🟢 2023: The Year of AI

Reprinted with citation; will remove if copyright issues arise. Original source

AI has undoubtedly made waves in 2023 and here we spotlight the most significant stories of the year poised to shape the future of this groundbreaking industry:

Correction: In the original blog post published on December 22, 2023, the title “AI Releases” caused confusion as the content encompassed announcements and updates in addition to releases. We clarified the title of the text and infographic. The mention of Stability AI open-sourcing its LLM was excluded from the infographic but left in the article, underscoring its significance in promoting accessibility rather than focusing on tech improvement. The infographic initially featured the establishment of the xAI startup, now removed because of irrelevance. Additionally, the mention of Apple Vision Pro was excluded as the article focuses on software. We also included Midjourney V.6 in the list as it is a very recent release. These adjustments aim to improve accuracy and coherence. We apologize for any confusion and appreciate your understanding!

AI Advancements

In the landscape of AI advancements this year, notable progress was made, refining existing technologies rather than introducing groundbreaking innovations akin to the ChatGPT or image generators of the previous year. While there was no wow effect and the real Artificial General Intelligence (AGI) is still far away, this year marked an intermediate stage between prior breakthroughs and something even more powerful to come. To showcase this evolution, we crafted a visual timeline, highlighting the most remarkable AI advancements that have shaped this year of AI:

Image Generation

• Adobe Firefly: Adobe’s Firefly and Generative Fill empowered diverse visual content creation, including illustrations, art concepts, and photo manipulation. Integrated into Photoshop, Adobe Firefly democratized AI, extending its power to a broad user base at once. The release of the Text Effect feature also marked a significant stride, allowing users to apply styles or textures to words and phrases.
• Midjourney: Midjourney’s V.5 model marked a milestone in image generation, showcasing improved efficiency, coherence, and higher resolution. The latest alpha-version, Midjourney V.6, brought additional enhancements such as more accurate prompt following, increased model knowledge, and minor text drawing ability.
• DALL·E 3: Built on ChatGPT, DALL·E 3 simplified image generation, eliminating the need for complex prompt engineering. In addition, ChatGPT introduced a feature to help users refine prompts and make image adjustments based on feedback.
• Shutterstock.AI: The stock image giant integrated AI capabilities, allowing users to transform prompts into license-ready imagery. Recognizing and rewarding contributing artists, Shutterstock made the first step in ethical AI.

The Evolution of Text-to-Image Algorithms, 2007 vs 2023

Video Generation

• Stability AI: Stability AI introduced Stable Video Diffusion, a groundbreaking model for generative video, with open-source access on GitHub. Drawing a parallel to AI image generation trends, it’s highly possible that the Stable Video Diffusion model will play a pivotal role in the creation of a significant portion of AI-generated videos.
• HeyGen: AI startup unveiled a tool for voice cloning, lip movement adjustments, and language translation in videos.
• Runway Gen-2: Runway launched the Gen-2 model, enabling users to effortlessly generate full-blown videos from just text prompts, images, or other videos. Just have a look at the example below. 
• Pika and Pika 1.0: With its initial release, Pika garnered half a million users, generating millions of videos weekly. Then upgraded AI model in Pika 1.0 empowered users to create and edit videos in various styles, including 3D animation, anime, cartoon, and cinematic.
• Codec avatars by Meta: Meta’s Pixel Codec Avatars (PiCA) model for 3D human faces in videos brought us closer to photorealistic telepresence.

