TU Wien
///////
The Grandfather should publish the method and apparatus used for computing the exact probability of 20%. Can he send the logistics and the training run? Maybe it was trained on a dataset with histories of civilizations, creating AI (1233055 images, collected by Fei Bei Mee from MIT via transgalatic connection by SETI with the aliens in 2452 Galaxies; ithe "Astro turk" service and student Klingons helped with the data labeling, under the supervision of captain Kirk; the model architecture was designed and implemented by Ilya Alexeev, a recent PhD student, and run for a week on his gamer's PC with a single 5090 GPU); Read more in the preprint at Arxiv 2602.12443.
///////
Dan Ivan Thanks for your comment and to the admins at TU Wien for tolerating critical and satirical comments. Speaking "more seriously", IMO any probability estimation is not serious or "scientific" and researchers talking with such figures are ridiculous (Bostrom as well). The problem is also ill defined. Even in gambling there is rigor, empirical data, attempts etc. These are random guesses based on no data, made by masters of statistical methods and proponents of empirical methods, working with "big data" from nature. In addition there is a Credit assignment problem. Logically "it could happen", there is a chance of a virus, a "dictator" pressing the "red button", asteroids, aliens, changes in the laws of nature etc. Another irony is that the art of operation with logic and little data to compute discrete and deterministic results belongs in theory from an "enemy camp" to his "connectionist" Bible (actually "Neural networks are also symbolic", Arnaudov, 2019 )
///////
Todor Arnaudov they are usually very transparent about the fact that these are nothing but their personal gut estimates.
It's a subjective indicator; they communicate subjective estimates, not testable hypotheses. Nothing dishonest or unprofessional about that, unless it is misrepresented as the latter.
///////
But why are they invited? Why don't you invite in some top university "Extra sensory masters", talkers with the outer galaxies - how big is the chance an Alien civilization contacting us in the next 3 years, 5 years, 10 years? How should we prepare, depending on different probabilities? The extra sensory masters may also have gut estimates. Hinton and others are taken and "sold" as authorities and given the right to speak, to fear monger or to be promoted, because their image is of scientists, their work is reproducible, based on "empirical data" etc., and not due to their skills in gambling or gut feelings. Yet this event includes the exact percentage as a headline in the title for clickbait, acting like bookmakers and emphasizing the "Extinction", not "Discussing the dangers" etc., which is "scientific". Note also that Hinton "et al." "Godfathers" were surprised and scared by the speed of development of the direction of "their" own research, i.e. they had *wrong* predictions by many decades, e.g. Bengio, i.e. they didn't really understand their own work so well. Ones who did predict it correctly are "cranks". Now the professors who didn't have a clue about their own work are authorities about futurology, where they were provenly completely off, among many other fake experts who are now monetizing trivialities, obvious and already banal and visible trends, on social media and social events.
///////
Todor Arnaudov 20% aka 1 in 5 is not that difficult to predict, that is one thing. But the situation becomes ironic that those who created/fabricated/developed now crying "Oh, we should be careful, this is definitely dangerous if developed at a higher level... we were not aware how fast and unpredictable developed itself after initial start-up!"... Go figure...
----
--- END OF QUOTE ---
///////
The Grandfather and the other aligners should read about the "Credit Assignment Problem", addressed in other posts on AGI List, including recently, and elsewhere, e.g. Minsky, 1960.
See the Minsky's 1950s and 1960 cited work for the original list of references:
As summarized in: https://github.com/Twenkid/SIGI-2025/blob/main/AGI-The_Prophets_Of_The_Thinking_Machines-Arnaudov_2025.md
* "The Prophets of the Thinking Machines: Artificial General Intelligence & Transhumanism: History, Theory and Pioneers; Past, Present and Future", Todor Arnaudov, SIGI-2025, p.478-479 (in the edition from Jan 2026):
[Translated by Google Translate: with minor formatting]
...
Prophets and pioneers. Scientists and Schools: The Prophets of Thinking Machines
(...)
Marvin
Minsky – another participant in the Dartmouth "workshop"
and a leading figure in classical AI [393], offering insights and
generalizations during its early decades.
* Steps
towards Artificial Intelligence, M.Minsky, 10.1960
https://web.archive.org/web/20231108233913/https:/web.media.mit.edu/~minsky/papers/steps.html
https://courses.csail.mit.edu/6.803/pdf/steps.pdf
"An interesting overview of the state of AI at the time, identifying
directions, concepts, methods, and generalizations that remain valid to this day —
such as reducing problem-solving to search (verifying proposed solutions),
gradient methods (hill-climbing), and local optimization (applicable when
the task can be represented as a sufficiently smooth function of coordinates).
The connection between gradient and heuristic approaches; the "predictive"
school of thought — mentioned towards the end in the context of Ray Solomonoff’s work;
and the importance of hierarchical modeling. Planning: the idea that any
form of planning is essentially predicting the future; systems with a "goal";
methods used in early game-playing programs like Samuel’s checkers, etc.
