AI won't replace your creativity – or will it?

[German version of this issue | Subscribe to the German edition]

There is a debate about the extent to which AI tools are creative themselves, or whether they are "just" rehashing what others have already created.

The much more important question for me, as a practical person who makes a living from creative work, is: Does it makes a difference to the audience whether a human brain or an AI is behind a piece of work?

Because the answer determines whether creative people will still be able to make a living from their work and thus from their creativity in the future.

The answer is complex. Personally, I hope that a work created by a living, breathing human being is still worth more, even if an AI can create something very similar in a few seconds.

After all, a person has an individual perspective, has experience, and brings something new to the table.

However, anyone who works in a creative profession knows that not every work is an earth-shattering, never-before-seen masterpiece.

Sometimes it's a transaction: You are paid to do a job well enough. Your work is supposed to serve a specific purpose. Your artistic expression is secondary.

A store needs some product descriptions. A new game needs sound effects. A commercial needs the right music.

Such tasks are not always mentally or emotionally fulfilling. But they do bring in money to pay the rent, buy groceries, and pay the electric bill.

For many creative people, such jobs are an important foundation of their livelihoods. They give them the freedom they need to do the things they enjoy, that allow them to express themselves, and that help them grow artistically. Freelance platforms such as Upwork or Fiverr, for example, build a business around arranging such jobs.

I wonder more and more to what extent this will still exist in the future. Because that depends a lot on whether it will still be worth it for people to spend the money to have a human do it instead of a fast, cheap machine.

Or whether they will let a cheap creative novice operate the AI instead of spending the money on an experienced specialist.

As creatives, we may be able to see the difference between a well crafted, special piece of copy and a standard piece of copy from a machine. But clients don't always.

Do you have an opinion about this? Let me know in the comments!

T O O L S

Meta releases open GPT-4 competitor Llama 3.1 405B

Meta has released the most powerful model of its Llama series: Llama 3.1 405B. With an impressive 405 billion parameters, it is believed to be the first openly available AI model that can compete with leading closed models such as OpenAI's GPT-4. The latest version of Llama offers an extended context length of 128,000 tokens and now supports eight languages, including English, German and Spanish.

With this release, Meta is also making a clear commitment to open accessibility for AI. The models are available on platforms such as AWS and Hugging Face, facilitating development and customization for a broad developer community. Meta's CEO Mark Zuckerberg emphasizes that open AI should benefit everyone and not be concentrated in the hands of a few companies. Read more in the "Good Reads" section below.

Sources: Meta, Ars Technica, VentureBeat

OpenAI GPT-4o mini attracts with incredibly low price

OpenAI has introduced their new GPT-4o mini AI model, which is now the most cost-effective closed option on the market. Priced at only $0.15 per million input tokens and $0.60 per million output tokens, GPT-4o mini is 60% cheaper than the previous GPT-3.5 Turbo model. Despite the lower price, it still appears to outperform many competing products in key benchmarks, offering improved performance for text and image-based tasks.

According to Olivier Godement, Head of Product at OpenAI, GPT-4o mini opens up new applications in areas such as customer service, software development and creative writing. It supports textual and visual input, and will support audio and video output in the future.

OpenAI has also announced free fine-tuning for GPT-4o mini. This allows the model to be optimized for specific use cases. Until September 23, users can use 2 million training tokens per day for free.

Last but not least, the new model will feature a new security measure to prevent a typical attack vector. Currently, chatbots can often be freed from their predefined limits by being asked to "ignore all previous instructions". The new "instruction hierarchy" should ensure that user commands can no longer override the developer's instructions.

Last but not least, OpenAI is considering making its models available locally in the future.

Sources: OpenAI, VentureBeat, Simon Willison’s Weblog, ?Wired, The Verge, VentureBeat

Mistral Large 2: New AI model from France

Mistral has launched its latest AI model, Mistral Large 2, with 123 billion parameters. It offers improved capabilities in code generation, mathematics and multilingual applications. With a context window of 128,000 tokens and support for over 80 programming languages, Mistral Large 2 looks like a powerful offering.