Text Generation

• Bard and Gemini: Google’s Bard added human-like emotion and sentiment to the chatbot landscape. Introduced into Bard chatbot and trained on a multimodal dataset, Google’s Gemini emerged as the “most capable” AI model and the closest competitor to OpenAI’s ChatGPT.
• Grok: Elon Musk’s startup xAI signaled a commitment to AI development, potentially competing with OpenAI, by unveiling “Grok” — a chatbot with humor, rebelliousness, and real-time knowledge via the 𝕏 platform. The xAI promised that Grok was designed to answer provocative questions rejected by other AI systems.
• OverflowAI: Stack Overflow’s OverflowAI enhanced knowledge curation, enabling AI-powered search for relevant answers in Visual Studio Code and Slack.
• Llama 2: Meta released Llama 2, the next generation of its open-source large language model, showcasing enhanced efficiency. Meta’s fine-tuned LLM was also optimized for dialogue use cases and outperformed other open-source models on most benchmarks.
• GPT-4: OpenAI’s GPT-4 now handles image input, generates captions, classifications, hears, and responds in a back-and-forth conversation, and supports real-time web browsing. OpenAI also extended support for plugins, fostering a landscape enriched with open-source competitors. GPT-4 is the next step in OpenAI’s journey to develop AGI.
• Mistral 7B: Mistral AI, valued at around $2 billion this year, released Mistral 7B, a large language model challenging GPT-4 and Claude 2. Emphasizing an open technology approach, Mistral AI offered its model for free download.
• Mixtral 8x7B: Mistral AI also introduced Mixtral 8x7B, a high-quality sparse mixture of expert model (SMoE) with open weights, featuring 46.7B total parameters, pioneering openness in models with enhanced truthfulness and reduced biases.
• Yi-34B llm: Valued at $1 billion this year, Kai-Fu Lee’s startup 01.AI released Yi-34B — an open-source neural network that outperformed competing models with significantly higher parameter counts, emphasizing its cost-efficiency.

Other Advancements:

• Segment Anything Model (SAM): Meta AI presented SAM, a segmentation model capable of “cutting out” objects in images without additional training, underscoring its adaptability. SAM was trained on a vast dataset, showcasing its robust performance in object segmentation.
• Direct Preference Optimization (DPO): DPO emerged as a stable and efficient method for fine-tuning large-scale unsupervised language models and teaching text-to-image models. It achieved precise control without complex reinforcement learning from human feedback (RLHF).
• Zephyr Direct Distillation of LM Alignment: Zephyr-7B, a result of distilled direct preference optimization (dDPO), set the benchmark for chat models with 7B parameters, enhancing intent alignment without extensive training.
• Autonomous AI Agents: Autonomous AI agents emerged as a notable trend, showcasing a transformative shift toward advanced and autonomous AI systems. AI Agents are considered a first glimpse of AGI as they can generate self-directed tasks and instructions based on a user’s goal, and work on them autonomously until the goal is achieved.
• EvoDiff: Microsoft’s EvoDiff, an open-source AI framework for fast and cost-saving protein generation, promised advancements in therapeutics and industrial applications.
• Stable Audio: Stability AI launched a tool for generating short high-quality audio clips from simple text prompts.
• GPT Store, Copyright Shield, ChatGPT Bot Constructor: OpenAI introduced the GPT Store to sell custom GPT bots, Copyright Shield to cover legal costs related to copyright infringement claims, and a no-code platform for custom ChatGPT versions.
• Stability AI Open-Sourced its LLM: Stability AI has open-sourced its models, StableLM-Alpha and Stable Vicuna, renowned for their impressive performance in generating text and code. Stable Vicuna is the first open-source chatbot trained using reinforcement learning from human feedback (RLHF). Furthermore, Stability AI unveiled SDXL Turbo, a real-time text-to-image generation model.

Partnerships

In the dynamic realm of 2023, significant collaborations have surfaced among industry leaders, shaping the trajectory of the future. Here are the top merges and partnerships that were defining the AI landscape in this year 2023:

Stability AI and Init ML

Stability AI has made a significant move by acquiring Init ML, the brains behind the popular editing app ClipDrop. The objective was clear: integrate Stability AI’s advanced technologies into ClipDrop’s ecosystem. The collaboration has already resulted in the development of SDXL Turbo.

Runway and Getty Images

Runway has joined forces with Getty Images in a strategic partnership to introduce a new video generation model RGM (The Runway and Getty Images Model). The model combines Runway’s AI capabilities with Getty Images’ licensed creative content library. The collaboration aims to revolutionize content creation workflows, enabling companies to generate high-quality, customized videos tailored to their brand identities.

Snowflake and Neeva

Snowflake, a major player in the data warehouse platform, has acquired Neeva, a startup known for using generative AI to enhance the search experience. Neeva had recently closed its subscription-based, ad-free search engine. The founders of Neeva also acknowledged the challenge of convincing users to try a new search engine.

