- Loop
- Posts
- ❌ OpenAI abandons work on new Arrakis model
❌ OpenAI abandons work on new Arrakis model
Plus more on Amazon’s humanoid robots, OpenAI’s image detector, and robotaxis in Tokyo.
Liam
October 23, 2023
Hello,
Welcome to this edition of Loop! We aim to keep you informed about technology advances, without making you feel overwhelmed.
To kick off your week, we’ve rounded-up the most important technology and AI updates that you should know about.
In this edition, we’ll explore:
- Amazon’s humanoid robots
- OpenAI’s setback with the new Arrakis model
- Positive Carbon’s use of computer vision to reduce food waste
- … and much more
Let's jump in!
Top Stories
1. OpenAI may release a new image detector [Link]
The company has created a tool that can detect images that were made using its DALL-E 3 model. While the tool has a high accuracy rate, it hasn’t yet met OpenAI’s standards for release and there are internal debates about what constitutes an AI-generated image. Following the release of their AI-text detector earlier this year, which had to be pulled just months later, OpenAI will be treading carefully before releasing another.
2. Amazon are trialing humanoid robots in their warehouses [Link]
Digit can walk around the warehouse and place boxes onto a conveyor belt. While Amazon already use wheeled robots to move items around their fulfillment centres, these human-like robots can operate in environments that were originally designed around humans.
As a result, the trial could have a huge impact on the wider robotics industry - since the robots can work in factories and warehouses that were built decades ago. Amazon is under pressure to reduce costs and deliver items faster than they do now, which has led to a major push into the robotics industry.
3. Cruise, General Motors & Honda to launch robotaxis in Tokyo [Link]
The service will use custom-made Cruise Origin vehicles and is anticipated to launch in central Tokyo by 2026. Initially, it will start with a few dozen vehicles and later grow to 500. The joint-venture aims to eventually broaden the service area and expand beyond central Tokyo. Cruise is facing increased scrutiny in the US at both a state and federal level, where the NHTSA have recently opened an investigation.
4. Microsoft’s $69 billion acquisition of Activision has finally been approved [Link]
Microsoft has successfully acquired Activision Blizzard for $68.7 billion after a 20-month regulatory battle in both the UK and US. This is by far Microsoft’s largest acquisition to date and is part of their strategy to increase the number of titles that are made for Xbox Game Pass.
The deal also includes King, which is the mobile developer behind Candy Crush, after it was bought by Activision in 2016. The FTC remains worried about the impact this deal will have on competition.
5. Eureka! NVIDIA unveils a new AI agent that can train robots [Link]
NVIDIA’s research team has unveiled Eureka, which can teach robots how to accomplish complex tasks in a simulated environment - from spining a pen in one hand, to opening drawers, or using scissors. Their AI agent uses GPT-4 to write code that rewards robots for reinforcement learning.
Closer Look
OpenAI abandons work on new Arrakis model
At the end of 2022, researchers at OpenAI started work on a new AI model called Arrakis - with the aim of using it to power ChatGPT and make it a lot cheaper to run. Microsoft executives were also interested in using Arrakis, following their $10 billion deal with OpenAI, and wanted to use it within Bing and for new AI features in Office 365.
However, that work has now been scrapped, after the model was underperforming and failed to meet the leadership’s expectations. In recent times, both Google and OpenAI have been experimenting with making their models “sparser”. This machine learning technique means that when the model is sent a task, it will only use a number of parameters within it - not the entire model. As a result, tasks will use less computation and therefore reduce costs.
This is a clear setback for OpenAI, who are competing against Google and Meta in a fierce generative AI race. Although, it’s worth remembering that these companies are operating in completely unknown territory and the lessons learnt from this failure can be applied to future GenAI models.
You can read more here.
Startup Spotlight
Positive Carbon
This an interesting startup based in Dublin, which is focused on reducing food waste in kitchens and offices around the world. Their solution works by placing cameras above the bins and then using computer vision to categorise the objects. Kitchens can then use this data to clearly identify ingredients that they are ordering too much and plan accordingly - therefore, reducing both food waste and costs.
I saw their demo a few weeks back at TechCrunch Disrupt and it was certainly impressive. They are working with a number of companies, such as X (formerly Twitter), Dublin City University, and multiple hotel chains.
If you want to see more, you can check out their website.
Analysis
While OpenAI have scrapped work on the new Arrakis model, we’re seeing a wider move towards sparse models that can more efficiently solve tasks than what we have today. The costs of running Large Language Models (LLMs) are high and it’s been reported that Microsoft are losing up to $80 per user with their GitHub Copilot tool.
The Microsoft Research team have since been asked to build smaller, less performant models that can be used in their growing list of new AI features. This is likely off the back of OpenAI’s struggles with Arrakis and their subsequent order to stop work on it. We should expect to hear more about the topic of sparser models at OpenAI’s developer conference in November.
The other big news came from Cruise, where they announced a partnership with Honda to launch robotaxis in Tokyo. Just a few weeks back, I was at a talk in San Francisco where Kyle Vogt (Cruise’s CEO) spoke about his ambitions for the company.
To date, the self-driving company has been operating in the sun-belt area of the US, since they’re less at risk of snowy weather that can impact their vehicle’s sensors. But there’s a clear ambition to expand rapidly in the next few years, as they wrap-up development on new vehicle sensors that can handle more challenging weather conditions.
However, Cruise is under pressure in San Francisco following a number of accidents in recent months. Locals and politicians have been angered by incidents involving fire trucks, ambulances and pedestrians. The company has had to reduce its fleet of vehicles in the city by 50% as a result, but it hasn’t put a stop to their expansion plans - Vogt spoke about their partnership with Walmart to explore driverless deliveries.
While launching vehicles in central Tokyo is certainly good news for the company, they have found themselves coming under increased scrutiny in the US - and if this snowballs into a much bigger story around safety, it could scupper any hope of expanding into the Japanese market.
End Note
It’s been another busy week. We’ve got a glimpse into OpenAI’s research on sparser models, their internal debate about whether to release an AI image detector, Amazon’s pilot scheme for humanoid robots, Cruise’s plans to launch robotaxis in Tokyo, and Nvidia’s new AI agents that can train robots much quicker than before. It’ll be interesting to see how researchers start to use Eureka in the coming months and what’s built off the back of it.
Have a good week!
Liam
Share with Others
If you found something interesting, feel free to share this newsletter with your colleagues.