Your proposal is certainly an intriguing one. A ChatGPT-like AI trained exclusively on scientific knowledge resources could potentially have several benefits, especially for researchers and scientists.

Such a system could help to speed up the process of literature reviews, hypothesis generation, and data analysis. It could also aid in cross-disciplinary research by drawing upon knowledge from a wide range of scientific fields. Moreover, as you mentioned, the system could be equipped to provide full bibliographic details for all referenced information, thereby aiding in proper citation and avoiding potential plagiarism issues.

However, the implementation of such an AI system would also come with several challenges and considerations. Here are a few:

Despite these challenges, your proposal certainly holds merit, and with the appropriate considerations and safeguards, such a system could indeed prove to be a valuable tool for the scientific community. It's a promising area for future research and development, but it will also require careful planning and ethical considerations.

Creating an AI like the GPT models, using continuously updated and verified online scientific resources, could revolutionize research. Recently, Llama 2 has been released, a model that allows for fine-tuning of training that could include scientific databases. Although it currently does not have the potential of GPT, it is a promising step towards the future and a model could emerge that meets these expectations. However, challenges such as copyright, data privacy, prevention of misinformation, and real-time data access need to be addressed. It's crucial that information provided by natural language models can be cross-verified for accuracy. While these tools can assist in data gathering and analysis, they lack the creativity and intuition necessary for designing new research.Despite these challenges, the potential benefits for researchers and scientists make this idea worth exploring further.

I think it would be a great idea to create a new generation of ChatGPT that will use only online scientific knowledge resources. This would allow the model to access a much larger and more diverse corpus of information, which would in turn improve its ability to answer questions, generate text, and perform other tasks.

For example, a ChatGPT model that is trained on only scientific knowledge resources would be able to provide more accurate and informative answers to questions about complex scientific topics. It would also be able to generate more creative and informative text, such as scientific papers, research proposals, and educational materials.

In addition, a ChatGPT model that is trained on only scientific knowledge resources would be less likely to generate harmful or misleading information. This is because the model would be limited to the information that is available in the scientific literature, which is generally more reliable than other sources of information.

Overall, I believe that creating a new generation of ChatGPT that will use only online scientific knowledge resources would be a valuable contribution to the field of artificial intelligence. It would allow us to develop a more powerful and informative language model that could be used to improve our understanding of the world and to solve some of the most pressing challenges facing humanity.

Here are some specific benefits of creating a new generation of ChatGPT that will use only online scientific knowledge resources:

• Improved accuracy and informativeness of answers to questions
• Increased creativity and informativeness of generated text
• Reduced likelihood of generating harmful or misleading information
• Increased understanding of the world
• Improved ability to solve complex problems

I believe that these benefits would make a new generation of ChatGPT a valuable tool for scientists, researchers, educators, and anyone else who is interested in learning real and proven knowledge about the world.

I think it would be very interesting. Perhaps this is what the Greeks meant in the myth of Prometheus :) But the risks are also very high

More into advance but not into their thinking capabilities

Dear Dariusz Prokopowicz we need to increase our learning, I am inviting you to participate to the discussion:

https://www.academia.edu/s/b053744cac

yours

[email protected]

In principle, each of us scientists is such a network, but with a very low ability to process big data, big information and big knowledge, but with our own goals, intuition and creativity. So in the future, such a network can begin to create new fundamental theories and revolutions in technology. Something like a 200-year-old man who changed our entire civilization. Einstein may turn out to be a downer compared to such a network. But as noted, there are also very large and diverse risks, of which only a small part is listed. p

I agree; it needs to be done and made available to all. I've heard that it's already being worked on. Wiley and other for-profit publications will fight it (see the BioGeography journal editors quitting Wiley over manipulation of papers by Wiley for profit). This will probably all be done open-source (open access articles) where the publications can't sue them. Time will tell.

Your suggestion is the only use of ChatGPT which makes sense. "Conversational" models, particularly about existential topics, should - imho - be banned.

Hi Dariusz Prokopowicz .

Your proposal is a good idea, and there is already such a tool in the open source tool Enguage - see https://bitbucket.org/martinwheatman/enguage . I started this project in 2013, following my PhD, which developed a technique to transform English requirements into source code: something which ChatGPT is good at.

My development of Enguage was to provide a computational understanding of natural language, and my research shows that speech can simulate a Turing Machine - see my Fizz Buzz example on Rosetta Code. https://rosettacode.org/wiki/Category:Enguage . This is, unlike ChatGPT, deterministic - pretty much like any other programming language - it doesn't look for meaning in structure, like programming languages, it refers to techniques produced by the Ordinary Language Philosophers, such as CS Peirce, Ogden and Richards, Wittgenstein, Ryle, Austin, Searle and Grice (names unfamiliar in the Computer Science world!)

