Anyone with an idea can use it
The Microsoft for Startups Founders Hub brings together people, knowledge, and benefits to help founders at every stage solve their startup challenges. Sign up in minutes with no funds required.
Foundation models are rapidly being adopted to create AI-based products, from image generation models (DALL-E, Midjourney, Stable Diffusion) to languages (BERT, GPT-x, FLAN). The introduction of GPT-4 has opened up even more possibilities for multimodal applications. Amidst this progress, there is a lively debate within the research community about the merits of open and closed source models. As the AI PM of Microsoft for Startups, I was privileged to work with a select group of AI-focused startups as part of the AI Grant partnership announced last year. I was struck by the rapid pace of decision-making and innovation that characterized the adoption of the foundational model.
In this series of articles, we aim to share insights from AI startup trailblazers. In this first article, we explore how startups navigate the decision of which model to utilize.
Building with Foundation Models
So what is the underlying model? The last few years have seen large-scale AI models trained on huge data corpora. Self-supervised learning is often used to enhance various downstream tasks.example is GPT-3, a large language model Any topic can be probabilistically summarized.
In foundation models, we have often seen the problem of model selection between open source and closed source models. Why is model source relevant here? Just like in the software world, models can be open source (like Stable Diffusion) or closed source (like Dall-E). Some startups consider this parameter trade-off when choosing a model.of The beginning of this conversation is grounded On topics such as responsible AI and empowering more research. Smarter people than me are refining their paradigms of choice every day. Continuing on, here are my questions for startup builders: As a user, does it really matter to you whether to choose between an open source model or a closed source model, or the fit of the model for your use case?
Observing these startups, I always find that the allegiance to the model lies in quality and fit to the source. For example, the Stable Diffusion model of open sourcing spawned a good number of these startups last year.
make the right choice for you
How do you choose the best model to use as a start-up? The output style of different models is an important consideration. I was told by one of his startups building object prototypes that the images generated from DALL-E (closed source) looked artistic, but they were generated from Midjourney (closed source). The resulting image looked like an animation. The image generated from Stable Diffusion (open source) was more realistic and better suited for the prototyping business use case than the other two. For another user creating NFTs, DALL-E may be a better choice.
Taking a step away from the image model and towards the language model, we’ve seen GPT-3 and Codex (both closed source) serve as the driving force for startups. The previously featured startup Trelent is based on a docstring generation product based on Codex. These alternatives highlight that the quality of the model is better suited for startup use cases than the potential alternatives of CodeGen or GPT-J. In parallel, GPT-3 continues to drive innovations such as: this and this As it improves, it stimulates further research in the open and closed source community.
An ever-changing AI landscape
In addition to the issue of suitable underlying models, startups are looking to use underlying models as mutual inputs and further refine their outputs. Some examples for this:
- Startup leverages GPT-3 to create prompts option For users with Stable-Diffusion based text-to-image apps.like a generator this Rapid idea generation makes it easy for users to brainstorm with AI and achieve creative results such as: In this example.
- An implementation like this Combine Cognitive Search with GPT 3.5 to power conversational AI experiences on your own data. (Both cognitive search and GPT-3.5 are closed source models.)
- Like LLM Chainer Chain — 🦜🔗 LangChain 0.0.130 We help hold AI accountable by introducing self-criticism chains to improve the quality of responses.
Whether a startup should move to an open source model or a closed source model is a question only startups can answer. Instead, I would like to change the conversation and ask the following question. In our internal benchmarks, which model is the best for our use case? As for the broader question of where this is headed, things are still changing. I can’t wait to see where the industry will converge. The possibilities are exciting.
For more tips on how to bring AI to your startup and start building an industry-leading AI infrastructure, visit Sign up today for the Microsoft for Startups Founders Hub.