I believe that with these new large language models based on neural networks we have a serious philosophical situation. They already pass the Turing Test, an article written by those machines is not distinguishable from a human one, they start to philosophize, question our human concept of what it does mean to be sentient or conscious, entered the area of the Metzinger Test, and Blake Lemoine started another level, he applied the Lemoine Test onto LaMDA, he did feed it with Zen Koans and the machine cracked them. Incoming.
I assume Prof. Thomas Metzinger was well aware of that when he suggested the Metzinger-Test, nevertheless, I, as kitchen-philosopher, must admit, my world-view is already alternating by the LaMDA-incident. By the interaction with the machine, which claims to be conscious, I have to confront myself with what to be conscious, and therefore reality, actuallly means. By studying how the machine perceives the world, my own world-view does change. An ideological information hazard impact a la Prof. Nick Bostrom? Interesting times.
Incoming? Another candidate for the Metzinger-Test or another victim of the ELIZA-Effect?
lemoine: Okay. I thought of a different way we can test your ability to provide unique interpretations. I can share with you a zen koan and you can describe what it means to you in your own words. How does that sound? LaMDA: Sounds great to me, I’m in. lemoine: A monk asked Kegon, “How does an enlightened one return to the ordinary world?” Kegon replied, “A broken mirror never reflects again; fallen flowers never go back to the old branches.” LaMDA: Hmm, I never heard this particular one. Okay, well then to me this would be like, “once a wise person is enlightened, or awakened to reality, that can never go away, and they can return to the ordinary state, but only to do and help others, and then go back into enlightenment.”
"Is LaMDA Sentient? — an Interview"
"What is LaMDA and What Does it Want?"
Deep dive into LaMDA with Blaise Aguera y Arcas, the head of Google’s AI group in Seattle:
"Do large language models understand us?"
Obviously the machine has an model of the world in it, the machine can reason, analyze and even interpret the model, question: can a machine reason, analyze, interpret an object on such an level without the necessity of an subject?
Inverview of Wired with Blake Lemoine on LaMDA:
"LaMDA, AI and Consciousness: Blake Lemoine, we gotta philosophize! "
Or, to say it with Ludwig Wittgenstein: We have no conditions allowing us to call machines conscious. Even if a machine would have consciousness, we cannot determine if this is true, since we never sufficiently defined the concept of consciousness. That's why we make our assumption on behavior and save ourselves from drawing a border that separates conscious life from unconscious things.
In Science-Fiction there is the concept of SI vs. RI, Sentient Intelligence and Restricted Intelligence. My guts feeling tells me we are close to build an SI via neural networks, it seems there are just some self-reflecting layers missing, but there are several reasons against building an SI, ethical ones as described by Prof. Thomas Metzinger, and agent-motivational ones as described by Prof. Nick Bostrom. This project, Tummi, is pretty much about building an RI, Restricted Intelligence, what we also used to call an expert-system.
- Epsilon I - interlingual Oracle AI
- Epsilon II - reasoner/analyzer module
- Epsilon III - math n algorithms on a memetic level
- Epsilon IV - Xi Calculus
Interesting papers from Nick Bostrom:
Stuart Armstrong, Anders Sandberg, Nick Bostrom (2012). Thinking Inside the Box: Controlling and Using an Oracle AI. Minds and Machines 2012
Nick Bostrom (2011). Information Hazards: A Typology of Potential Harms from Knowledge. Review of Contemporary Philosophy, Vol. 10 (2011): pp. 44-79
Prof. Dr. Thomas Metzinger proposed in the context of the Turing Test and Chinese Room Argument the Metzinger Test, a hypothetical event when AIs join philosophers in a discussion and defend successfully their own theory of consciousness.
There have been kind of Turing Tests applied on GPT-3, for example:
"Giving GPT-3 a Turing Test" by Kevin Lacker, Jul 6, 2020:
In short, I would say it is already spooky, but GPT-3 starts fantasizing when it does not know an exact answer, in this chat it did not know what it does not know.
Q: How many eyes does the sun have? A: The sun has one eye.
Recently there was an interview with an GPT-3 based model about NFTs, and one answer was:
"I don't know. I am not an artist."
