Discover more from Artificial Ignorance
OpenAI's top competitors
A look at OpenAI's dominance and the rivals looking to dethrone it.
Four months ago, I had the opportunity to write a guest post foron the organizations going up against OpenAI. Here’s the original post, and I highly recommend subscribing to Michael’s newsletter.
But a lot can change in four months. One of OpenAI’s best-funded competitors wasn’t even public when I originally wrote the post, and others have made huge announcements on product and funding. So I’m revisiting the article again today, with updated stats and a more nuanced perspective.
In a recent interview, Sam Altman, the CEO of OpenAI, described the company’s struggles early on.
“We have been a misunderstood and badly mocked org for a long time. When we started … people thought we were batshit insane,” he said. “We don’t get mocked as much now.”
OpenAI and its flagship product, ChatGPT, have astounded the world. After becoming the fastest-growing consumer product in history, ChatGPT kicked off an AI arms race and is now in the crosshairs of governments worldwide.
But OpenAI is a lot more than ChatGPT. Few people are familiar with the company’s history or its diverse offering of machine learning models. Even fewer people are familiar with the company’s competitors. So let’s take a look at OpenAI’s history, its current lead in AI, and the companies working to catch up.
A brief history of OpenAI
OpenAI began in 2015 as a non-profit with several founders, including Sam Altman and Elon Musk. Even before starting OpenAI, both Altman and Musk had publicly expressed concerns about AI. Musk called it humanity’s “biggest existential threat,” though he would later leave OpenAI due to “potential future conflicts” with Tesla.
In the year that followed, OpenAI released its first two products without much fanfare. But in 2018, the company released a paper introducing a new kind of model: a Generative Pre-trained Transformer, or GPT. While not obvious at the time, the GPT model radically changed OpenAI’s trajectory.
After the GPT paper, OpenAI built several models based on the architecture. With the second release, GPT-2, OpenAI decided to withhold the model code and weights, in contrast to its previous releases. The move surprised the AI community, but it foreshadowed the company's future restrictions. Today, OpenAI’s chief scientist has completely soured on their original research-sharing approach. “We were wrong. Flat out, we were wrong,” he said.
Shortly after GPT-2, the organization announced it was creating a "capped-profit" company. The new entity would allow investors and employees to earn a “capped return” up to 100x. The creation of the “capped-profit” company soon led to a partnership with Microsoft. In 2019, the tech giant announced a $1 billion investment, as well as several joint initiatives. And by 2023, the companies announced a new multiyear, multibillion-dollar investment.
Fast forward to today: the company has almost two dozen AI models/products across text, images, and audio. It is firmly in the lead in generative AI and is on track to impact hundreds of millions of users. Its partnership with Microsoft has led to AI upgrades for Bing, Github, and Office 365. ChatGPT is the first serious threat to Google's search dominance in years. GPT-4 is so advanced it triggered a call for a pause on continuing AI development. And multiple governments are now scrutinizing the potential impacts of OpenAI and ChatGPT.
A look at OpenAI’s products
Before looking at OpenAI’s competitors, it’s useful to understand OpenAI’s products. At this point, OpenAI has released a wide array of models. Taken together, they are one of the most advanced collections of AI products from a single company. The common thread is language - using text to generate conversation, code, images, and insights.
ChatGPT, GPT-4 & Plugins
ChatGPT is OpenAI’s most successful product to date by a long shot. The product builds on the company's previous large language models, GPT-2 and GPT-3. But ChatGPT is designed for conversation, not general text completion. To achieve its high-quality responses, OpenAI invests significant resources into RLHF (reinforcement learning from human feedback). RLHF involves humans giving the AI feedback on its responses to improve it little by little.
The first version of ChatGPT used an internal model named GPT-3.5, more advanced than GPT-3. Shortly after releasing ChatGPT, OpenAI released GPT-4, its most sophisticated LLM to date. While still in beta, GPT-4 is extremely powerful as a language model. It has a much more nuanced understanding of language, scoring in the top 90% percentile on mock bar exams. GPT-4 is also multi-modal, meaning it can work with image inputs. In an impressive demo, GPT-4 explains a funny image, then builds a working website from a napkin sketch. It's still early days, and we're still scratching the surface of what's possible with GPT-4.