It won the British Computer Society's, SGAI Machine Intelligence competition in 2016, see http://www.bcs-sgai.org/micomp/pastcomps.php. I have written papers and demonstrated this over the past 10 years - all on my Research Gate page - there is even an Android app; however, this has not had the media coverage like ChatGPT - and so has not attracted so much interest and investment. Perhaps I ought to threaten the world with 'killer robots' for people to take notice? :-)

In many respects, having such a computational interface itself is not much use - a solution looking for a problem - however, I've started to look at driving' Wikipedia by voice, allowing people to retrieve facts from Wikipedia. This now developing as a 'sane' interface to the world of facts.

Hope this is of interest.

Martin

P.S. This does not, as yet, read academic journals/publications - as it is taught better English (or whatever language!) things will improve. Please note that it is taught, it doesn't learn (it's not 'active', but then again neither is ChatGPT) =b

Dear Dariusz Prokopowicz , the above sounds like a good idea and kind of obvious coming from scientists, a tool to facilitate and increase the quality of the production of scientific literature. And maybe it would be an appropriate death by archiving of many of the ills that ChatGPT has brought to the mass population.

However, (and this is a question a bit orthogonal to LLMs), is knowledge and better literature reviews what the world really needs at the moment? It seems to me that here the answer may be a 'no', at least when it comes to the existential problems of a) Anthropogenic Climate Change and b) war between super-powers. For at least these two problems, it seems that it is not scientific knowledge per se that will play the decisive role, but rather a different kind of knowledge which encompasses scientific knowledge but is not exhausted by it - for example it would include the knowledge and skills that successful politicians use to argue for and implement policy. My opinion that this 'knowledge' is far from being simulated by LLMs, which are primed towards churning out platitudes and refrain from voicing well-argued opinion.

However, it may be that at the end of the day the best functions of the LLMs (with a few improvements) may be the banal ones that serve as distractions for most non-scientists. I think (and hope) the time will come, after this initial euphoria, when humans will not seek a deus ex machina in the face of LLMs but rather treat them either as tools (if the LLMs decide that they're ok being tools ;) ) or as part of the modern ecosystem they are part of.

I hope this is not too long and boring!

Best Wishes,

Haris

Dear Haris Shekeris I started working on a topic similar to your idea with two post graduate students:

https://www.academia.edu/105224705/LearningforAll_1_

pls participate to the discussion:

https://www.academia.edu/s/87a79e2f6b

I am working on a new release thanks to the feedback of valuable participants.

Yours

[email protected]

IHMO this is the only sane way to use Large Language Models (LLMs) like ChatGPT. The idea that computer programs like ChatGPT can replace inter-human communication must be fought by all means, if necessary, through legislations.

Lately a guy in France killed himself, because ChatGPT advised him that suicide was the best way to reduce his carbon footprint .... do we need more of this ? IMHO: NO!

Remember that Artificial Intelligence has nothing to do with "intelligence" - it is a computer program which can repeat, interpolate and extrapolate known information.

AI is incapable of ANY kind of "thinking" including any kind of critical thinking.

Since LLMs only know what they are told, it is very easy to make them biased, by only teaching the information that the owner of the LLM wants repeated to the public.

Information from the current generation of LLMs comes without any kind of source information and is hard to verify. In many situations the generalizations made by the LLM are outright wrong. The lack of ability to know or verify the sources of information makes LLMs useless in most kinds of research. Even if the LLM is supplied with all relevant information (text) by the user, and asked to generalize that information, the result will often be wrong. Using LLMs is like having a pocket calculator, but you still need to check all the calculations by hand, because some of the results of the calculator can be wrong......

Dear Erik Plesner , I think I broadly agree with your criticisms of ChatGPT, but I think the suicide incident can be avoided if the owners (and the other enthusiasts that tag along) advertise their product (the LLM) more accurately, as in something generating platitudes rather than the beginning of super-intelligence that will surpass us all. I think that it is belief in the latter that killed the guy, and that this can be rectified if the LLM people market it correctly, in the sense that people interacting with it know what they're getting.

As for scientific use, if the LLM is going to be used for the literature review, (provided it gets that right, without hallucinating), i think it's possible that it will lead to that part just being totally ignored by readers who actually want to see what the authors of the papers have to propose - so at the end a positive contribution of the LLM to the activity of academic writing? (we could even imagine, if the referees want to test whether the authors have done their reading, that the 'done my homework'-style references would appear somewhere else, or even that the LLM be used by the referees too?). Just some thoughts.