It has been specially trained to minimize hallucinations and provide accurate, reliable answers. This could make the model particularly suitable for complex tasks in business applications.

Use of Mistral Large 2 is permitted under the Mistral Research License for non-commercial purposes, while commercial use requires a separate license. The model is available through the Mistral platform and various cloud services such as Google Cloud and Amazon Bedrock.

Sources: Mistral, VentureBeat

AI trend: smaller is better

A current trend in AI is towards smaller language models that are still suitable for many everyday applications.

For example, Mistral NeMo, a new model from Mistral AI in collaboration with NVIDIA, offers advanced features for businesses with just 12 billion parameters and a context window of 128,000 tokens. This model is specifically designed for local applications on desktops and laptops, enabling powerful AI solutions without the need for extensive cloud resources.

In parallel, Hugging Face has introduced the SmolLM models that can run on mobile devices. These compact offerings perform well in specific tasks and provide a cost-effective and privacy-friendly alternative to larger models.

The increasing willingness to invest in smaller models, such as the $24 million for Arcee AI, illustrates the growing interest.

Sources: Mistral, VentureBeat, VentureBeat, VentureBeat, VentureBeat

More tools in brief

Adobe has introduced new generative AI capabilities for its Illustrator and Photoshop design software. For example, the new Generative Shape Fill tool in Illustrator lets users fill shapes with scalable, generated images based on text descriptions.

The German AI startup DeepL has introduced a new translation model that has been specially developed for business applications. It is designed to help companies translate internal communications and marketing materials more efficiently and with fewer errors.

Anthropic released the official Android app for its Claude chatbot, bringing its AI capabilities to a wider audience.

Patronus AI has released Lynx, an open-source model that can detect and reduce hallucinations in large language models such as GPT-4, which is particularly important for companies in highly regulated industries such as finance and healthcare.

Traceloop provides a monitoring platform that helps LLM applications detect hallucinations and incorrect responses in real time.

Microsoft's "SpreadsheetLLM" is a new AI model for analyzing spreadsheets.

The new video AI Haiper 1.5 can generate eight-second clips from text, image and video input. The startup also plans to add image generation and offers a built-in upscaling tool to improve the quality of the content created.

Microsoft's Designer app for iOS and Android uses AI editing capabilities to help users create and customize images, greeting cards and more. The app is free for users with a Microsoft account and offers 15 daily "boosts" for image creation.

Japanese startup Sakana AI has released two new AI models that create images in the style of traditional Japanese ukiyo-e art. The Evo-Ukiyoe and Evo-Nishikie models were trained on a dataset of over 24,000 ukiyo-e images and understand both text and image prompts.

Stable Video 4D from Stability AI can generate eight new views from a video, which could be interesting for movies, games, and augmented reality applications.

Microsoft's Bing search engine now displays AI-generated answers alongside traditional search results.

Nvidia's new AI Foundry service is designed to help companies develop custom AI models. The company promises to significantly improve the performance of these models.

N E W S

OpenAI faces more criticism ...

The ChatGPT company OpenAI has come under fire for apparently not keeping its promises regarding AI safety. Last year, the company promised the US government that it would thoroughly test new AI models for potential problems. However, the safety team reportedly felt pressure to rush the testing of the new GPT-4o model in order to meet a set launch date. This raises questions about safety and how to deal with potential risks.

Former employees are expressing concern about the current corporate culture, which they say is increasingly putting commercial interests ahead of security considerations. Jan Leike, a former OpenAI manager, and others have criticized the company's security practices, comparing it to the Titanic.

In response to these concerns, the US government has stressed that new laws are needed to protect the public from the risks of AI. OpenAI itself has stated that it has not cut any corners in its security processes, but acknowledges that the timeframe for testing was not ideal and that more time should be allocated in the future.

