OpenAI is following in the footsteps of AWS
Formula: Outpace the competition, constantly lower costs, remove risks to spur customer adoption
OpenAI’s new capabilities, announced at their OpenAI DevDay, are game changing in multiple ways. They are taking a very similar approach to what AWS did in the early days of cloud, releasing new capabilities at a blinding pace, while continually driving down pricing on their core AI services. Moreover, they are positioning themselves to become the go-to AI platform for businesses by offering increased protection and simplifying the creation of custom models.
First, let’s dive into the five key things they announced:
1. New models and lower pricing
GPT-4 Turbo is a new model that includes updated content through April 2023. It enables prompt inputs that are much larger than older models, allowing around 300 pages of text. The model is available through APIs and Chat-GPT. The API also includes a version of GPT-4 Turbo that supports images as inputs. GPT-4 Turbo also has the advantage of being 3x cheaper than GPT-4 for inputs and 2x cheaper for outputs. A new GPT-3 Turbo model also enjoys these price reductions versus the previous version. The API also now includes a new text-to-speech model, generating human-quality speech from text, available in six voice settings.
2. New Assistants API
The new API is designed to allow people to build true agent experiences, where the agent “has specific instructions, leverages extra knowledge, and can call models and tools to perform tasks”. Think of this as enabling anybody to build better, customized, and more functional versions of Amazon’s Alexa or Apple’s Siri. The new API is loaded with new capabilities: It can write and run Python code in a sandbox to help the assistant to solve complex problems programmatically. It enables native function calling for functions pre-defined by the user and simplifies data retrieval by eliminating the need to pre-compute embeddings. It also supports ‘persistent and infinitely long threads,’ which allows the API to deal with state management of the prompt threads instead of developers having to do it in their code.
3. Better support for instruction following
The new models (both GPT-4 Turbo and GPT 3.5-Turbo) provide much better adherence to detailed instructions within complex prompts, likely driven by the development of the Assistants API. OpenAI claims a 38% improvement in following format task within GPT 3.5 Turbo compared to the prior version. These features are important for making the assistant work well, and for enabling function calling to work seamlessly. GPT-4 Turbo also enables reproducible outputs, so that a given prompt will return the same output each time, which is very useful for testing.
4. Model customization
OpenAI is enabling GPT-4 Fine Tuning, but they are careful to label it as experimental, warning, “preliminary results indicate that GPT-4 fine-tuning requires more work to achieve meaningful improvements over the base model.” This capability essentially provides the option to apply any fine-tuning you have done to GPT-3.5 against a GPT-4 model within the fine-tuning console. OpenAI also introduced a Custom Models Program, which provides a dedicated team of researchers to customize a model to a specific domain. This includes “modifying every step of the model training process” - it says it is expensive, and I believe it. However, for many companies, this provides a realistic way to fine tune a proprietary model without having to do it themselves from scratch using an open-source model, which is also a very expensive approach that requires deep expertise.
5. Copyright Shield
Perhaps the most important part of the announcement was mentioned almost as an afterthought. In response to the growing market noise around copyright challenges to generative AI, OpenAI introduced Copyright Shield. Essentially, OpenAI will defend customers and pay the costs incurred if they face legal claims around copyright infringement. This is huge for any company that was sitting on the AI sidelines due to legal concerns. Between this and the repeated mantra that API and prompt inputs will never be used to train OpenAI models, a lot of the objections to using OpenAI have been rebutted.
Okay, so these are all cool, but what makes this game-changing? Well, as mentioned, OpenAI is operating very similarly to how AWS operated in the early days of cloud. In a very similar way, AWS faced concerns from customers over privacy and compliance, while locked in a constant battle against Cloud competitors with comparable platforms. On one side, there’s the risk of adoption paralysis, and on the other, the risk of commoditization, placing them in a challenging environment. AWS's response? To outpace the competition by rapidly introducing new capabilities, consistently reducing prices to undercut competitors, and enhancing security and compliance protections. They provide these assurances - often contractually - to convince corporate risk and compliance teams of their platform's safety..
OpenAI is following this exact same formula. The capabilities in this announcement leapfrog other foundation model providers and disintermediate technical platforms that wrap around generic LLM APIs. The price reductions reassure customers that the pricing structure will only continue to drop over time, making them feel more comfortable about a proprietary investment. Meanwhile, the privacy assurances and Copyright Shield provide a buffer for corporate risk.
I don’t know if OpenAI will be as successful at dominating the early stages of this next wave of computing as AWS was at dominating the early stages of the last wave, but they are certainly doing their best to repeat the patterns that worked previously. One thing that could challenge them, assuming they retain this approach, would be the ultimate lesson learned along the same historical path that OpenAI is now traveling. Early adopters of the cloud platforms learned that proprietary features can lead to vendor lock-in over time. If customers fear that same repeating pattern, they could trend toward open source alternatives, even though OpenAI currently offers a superior and well-packaged set of capabilities Only time will tell, but for now, it has never been easier or cheaper for businesses to get started with Generative AI capabilities.