2023-11-24

Q*

Researchers working on a project at OpenAI called Q* sent a letter to the board informing them that they made a discovery that threatens humanity, just days before Sam Altman was fired. Few details were provided about what this breakthrough did, other than a ton of compute was thrown at a model to solve math problems. Solving Math problems requires the ability to generalize ideas and connect them to form new information, and this is something that Transformers in LLM’s have been argued to not do very well. LLMs are very poor at math so if OpenAI has a model that can do math they may have uncovered the infrastructure that succeeds transformers and opens up the path to AGI. And apparently the OpenAI board didn’t trust Sam with that knowledge.

Is RLHF solved?

Reinforcement learning through human feedback was considered the data bottleneck to evolving LLM’s. But Orca2 was trained with synthetic data provided by GPT-4. Researchers at Microsoft published a paper concluding that GPT-4 provided better data than humans for training. If so, Orca2 is an OpenSource model demonstrating that we may not need humans in the loop as much for some training techniques and models. If models can train themselves with their own data and they learn to reason like humans – we may be in for some unexpected surprises.

 paper here: https://arxiv.org/abs/2311.11045

Lithium Silicon Batteries

Lastly finding news that’s not about AI is definitely tough right now, but interesting things are happening with silicon batteries. Silicon anode batteries are not a new idea, they are proven to be 24x more energy efficient and 10x more energy dense than graphite or carbon anode batteries. It’s just that that silicon is so lithium hungry that it swells significantly during charge and discharge eventually cracking the anode and destroying the battery. But now some creative nano technology innovations may have solved this problem.

IEEE Article: https://spectrum.ieee.org/silicon-anode-battery