture.

The update also includes AI curves in masking and a DeNoise tool to remove “grain” from photos.

Stability AI unveils the StableLM suite of large language models

Stability AI has unveiled a suite of open large language models (LLMs) named StableLM. They are available on GitHub and Hugging Face.

The models are designed to generate text and code. Currently available are LLMs with 3 billion and 7 billion parameters, with others in the pipeline ranging from 15 billion to 65 billion. They were trained on an experimental dataset based on The Pile, containing 1.5 trillion tokens of content from sources such as Wikipedia, Stack Exchange, PubMed and others.

The company warned that the LLMs may generate biased and toxic content.

Atlassian unveils AI assistant for team collaboration

Atlassian, the maker of Jira and Trello, unveiled “a virtual AI team-mate” called Atlassian Intelligence.

The system is built on the company’s own algorithms and OpenAI’s large language models. The assistant can create custom charts for team collaboration, generate summaries from the internal Confluence wiki, draft test plans in Jira Software and respond to customers in Jira Service Management.

The AI has been trained to summarise meetings and transcribe conversations, and then to extract a list of necessary tasks. It can also search terms in the enterprise database and answer work-related questions.

The assistant can process natural-language queries and translate them into JQL and SQL. It will appear in the company’s cloud products later, and for now Atlassian has opened a waitlist.

Snap’s My AI chat bot available to all users

Snap has opened access to its My AI chat bot to all users. Previously it was available only to Snapchat+ subscribers.

The developers taught the bot to recommend restaurants and other popular activities on Snap Map, suggest AR lenses and respond to content. They say that soon My AI will respond to images with AI-generated pictures.

The company also announced other features for Snapchat:

• AR lenses powered by generative AI;
• 3D models of landmarks and recommendations of places of interest in Snap Map;
• a lens for video calls in VR environments;
• AR Mirrors for deploying AR on real-world screens.

Snap also expanded its partnership with the entertainment company Live Nation, which is deploying AR technologies at 16 music festivals.

Japanese authorities used ChatGPT to draft documents

The Ministry of Agriculture, Forestry and Fisheries of Japan used ChatGPT to draft regulatory documents.

The ministry is testing the chatbot to simplify official government directives and make them more understandable. It said the AI processes only publicly available information, as “there is always a risk of leaking confidential data”.

Earlier, Asahi reported that the ministry planned to use ChatGPT to update online guides for filling out subsidy applications and other government support.

Google’s robots learn to sort waste

Google trained a fleet of robots to sort waste.

Researchers tested the technology over two years. They used a reinforcement-learning approach based on real data, in combination with simulations.

To collect the dataset, engineers programmed 30 robots to search the office for waste-sorting stations and allocate objects to bins in a specified way. For example, cans and bottles should be taken to the recycling container, while paper cups to compost.

Another 20 robots sorted waste in specially equipped ‘classes’.

In total, the researchers conducted around 572,500 tests. They said that performance increased as the amount of data grew.

As a result, the system learned to sort 84% of objects correctly.

Chinese AI used a satellite to spy on neighbours

Researchers at Wuhan University allowed AI to control an experimental satellite Qimingxing 1.

A ground AI controlled the satellite around the clock without any human intervention. The aim was to study the decisions made by the algorithm in autonomous conditions.

The researchers said the AI selected several locations on Earth and ordered Qimingxing 1 to study them more closely. Among the targets was Patna, home to the Bihar Regiment of the Indian Army, which fought Chinese troops in 2020.

The AI also directed the satellite to the Japanese port of Osaka, where US Navy ships sometimes dock.

It is not known why the AI chose these targets for observation.

AI-generated image wins photo contest, but entrant declines prize

An AI-generated image won the Sony World Photography Awards 2023 in the Creative category. The author declined the prize.

German photographer Boris Eldagsen used image generators to create the piece PSEUDOMNESIA: The Electricia. He says this work is the result of a ‘complex interaction of prompt-engineering, painting and repainting, drawing on rich photographic knowledge’.

Eldagsen wanted to know whether AI-generated art would be acceptable for competitions. By declining the prize he aimed to draw the photo world’s attention to the future of art amid rapid development of generative technologies.

A spokesperson for the World Photography Organisation said that while judges were aware of AI usage, they had anticipated ‘collaborative creativity’. Eldagsen said he had deliberately misled the jury.

Deepfake track with Drake and The Weeknd voices goes viral

TikToker Ghostwriter, with the help of AI, created a song with deepfake voices of Drake and The Weeknd. Over the weekend the track Heart on My Sleeve racked up more than 10 million views on TikTok and 250,000 streams on Spotify.

According to the author, he wrote and produced the song, then replaced his own voice with the performers’.

Drake and The Weeknd have yet to respond to the deepfake. Universal Music Group confirmed the track is an unauthorized work, likely AI-generated, and voiced its displeasure with Ghostwriter.

A few days later streaming platforms removed the track.

The week’s most significant AI deals

From 16 to 22 April 2023, AI startups attracted over $290 million. Here are the most notable deals.

• CoreWeave привлекла $221 million to expand its specialized cloud infrastructure for compute-intensive workloads, including artificial intelligence and machine learning.
• Trullion получила $15 million.
• Fletch собрал $12.5 million.
• Quantum Source получила $12 million to propel continued development in photonic quantum computing.
• Groundlight привлекла $10 million to support a computer vision platform.
• Cortical Labs собрала $10 million to develop organoid intelligence.
• Mediwhale получил $9 million.
• Loopin привлекла $1.9 million to develop a platform for measuring, managing and boosting employee morale.

Also on ForkLog:

• Elon Musk founded an AI company to compete with OpenAI. Also the billionaire announced a “most truthful” alternative to ChatGPT.
• Sam Altman denied reports of training GPT-5.
• Media: Hikvision knew about the use of its tech for monitoring Uyghurs.
• The Information: Microsoft is developing its own AI chip Athena.
• Nvidia unveiled a video generator from text descriptions.
• Google will merge Brain and DeepMind into one team.
• The Schumacher family will sue a tabloid over an “AI interview” attributed to the racer.

What to read this weekend?

ForkLog has updated its coverage of AI-generated imagery. The selection includes five free services that allow anyone to test the algorithm and create unique artworks.