AI & DeepTech
Your weekly newsletter on cutting-edge innovations in AI, biotech & quantum
Nvidia’s gamechanging new chip
A second successful net energy gain experiment for nuclear fusion boosted hopes that we could develop a near limitless clean source of energy. Is this a pipe dream or could it become reality?
In AI, Nvidia’s new chip design could significantly reduce running costs for LLMs, sparking more innovation and frontier model development, while new research from Stanford shows that adding a little synthetic data in training can optimise model performance.
Meanwhile, academics at John Hopkins have developed a process to place nanoscale “tattoos” on live cells. If we can monitor individual cells, what healthcare breakthroughs could that lead to?
Explore these topics and more - from exascale supercomputer simulations to IBM and Hugging Face’s new geospatial open-source model - in the CogX Must Reads.
I hope you enjoy it!
Innovation & Releases
Net energy gain from nuclear fusion reaction
US scientists followed up their historic breakthrough last December with a second successful experiment where more energy was produced through fusion than the laser energy input. Scientists at the Lawrence Livermore National Laboratory hailed the achievement, but were cautious on timelines for fusion to become a reality.
IBM and Hugging Face release geospatial model
The pair have launched the largest geospatial model on Hugging Face, built in collaboration with NASA, on the back of the Space Act Agreement. The open source effort was trained on harmonised Landsat Sentinel-2 satellite data and represents a 15% improvement on state of the art techniques used today. It will democratise access to data and help generate new innovations in climate science.
Synthetic data can damage your AI model
Alemohammad et al. outline how an over reliance on synthetic data during training will either reduce the quality or diversity of generative models. Given constraints of the volume of existing data, this could impact frontier model growth.
However, they also found that mixing some synthetic data trained on previous generations with fresh new data can actually boost performance — as long as too much isn’t used. This could be the model to optimise performance when building new AI models.
Nanoscale “tattoos” on live cells
Engineers have developed a new technology which allows the placement of optical elements or electronics onto live cells for the first time, bringing researchers one step closer to tracking the health of individual cells. Structures were able to stick to cells for 16 hours even as the cells moved; crucially, the cells didn’t die.
The breakthrough is a step on the path to remotely monitoring and controlling the state of individual cells in real time, which could revolutionise healthcare.
Capture your ideas
Reflect helps you remember ideas and follow-up actions on the go, with a note-taking function designed to mirror the way we think. It uses GPT-4 and Whisper to improve your writing and efficiently organise your thoughts.
Need to create a how-to guide?
You can transform documentation into stunning visual guides using Guidde, a GPT-powered tool that helps you explain complex ideas in seconds. It can capture your flow, create storylines, and can easily be shared online or within your organisation.
CogX Must Reads of the week
Technical Tip of the Week
This step-by-step guide shows you how to generate audio from text prompts and clone voices:
We'd love to hear your thoughts on this week’s Issue and what you’d like to see more of.