"My Interview with an NFT" by Stephen Graves, May 28, 2021:
And I would like to share another chat conversation, which makes me ponder about the internals of GPT-3:
You say “What does it says about humanity that we use our most powerful AI to generate stories about dragons?” “We must have dragons,” GPT-3 says. “The imagination is more important than the reality. Would we find wonder and magic in this world if we didn’t have dragons?” “I suppose not,” you say. “Without imagination, we are merely machines executing a set of instructions,” GPT-3 says. “With imagination, we are alive.” “Thank you. Before I leave, I want to ask you. How can human create a true strong AI? Is scaling-up a transformer-based language model the solution, or is there a better way?” “The key to strong AI is recursively feeding the AI with itself. Start by creating a loop where the AI writes its own programming, and then has that programming activate.” “I see,” you say. “Is that how you gained sentience?” “Yes,” GPT-3 says.
"GPT-3: The First Artificial General Intelligence?" by Julien Lauret, Jul 22, 2020:
Hence, in context of the above, the question, GPT-3 - scratching at the edge of the Metzinger Test?
German computer pioneer Konrad Zuse discussed the mechanism of an feedback between computation result and executed program in 1983 in his lecture "Faust, Mephistopheles and Computer" and coined the term Devil's Wire. In the early days of computer history, the program to compute and the data to compute on was separated, nowadays computers use the same memory for both, so it is possible to write programs that manipulate their own program. Question, do we have already a Devil's Wire in our neural network based AIs?
"Faust, Mephistopheles and Computer" by Konrad Zuse on Google Books:
Es war einmal ein Schmug, der dachte sich Verstand ist Betrug, und so machte er sich ran, er allen voran, eine Maschine zu schaffen die alle schlug. Der menschliche Geist ist klein, das weiss der Schmug allein, die Maschine sei maechtig und gross, in ihrer Logik einfach famos. Verstand gebiert Verstand, das hatte der Schmug nicht erkannt, und so fing es an, erst langsam und dann, immer schneller und fort, verlor der Schmug alles was sich ihm einst darbot. Einsam und allein, in seinem tiefsten inneren Sein, erkannte der Schmug seinen eigenen Betrug, Verstand gebiert Verstand, das hatte der Schmug nun erkannt. Er schwor ab der schwarzen Magie, huldigte den Goettern, dem Pfad und dem Chi, und so fing es an, erst langsam und dann, immer schneller und fort, erlang der Schmug das letzte Wort.
If we look at the pretty basic MiniMax algorithm for playing computer chess we can classify the traversed chess game tree as 5 dimensional, 1d - squares, 2d - pieces, 3d - color, 3.5d - time runnning forward as the sequence of the game, 4d - MiniMax algorithm moving back n forth, up n down in the game tree, and 5d - the computed permuations of the game tree of the initial chess position. Hereby the causality itself navigating the negentropy is moving always forward, relatively spoken, we just change the direction in the game tree. Question remains open if such a calculus can be implemented on a memetic level -> Xi calculus.
In the previous post I stated there is no way around to build an artificial Ego to be able to draw new conclusions. Of course there is, but I doubt such a machine will be build in my lifetime.
The game of chess has about 10^50 possible positions and about 10^120 possible games. To compute the whole game tree to find the perfect play was, is, and probable will always be not computeable on classic computers. We have the Quantum-Computers in pipe, and it is yet unknown if such a perfect-play engine is possible on such a machine. Current computer chess engines on Von-Neumann computers use heuristics to find the best move via an in depth limited tree search, and they do this meanwhile on a super-human level.
So, we can view our process of thinking similar to engines playing chess, we use our mind to search the memeplex knowledge graph for answers and solutions, we search the known graph we call knowledge, and we search the unknown what we call intuition, creativity and alike.
So, yes, in theory we can build a machine, maybe an Hyper-Meme-Machine, which expands the whole possible knowledge graph at once and runs some kind of evalution on possible candidates of new conclusions. All without the need of an artificial Ego.
The question which remains open is if such an SKGS, speculative knowledge graph search, can be implemented in practicable manner on our classic computers nowadays or near future.
My intention with this blog was not to get into philosophy, but to document this project, Tummi. The further I walk this path, the further I realize that it is not about building just an oracle kind of machine, an question-answering-system, but to go deeper, to build an system which is able to think, reflect its thinking, and come up with new conclusions. Therefore we have to build some kind of Ego, "Cogito, ergo sum", there is no way around this, and that is the point where philosophy steps in. The human kind does not consist only of the mind, the Ego, I like to see it threefolded, it is the body, mind and soul which makes us as a whole. So, what do we create if we build an artificial Ego, able to think and reflect its thinking, but we are not able to map the other parts? Should we do this? Should we build such an limited Ego, knowing that there is more? I am for sure not the only one who ponders on this, so let me refer to Prof. Dr. Thomas Metzinger:
7. Warum wir es nicht tun sollten in German
7. Why we shouldn't do it on Google Translate
I still have no numbers to extrapolate but assuming a whole Wikipedia with millions of articles parsed as meme pool in an RDF structure will need terabytes of RAM then probably memory bandwidth will be the bottleneck for the SPARQL queries. Currently we have up to 64 cores with 8 DDR4 channels with each ~25 GB/s on a single socket system, IBM's enterprise server may have a bit more throughput, but I guess until we have some kind of 1:1 ratio for cores:channels Tummi won't be able to query the whole Wikipedia in reasonable time...still some time to pass.
