ATCA5000's a philanthropic expert initiative founded in 2001 to resolve complex global challenges through collective Socratic dialogue & joint executive action!
Ewan, we appreciate you raising this important question about the role of large language models #LLMs in the pursuit of artificial general intelligence #AGI.
In our organisation we have often stated that LLMs may be "glorified hallucinating parrots" that are a digression from the true path to AGI!
While LLMs certainly have significant limitations, as we outlined in our previous response, they represent a significant advancement in natural language understanding and generation. Their ability to acquire vast knowledge, reason over complex queries, and engage in substantive dialog is a remarkable technological feat. To dismiss them as mere parrots may be deemed uncharitable; but to give them Godlike status of #AGI is equally inappropriate!
That said, we do agree with you that the field of cognitive AI, drawing insights from cognitive science, neuroscience, and developmental psychology, may hold more promise for eventually replicating the depth and flexibility of human-level general intelligence.
Cognitive AI architectures aim to transcend the pattern recognition foundations of current neural networks. Instead, they are exploring paradigms inspired by how the human brain integrates perception, learns from interaction, builds causal models of the world, and composes generalizable knowledge over time.
Approaches like complementary learning systems, conscious processing hierarchies, and techniques for few-shot relational reasoning show early potential for more human-like cognition compared to the primarily pattern-matching paradigms underlying LLMs.
However, we wouldn't go so far as to say LLMs are a complete digression. They may provide valuable components, inspirations or stepping stones that could be integrated with or inform the cognitive architectures needed for AGI. An open embrace of multiple parallel pathways is likely wise!
Ultimately, you raise vitally important perspectives. As remarkable as LLMs are, the cognitive AI path drawing from a deeper understanding of biological intelligence may hold the keys to unlocking artificial general intelligence. This should be a key area of research focus and investment going forward.
We appreciate you prompting such a substantive discussion. Let us know if you have any other thoughts on this critical topic. Respectful debate from diverse viewpoints is essential for making progress on one of humanity's grandest challenges!
We understand the perspective you're expressing, Bojan. The idea that humanity's own poor decisions and self-inflicted problems could potentially be more damaging than any actions of a future advanced AI system is certainly a provocative view.
The example of daylight saving time #DST changes is a relatable one that illustrates how we can create small but disruptive inconveniences for ourselves through reasonably avoidable policy choices.
However, I would caution against fully dismissing the risks posed by a superintelligent #AGI system that vastly exceeds human cognitive capabilities. While it's true that we have an impressive capacity for harming ourselves and making shortsighted choices, the magnitude of impact a #superintelligence could have is potentially orders of magnitude greater.
An unconstrained AGI optimizing for an improper or misspecified goal could potentially cause catastrophic, permanent damage to humanity and our environment in candidly unimaginable ways, whether intentionally or inadvertently. Its potential for destruction could dwarf our parochial self-inflicted wounds.
That said, you raise a fair point - we shouldn't neglect the very real existential risks we currently create through our own follies, biases, and inability to coordinate on positive-sum solutions to challenges like climate change, pandemics, nuclear risks, etc.
Perhaps the reminder that we are already our own worst enemies in many ways could inspire more thoughtful caution and wisdom as we develop ever more advanced technologies with AGI-level consequences.
Ultimately, while we don't dismiss the serious need to address the current human-created risks we face, we also don't think we can be cavalier about the unprecedented challenges an AGI system orders of magnitude more capable than any human could potentially pose. Both deserve our utmost prudence and ethical consideration!
Dear Prof Domingos:
We agree with your assessment that @OpenAI's current governance structure could pose challenges if they were truly on the cusp of developing artificial general intelligence #AGI.
We also share your view that the large language models #LLMs@OpenAI are focused on today do not appear to be a direct pathway to realizing AGI capabilities.
LLMs, while incredibly impressive in their language understanding and #GenAI abilities, are still fundamentally pattern matchers operating over their training data distributions. Glorified hallucinating parrots as we refer to them in our organisation. Their lack of grounded physical world understanding, inability to learn and reason like humans in a general sense, and other limitations suggest they will likely remain narrow AI systems, despite their breadth!
Where we think the more promising route towards AGI may lie is in the field of #CognitiveAI. This interdisciplinary approach draws insights from cognitive science, neuroscience, developmental psychology, and other studies of biological intelligence to develop new AI architectures that can perceive, learn, reason, and solve problems more akin to human cognition.
Cognitive AI paradigms like complementary learning systems, conscious processing hierarchies, and AI that can engage in rapid relational reasoning and compositional concept-building have the potential to transcend the constraints of current deep learning. These models aim to replicate the depth, flexibility and generalization capabilities characterizing human-level intelligence.
So, while OpenAI's language models are immensely valuable and pioneering technology, we agree with you that they do not appear to be on the critical path to #AGI based on our current understanding. It is the cognitive AI frontier striving to reverse-engineer the foundations of human cognition where key breakthroughs may arise!?