The third ChatGPT-related product is the plugin platform, which is currently in alpha. Plugins give ChatGPT additional abilities such as browsing the web or running code, as well as access to third-party APIs from hundreds of companies. OpenAI has said it plans to build ChatGPT into the ultimate smart assistant, and plugins are clearly a way to grow its capabilities.
GPT-3 & Embeddings
ChatGPT and GPT-4 have stolen the spotlight, but OpenAI has several other text-related products. GPT-3, the previous flagship LLM, is a general text-completion model: start a sentence, and watch GPT-3 finish it for you. That sounds simple now, but GPT-3 produces far more advanced text than any other LLM that came before it. It can also be fine-tuned on specific examples, which isn't yet possible with ChatGPT and GPT-4.
OpenAI also has an embedding creation model. Embeddings are numerical representations of text - they're a way to measure the "relatedness" of documents. They have several practical use cases, such as search, recommendations, clustering, and classification. Once created, embeddings can be stored in a vector database, such as Pinecone or Weaviate, for further analysis.
DALL-E 2 is a model capable of generating images from a natural language description. While the first version was slow and produced grainy images, V2 has had much more success. At the time, V2's launch was quite impressive in the text-to-image space. But since DALL-E 2's release, competitors like Midjourney and Stable Diffusion have taken the lead on AI image generation.
On the audio side, Whisper is a speech recognition model trained on over 600,000 hours of audio data. OpenAI offers APIs for Whisper transcription and translation.
If this wasn’t enough, OpenAI has released several different models that haven't yet made it into finished products. They include:
Codex: a GPT-3 variant fine-tuned to generate code. Codex powers GitHub Copilot, the coding assistant and autocomplete tool. While Codex used to be available as a product, it is being discontinued by OpenAI.
MuseNet: a neural network to compose short musical pieces with up to ten different musical instruments in any genre. It doesn’t “understand” music but rather uses a GPT architecture to predict the next note in a MIDI file.
Jukebox: another neural network meant to generate music. In contrast to MuseNet, Jukebox generates music (and vocals) as raw audio data, not MIDI files.
That's a lot for one company. But OpenAI is far from alone in developing language models.
Artificial Ignorance is reader-supported. If you found this interesting or insightful, consider becoming a free or paid subscriber.
Who is competing with OpenAI?
There’s no one company that does everything that OpenAI does. While there is plenty of competition for each product, few companies compete with OpenAI across the board.
The closest competitor to OpenAI is Google DeepMind. Founded in 2010, DeepMind is an AI research lab focused on building general-purpose learning algorithms. The company has repeatedly made headlines over the years, including when its AlphaGo AI beat world champion Go player Lee Sedol. Another big release was AlphaFold, a protein-folding AI that has predicted over 200 million protein structures to date.
Besides AlphaGo and AlphaFold, DeepMind has an impressive list of models (though few finished products):
Flamingo: a visual model that is being used to help generate descriptions for YouTube Shorts.
Gemini: a model using both GPT and AlphaGo techniques that is meant to rival ChatGPT.
WaveNet: a text-to-speech model. While originally too difficult for consumers, it became WaveRNN, which now powers Google Assistant and GCP’s Cloud Text-to-Speech.
AlphaStar: an AlphaGo descendant that plays Starcraft 2. In 2019 it reached Grandmaster level on the public ladder.
The company has as much experience as OpenAI, if not more, with developing practical machine learning models. But a big difference is the types of products that it builds with them - unlike OpenAI, very few are available as consumer products. This likely has something to do with the company's acquisition history - in 2014, Google acquired DeepMind. Until a few months ago, the team operated separately, building its models for Google products. But ChatGPT's success triggered the merging of Google Brain and DeepMind and a new focus on consumer products.