Shutterstock and OpenAI

Shutterstock and OpenAI have committed to an extended 6-year partnership. OpenAI gained access to high-quality data from Shutterstock, enriching its model training datasets with a diverse range of images, videos, and music libraries. Shutterstock continued to leverage OpenAI’s technologies, leading to the launch of Shutterstock’s AI image-generating tool.

Legal Landscape

In the ever-evolving legal realm of AI, 2023 finds itself amidst a landscape filled with uncertainties and ongoing debates. As new challenges emerge, discussions surrounding copyright, corporate policies, and the broader regulatory framework continue, shaping the contours of AI’s legal landscape. Here are the most important legal issues of the year 2023:

European AI Act

The European Union introduced the AI Act, the world’s first comprehensive law, to regulate the use of AI. The act classifies AI systems based on the risk they pose and sets forth regulations accordingly. Although the AI Act has been provisionally agreed upon, its implementation faces delays, and the enforcement won’t commence until 2025.

U.S. Copyright Office Stance on Registration of AI-Generated Content

The U.S. Copyright Office took a decisive stance, denying copyright registration for images created by the AI algorithm Midjourney. The rejection set a precedent, asserting that AI artworks solely created by AI, without human involvement, are ineligible for copyright protection. In the same vein, the U.S. Copyright Office issued guidance on AI-assisted works, clarifying that works created by humans using AI tools may be eligible for copyright protection. The guidance confirmed that works created by humans using AI tools should be evaluated based on whether the human role in the creation of those works was determinative.

“Currently, the existing legal system is not prepared to acknowledge copyright for works created with AI, given that AI learns from existing data, the rights to which belong to other people, challenging the attribution of ownership. The practice for addressing this issue is expected to develop next year, facilitated by public participation through state-conducted surveys. Resolving this matter independently is now difficult without broader public engagement.”

Daria Kuznetsova, Corporate Lawyer of Everypixel

McKinsey also released a comprehensive graph capturing the most important AI governance-related policy and regulatory efforts in 2023. The visual representation highlights the significant contributions of 2023 in shaping the legal landscape of AI.

Source: McKinsey

Debates

The year 2023 was abuzz with intriguing debates and discussions, grappling with uncertainties and the evolving norms of the AI landscape. As the industry shapes its course, these debates become inevitable, promising more thought-provoking dialogues and challenges on the horizon. Here are some of the most noteworthy debates that defined the year:

Corporate Restrictions on ChatGPT

Major financial institutions, including JP Morgan, Citigroup, Bank of America, Deutsche Bank, Goldman Sachs, and Wells Fargo & Co, have restricted ChatGPT usage due to security and privacy concerns. This reflected a broader trend where companies were issuing warnings to employees about the legal considerations associated with AI applications in corporate environments.

OpenAI’s Use of Low-Paid Workers

Time’s investigation exposed OpenAI’s collaboration with Sama, employing low-paid workers in Kenya to sift through sensitive content for ChatGPT. The revelation raised ethical questions about the treatment of workers and the impact of content moderation on mental well-being.

Leadership Transition at OpenAI

Sam Altman’s departure and quick return made headlines last month. A leadership transition unfolded at OpenAI as Sam Altman stepped down amid communication inconsistencies with the board. Interim CEO Mira Murati, along with a majority of staff, advocated for Altman’s return. This unprecedented situation attracted widespread attention, leaving questions about the true reasons behind the transition and future implications.

Adobe and Figma

Adobe’s $20 billion acquisition plan for Figma encountered regulatory hurdles, prompting investigations by the European Commission and the UK Competition and Markets Authority over potential antitrust issues. The proposed deal’s impact also extended beyond design considerations, as Adobe’s dominance in customer data platforms raised concerns among Chief Information Officers (CIOs) about its potential influence on cloud software spending. However, Adobe abandoned the deal due to challenges in securing antitrust approvals in Europe and the UK, resulting in a termination fee of $1 billion to Figma.

Photographer Hacked the World Photography Awards

Photographer Boris Eldagsen disrupted the Sony World Photography Awards by submitting AI-generated artwork. Eldagsen’s refusal to accept the prize sparked a debate on the place of AI-generated images in traditional photography competitions, challenging perceptions of authenticity and creativity.