- Starter AI
- Posts
- OpenAI’s “Strawberry” project, Robot gets Google’s AI power, The five stages of AI
OpenAI’s “Strawberry” project, Robot gets Google’s AI power, The five stages of AI
How close are we to achieving AGI?
Hello, Starters!
A world where a whole company can be managed by AI completely is still very far away in the future. Yet, every advancement in big tech inches us closer to that AI-powered destiny... Exciting, right?
Here’s what you’ll find today:
OpenAI is working on “Strawberry”
Google mixes its Gemini model with robotics
Introducing: The five stages of AI progress
Meta’s upcoming Llama model
Are models better at memorising than reasoning?
And more.
🍓 OpenAI is working on “Strawberry” (4 min)
Last year, OpenAI made a round of buzz with insiders claiming that they were close to reaching AGI with their top-secret project "Q*. "A lot has happened since that time, but we know for sure that the company hasn't stopped working toward that goal, just that they've changed a few pathways, and now its mysterious project is allegedly called Strawberry.
The insights are minimal; we only know that models supported by Strawberry will be able to navigate the internet autonomously and that they excel at solving complex maths tasks, but there's no information about its launch or specifications.
Google DeepMind has found a new field where AI and robotics make a good match, and that is navigation. Researchers have recently released a paper called "Mobility VLA: Multimodal Instruction Navigation with Long-Context VLMs and Topological Graphs," where in simple terms, they've harnessed Gemini 1.5 Pro's capabilities to teach a robot from a former Google project to navigate and perform tasks through voice commands.
In the footage shared by DeepMind, we can see how the robot announces its "thinking" with Gemini before performing the task, which includes roaming around an office space and finding a place to draw things.
As we're stating, reaching AGI is a major goal for OpenAI, and according to Bloomberg, the company has been sharing with employees a list that includes five levels to measure AI advancements. These "five stages of AI" are mostly hypothetical as the technology is still in very early stages, but it gives a clear outlook on what things will be like in the future given its fast-paced development.
The framework shared by OpenAI looks like this: Chatbots sit at level 1, followed by Reasoners and Agents at levels 2 and 3 respectively, Innovators, which can aid with inventions, are at level 4, and the milestone is Organizations at level 5, which consists of AI that could do the work of an entire organisation. Employees have been told that AI is currently at level 2, as reasoners are related to human-level problem-solving.
🦙According to reports, Meta is launching a larger version of its Llama 3 model on July 23. The model, which is expected to boast 405 billion parameters, will be multi-modal, shifting from previous versions limited to text generation. Also, as their commitment to open source grows, they're releasing the weights of the model too.
🤖Some models tend to work through memory. A study made by MIT and Boston University, based on creating counterfactual variations in common model tasks, allowed them to conclude that the models had better performance on tasks they had memorised, instead of figuring out how to act through reasoning.
⚡️Quick links
What did you think of today's newsletter? |