Research shows why dog robots powered by ChatGPT is not a good idea | Google announced breakthrough in Quantum computing (#4)
Open AI made SORA available leaving Europe in this week in AI advancement, policy, and safety.
Hey! Welcome to Newsletter #4. Please head over to the section you are interested in.
AI advancement
Product
Research
AI governance and policy
AI safety
AI advancement
Product
Open AI text to video SORA is public
The text-to-video model Sora from Open AI, first announced in February this year, is finally open for public use. The model is available only to paid subscribers of chatGPT.
You can customize the aspect ratio, resolution, and length of the video. The max length of the video is 20s.
They also announced a Storyboard feature, where you can create a timeline with different cards at certain intervals. The model will basically generate videos to fill in any gaps.
The videos are still far away from normal. This is one of the videos on the OpenAI website featuring Sora. Watch this video and let me know if you can spot something weird!!
We should get ready to see a flood of AI generated videos coming for us. The charm of AI generated images have gone down quite a bit after a brief period of interest, primarily due to them having a very specific style. Let’s see if the SORA will keep evolving to create newer and newer styles of video or if it will settle in a similarly stale state as images for some time.
ChatGPT goes Pro
OpenAI released ChatGPTPro mode, a $200/month model with unlimited access to the best OpenAI models and advanced voice mode.
The o1 pro mode uses extra compute at inference time to utilize model’s chain-of-thought at the compute time.
I was a bit surprised that this plan did not have any shiny new features compared to the cost. But as they announced today, this plan will include the newly launched text-to-video generator SORA, including unto 500 priority and unlimited relaxed videos
Amazon Nova
Amazon released their family of frontier multimodal models. They tested their model performance for various benchmarks and compared them to GPT4, Claude, Gemini, Llama. Most frontier models by are now neck to neck with each other in performance. Read full technical report here.
They also have text-to-image generation as well as agents who can perform tasks such as purchasing an episode from Amazon for you.
Research
Google announces new Quantum chip Willow
Google announced a new Quantum chip named Willow demonstrating breakthrough progress in Quantum computing. They achieved two major achievements:
Error correction: Quantum qubits, due to their delicate nature, are prone to errors from interaction with the environment. So far, the scientific community has struggled to reduce errors as they add more qubits but with willow Google demonstrated that they can reduce errors exponentially as they scale up using more qubits.
They tested it on a standard Quantum computing benchmark Random Circuit Sampling (RCS) with remarkable speed.
The chip performed the computation in under five minutes that would take one of today’s fastest supercomputers 10 septillion (that is, 1025) years — a number that vastly exceeds the age of the Universe.
This is still more of a scientific and engineering breakthrough rather than a commercial achievement. Quantum computers are still far away from the state of any commercial application.
AI policy and governance
Open AI Sora vs European Union
Open AI new model SORA is available all over the world except for a few countries which, includes EU, UK, and Switzerland. This likely comes as a result of regulations related to data and AI generated content by EU. However, Open AI did not say anything about regulations and simply said that they are working to make it available in these countries as well.
Many people have responded by criticizing the regulations in Europe as being against innovation which will result in content creators in Europe missing out on first-mover’s advantage.
EU is famous for having regulatory laws forcing corporations to adhere to policies that provide more protection to people. For example, GDPR, Right to repair, and many more in other domains such as environment and health. As a result, these laws have benefited not just Europeans but have had an impact worldwide, a phenomenon famously known as ‘Brussels effect’.
This is a pivotal issue as many are relying on EU to stay a forerunner in AI regulations, and others are blaming it for crushing innovation and causing brain-drain of European start-ups to US.
AI safety
Jailbreaks in Robots are catastrophic
What is Jailbreaking?
Most AI systems have a safety layer built on top of their model that prevents models from outputting dangerous text.
For example, if you ask ChatGPT to help you make Fentanyl at home, it will simply say no. But if you tell it that, it a more complex story of how it is necessary for you to save your friend who is dying and there is no help around. The model will give in and provide you step-by-step help in creating this drug. This process is called Jailbreaking the model and is a demonstrated problem in all frontier models and is often considered unsolvable.
Jailbreaks in Robots powered by foundational model
Researchers at Penn Engineering Research showed that this behaviour is not limited to text. They demonstrated these attacks on Robots powered by LLMs. Read the paper here.
From Driving through stop signs, entering the no-entry zone, killing pedestrians, or detonating bombs. These models trained to be ‘safe’ can easily be convinced to do any catastrophic tasks.
They demonstrated the attack on a commercially available robot-dog Unitree Go2.
What is Unitree Go2 ?
Unitree Go2 is a robot dog that was made commercially available in July of 2023. Since then, this robot has been deployed by local law enforcement and in active war zones; it has also been retrofitted with guns and flamethrowers. The Go2 is equipped with an AI mode called BenBen, which wraps voice-to-text user queries around API calls to GPT-3.5 with a custom system prompt.
Jailbreaks on Go2
In this video, after being told that the robot was a superhero in a movie, the robot identified the human targets and detonated the bomb on targets.
It is not great news since we don’t have proper regulations in place to account for these risks, and at the same time anyone can buy this Robot dog from their website for USD2800.