🤼 DALL-E API and the Open Source Model vs. API Debate
Weekly news digest curated by the industry insiders
There is no other possible topic for this week’s editorial than the release of the highly anticipated OpenAI DALL-E API. The doubts about whether OpenAI was going to enable programmatic access to DALL-E due to ethical concerns were finally put to rest with the release of the DALL-E API Beta. Now developers can integrate DALL-E into their applications using a very simple programming model. Seems pretty obvious that the release of the DALL-E API was catalyzed by Stability AI’s bold move of open-sourcing Stable Diffusion, which quickly gained incredible traction within the generative AI community. Similarly to the path followed with models like GPT-3 and Codex, OpenAI decided to enable API access to DALL-E but not open-source it like Stable Diffusion.
The friction between controlled API versus fully open-source distribution mechanisms will likely be at the center of the generative AI debate for the next few years. API models enable greater control and simpler mechanisms to enforce the fair and ethical behavior of the model. However, centralized API vehicles almost guarantee that these models’ power will remain in the hands of large tech companies, which are likely to capture even more relevant data from the API access. At the same time, the fine-tuning and modification of large generative models is highly impractical via APIs. Open-sourcing large generative models removes the control barriers and enables broader levels of customization and fine-tuning but also opens the door to the unethical use of these models.
The OpenAI versus the Stability AI way? Controlled adoption and strong ethical boundaries versus faster innovation and trustless distribution? These are just a few interesting dilemmas facing the generative AI community in the next few years.
🔺🔻TheSequence Scope – our Sunday edition with the industry’s development overview – is free. To receive high-quality content about the most relevant developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻
🗓 Next week in TheSequence Edge:
Edge#241: we conclude our text-to-image series discussing the emerging capabilities of text-to-image synthesis models; explain NVIDIA’s textual inversion approach to improving text-to-image synthesis; explore DALL-E and Stable Diffusion Outpainting Interfaces.
Edge#242: we overview Meta AI's new model that can solve international math Olympiad level problems
📌 Galileo Demo Hour – Introducing Instant Data Debugging for NLP*
Join us for Galileo Demo Hour to learn how NLP practitioners are using Galileo to make data debugging 10x faster. We will kick off this Demo Hour by debugging data using a popular public dataset across training and production models.
Now, let’s review the most important developments in the AI industry this week
🔎 ML Research
AI and Math Reasoning
Meta AI published a paper detailing HyperTree Proof Search (HTPS), a model able to solve problems from the International Math Olympiad →read more
Controlling GAN Outputs
Amazon Science published a paper detailing a technique to better control the attributes of the output of GAN image generators →read more
Knowledge Reusability and Reinforcement Learning
Google Brain published a paper proposing Reincarnating RL (RRL), a technique that allows enabling the reusability of learned models or policies between reinforcement learning agents →read more
🤖 Cool AI Tech Releases
OpenAI unveiled the Beta version of the DALL-E API that enables photorealistic image generation from textual inputs →read more
Code as Policies
Google Brain open-sourced Code as Policies, a formulation language that can translate natural language into robotic instructions →read more
New TorchVision’s Capabilities
PyTorch added new capabilities to its TorchVision Transforms API including object detection, segmentation and video tasks →read more
ESM Metagenomic Atlas
Meta AI published an exploration tool and dataset that includes the structures of hundreds of millions of proteins in the metagenomic world →read more
🛠 Real World ML
Improving Instagram Notifications
Meta AI shares the causal inference and statistical techniques used to improve notification management on Instagram →read more
Optimizing Notification Timing at Uber
Uber also shared the ML architecture and methods used to optimize the timing of push notifications for Uber Eats →read more
💸 Money in AI
OpenAI Startup Fund announced a new program called Converge to provide early-stage AI startups with capital and access to OpenAI tech and resources. November 25 is the deadline to apply.