Sources: Washington Post, Windows Central

... and continues to work hard on new products

OpenAI has announced the prototype of its search engine SearchGPT. SearchGPT is designed to pre-sort search results and present them in a summary. Users can then ask questions or get relevant links. The service will initially be available to 10,000 test users. Source: The Verge

OpenAI is also developing a new AI project, codenamed "Strawberry," that aims to significantly improve the capabilities of its models. For example, the AI will be able to not only provide answers, but also navigate the web and conduct extensive research. This will enable the AI to plan complex tasks and pursue long-term goals – in the spirit of an "AI agent". The company plans to present its progress in this area to the public "soon".

OpenAI also anticipates that the cost of AI models, including its own GPT family, will continue to decline as a result of optimizations and an increasing number of users. This should enable new use cases and make existing ones more cost-effective, Olivier Godement, head of API product at OpenAI, told VentureBeat.

EU AI Act enters into force

The EU AI Act, a comprehensive set of rules to regulate artificial intelligence in the European Union, has been published in the bloc's official journal and will come into force on August 1, 2024. The Act imposes different obligations on AI developers depending on the use case and risk classification, with certain high-risk applications subject to strict requirements.

The new EU regulation is seen by many startups as a potential burden, as the cost of compliance could be significant. Critics warn that the regulations will limit innovation in Europe and increase the risk of falling behind competitors such as the US and China.

Meta, meanwhile, will not offer its upcoming multimodal AI models in the European Union, citing concerns about the unclear regulatory framework. This decision could also have an impact on European companies, which will therefore not be able to access the new technologies, while Meta will remain active in other regions, such as the UK.

Training Data and Copyright: AI companies under scrutiny

The Runway Gen-3 AI video tool appears to have been trained on thousands of videos from popular YouTube creators, brands, and copyrighted films. The data includes content from channels of major media companies such as Disney and Netflix in addition to well-known influencers' videos. A former employee reported that the videos were downloaded using open source software, specifically YouTube-DL, to avoid being blocked by YouTube. Links to piracy sites are also included in spreadsheets leaked to 404 Media. Runway has not yet responded to requests for comment.

Companies such as Apple and Nvidia have also apparently used thousands of YouTube videos to train their AI models without the consent of the creators, which violates YouTube's terms of service.

Meanwhile, Apple has clarified that its "OpenELM" model is not being used to develop Apple Intelligence or other AI capabilities. It is used for research purposes only. The training data for Apple Intelligence's AI models is derived from licensed and publicly available sources.

More news in brief

AI search engine Perplexity is launching a revenue sharing program with web publishers next month. According to Chief Business Officer Dmitry Shevelenko, the program will be funded by ads placed next to search queries on the Perplexity platform. Publishers who help answer questions will receive a share of the revenue generated by the ads. Perplexity plans to launch with high-profile partners, including not only media organizations, but also individuals with WordPress sites or newsletters. Despite criticism over copyright infringement and other issues (see Smart Content Report #13), Shevelenko stresses that the company has been careful about attribution from the beginning and was working on the program before the current allegations.

EU competition authorities are investigating whether the multi-year AI agreement between Google and Samsung hinders competition from chatbots on Samsung smartphones. A questionnaire will determine whether the pre-installed Gemini Nano software restricts the ability of other generative AI systems and whether this leads to interoperability problems with other applications.

A new bipartisan bill in the U.S. Senate, the COPIED Act, aims to protect artists, songwriters and journalists from the unauthorized use of their content to train AI models. The bill would require companies to provide information about the origin of digital content within two years, creating transparency and giving creators more control over their work.

The number of data sources for training AI models is rapidly decreasing as more website operators protect their content with restrictions and terms of use. This is according to a recent study by the Data Provenance Initiative. This could pose a challenge not only to large AI companies, but also to smaller developers and research institutions that rely on public datasets.

Universities are increasingly falling behind in AI research as they lose out to large technology companies in the competition for computing resources and research funding. This is forcing them to shift their focus to less computationally intensive areas of AI. As educational institutions seek to expand their infrastructure and forge partnerships with industry, there is concern that the loss of talent to the private sector could further undermine their role in AI research.

G O O D ? R E A D S

Automated, Ad-supported Blogs

This article outlines a technologically fascinating but content-wise depressing vision of the future: a blog that delivers AI-generated content at the push of a button.