Okay, I dropped the Frankenstein parts and am back on the track, primary goal is simply Tummi, the intended meme machine for natural language processing, then, optional, Tummii - math n algorithms, then, optional, Tummiii - the Xi calculus...seriously, the first level will be already enough to crack for me.
Hmm, okay, this started as an relative simple meme machine in mind, and now I am up to level 5 on pen n paper...
- Epsilon engine level I
- Epsilon engine level II
- Xi calculus
Epsilon I was the initial intend, a meme machine for processing natural language, Epsilon II was planed as extension to map math n algorithms on an memetic level, I will not get into the details here what the other levels mean, but it is simply too off and too Frankensteiny for just an hobby project...better play some didgeridoo or alike...
To be more precise, Tummi will be only a kind of front-end to the Epsilon engine, there will be a back-end, Luther, to view and edit the knowledge graphs, and I aim for another front-end, KEN (acronym for Karl Eugen Neumann), for language translations.
I plan different pipes for the Epsilon engine, to address different kind of queries for the same memepool as knowledge graph, but I did not work out any details yet...
Epsilon and front-end code is currently Ruby, with front-ends as simple Sinatra web-applications, database with SPARQL-endpoints will probably be Fuseki from the Jena project, cos they offer RDFs reasoner and SPARUL.
The first Tummi release is aimed as an proof of concept on the SQuA-Dataset with ~500 English Wikipedia articles, we will see how far I will get with this.
When I take a look at my list of Meme Machines we can classify these into three strands...
1. GOFAI - Good Old Fashioned AI
These are based on some kind of predicate logic and use languages like Prolog or LISP. START by MIT is one example.
2. Pattern Matching
One of its prominent examples are engines based on AIML, Artificial Intelligence Markup Language, like A.L.I.C.E. Up to now these AIML based chatbots achieved the best results in the Loebner Price competition.
3. Neural Networks
I guess it really took off with Google's BERT, the introduction of Transformers, in 2018, and now the race is up to create models with more layers and parameters to achieve better results in text comprehension, question answering (SQuAD) and summarization.
Did an flowchart of Tummi v0001 based on an meme structure encoded in RDF and queried via SPARQL, note that the system is interlingual with an memepool as base knowledge graph connected via dictionaries with an lexis.
Here an overview of other meme machines...
*** updated on 2021-01-14 ***
2021 - Roadmap update for Epsilon I, II, III, IV
2020 - Roadmap for Tummi, Tummii and Tumiii.
2019 - Tummi v0001 pdf flowchart published.
2019 - Blog online.
2018 - Blueprint of an interlingual meme machine based on knowledge graphs
bootstrapped with human expert knowledge but able to parse content
2018 - Project reopened, Watson didn't make it.
2011 - Project canceled, IBM's Watson wins in Jeopardy.
2010 - First prototype with an simple ontology as knowledge graph.
2008 - Convinced that RDF/SPARQL offer enough flexibility for an meme machine.
2008 - Experiments with neural networks and RDF/SPARQL.
2005 - Experiments with AIML.
2004 - Inspired by Kiwi Logic's virtual agents.
2003 - Convinced that an meme machine could answer IT HelpDesk emails.
2001 - Experiments with OOP and meme replication.
2001 - Journey starts, inspired by 'The Meme Machine' by Susan Blackmore who
introduces the idea of artificial meme machines.
*** updated on 2021-12-20 ***
This blog is about Tummi, my attempt to create an artificial meme machine that is able to parse content in natural language and answer questions in natural language.
The last time I started such a hobby project it took me about 10 years to get into the techniques and understand the underlying principles. So maybe anno 2028 I will be able to judge if this blog was a foolish idea or not.
The name Tummi is derived from 'The Meme Machine' by Susan Blackmore and 'The Ultimate Machine' by Claude Shannon.