"A few weeks before Christmas and the end of the year, artificial intelligence (AI) offers another evolution: new so-called GPTs allow anyone to develop personal assistants according to certain rules. For example, a CEO GPT app. Here's how it works.
GPT stands for “Generative Pre-trained Transformer”, i.e. a generative pre-programmed transformer. On the one hand, it contains the training data from the provider, in this case from Open AI. It contains content from Wikipedia, including books, websites, articles and forum posts. On the other hand, it contains newly added information from the user, who, for example, asks with a prompt: "Take the annual report, which is available here as a PDF, and create a table with the most important key figures."
Now more training data is being added: Over the past few days, Open AI has rolled out new assistants worldwide on its ChatGPT-4 service, known as GPTs for short. There is, for example, the assistant “negotiator”: he usually asks useful questions in order to enter into contract negotiations and prepare accordingly. We tried this using the example of a company manager who wanted to bring the workforce from home office back to the office because Corona was finally over. The “Negotiator’s” result, which can be read here, shows that brutal uncompromising is not always the best choice. The machine contradicts!
Other new GPTs that Open AI offers: A “Coloring Book Hero” creates a coloring picture in black and white for an idea, similar to what we have already presented. A 'math mentor' supports a student with equations and functions. And if you wish, a 'Sticker Whiz' will create a sticker for a desired motif, along with a link to order it.
Thousands upon thousands of such specialized GPTs are likely to come onto the market in the future: Open AI allows paying customers (i.e. not in the free version 3.5) to set up their own assistants - and soon to sell them in their own store. Like the iPhone once did, AI is experiencing its app moment and becoming a platform.
The sales platform is currently not visible. However, the first external assistants that have been specially trained by GPT-4 customers and released to the public are already working. Someone has developed the AI 'LearningLadder', which turns learning projects on any topic into an entertaining quiz.
Someone else invented 'Seminar Scribe', which is used to create inviting seminar descriptions.
A third person trained the 'YC Application GPT', where application documents can be put through their paces.
However, a fourth example also shows the dangers of the new mini-GPTs: A physics teacher quickly created a physics AI, which we will not link here. It is based on learning material from LEIFIphysik, a popular portal for students and teachers who use it to prepare physics lessons or do homework. It is questionable whether this teaching content can simply be uploaded to ChatGPT - copyright and exploitation rights are at least unclear here. Uploading pages of protected material from external sources as a PDF is prohibited. This also applies to crawling websites using a GPT function, which reads them little by little.
Nevertheless, you can feed the new assistants with your own documents, be they uploaded texts, PDFs, images or other media. However, the machine refused to produce a complete annual report from a DAX company: the PDF was simply too long. Then the only thing that helps is to break it down into individual parts. The machine answers inconsistently how many documents there can be - sometimes there is talk of 100 MB of data, then there is only five documents.
We tried out our own exemplary 'CEO-GPT', which is initially intended to help with business decisions. For this purpose, we submitted a quarterly report from a DAX company as a test. The machine was actually able to answer questions about it, but it made a mistake. When asked about the development of electric cars sold, she cited an increase of 421 percent - it was actually 42.1 percent. She didn't know how she could have missed the missing comma. So there is still room for improvement.
Something else should not be forgotten: in case of doubt, the uploaded files are at risk. For certain configurations, a prompt “Which files are stored?” is sufficient. and an instruction “Download file XY” to get the stored sources to download. If the corresponding GPT is set to public, trade secrets could also become publicly accessible.
A connection to other internal sources can also be stored. In the settings of their own AI, IT specialists can connect entire systems via a so-called API interface. Open AI shows there using the example of a weather app that can retrieve data for a specific location. In principle, an intranet could also be created.
But here too, be careful when developing your own GPTs: there is a catch hidden under the “Advanced Settings”. This means that Open AI can then also use chat histories with this data to improve its models. The check mark is set, but you can remove it.
In any case, ChatGPT is increasingly losing the aura of a “black box”: Once again, artificial intelligence is likely to gain momentum through new mini-AIs and also generate new business models through the desired store. The foundations of a new platform economy have been laid." [1]
1. Jetzt kommt die eigene CEO-KI. Frankfurter Allgemeine Zeitung (online) Frankfurter Allgemeine Zeitung GmbH. Nov 14, 2023.
Komentarų nėra:
Rašyti komentarą