Google, as a whole, is a fascinating competitor to OpenAI. On the one hand, the company is an absolute AI powerhouse. It's released thousands of AI research papers, including the Transformer paper that was directly responsible for today's GPT architecture. It has several internal teams working on AI. It has billions in resources (both cash and compute), meaning it can afford to build state-of-the-art models to compete with OpenAI. And it's clear Google is trying to compete - this year, it released Bard, is rolling out Duet inside of Docs/Sheets/Gmail, and is developing a laundry list of generative AI features from Search to Android.
On the other hand, the company is gun-shy when it comes to releasing AI products to the public. There are likely several reasons why, but Google has a history of taking its responsibilities seriously as a steward of AI technology - its AI policy site has years worth of principles, guidance, and policy recommendations for AI safety. There may also be some scar tissue at Google regarding AI. In 2018, it showcased Google Duplex, a language and voice technology years ahead of its time. People freaked out. And while it did launch eventually, the rollout was slow and limited in scope. Duplex was discontinued in 2022 - 6 months before ChatGPT launched.
Regardless of the reasons, Google still represents the biggest competition to OpenAI. On paper, it has everything it needs resource-wise. Time will tell whether the company can overcome its organizational challenges and keep up.
One of the most often mentioned OpenAI competitors is Anthropic, despite having only been founded in 2021. It's often cited for two reasons: first, the founders are ex-OpenAI employees who disagreed with OpenAI's approach to AI safety. Second, it built one of the first ChatGPT competitors: Claude.
Claude, which is still in beta, puts a strong emphasis on ethics and safety. The chatbot is trained as a "Constitutional AI," which is a method of helping language models against adversarial prompting. Constitutional AI training involves several rounds of feedback using both humans and AI. Claude (which just launched V2 this week) is not quite as advanced as GPT-4, but does offer a version with a 100K context window.
After Google, Anthropic is OpenAI’s best-funded competitor. The organization has raised over $1 billion to date, including $300 million from Google. And it's not slowing down - it plans to raise as much as $5 billion over the next two years to take on OpenAI and enter over a dozen major industries. It wants to build a model (codenamed “Claude-Next”) ten times more capable than the current state-of-the-art.
And the company also routinely punches above its weight on regulation and policy, despite being such a new company. The CEO has engaged in discussions with the Biden Administration, the Prime Minister of England, and Congress on AI safety.
When I first wrote this post, Inflection AI wasn’t on my radar - it wasn’t on anyone’s radar. Helmed by LinkedIn founder Reid Hoffman, DeepMind founder Mustafa Suleyman and DeepMind researcher Karén Simonyan, Inflection AI came out of stealth in May 2023 and debuted its chatbot Pi - another ChatGPT competitor.
Pi is more of a conversation partner than personal assistant, but Inflection is planning on improving its capabilities over time. And it’s got the resources to do so - in the two months since launching, the company has raised $1.3 billion. What’s remarkable about this funding, though, is how much of it is explicitly earmarked for GPUs. The company is building the largest known GPU cluster for AI applications (so far), with 22,000 H100s in a single cluster.
It’s too early to say whether Pi will prove to be a significant rival to ChatGPT, but it’s hard to overlook Inflection AI given their founding team and gargantuan war chest.
Cohere is an AI research company building language models for companies, not consumers. It was founded in 2019 by AI researchers, including one of the authors of Google’s Transformer architecture paper. The company has several different products for working with text:
Summarization, to extract insights from text and documents.
Generation, similar to GPT-3, to write ads, blog posts, and product descriptions.
Classification, to group text into categories or identify hate speech.
Embeds, similar to embeddings, to search and cluster input text.
Neural search, to perform semantic search across text content.
Command prompting, similar to GPT-3.5, to turn instructions into generated text.
Unlike nearly all of the other competitors here, Cohere does not offer a chatbot competing with ChatGPT. However, it does have a diverse, advanced suite of products. Unlike OpenAI, it’s targeting enterprise users and, as such, is focusing on high-performance, secure language models. The company has raised $445 million to date and is in talks to raise additional funding at a $2+ billion valuation.