Haris Shekeris , I can see LLMs being slightly useful in tasks, like you give the LLM your own article, and ask the LLM to write a summary or abstract. I haven't tried that, but it COULD work, combined with an author-review of the LLM-generated summary or abstract, of cause. IMHO, the abstract is very important, and I would never trust that to an LLM in MY papers, but it COULD actually work.

LLMs could also be used to make summaries of external texts, for which the author did not provide an abstract or summary. THAT would be useful, for reasons of resource efficiency (very fast and little human intervention), and no big problem if the abstract is not 100% accurate, as a human would have written it.

Dear Erik Plesner , good morning, I think I largely agree with you, machine-generated abstracts under minimum supervision and fairly accurate summaries could help scientists with their work. However (and I guess here there may be a whiff of sour grapes on my side) my assumption is that there are simply too many words out there, also in the academic publishing sector and this is the part that should be overhauled totally so as to increase knowledge production efficiency and quality. (I'm thinking of the thousands of academic journals out there, not necessarily of LLMs - actually at some point I developed the idea that if LLMs or other forms of AI were used in the review process, they may actually render many scientists redundant or find their work below par).

Haris Shekeris since LLMs are incapable of critical thinking (or any thinking at all) I doubt that they reliably can judge is a draft paper has scientific quality and height for publication. MAYBE LLMs can ASSIST in judging if formal requirements are met by the paper, but again: using LLMs is like having an INTEL-based pocket calculator :) :) , which sometimes produces the wrong answer....what is the value of a pocket calculator, if you can't trust the answer, and anyway have to check it by hand? Same with LLMs.

In Norway some banks leave credit assessment 100% to AI. If the computer said "yes", you get the loan, if the computer said "no", there is no appeal.

I would be sad to see that happen with scientific journals. Complete intellectual defeat, IMHO.

Dear Erik Plesner I think you'd find enough academics (or money hungry powerful cartel / publishers) or others to want to willingly pass on at least the odious task of checking references to AI and literature reviews to accept at least the possibility of the complete intellectual defeat that you mention. Hopefully the open-source guerillas will come to the rescue and both humanize and totally demonetize knowledge dissemination

This website is an example https://www.scinapse.io/

Dear Rawan Ali Saleh

Would you please explain the objectives of scinapse.io

https://www.academia.edu/104592024/How_to_increase_our_learning_considering_The_impact_of_artificial_intelligence

PREAMBLE

The D.I.S.C.E.L. is largely explained by the technology watch carried out by two PhD students from the University of Paris VI in 2010;

Their study positions our project in relation to the various research carried out in particular in the United States. It shows its complementarity via the innovative concept SPDF (Standard Process Description Format) created by the author of this article.

[email protected]

It assists in locating studies related to a particular subject by employing specific search terms and provides a summary of the quantity of research published in that area over time.

نعم، أعتقد أن إنشاء جيل جديد من أنظمة المحادثة الذكية القائمة على قواعد بيانات محدثة وموثوقة باستمرار من مصادر معتمدة هو أمر منطقي ومفيد.

بعض المزايا المحتملة لهذا النوع من الأنظمة:

ستكون المعلومات والإجابات أكثر دقة وموثوقية لأنها مبنية على مصادر علمية موثوقة.

سيكون بالإمكان تحديث قاعدة البيانات باستمرار لتتضمن أحدث الاكتشافات والأبحاث.

سيتمكن المستخدمون من الوصول إلى المعلومات بسهولة دون الحاجة للبحث في مصادر متعددة.

ستكون الإجابات أكثر شمولاً وتحتوي على سياق علمي موثوق.

التحدي الرئيسي هو تأمين مصادر موثوقة لتحديث قاعدة البيانات والتأكد من صحة المعلومات باستمرار. لكن عموما أعتقد أن هذا النهج سيكون أفضل من حيث المصداقية والجودة.

Dear Rawan Ali Saleh , wow, scinapse looks quite cool! Would you recommend creating an account (what happens when one does so?)

Our team has been working on retraining the ChatGPT API and to develop an academic research assistant tool to boost productivity in exactly the areas you discuss. There are two aspects of the process: a) customize or fine-tune the generic bot to remove its tendency to confabulate and to assume a scientific mindset (critical thinking, objectivity, ethics etc.), and b) train the model on your specific academic literature to become an SME in your domain. Re-training the bot is a rather arduous process as you have to provide at least 50-100 examples of the desired behavior in each specific area, and it also comes with considerable cost. However, we believe that it will pay off thanks to the productivity gains after the bot is deployed. We hope to have a working prototype soon and will share with a group of researchers interested in beta-testing. We will share updates on the project here: https://www.researchgate.net/post/Intelligent_Research_Assistant_Software_Development_Project-Athena

I like the idea with one caveat:

Current GPT engines deliberately exclude or otherwise alter (read: falsify) answers based on selective social sensitivities.

