Back
RCreddit.com

Signs of AGI? GPT-6 Astra scored 95% on a robot control task vs Fable 5.1's 40%

OpenAI
Time & source
Published
09/05, 13:13
Ingested
09/05, 17:00
Source type
Dev community
Tier
Community
Source status
Healthy
Tier is a per-source editorial setting, not a per-item score.

"GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost. On harder tasks involving precision, Astra has the same success rate as Fable 5.1 (2/20), but uses 3.9x fewer output tokens and is 1.6x cheaper. If trends hold, LLMs could control robot arms in real time as soon as the end of this year, or by 2029 at the latest. We urge the research community to look into the implications if general-purpose robots arrive within the next two years. SOTA LLMs are showing surprising generalizability to robotics and their latency is improving rapidly."

Imagine a future model or system that isn't just a digital AI controlling a computer, it can also input/output practically anything a human can because its cognitive ability matches humans.

So it can control robotics, drones, cars, fighter jets, etc. My definition of AGI is similar to what we had in cinema / science fiction for decades. Jarvis / Friday for example not only ran the entire stark industries, it also could run scientific experiments and simulations, it also ran and controlled tony's entire house, while also being able to control his computer to do computer tasks and finally ability to interface with his robotic suits to control them.

But for something to be AGI. I should be able to tell a future version of astra / gpt-6 to:

- Ability to drive a car to a certain performance metric, doesn't mean it has to drive ~25k without a minor accident which is actually the rate for humans non police reported accidents, or 500k with police reported accident or beat humans million miles between fatality. It just has to be able to drive well. So the benchmarks can be 50 or 100 hours between disengagement.

Astra has clearly made huge progress on the Computer Use and spatial reason. But alot needs to be done to get the latency to real time on all input/output and getting spatial reasoning to match human's visuospatial.

The definition of AGI has always been in times past a AI system that can match or exceed human cognitive abilities across any intellectual task or domain. Notice its not matching human performance nor exceeding it. For example there is only 1 human on earth that can shoot as good as steph curry out of 7 billion.