Each week, we bring you 5 stories that resonated the most in our internal Slack channel #AI-news. We write the newsletter using various AI tools because we're an AI company and our marketing wants to move with the times too. π
Today you're reading the 52nd issue in a row.
#1
Models o3 by OpenAI π: one step closer to AGI π€π€?
OpenAI has unveiled its new o3 models, the successors to the o1 π€ "reasoning" models.
These AI models can "reason" through problems step by step. In benchmarks, o3 excels: it excels at math, programming, and science π¬, but still burns at simple games like Mouseketeers ββ. Where did they leave out the o2 π€ you ask? Because OpenAI skipped that designation due to trademark disputes with the telecom company ππ«.
OpenAI claims that under certain conditions, o3 approaches AGI - a level of intelligence that could surpass humans in most economically valuable work π§βπΌπ‘. Although experts warn that AGI isn't here yet, the latest results show that we're not as far from AGI as we think π.
And what about the availability of this model ? The o3-mini preview is already open to security researchers, with public releases planned for 2024 πβ¨.
Β
#2
Genesis π: New physics simulator "trains" robots 430,000 times faster than real time π€β³.Β
Training robots in the real world is very tedious-it can take years, maybe even decades π€β³. Here comes Genesis, an open-source platform that speeds it up considerably π.
It lets you run up to 100,000 simulations at once, giving robots decades of virtual training in just a few hours π. Scientists can use this platform to teach complex skills such as walking, using tools or even acrobatic stunts π€ΈββοΈ, all in a virtual environment π₯οΈπ. Genesis also presents dynamic 4D worlds generated by simple text commands, eliminating the need for tedious manual programming π.
With its free, Python-based design π, Genesis makes advanced robotics research faster, cheaper, and more accessible πΈ. And what do you think of it π€?
Β
#3
Gemini 2.0 Flash Thinking π€: a new competitor to OpenAI's o1 π€?
Google has introduced Gemini 2.0 Flash Thinking, an experimental AI model that answers complex questions while explaining its thought process step-by-step π οΈ.Β
Unlike conventional AI models, it breaks problems into smaller parts, improving accuracy π―. It runs on a faster version of Gemini Flash 2.0, which combines speed with a new "reasoning" capability that helps better solve problems π. As a result, it handles challenging physics and logic problems and can work with both images and text π.
The model is meant to be a competitor to OpenAI's o1 reasoning model and is part of Google's effort to create smarter AI π€. Why is this interesting? Better reasoning in AI can yield more reliable and useful answers, and this technological rivalry is accelerating advances in AI β‘.
Β
#4
OpenAI co-founder Sutskever announces that AI models no longer have anything to "learn from" π§ : What does this mean for the future of AI ππ»?
OpenAI co-founder Ilya Sutskever announced that we have reached "peak data" π, which means that there is no longer new content available on the internet to learn current AI models.
This will require a change from today's approach which uses the huge amount of data available π. Future AI will instead focus on thinking and decision making π§ , making systems more autonomous and able to function with less information. Unlike current systems, which mostly just combine known patterns, new generations of AI will explore problems step by step π and solve them more efficiently π οΈ.
Sutskever claims that these thinking systems will be less predictable π€ but much smarter, which will completely change the world of AI π.
Β
#5
AI with cognitive gaps π§ ? Chatbots struggling with tests for dementia π¬π€.
A study in The BMJ has found that leading AI chatbots, including ChatGPT and Gemini, show mild cognitive impairment on tests for dementia designed for humans π§ .Β
While they excelled at tasks like language and attention, they struggled with visuospatial and executive tasks like drawing clocks π°οΈ or organizing sequences π. Gemini models performed the worst, scoring only 16 out of 30 on the Montreal Cognitive Assessment (MoCA) π¬. The researchers noted that even the best model, ChatGPT 4o, barely reached the threshold of normal cognitive function β οΈ.
These gaps may limit the use of AI in medicine and suggest that replacing doctors with chatbots is the music of the future for now πΆ. What do you think about this π€?
Β
