We learned quickly from speaking to clients that many names we take for granted aren’t always known outside our small world. In comparing the finer points between Google and Anthropic, we’ve been asked, “who is Claude?”
A business owner should be aware of the major players and their offerings when navigating the landscape. Decisions with real, financial consequences shouldn’t be made in darkness. Here, we will distill the major players in the AI world and offer a brief history of how they got there.
Anyone who was online in the early-2010s witnessed the rise of CAPTCHAs––those little squares you have to click on asking you to identify which parts of a photo contains a motorcycle. The name gives away its purpose: Completely Automated Public Turing test to tell Computers and Humans Apart. The idea goes back to the 1990s, but they really took off with Google’s acquisition of reCAPTCHA in 2009.
Their primary purpose was security. Bots, sans modern AI, were already rampant back then and CAPTCHAs provided an automated defense against them when submitting forms because bots of that era couldn’t do optical character recognition (OCR).
They also hid a secondary objective unbeknownst to the average user: it was helping to train AI. Humans were labeling data from Google Books and Google Street View. That labeled data formed the foundation for future pre-training of AI models.
In 2017, Google scientists wrote the landmark Attention Is All You Need research paper that introduced a new deep learning architecture, the transformer, that led to LLMs as we know them today.
Google’s current offering is Gemini.
ChatGPT
Transformers shook the scientific world, but it was ChatGPT who put AI as we know it on the map.
OpenAI was founded in 2015 as a nonprofit research lab focused on artificial general intelligence (AGI)––the kind you think of from sci-fi. It was founded by names that may sound familiar: Elon Musk, Sam Altman, Ilya Sutskever, et. al. They wrote many foundational papers around robotics, game AI, and safety.
They spent years collecting public datasets of Web data for their initial training data. GPT-3 was released in 2020 based on 600GB of public data. GPT is the model; ChatGPT is the product.
Today, ChatGPT is the largest player in the AI world with 70% market share and its name is virtually synonymous with AI itself.
Claude
Former OpenAI employees founded Anthropic in 2021 over concerns about OpenAI’s commitment to safety. Claude followed a similar trajectory to GPT, having been created from a corpus of public Web data and then growing into the models we know today. The Claude 3 family was introduced with three tiers: Haiku, for speed; Opus, a reasoning model; and Sonnet, a balance of the two.
By 2024, word was spreading throughout the software development community that Claude was a standout performer when it came specifically to writing code. It was quite possible to ask ChatGPT to generate code––many developers did, but it was often lacking in context that it needed to produce more holistic output.
Anthropic launched Claude Code as a coding assistant that autonomously plans and iterates on code by considering the entire project at once. Every AI model can produce code reasonably well, though this is hotly debated, but Claude is considered the best in class for this purpose.
Open source
None of the aforementioned models are open source; you can’t download them and run them on a server in your own office. Any data you send them is theirs. For our clients with privacy concerns, which are the majority of them, this is a non-starter for any tasks more complex than writing a spreadsheet formula.
To run an AI model entirely on-prem, you need to provide the model and the infrastructure that surrounds it entirely for yourself. This is where FindAura comes in.
There are many models in the open source space that do the job admirably. The big models try to be everything to everyone, which means they live and die by the quality of their training. Open source models can be tailored to particular business needs and require much less power to run; it’s much easier for an AI system to analyze your data when it’s not also expected to make videos.
Facebook’s LLama 3 is the most widely used family of open source models, available in many parameter sizes. Chinese models are also common in this space. Alibaba’s Qwen models, used by AirBnB, have been downloaded the most according to The ATOM Project. Kimi K2 is one of the largest models supporting the most parameters. DeepSeek-V3 made headlines in early 2025 by providing that open source weights can deliver excellent performance.
Business applications
Some small businesses and sole proprietors are fine with the big, hosted AI providers. Many are not, and are unwilling or unable to use such services given their privacy concerns. On-prem AI is also very practical; everything happens quickly and transparently. It’s also much easier to modify the product to your particular needs when you control the entire application.
We’ve worked with clients in America and Europe to bring them into the AI world and we hope to do the same for you. Stay tuned as we continue our series on AI engineering for businesses.