First up is AI chips, since Groq made waves this past month with their LLM optimized chips with much faster inference than anything we’ve seen before. Etched is another player in this domain, but we haven’t seen a public demo yet. AI chips aren’t necessary new (eg. Graphcore, Ceerebras), and a question is, will specialized chips maintain value as model architecture evolves?
AI Video was a focus when OpenAI dropped Sora, their text-to-video model. This has blown people away by its internal ability to essentially understand complex physics. Pika, another leading text-to-video tool continues to add new features, and Alibaba also showcased an impressive demo called EMO, which adds natural body movement to lip sync (audio file > AI video). AugXLabs is the type of solution we need now to stitch these AI generated videos together in a cohesive manner.
GPT-4 is an incredibly fast way to prototype AI tools, but often more costly and slower than necessary. We’re seeing an increasing number of developers start to swap out GPT-4 calls with other models (eg. Claude for writing), and fine-tuning smaller models for more specific tasks. In this realm, two companies that seem to be picking up steam here are Together AI and Anyscale, which allows for easily fine-tuning models via API. We explored this realm with AutoFineTune, which generates synthetic data with GPT-4 based on a user input, and fine-tunes a smaller model with Together AI.
One bottleneck we’ve identified in enterprise AI is adoption – specifically having people remember to leverage AI when beneficial. One solution we’re excited about is what we’re calling Passive AI, or AI that runs in the background. The idea here is that you can set-and-forget these tools, and they will continue to create value for the users. Some examples include Cognosys, which allows users to set up AI automations with natural language, Joyn, which focuses on collecting and bucketing enterprise app data into projects, and Broadcast, which is more focused on taking unstructured data (eg., meetings) and pushing the correct updates to relevant tools (eg. Tasks added to task management system).
We’ve noticed excitement in neural interfaces the last few months. Neuralink embedded their first chip into a human. Meta has started releasing more information about their wrist based neural interface, which captures EMG signals sent from brain to hand at the wrist, which allows the system to detect hand gestures without the use of cameras. Prophetic launched MORPHEUS, which is neurotech that sends signals back into the brain to induce ludic dreams.
Big news today (adding before I send this out) was seeing the launch of Devin, an AI software engineer that seems to handle complex multi-step tasks, fixing errors, and more. One of the more impressive autonomous agents we’ve seen. Looks very impressive!
Leave a Reply