Readers of this fictitious service would a question and receive a complete blog post as an answer. What sounds absurd could become reality thanks to OpenAI's new GPT-4o mini-AI model: It is so cost-effective that such a fully automated blog could theoretically be refinanced solely by advertising revenue via Google Ads, according to the author's calculations.

He also takes a critical look at the future of online publishing: Will the Internet soon be flooded with AI content? "Websim" – an AI-generated alternative Internet environment – is presented as an example. It is intended as a fun project, but could also be seen as a dystopian preview.

See also: Dead Internet Theory

More good reads in brief

The case for Open Source AI: Meta CEO Mark Zuckerberg calls for open source AI as the new industry standard, comparing the development to the rise of Linux. With the release of the open source AI model Llama 3.1, Meta wants to enable developers to train, optimize and control their own models without being dependent on individual vendors. Zuckerberg emphasizes the benefits of open source for developers, Meta, and the world, including faster innovation, better security checks, and broader participation in the benefits of AI. However, his use of the term "open source" is controversial among experts, as there is no access to the training data, for example.

AI assistants between hype and reality: Microsoft promises to revolutionize office work through automation and content generation with AI assistants like Copilot, but the reality is different. While Copilot is good at summarizing information, it struggles with complex tasks and contextualization. In addition, there are security concerns and high implementation costs for companies that must first adapt their data structures. Despite the challenges, experts see great potential in such assistants and predict high revenues for Microsoft in the long run.

Advice on AI in content marketing: Generative AI is a hot topic in content marketing right now. But what advice on using the technology is really helpful, and what is more likely to cause harm? This article summarizes the perspectives of more than 20 experts and offers concrete insights into the do's and don'ts of using generative AI in content marketing.

Why AI fails in the enterprise: Capgemini examined why companies struggle to put AI prototypes into practice. According to the study, the biggest hurdles are a lack of data quality, unclear digital limits for AI applications, and a lack of integration of AI into the corporate structure. According to Capgemini, companies are relying on AI solutions without defining which decisions AI should influence and which data is relevant.

From art to content to slop: In his article, Nick Heer takes a close look at Ryan Broderick's essay on the phenomenon of "slop". Broderick defines "slop" as worthless, mass-produced content forced upon the consumer. Heer shares Broderick's analysis of a media landscape characterized by soulless blockbusters and content filler. Both authors see this development as a devaluation of art to an interchangeable product and a reduction of the consumer to a mere data set.

Recognizing and understanding the limits of AI: In the latest episode of The Joy of Why podcast, computer scientist Steven Strogatz discusses the limits of artificial intelligence with AI expert Yejin Choi. Despite their impressive capabilities, large language models like ChatGPT lack common sense. Choi explains how these models are trained and why they struggle with everyday knowledge. She argues for more transparency in AI development and warns of the dangers of disinformation and the potential loss of human literacy in the age of artificial intelligence.

AI does not work well as a creativity booster: A new study shows that while artificial intelligence can boost the creativity of individuals, it also reduces the creativity of groups as a whole. In the study, participants were asked to write short stories, some of which were provided with AI-generated ideas. People with low baseline creativity benefited from the AI suggestions and wrote better stories. Those with high baseline creativity, on the other hand, wrote the best stories without AI help. The AI-generated stories were also more similar and less varied overall. The researchers warn that the increased use of AI in creative fields could lead to a decrease in originality and a standardization of art, music, and literature.

C U R I O U S ? F I N D

Poof

... is the title of a funny little video created with AI tools like Midjourney, Luma Labs Dream Machine, Runway Gen-3 and Udio AI. Typical professional tools like Adobe After Effects and Adobe Audition were also used. So this one-minute clip was not just made on the side. You can learn more about the background in this "Making of".

G L O S S A R Y

Benchmark

A benchmark is a standardized test or set of tasks used to compare the performance of different AI models as objectively as possible. In the same way that a standardized test in school measures the performance of students, a benchmark enables the evaluation and comparison of the performance of AI models based on the same tasks. A benchmark includes both the tasks to be solved and a defined evaluation procedure to make the results transparent and comprehensible.