Classification (pattern recognition, model-schemas)
based on heuristically extracted key features of the objects under study —
features that must remain constant despite various forms of distortion
(invariance). The challenge of discovering new, useful recognition features
and combining them to form a recognition system. Decomposing complex
objects into sub-objects and describing the intricate relationships between
their parts. Recognition methods from that era. Mention is made of
learning Bayesian networks and maximum likelihood estimate.
Parsing images composed of nested geometric
figures: a specialized language/code for describing their types and
relationships (inside, to the right, to the left, below, …); nesting and
recursion — describing arbitrarily complex objects from a small number of
component parts; note the "ability to parse" — to "devote full attention"
to selected parts of the image and apply all available resources
(to attend fully — compare with *transformers* ) — when the current
description is insufficient to achieve the immediate goal. Etc.
Here, Minsky points to the "credit assignment problem" [394]
— a challenge that continues to be studied in machine
learning as of 2024; see the 2024 overview of approaches to it
in the footnote.
* [393] Also known as GOFAI – Good Old-Fashioned AI.
* [394] A Survey of Temporal Credit Assignment in Deep Reinforcement Learning, E. Pignatelli et al.,
July 2024 https://arxiv.org/pdf/2312.01072 Minsky's page on the MIT server was
accessible until 2023–2024, but as of December 20, 2024, it no longer loads.
Original in Bulgarian:
"
Пророци и пионери. Учени и школи: Пророците на мислещите машини
• Marvin Minsky – Марвин Мински
Друг от участиците в Дартмутската „научна работилница“ и водеща личност в класическия ИИ [393] с прозрения и обобщения в първите десетилетия. * Steps towards Artificial Intelligence, M.Minsky, 10.1960 https://web.archive.org/web/20231108233913/https://web.media.mit.edu/~minsk y/papers/steps.html https://courses.csail.mit.edu/6.803/pdf/steps.pdf
Интересен обзор на състоянието на ИИ тогава, с разпознати направления, понятия, методи и обобщения, валидни като цяло и до днес, като свеждане на решението до търсене – проверка на предложените решения, градиентни методи (hill-climbing) и локална оптимизация, когато задачата може да се представи като достатъчно гладка функция на координати. Връзката между градиентни и евристични подходи; както и „предсказващата“ школа, спомената в края с работата на Рей Соломонов; важността на йерархичното моделиране. Планиране: всяко планиране също е предвиждане на бъдещето, всяка система с „цел“, методите в ранните програми за игри като „шашки“-те на Самуел и т.н. Класификацията (разпознаване на образци, модели-схеми, шевици) чрез евристично извлечени важни особености от разглежданите обекти, които трябва да бъдат неизменни под различни форми на изкривявания („invariant”, invariance). Задачата за откриване на нови полезни за разпознаването особености и за съчетаване на множество от тях за да се образува система за разпознаване. Разделяне на сложните обекти на подобекти и описване на сложни отношения между частите им. Методи за разпознаване от тогава. Споменават се обучаващи се мрежи на Бейс и 393 Познат и като GOFAI – Good Old Fashioned AI – Добрият стар (старомоден) ИИ определяне на максимална правдоподобност (maximum likelihood estimate). Отчленяване на изображения, съставени от вложени геометрични фигури: език/код с тясно предназначение за описание на вида и за взаимоотношенията им (във, отдясно, отляво, под, …); вложеност и рекурсивност – описание на произволно сложни обекти от малък брой съставни части; забележи – „способност да отчлени – да „отдели пълно внимание“ на избрани част от картината и да приложи всички налични средства“ (to attend fully – сравни преобразители (transformers)), когато текущото описание не е достатъчно за постигане на настоящата цел. И пр. Тук Мински посочва задачата за „приписване на значимост“ 394 – credit assignment problem, която продължава да се изследва в машинното обучение и през 2024 г. – виж обзор на подходите към нея от 2024 г. в бележката под линия. Виж също препратките от работата към авторите от времето, напр.:"
394 A Survey of Temporal Credit Assignment in Deep Reinforcement Learning, E.Pignatelli et al., 7.2024 https://arxiv.org/pdf/2312.01072 Страницата на Мински на сървъра на MIT работеше до 2023-2024 г., но към 20.12.2024 вече не се отваря.
(...)
[393] Познат и като GOFAI – Good Old Fashioned AI – Добрият стар (старомоден) ИИ
[394] A Survey of Temporal Credit Assignment in Deep Reinforcement Learning, E.Pignatelli et al., 7.2024 https://arxiv.org/pdf/2312.01072 Страницата на Мински на сървъра на MIT работеше до 2023-2024 г., но към 20.12.2024 вече не се отваря.
...