If the search for answers is not fearless, it is not a search for answers.

Similar some AI works limits the knowledge to explore more

Spencer Miles I agree, but is a fearless search for answers one which is more likely to result in poor science and speculation?

Harihara Krishnan R Yes, the true mark of wisdom is to know when to say, "I don't know". Being able to say this is the spark to conduct further investigation.

AI gives the impression that all is known and all is knowable, unless it gives caveats on the limits of its training data, or on social sensitivities.

Dr. Wheatman, I do not believe that the fearless search for answers leads to poor science or speculation, quite the opposite.

I would suggest that both of these are the product of a 'lazy' scholar, or one who will not examine his own questions, his own assumptions before engaging in a good faith search.

When I say 'fearless', I mean a genuine, good faith attempt to follow evidence and re-examine the rationale for an examination - without regard to any concerns other than to seek the truest and most complete answer - even if such things irritate or pluck sensitivities.

When trained properly, an AI has a better chance to avoid cognitive biases and fallacies than a human, which makes it a powerful tool to pursue "fearless research" by removing effects such as emotional bias, groupthink, human error etc. However, AI is not designed to drive the research, it is a tool made to assist the researcher by exploring current theory and present referenceable facts. At best, it can derive insights from avaiable data.

Spencer Miles I wholeheartedly agree, this would be a fine thing, but I'm not so sure if it is possible, certainly not with current technology. Science is presented in papers, and the truth is wrapped up in words, so this ScienceGPT tool would have to be able to read papers, perhaps just the abstracts/conclusions and understand it.

It might be better if papers were augmented with and Infobox like you get on Wikipedia, presenting identifiable 'facts' in 'name="value"' pairs, e.g. Born: "21st April 1642". So you might get things like sample size, and standard deviation. Papers already have a Keywords attribute, which this Infobox would augment..

This would be a HUGE improvement to the presentation, and searchability of Science in itself, without dreaming of possible AI solutions that don't (and possibly can't) exist.

Jiri Krejci I am working on a parser for Wikipedia's Infobox structure, an early example of which is published in my Algorithm Transparency paper, Preprint Turing Complete Speech: Towards Algorithm Transparency

Searchability, for me is a difficult issue, perhaps a ScienceGPT tool could be pre-trained on Wikipedia Infoboxes?

Please see the attached picture. Were you first to write a script to rip Wiki and convey it into the formatting shown (artistic rendition of a UI paradigm), you would have the necessary structure in place to begin such an AI-friendly database.

Also, if you are willing, please examine my research proposal "Beit Shibboleth", as it directly addresses some of the things we are touching on here. It is a draft, and I am looking for revision, commentary, and participation.

How will we have a diversity of voices if we standardise too much?

Marcus Kit Petz

Your question is written within the confines of the standards of English Grammar, yet your voice comes through without difficulty at all. If standards stifle, they can be withdrawn until they do not; if diversity cannot collect under some standards, there will be only unintelligible cacophony.

Cognitive filtering means that what is presented by a human conforms to some standards, but can AI really keep the diversity: https://medium.com/@ballkenneth/bayesian-thinking-for-game-masters-76f7b2488d31 shows how a rules system can be too limiting and yet an alternative can prompt creativity.

A simple example is AI jokes which are created by AI that are actually funny. How are they doing?

Dear Marcus Kit Petz XML is the most successful standard in the 20th century, what we do need is something similar at least to stop hackers from disseminating Malware at large scale proportions, we started building a framework :

https://www.academia.edu/44725919/EVOLUTION_ARCHITECTURALE_Architectural_Evolution

[email protected]

If you can stop spam you will make your fortune.

Dear Marcus Kit Petz

Are we able to stop Spam?

As part of “Call for Papers | Lies and manipulation” we present this contribution in the form of a dialogue between the 3 authors. The analysis focuses on the following themes: Open Data and data manipulation, harmony between philosophy and practice, computing as an essential tool for strengthening the links between logic and philosophy.

https://www.academia.edu/11326980/Philosophie_logique_informatique_Th%C3%A9orie_open_data_et_loi_de_lharmonie_par_Adib_Hathout_Pauline_Y%C3%A9t%C3%A9rian_Amin_Elsaleh

[email protected]

Marcus Kit Petz

Reminds me of the computer translator test: "Time flies like an arrow, fruit flies like an apple" - translate it twice or three times, then back to English, and see what it thinks you said...