While OpenAI has stopped open-sourcing its models, other companies are continuing that approach. One of the most prominent is Stability AI, which released the image generation model Stable Diffusion (and later Stable Diffusion XL). Stable Diffusion is one of the most advanced text-to-image models available, beating DALL-E in image quality. And the model is fully open source, allowing anyone to download and run it locally. The company has also released DreamStudio and StableStudio, paid and open-source versions of its text-to-image UI.
Stability AI has also released StableLM, an open-source ChatGPT alternative. StableLM is a family of LLMs, with alpha releases available on GitHub. StableLM is a great step toward accessible language models, but the model still needs refinement and testing. And lastly, the company has said it plans to release video-generation models later this year.
In the last few months, Stability’s future has become somewhat uncertain. The company has had a string of bad press and executive departures, and has reportedly been having difficulties securing additional funding (which the company denies). Previously, the Stability AI had raised over $100 million.
EleutherAI is a nonprofit AI research lab and perhaps the true heir to the “Open” AI name. The organization grew out of a Discord server for researchers and enthusiasts to discuss GPT-3 back in 2020, and the researchers quickly decided to build an open-source alternative to GPT-3. In the years since, they have released several significant open-source datasets and machine-learning models:
The Pile: an 825GB language modeling text dataset, mostly from academic and professional sources.
GPT-Neo: EleutherAI’s first attempt at an open-source GPT-3. GPT-Neo is a series of LLMs trained on the Pile dataset, with 2.7 billion parameters (vs. 175 billion parameters for GPT-3).
GPT-J/GPT-NeoX: additional publicly available model built to approximate GPT-3, with the largest versions using between 6 and 20 billion parameters. Upon release, the models still required fine-tuning, and EleutherAI stated they should not be used for human-facing interactions.
In its early days, EleutherAI relied on donations and contributions from cloud platforms in order to build and train its models. However, the group eventually ran into the harsh realities of modern LLM development - the scale of resources needed far outstripped what an amateur group of researchers could assemble. And in 2023, EleutherAI announced its transition to a nonprofit research institute, along with sizable donations from Stability AI, Hugging Face, and others.
Four months ago, Hugging Face was not the most obvious challenger to OpenAI. They’re primarily known as the GitHub for machine learning models - a platform for hosting, training, fine-tuning and deploying models. The company also offers over 200,000 open-source models and 40,000 datasets to help train them.
Previously, I said:
Right now, the company is focused on AI infrastructure, not products. But as open-source models get better, Hugging Face will be well-placed to offer more than just hosting and tooling.
And in fact, the company has announced a series of in-house models and a ChatGPT clone, HuggingChat. There’s already a big debate on whether open-source models or proprietary ones will win out in the end, and Hugging Face is one of the leading companies on the open-source side.
But even without its own products, Hugging Face will launch the next wave of OpenAI rivals by lowering the barriers to train ML models. To date, Hugging Face has raised $250 million in funding, and has secured a partnership with AWS.
Where we're headed
These are the biggest competitors to OpenAI today - but the AI landscape moves at a breakneck pace. ChatGPT showed the world generative AI's potential and opened a Pandora's box of AI products and companies.
The generative AI space has gotten incredibly competitive, with Microsoft, Google, Meta, Amazon, Apple and Snap announcing various AI products and models. And on the startup side, investors are pouring billions into generative AI startups.
It's still early days - we've barely scratched the surface of general models for text, image, and video generation. And more research will lead to more ChatGPT and Stable Diffusion alternatives. Not to mention the thousands of use cases where tailored models will make the most sense, from finance to construction to design. The current Cambrian explosion of AI startups will most likely continue for a while.
All that said, OpenAI is in a good spot to maintain its lead. Its biggest threats are from big tech companies, though it has years of experience and momentum at this point. And its partnership with Microsoft makes it unlikely to be outgunned by a new startup. Whether OpenAI gets acquired or becomes a tech giant in its own right remains to be seen. But we're in for a lot more technological progress in the coming months and years.