Being embedded in work on large language models applied to scientific literature, combining that with more classical machine learning approaches on existing structural models and databases, then linking both with an AI approach, I think care has to be taken with assuming that models such as ChatGTP are necessarily the best AI approach to use.
There is a very clear difference between context specific trained models and more generalist ones such as ChatGTP. Generalist ones are optimized for breadth and languange competence. Context or domain-trained ones are optimized for more precision and task specific interactions. Different models perform differently in different contexts. ChatGTP is trained on as much content as possible, and as such many of the correct technical details, but also every other possible source of noise which could cause it to wander from that detail. It will have conflicting conventions (different frequencies of electrical grids in your example), outdated information, fictional information (literature, scripts etc.), variable sources of quality, and any implicit bias that is out there. Some of this can be overcome. There are techniques such as Retrival Augmented Generation which grounds outputs in evidence, gives traceability, reduces hallucination, and enables validation (and even better approaches on the way) that can provide the reasoning behind the AI model answer. I think of this as the scientific presentation style of making a statement, then backing it up by the evidence that supports that statement. AI is a very useful tool. It accelerates a lot of processes but I think there has to be a 'buyer beware" approach. It requires knowledge of the training set that the AI is basing its probabilistic response on, an understanding of the model selection (there are many AI models available), the retrieval strategy, and good training in prompt engineering to produce the best outcome. Some things such as programming languages, or structural models, have readily accessible, well-curated training sets. Others are more suspect. A note of caution should be sounded, particularly with cloud-hosted models. While I believe thaty user inputs are not routinely used for further model training, sensitive information may still be processed, logged, or retained outside of your control. There are cases where pre-publication material has been exposed. For this reason, many industrial and academic organizations are developing internal AI systems to reduce this risk. The field is advancing so fast that publications are almost out of date as soon as they are released! It is very invigorating to see the developments and the democratization of access to them, but the bottom line is that it still needs real intelligence to make most effective use of artificial applications. Let's hope that continues 🙂 Best, Eddie (PS. note new address and contact details) Edward Snell Ph.D. Professor and Chair, Materials Design and Innovation | University at Buffalo, SUNY Director | NSF BioXFEL Science and Technology Center | Sunset 2025 Fellow of the American Crystallographic Society – The Structural Science Society p: +1 716 881 7573 (new) | c: +1 716 989 9128 | f: +1 716 898 8660 e: [email protected]<mailto:[email protected]> Department of Materials Design and Innovation School of Engineering and Applied Sciences University at Buffalo 120 Bonner Hall, Buffalo NY 14260-9640 Department website: Laboratory Website: https://snelllab.website/ ________________________________ From: CCP4 bulletin board <[email protected]> on behalf of Jeroen Mesters <[email protected]> Sent: Sunday, January 25, 2026 10:12 AM To: [email protected] <[email protected]> Subject: Re: [ccp4bb] AI AI sir! You don't often get email from [email protected]. Learn why this is important<https://aka.ms/LearnAboutSenderIdentification> This morning I asked a browsers AI, and in parallel ChatGPT, a question about certain frequencies used in the cities electrical grids to control wallboxes, heat pumps and PV systems. Since I got different answers I confronted the AI with ChatGPTs result’s and received the following answer: "The specification of 210 Hz is incorrect in this case. ChatGPT often tends to hallucinate when it comes to the technical parameters of local distribution network operators, or confuses the values with neighbouring networks or standard values.“ I then confronted ChatGPT with the latter AI statement and it agreed to some extend and finally came up with yet another frequency… But serious, as far as computer code is concerned, AI can help to generate and/or optimise code but you better understand the code yourself to further cure and optimise certain parts of the AI generated algorithm ... Cheers Jeroen __ Dr. math. et dis. nat. Jeroen R. Mesters Biological Safety Officer (BBS) Deputy, Lecturer, Program Coordinator Infection Biology Visiting Professorship in Biophysics - University South Bohemia [attachment.png] University of Lübeck Center for Structural and Cell Biology in Medicine Institute of Biochemistry Ratzeburger Allee 160 23562 Lübeck Tel +49 451 3101 3105 https://orcid.org/0000-0001-8532-6699 Am 25.01.2026 um 14:17 schrieb Hughes, Jon <[email protected]>: hi, there has been much talk of using AI to write code for us and of it making the world better. people in this group have their own opinions regarding alphafold, for example, but at a much simpler level, i just asked chatGTP something about electrical power generation: his/her/their answer finally included, ""Interpretation per joule: 4–12 €cents per kWh equals 4–12 × 10⁻⁶ € per joule, since 1 kWh = 3.6 MJ". well, we all make mistakes, right?! cheers, jon ######################################################################## To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 This message was issued to members of www.jiscmail.ac.uk/CCP4BB, a mailing list hosted by www.jiscmail.ac.uk, terms & conditions are available at https://www.jiscmail.ac.uk/policyandsecurity/ ________________________________ To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 ######################################################################## To unsubscribe from the CCP4BB list, click the following link: https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1 This message was issued to members of www.jiscmail.ac.uk/CCP4BB, a mailing list hosted by www.jiscmail.ac.uk, terms & conditions are available at https://www.jiscmail.ac.uk/policyandsecurity/
