OniGalore

Talk:Restless Souls/Technology: Difference between revisions

Page Discussion

Talk:Restless Souls/Technology (view source)

Revision as of 14:10, 14 March 2026

1,136 bytes removed ,  Yesterday at 14:10
m
no edit summary
mNo edit summary
Line 564: Line 564:
*** Unique training data can also sometimes be returned. Even if a piece of data appeared only once, if it is highly distinctive or contextually reinforced by similar patterns, the model may reproduce it with surprising fidelity. This occurs because the model has overfitted locally to these rare signals, making them more "retrievable" than typical generalized content.
*** Unique training data can also sometimes be returned. Even if a piece of data appeared only once, if it is highly distinctive or contextually reinforced by similar patterns, the model may reproduce it with surprising fidelity. This occurs because the model has overfitted locally to these rare signals, making them more "retrievable" than typical generalized content.
** Reinforcement learning from human feedback (RLHF) and (its successor RLAIF) can be named as another important feature that added a reward model for higher quality and alignment.
** Reinforcement learning from human feedback (RLHF) and (its successor RLAIF) can be named as another important feature that added a reward model for higher quality and alignment.
** Other features like chain of thought (reasoning), mixture of experts (MoE), context expansion and the use of external tools via model context protocol (MCP) to compensate own shortcomings are better described as incremental improvements in the evolution of GenAI.<!--
** Other features like chain of thought (reasoning), mixture of experts (MoE), context expansion and the use of external tools via model context protocol (MCP) to compensate own shortcomings are better described as incremental improvements in the evolution of GenAI.
* GPT = Generative pre-trained transformers (Large Language Model with the actual "learning" part de facto outsourced to humans: Reinforcement learning from human feedback (RLHF), in best case GPTs have a ''transplanted base intelligence'' but they lack the important feature to really learn for themselves. Low quality "synthetic data" can even worsen the models.
:: After the obvious slowdown in advancement through scaling, chain of thought (COT) was introduced. It is also known under the marketing term "reasoning [AI]". (See: Gemini 2.0 Flash Thinking and ChatGPT o3, the later one especially [https://techxplore.com/news/2024-12-ai-human-general-intelligence.html trained to cheat … erm ... score high in the ARC-AGI test].)
:: This pushes ANI somewhat more into direction of AGI with the drawback of slower response times and higher energy costs.-->
<!--
<!--
Autonomous AI (older ANI) vs. newer Agentic AI
Autonomous AI (older ANI) vs. newer Agentic AI
Line 573: Line 570:
* '''Agentic AI''' = autonomously working AI, typically utilizing increased reasoning and planning.
* '''Agentic AI''' = autonomously working AI, typically utilizing increased reasoning and planning.
: It is essentially an LLM with extended write access that can use external tools via MCP. By design, agentic AI systems (including their cloud-hosted LLM) still lack sufficient alignment. In other words, they are effectively beta software and can be dangerous both for one's own production environments and for others.
: It is essentially an LLM with extended write access that can use external tools via MCP. By design, agentic AI systems (including their cloud-hosted LLM) still lack sufficient alignment. In other words, they are effectively beta software and can be dangerous both for one's own production environments and for others.
:: It can therefore be seen as the ironic embodiment of the tech industry's saying: "Move fast and break things." Agentic AIs are, in many ways, public betas released into the world. People were impressed by demonstrations such as moltbook, which fuel the illusion that these programs possess true intelligence.
:: It can therefore be seen as the ironic embodiment of the tech industry's saying: "Move fast and break things." Unfortunately, people were impressed by demos such as moltbook, which fuel the illusion that these programs possess true intelligence.
:: As probabilistic systems - a.k.a. "statistical parrots" - LLMs tend to treat many possible outputs as valid solutions unless they are explicitly prohibited. [https://www.golem.de/news/unkontrollierbares-fehlverhalten-ki-agenten-werden-zu-immer-groesserem-insider-risiko-2603-206491.html Because hacking is fundamentally a creative act, even simple and seemingly harmless directives such as "be more creative" can lead to unintended or even catastrophic outcomes.]
:: As (still) probabilistic systems - a.k.a. "statistical parrots" - LLMs tend to treat many possible outputs as valid solutions unless they are explicitly prohibited. [https://www.golem.de/news/unkontrollierbares-fehlverhalten-ki-agenten-werden-zu-immer-groesserem-insider-risiko-2603-206491.html Because hacking is fundamentally a creative act, even simple and seemingly harmless directives such as "be more creative" can lead to unintended or even catastrophic outcomes.]