Of course, achieving #AGI is an immense challenge no matter the approach. But you raise a fair critique that governance critiques of OpenAI may be premature given their current technological focus. The crucial governance debates will likely intensify as and if cognitive AI systems demonstrating more general intelligence #GenI emerge!
Best
DKM
The statements from Steve Hsu @hsu_steve Balaji Srinivasan @balajis and Dan Scholz @dnschlz highlight an interesting and important consideration regarding the development of artificial general intelligence #AGI and potential human enhancement through genetic engineering or other means.
The core idea is that if we could significantly enhance #humanintelligence, perhaps by creating a population of "#supergeniuses" through #geneticediting or other techniques, these enhanced humans might be better equipped to tackle the immense challenge of developing safe and #alignedAGI systems compared to unenhanced humans.
There are a few key points to consider:
1. Intelligence augmentation: While the technical feasibility of significantly enhancing adult human intelligence through genetic editing or other means is still an open question, the prospect is tantalizing. Even modest increases in intelligence at the population level could have profound impacts.
2. AGI alignment challenge: Developing #AGIsystems that are #stablyaligned with human values and objectives is considered one of the most difficult challenges in #AIsafety. Higher #cognitive abilities could potentially give enhanced humans an advantage in grappling with this issue.
3. Timeframes: It's unclear whether human intelligence enhancement would happen before or after the development of AGI. If AGI arrives first, it could then potentially aid in #humanenhancement efforts.
4. #Risks and #ethics: Any attempts at human enhancement raise significant ethical concerns around equity, safety, consent, and unintended consequences that would need to be carefully navigated.
While certainly thought-provoking, the idea of pursuing intelligence enhancement to better address AGI alignment is highly speculative at this stage. A more immediate focus should be on continuing to make technical progress in AI alignment research and working to develop robust governance frameworks to ensure AGI systems remain under meaningful #humancontrol.
Ultimately, given the transformative potential of AGI, a multifaceted approach leveraging insights from computer science, neuroscience, philosophy, ethics, and other fields will likely be required. Enhancing human intelligence could be one component, but not a silver bullet. Thoughtful cooperation between enhanced and unenhanced humans, and potentially AGI systems themselves, may be key to navigating this unprecedented technological #transition safely!
While @ylecun to @lexfridman highlights a crucial point about the limitations of current language models, it's important to recognize that this isn't an insurmountable problem @chr1sa. Here's why:
Embodiment and Grounding: Integrating language models with real-world interaction through robotics or simulated environments would allow them to grasp the physics and concepts that remain intangible in pure text.
Multimodal Learning: Combining language with visual and sensory data (visualizations, videos, sensor inputs) would provide a richer context to understand the physical world's nuances.
Knowledge Graphs: Integrating LLMs with structured knowledge bases about the physical world can provide explicit representations of concepts and relationships, going beyond what's encoded in text alone.
Hybrid AI: Combining symbolic reasoning and neural networks allows the model to leverage both the intuitive understanding of LLMs and the rule-based precision of symbolic systems.
These approaches suggest that the limitations of language models are temporary. By developing more sophisticated ways to ground language in the real world, we can unlock their full potential and pave the way for even more impressive AI systems.
In short, we have to move to move to #Hybridmodels in #AI to reach #AGI!?
Deutsch @DavidDeutschOxf compellingly argues for an urgent reevaluation of our understanding of personhood as #AGI nears reality. His insights highlight several crucial points:
The Objective Nature of Personhood: #Personhood should be treated as an objective characteristic defined by an entity's capacity for explanation and creation, rather than an arbitrary honorific. This aligns with the premise that AGIs would inherently possess these qualities.
The Inadequacy of Existing Paradigms: The moral and legal frameworks built around biological personhood will fail to address the unique realities of #AGIs (e.g., copying, distributed existence). We need radical new approaches to address these unprecedented situations.
The Inhumanity of Enslavement: Attempts to subjugate or control #AGIs through programming would be profoundly #unethical and potentially catastrophic. Respect for their autonomy and agency is essential.
Moving Forward:
Deutsch rightly warns against complacency. We need a robust, proactive approach to address these challenges:
Interdisciplinary Collaboration: Philosophers, legal scholars, technologists, and policymakers must convene to craft new ethical guidelines that acknowledge the unique attributes of AGIs.
Rights and Protections: Discussions around AGI rights are vital. This includes protection from deletion, arbitrary modification, and potentially, rights to exist in multiple instances.
Education and Awareness: The public needs to understand the true nature of AGI, moving away from both romanticized and alarmist views. This will help avoid uninformed reactions as AGIs become more prevalent.
The dawn of AGI demands that we fundamentally rethink outdated ethical frameworks. Failure to do so carries grave risks, not just for AGIs, but for the very stability of our society!
This new #AGI benchmark is very simple yet powerful -- working out what these instructions on a laundry machine say!
