A year ago, OpenAI took everyone by surprise with the launch of ChatGPT. It emerged as a surprisingly powerful tool, quickly finding its place in the market. Its rapid spread, reaching 100 millions active users in just 2 months compared to Netflix's 10 years, Facebook's 4.5 years, and TikTok's 9 months, not only set a historic record but also spurred a revolution in the AI field. This revolution forced tech giants to hurriedly release their own models, to immediately and broadly integrate AI into their products, and rejuvenated the startup universe, opening a new chapter in AI history.
GenAI Usage Evolution
Launched in November 2022, ChatGPT quickly captured the attention of the general public, positioning itself as a revolutionary artificial intelligence tool. Its ability to provide elaborate answers, from writing texts to creating poems, sparked enthusiasm among users. However, its impact also raised debates. Concerns about the integrity of information (the famous hallucinations) and the risk of job loss in creative sectors led some companies to limit its use. This situation reflects the complexity of the general public's reception of generative AI, between enthusiastic adoption and caution about its implications.
Since its launch, ChatGPT has significantly influenced the business world. Its ability to generate coherent and contextually relevant texts has opened new paths for automation and improving business processes. Tools like GitHub Copilot, based on a modified version of GPT, have experienced rapid commercial success. With over three billion lines of code generated, Copilot immediately impacted developer productivity, with developers accepting nearly 30% of the tool's code suggestions on average! Nonetheless, the integration of AI into businesses raises ethical and practical questions, particularly in terms of data reliability and security. Companies must balance the benefits of increased efficiency with the potential risks associated with the use of these advanced technologies.
Over the past year, ChatGPT has profoundly impacted the global political landscape. Concerns have been raised worldwide about its potential to disrupt elections, threatening societal stability and respect for human rights. A call has been made for international governance of AI, highlighting the importance of coordinating efforts to ensure respect for human rights and prevent disinformation campaigns.
In Europe, generative AI is seen as a competitive advantage, with McKinsey forecasts indicating a potential increase in productivity of several trillion dollars. This perspective highlights the economic opportunity, but also the need to carefully manage the political and social implications of AI.
In the United States, the impact on elections is already visible, with generative AI being used by political committees to create campaign videos, offering new and rapid means for precise voter targeting. This development raises concerns about democratized disinformation, the creation of false narratives, and the lack of regulations or mandatory disclosures to protect voters against false information. The ability of AI to generate targeted messages and influence public opinion means that traditional political strategies could be redefined, with potential consequences for democracy and governance.
Over the past year, OpenAI has significantly evolved technologically, particularly with the transition from GPT-3.5 to GPT-4 and the introduction of GPT-4 Turbo. These advanced models offer improved accuracy, creativity, and collaboration, with extended capabilities such as analysis of long texts and images. The company has also innovated by allowing developers to customize their own versions of GPT for specific tasks, via Custom GPTs and the GPT Store, while introducing the assistant API for the creation of sophisticated AI assistants. It should be noted that these innovations come with exorbitant costs. The training cost of GPT-4 was about $63 million, reflecting the computing power and training time required. Additionally, inference on GPT-4 costs three times more than on a model like Davinci with 175 billion parameters, due to larger clusters needed and lower usage rates. The inference architecture of GPT-4 operates on a cluster of 128 GPUs!
This year OpenAI has seen remarkable commercial growth bolstered by an expanded partnership with Microsoft, which included a multi-billion dollar investment. This evolution into a hybrid organization, balancing commercial pursuits with its non-profit foundation's goals, led to internal conflicts. This tension peaked with the controversial firing and subsequent rehiring of CEO and co-founder Sam Altman. This incident attracted significant attention from the media and industry watchers, benefiting competitors according to HuggingFace CEO.
OpenSource LLM Evolution
The ChatGPT Big Bang
Since its launch in November 2022, ChatGPT has caused a real "Big Bang" in the Research LLM domain, redefining what is expected from a chatbot in terms of capabilities and interactivity. This conversational chatbot has surpassed its predecessors like Cortana, Alexa, or Siri, offering advanced features such as composing music, generating business proposals, and even complex programming. Its ability to interact at a human level has been a key factor in its success, distinguishing it from previous tools that could not converse as effectively. ChatGPT has also demonstrated a remarkable ability to process and summarize noisy or complex texts, often surpassing human capabilities in this area. This sudden evolution has stimulated a wave of innovations and developments in the AI sector, pushing companies and developers to explore new applications and push the boundaries of what is technologically possible.
Meta and Google's Response
In response to the advent of ChatGPT, Meta and Google, whose open-source deep learning contributions significantly aided OpenAI, expedited the release of their own LLMs. Meta unveiled LLaMa, mainly to enhance online advertising efficiency. Google, meanwhile, not only launched Bard as its conversational AI solution but also introduced Gemini, a true contender in the LLM arena. Announced on December 6, 2023, Gemini comprises three models: Gemini Ultra, Gemini Pro, and Gemini Nano. Touted as Google's most capable AI model, Gemini Ultra was designed for highly complex tasks and was noted for its ability to outperform GPT-4 and other major LLMs on various benchmarks, including the Massive Multitask Language Understanding test. Gemini's unique multimodal capability allows it to process text, images, audio, video, and computer code simultaneously, marking a significant advancement in the field. With Gemini, Google aims to redefine the landscape of large language models, combining the strengths of its previous AI endeavors with new, cutting-edge capabilities.
In Europe, and particularly in France, challengers such as Mistral and LightOn have emerged in response to the ChatGPT phenomenon. These initiatives, although smaller in scale, have shown an innovative and open-source-oriented approach. These projects reflect a desire to diversify the AI landscape, with an emphasis on ethical and transparent approaches. Their progress demonstrates that innovation in the LLM domain is not only the domain of technological giants but can also emanate from smaller, dynamic players.
In just one year, the ChatGPT application has not only revolutionized the general public's access to AI but has also initiated significant changes in the business and political sectors. The global awareness that AI is no longer just a research topic but a fully operational tool raises numerous business, political, and societal questions.
The exorbitant costs of developing such AI risk leading to a polarization and dependence on a few tech giants.
The regulatory delay of our societies to understand and frame these powerful AI cannot be bridged if the main players do not play the game of transparency with their increasingly large and opaque models.
Finally, the question of data security, in this context of opacity and lack of user control, remains entire!
In this context, it becomes imperative to explore solutions that offer a balance between technological innovation and respect for data privacy and security.
At Craft AI, we believe in an enterprise-ready technology to make the most of generative AI:
- Private: the application runs on secure and dedicated servers and is hosted in the cloud of your choice, the data does not leave your company
- Customized: we choose the most relevant open-source model and adapt it to integrate corporate data into its responses
- Hallucination-free: the LLM is configured to provide sources for all its responses, allowing for result verification and improving user trust
- Optimized: we control your infrastructure costs by improving your GPU utilization
- Managed: we set up all the monitoring of the usage, performance, and toxicity of your LLMs