Let’s look at some things happening in December first:
In early December Google Deepmind launched its large language model Gemini in three different versions: Ultra, Pro, and Nano. The Ultra version is said to beat OpenAI’s GPT-4 in almost everything – it is also said to be better than human experts in language understanding. The one thing that GPT4 is still stronger at is commonsense reasoning for everyday tasks.
https://www.artificialintelligence-news.com/2023/12/06/google-next-gen-ai-model-gemini-outperforms-gpt-4/
What do you think about that and have you tried Gemini, Ather?
I was reading Fast Company, that the Elon Musk’s newish large language model Grok, had declined to modify some malware because it quote “goes against OpenAI’s use case policy” unquote.
Is Grok plagiarising OpenAI’s ChatGPT4?
Have you heard anything more on OpenAI’s progress, now when Sam Altman is back?
EU AI ACT
Let’s turn our eyes to Europe
On Friday the 9th, European Union policymakers agreed on a provisional deal on some landmark rules governing the use of artificial intelligence (AI). They include the use by governments of AI in biometric surveillance and how to regulate AI systems such as ChatGPT.
Many seem to be happy about it, when reading Reauters but in Sifted, The Act — which has yet to be finalised — is said to be perhaps hindering the progress of smaller startups, according to investors and AI founders, potentially putting Europe even further back in the global AI race.
What do you say, Ather?
(https://www.reuters.com/technology/eu-clinches-deal-landmark-ai-act-2023-12-09/)
The act would not apply to open source models (like Mistral’s), unless they are considered high risk or being used for banned purposes, according to a draft of the legislation that the European site Sifted has seen.
But some say that most large-scale AI models will be considered “high risk”.
What would that mean for European AI companies do you think?
But so far, AI-startups in Europe are still getting the attention from US investors. Paris-based generative AI startup Mistral AI managed a deal which they announced on 12th December – raising €385m Series A led by US investor Andreessen Horowitz and Lightspeed Venture Partners.
Robots making house-tasks
A couple of months ago in the AI podcast we spoke about Nvidia using an AI-enabled system to teach robots complex skills — like training a robotic hand to spin a pen between its fingers.
Now A new open-source system, called Dobb-E, has launched that teaches robots domestic tasks, according to MIT Tech Review. The biggest challenge for this kind of tasks is: a lack of training data.
The robot learns a domestic task by having the tasks recorded using a simple Iphone, then it learns the task in about 20 minutes, anything from how to open an air fryer, close a door, to straighten a cushion.
The researchers tested the robot in 10 homes in New York over 30 days, and it completed 109 household tasks with an overall success rate of 81%.
https://www.technologyreview.com/2023/12/14/1085231/new-system-teach-robot-household-task/
This is a step in the right direction, right?
Looking back at 2023
The past 12 months have represented a year of collective AI hysteria, with big advances like the release of OpenAI’s GPT-4 and Google DeepMind’s Gemini – so Generative AI is probably top of mind when it comes to what has happened in the AI space in 2023, isn’t it?
Are there any other developments that have been overshadowed by GenAI?
Ather: Certainly breakthroughs related to reasoning and planning, such as GFlowNets by Yoshua Bengio and his research group. But also, Nvidia’s AI system that can learn how to design “rewards” to train robots to perform physical tasks, such as the pen spinning example that was demonstrated. Google’s multi-modal Gemini is on par with GenAI, which is in fact more than GenAI.
Happy Holidays to all of our listeners!
view more