Founded by Deep Mind’s core developers, the AI Agent Startup Reflection AI (AI), which has been a hot topic, revealed its investment attraction and left the stealth state. They aimed to construct the Superintelligent Autonomous Systems through the AI coding agent.
Bloomberg reported today that the Reflection AI has been officially launched and officially launched a complete of $ 130 million. The corporate announced that it’ll officially open the homepage and recruit developers.
The funds mix $ 25 million within the seed round last 12 months, led by Sequoia Capital and CRV, and $ 150 million within the series A round led by CRV and Light Speed Enterprise Partners. As well as, famous investors akin to NVIDIA and Leadhofman Linked, co -founder and Alexander King AI CEO, participated. The corporate’s value is estimated at $ 550 million.
The corporate’s establishment was already released in August last 12 months with Sequoia’s investment announcement. On the time, it was not known that it was co -established by CEO of Missha Laskin, who joined the deep mind through the Google Deep Mind Veteran Researcher, Ioanis Antonogloo, Sergil Ozar, and Berkeley AI Research Institute.
Nonetheless, the name of the member attracted the eye of many individuals.
The co -founder of Antonoglu participated in Deep Mind’s ‘Alphago’ and can also be famous as considered one of the authors of ‘Deep Reinforcement Learning’, which developed Deep Mind to learn the Atari video game without human intervention in 2013. Until last 12 months, Google’s ‘Geminai’ development team is a key figure who led the RHLF through human feedback. Laskin CEO also participated in the event of Geminai.
Ozar can also be considered one of the authors of the ‘hostile creation neural network (GAN)’, which is as famous as ‘transformer’. He joined Tesla and joined the Reflection AI.
Particularly, the models that the members of the corporate participated in the event were ▲ in-depth reinforcement learning algorithm ‘DQN’ (2013), ▲ Alphago (2015), Alpha Zero (2017), and Muzero (2019), and thru a series of game theory models akin to Palm (PALM, 2022), GPT-4 ▲ Jeminai 2023) ▲ It could actually be seen because the history of the AI model itself, as much as 1.5 (2024).
For this reason, the proven fact that they left Google was quite noticed, and news that it had already established an organization since January last 12 months.
They aimed to develop ‘super -intelligent agents’ and said they’re developing autonomous programming agents as their first project. This agent can robotically perform tasks akin to cord vulnerability detection, memory use optimization, automatic code description documents, and application infrastructure management.
To this end, we’re designing a brand new AI system that mixes large language model (LLM) and reinforcement learning (RL). Even though it didn’t reveal details, it is thought that it’s in search of a brand new approach beyond the prevailing transformer architecture, and can also be considering using the choice architecture ‘MAMBA’.
“We have been fascinated about greater than a decade of problems that transcend transformers, and we’re confident that we’re now to mix LLM and reinforcement learning to construct practical super -intelligence and create a system that works on computers.”
Unlike the prevailing coding tools, the corporate is simply a developer’s instructions and squeezes the code as a substitute, the corporate is characterised by establishing a whole autonomous agent with the flexibility to make a software development plan and make decisions independently.
Laskin’s CEO compared it to the autonomous driving system. “If the AI agent of one other company is a cruise control that assists human driving, we aim to construct a ‘fully autonomous driving’ system like Weimo,” he said.
By Park Chan, reporter cpark@aitimes.com