* '''World model''' = World models get trained on multimodal data, especially videos.
* '''World model''' = World models get trained on multimodal data, especially videos.
:: These models build an internal world and can '''better understand spacial inputs and forecast physics'''. Therefore they are '''also named predictive intelligence''' and are '''suited for''' applications like video synthesis, 3D simulations, animations and robotic motion planning therefore '''physical AI'''.
:: These models build an internal world and can '''better understand spacial inputs and forecast physics'''. Therefore they are '''also named predictive intelligence''' and are '''suited for''' applications like video synthesis, 3D simulations, animations and robotic motion planning therefore '''physical AI'''.
Line 582: Line 579:
:::: https://www.nvidia.com/en-us/glossary/world-models/
:::: https://www.nvidia.com/en-us/glossary/world-models/
* '''Physical AI''' = Physical Artificial Intelligence. Basically AI used in robots, including self-driving cars.
* '''Physical AI''' = Physical Artificial Intelligence. Basically AI used in robots, including self-driving cars.
:: Since direct training in the real world can be dangerous, slow, and therefore ineffective, the AI is typically pre-trained in a simulation where the robot is represented by a digital twin. This setup naturally supports multimodal learning (MML) for robots. Like humans (or other real organisms), AIs benefit from having an "inner world" to improve understanding and reasoning. Alternatively, motion capture data can be used for pre-training. The use of large language models (LLMs) is optional but can be a useful design choice to assist humans in directing such systems. <!--Expecting [[wp:Symbolic_artificial_intelligence#Neuro-symbolic_AI:_integrating_neural_and_symbolic_approaches|some kind of hybrid approaches]] within physical AIs later when all other current ''cash cow'' approaches reached their end.--><!--Embodied AI-->
:: Since direct training in the real world can be dangerous, slow, and therefore ineffective, the AI is typically pre-trained in a simulation where the robot is represented by a digital twin. This setup naturally supports multimodal learning (MML) for robots. Like humans (or other real organisms), AIs benefit from having an "inner world" to improve understanding and reasoning. Alternatively, motion capture data can be used for pre-training. The use of large language models (LLMs) is optional but can be a useful design choice to assist humans in directing such systems.<!--Embodied AI-->
::: An older approach for physical AIs was direct training in real environments and therefore learning from real sensory inputs. As this method is slower it will become out of fashion, at least as primary training. '''Real world training will be kept for fine-tuning.'''   
::: An older approach for physical AIs was direct training in real environments and therefore learning from real sensory inputs. As this method is slower it will become out of fashion, at least as primary training. '''Real world training will be kept for fine-tuning.'''   
* '''Symbolic AI and neuro-symbolic AI'''
* '''[[wp:Symbolic_artificial_intelligence#Neuro-symbolic_AI:_integrating_neural_and_symbolic_approaches|Symbolic AI and neuro-symbolic AI]]'''
:: [...]
:: [...]
:: Real abstract thinking probably requires also a form a symbolic AI. Since LLM and world models are already available, symbolic AI might be combined at some point with these approaches.
:: Real abstract thinking might require a form a symbolic AI. Since LLM and world models are already available, symbolic AI might be combined at some point with these approaches.
<!--* Further hybrid systems, optimization for ultra compact AGI-->
* '''AGI''' = Artificial General Intelligence, also "strong AI" ('''on par with human thinking''', ''a real AI'' capable to fully self-improve and drive its own development)
* '''AGI''' = Artificial General Intelligence, also "strong AI" ('''on par with human thinking''', ''a real AI'' capable to fully self-improve and drive its own development)
:: In discussions AGI is often equated with super intelligence. The argument is that as soon as AGI is achieved ASI is just around the corner. This understates the '''wide scope of human intelligence''' and that AGI is achieved first by hyperscalers, making further improvement difficult through further scaling.
:: In discussions AGI is often equated with super intelligence. The argument is that as soon as AGI is achieved ASI is just around the corner. This understates the '''wide scope of human intelligence''' and that AGI is achieved first by hyperscalers, making further improvement difficult through further scaling.
Paradox-01
8,772

edits

Retrieved from "https://wiki.oni2.net/Special:MobileDiff/46756"