OpenAI has launched a new model OpenAI o1 with advanced coding capabilities. Advances in AI and the availability of such models may make traditional coding dead. It has struck the final nail in coding’s coffin.
OpenAI o1 is an advanced model with reasoning capabilities across coding, math and science. Its another version OpenAI o1 mini is also good at coding. Both cracked the recruitment interview for coding at a 90-100 per cent rate. It was an interview to recruit research engineers.
These latest models increase the risk of AI being used for the creation of biological weapons. These models have abilities to reason, solve hard math’s problems and answer scientific research questions.
It is a real breakthrough towards the creation of artificial general intelligence (AGI) enabling machines with human-level cognition.
OpenAI’s system card, a tool to explain how the AI operates, labels the new models as medium risk models as far as weapomisation is concerned. In future, such models can pose increased risk when they fall into the hands of bad actors. Youshua Bengio says the medium risk models do necessitate the urgency of regulatory framework.
Mira Murati, OpenAI CTO, says the company is cautious with how bringing o1 to the public on account of its advanced capabilities. The product will be accessible to ChatGPT’s paid subscribers and to programmers through an API. Murati informs that the current models performed far better on overall safety metrics than the previous ones.
The o1 model is trained with reinforcement learning (RL) to perform complex reasoning. OpenAI previously used transformer technology, and is now using reinforcement learning of DeepMind.
OpenAI o1 model has been code-named Strawberry. Though it is capable of reasoning, OpenAI has tried to keep how its thought process works off-limits. It has its ‘chain-of-thought’ reasoning and it arrives at an answer step-by-step. Mira Murati, the CTO, OpenAI calls this a ‘new paradigm’ for the technology.
A user invites trouble if he/she tries to get the reasoning trace. By not revealing, its raw, thought-process, the company maintains its competitive advantage. That robs the process of democratization of the language models. It is a step backwards. This way AI models remain more like an opaque blackbox.