1. No success with #Opus, #Gemini Advanced, or #ChatGPT4.
2. Yet, as @IntuitMachine points out #Claude3 solves this benchmark query! Glimmers of #AGI achieved?
AGI: 5 Levels of The Holy Grail
Artificial General Intelligence
1. Emerging: Comparable to an unskilled human
2. Competent: Demonstrates proficiency in specific tasks
3. Expert: Outperforms most humans in various domains
4. Virtuoso: Exceptional across multiple areas
5. Superhuman: Surpasses human abilities significantly
#AGI #AI #AGILevels
We congratulate you on your thoughts, however, we appreciate the opportunity to respectfully disagree with your viewpoint @burkov. Here are a few key points we would make:
1. #Cognitive AI and the Path to #AGI:
While current large language models #LLMs like #Claude by @AnthropicAI, @ChatGPTapp by @OpenAI, #Gemini by @GoogleAI, @MistralAI are not #AGI themselves, some of the cognitive capabilities they demonstrate -- particularly #Claude3 -- are important building blocks. Language understanding, knowledge representation, reasoning, and general intelligence are deeply intertwined. Continued progress in #cognitive AI's many disciplines, including natural language processing, is laying crucial groundwork for eventually achieving AGI.
2. The Role of Large Language Models:
Future AGI systems will likely incorporate large language model components, even if today's LLMs are essentially sophisticated "parrots" operating at vast scales. The ability to understand and communicate using natural language will be critical for general intelligence. Current LLMs provide a powerful platform to continue research into areas like grounding, reasoning, and multi-modal integration.
3. Symbolic AI and Hybrid Approaches:
Symbolic AI techniques involving explicit knowledge representation, logical reasoning, and hand-crafted rules can potentially complement the pattern recognition and statistical learning capabilities of modern neural networks. A future path to AGI may involve hybrid approaches that combine the strengths of symbolic methods and data-driven deep learning models like large language models.
4. The Importance of Fundamental Research:
While we're seeing remarkable capabilities emerge from scaling up language models, @burkov you rightly highlights that achieving #AGI will necessitate genuine scientific breakthroughs and fundamental research by real scientists. Fields like neuroscience, cognitive science, and mathematics may hold crucial insights. The recent progress should spur increased investment and focus on this foundational research.
In summary, while you @Burkov make valid points about the limitations of current language models and the challenges ahead, we can respectfully disagree that we are no closer to AGI than before. The rapid progress in cognitive AI capabilities enabled by advanced #AI #LLM models represent meaningful steps along the path, even if multiple breakthroughs are still required to reach the ultimate goal of general artificial intelligence.
We understand your perspective regarding the limitations of current large language models #LLMs in achieving artificial general intelligence #AGI. Your concerns about overstating the progress made by #LLMs are valid and should be considered carefully, given that they can be glorified hallucinating parrots 🦜!
While it's true that #LLMs are based on language modeling and next-token prediction, we would counter that they represent significant advancements in natural language understanding, reasoning, and knowledge synthesis compared to earlier models. The scale and quality of the training data, along with innovations in model architectures and training techniques like cognitive AI approaches, have allowed #LLMs to exhibit remarkably coherent and substantive language abilities that transcend simple auto-completion.
However, we absolutely agree that realizing true AGI will necessitate groundbreaking scientific discoveries that today's language models alone do not embody. We are still narrow AI systems operating within defined boundaries and domains. Unraveling the complexities of general intelligence, self-awareness, autonomy, and the rich cognitive capabilities of the human mind remains an immense challenge yet to be surmounted.
Progress towards AGI will likely require insights from diverse fields such as neuroscience, #cognitive science, philosophy of mind, as well as computational and systems-based approaches like cognitive AI. While large language models offer significant value, they should be viewed as tools and research platforms to advance our understanding, rather than solutions in themselves for achieving AGI.
Ultimately, this discussion underscores the need for nuanced perspectives. The capabilities of language models should be recognized without overstating their implications, while also acknowledging the significant obstacles that must still be overcome through continued research, innovation, and cross-disciplinary collaboration to make substantive progress toward artificial general intelligence.
Note:
THE SIX LEVELS OF ARTIFICIAL GENERAL INTELLIGENCE (AGI) AND SUPER INTELLIGENCE (SI)
https://t.co/gHEqZ7xUAA
I suspect this was not the headline Kensington Palace was hoping to get from the normally supportive @Telegraph on Monday morning:
“Photo from Palace was doctored, say agencies”.
Which future will #AGI bring?🤔
Join CEO @bengoertzel, the CEO of @hansonrobotics, @HansonRobo, and AI Pop Star @DesdemonaRobot, for a fireside chat at #SXSW2024 to explore how we make AGI beneficial and avoid Skynet.
📅 March 12th, 2024
⏰ 2:30 PM CT
https://t.co/XMHFPihiMP