Skip to main content

News

ChatGPT: go for tomato harvesting robot

Does ChatGPT design a better tomato harvesting robot than Syngenta or Metomotion? Or should the language model stick to what it’s good at?

Column Laura Thissen

ChatGPT recently designed its first robot, together with Swiss and Dutch universities. ChatGPT advised the researchers to go for a tomato-harvesting robot when they asked which robot would solve the world’s biggest problem. The large language model (LLM) also pointed the researchers towards a silicone or rubber gripper, as the tomatoes would be squashed otherwise. Makes a lot of sense, as any grower could have told them. Do we need a language model for this?   

Last January I consulted ChatGPT to write my New Years speech. In the end, I had to provide so much input to the model that typing it up myself would’ve been faster. And that’s precisely ChatGPT’s problem: it’s a good as the data it’s fed with. GIGO in jargon: Garbage in, garbage out.

ChatGPT is a large language model that determines the most likely answer – and formulates that answer into a human-like sentence. It doesn’t give sources as Google does, which would allow you to make up your own mind. No, it gives you the most probable answer. And potentially the biggest nonsense. Because who determines what’s the truth? If the model relies on open-source (web) data it’s been trained on, it fails the Dutch high school examination, starts hallucinating, and gets emotional.

No matter how advanced ChatGPT and its latest update GPT4 may be, its accuracy relies on the data it uses. If there are tons of accurate data behind it, GPT4 offers endless opportunities, for instance in protein design and other food & ag (bio)tech. ChatGPT can quickly combine loads of public climate and soil data with private sensor data from growers to support farm management decisions. Improving that annoying customer service chatbot is a piece of cake. Or use it to scan 100,000 scientific articles for their spin-off potential. Fantastic!

But what if your data isn't there yet? The best fermentation conditions for the highest yield of your innovative ingredient? Such an algorithm must be trained by loads of experimental data to be accurate. Benchmarking your satellite data algorithm to verify soil carbon storage? You need thousands of soil samples to get it right.

ChatGPT excels in analyzing loads of data at lightning speed and displaying it in understandable language. ChatGPT can do that. But in a lot of applications it is no more than an electricty-guzzling Google that is not more innovative than the data it is fed with.

Laura Thissen has over 15 years experience in business development for sustainable innovations contributing to the transition to sustainable food systems and a biobased economy. She writes on her own behalf on sustainability, innovative technology, and startups.