UNSW Computer Science and Engineering’s Post

Interesting summary of of Ilya Sutskever, OpenAI's co-founder, on the challenges facing superintelligent AI. I agree in the item on natural selection, but it warrants a much deeper discussion. Through an unfathomable amount of experimentation with organisms over a billion years, nature has honed motivation and capacity for all life forms, including intelligent life forms. Our AI has had very little "evolution" to develop motivation. I would add that the biggest challenge to superintelligence is that we have no good definition for what comprises intelligence and we also have no real idea about what might be limits to intelligence. I wonder if there are tradeoffs in AGI as there are in biological intelligence, like tradeoffs between energy, computability, stability and adaptability. Do/will we have enough energy for superintelligent AGI? Are some problems not reducible? Or is there just not enough information out there to warrant superintelligence? Does increasing an AGIs ability to learn undercut its stability and hence, intelligence, in a cooperative/competitive ecosystem? We tend to think superintelligence is infinite and stable near infinity. I think it may be chaotic. Or it might get incredibly bored with the paucity of information and problems nature provides and fall asleep. Perhaps the bigger question is will we even be able to know if an AGI is superintelligent? #ai #agi #superintelligence Stephen Wolfram https://lnkd.in/dh-QQ3ZG

Ilya Sutskever's powerful description of what it means for a neural network to learn has been stuck in my mind ever since I heard it. It is a beautiful, poetic expression of an impossibly complex mathematical process. I highly recommend checking it out. "The way to think about it is - when we train a large neural network to accurately predict the next word in lots of different texts from the internet - what it is doing is *learning a world model.* It may look on the surface that it is just learning statistical correlations in text. But it turns out that to “just learn” the statistical correlations in text - to compress them really well - what the neural network learns is some representation of the process that produced the text. What the neural network is learning is more and more aspects of the world. Of people, the human condition, their hopes, dreams and motivations, their interactions and situations we are in. The neural network learns a compressed, abstract, useable representation of that. This is what is being learned from accurately predicting the next word."

Ever wondered why AI models like DALL-E can't produce the exact same image twice, even with the same prompt? This stems from latent spaces. Latent Spaces are high-dimensional, abstract realms where data is transformed into compact, encoded forms. Each point in this space represents a potential image, and due to the stochastic nature of sampling, different points are selected with each generation attempt, resulting in unique variations. The complexity of latent spaces means that slight differences in sampling can lead to noticeably different images. The inherent randomness, combined with the way neural networks are trained, makes models like DALL-E less about replication and more about creativity.

OpenAI's latest research sheds light on decoding GPT-4's internal mechanisms by pinpointing 16 million features with sparse autoencoders. This breakthrough enhances our understanding of intricate AI models, ultimately bolstering trust. While advancements are evident, hurdles such as attaining complete coverage and tackling spurious activations persist. OpenAI generously shares these findings and resources openly, encouraging continuous exploration and progress in the field. https://lnkd.in/g7GYP7Bj

End of Transformers?! I've been working in the A.I. world for some time and always amazed how fast technology keeps changing. Meet newcomer Mamba, a Structured State Model (SSM) is similar to LLMs, that represent different "states" of a conversation with "mathematical models" instead of large neural networks. "The researchers, Albert Gu and Tri Dao, claim that Mamba is five times faster than Transformers and outperforms it on real data with sequences of up to a million tokens." Carnegie Mellon University Princeton University https://lnkd.in/gr9FXd8c #structuredstatemodels #transformers #mamba

Can neural networks be hacked? Can for example credit applications be automatically approved as they have been registered with certain specially constructed email addressses? When introducing AI into business, it's important to understand that they are only supposed to work with high reliability on typical input data. Not every part of a huge neural network needs to make sense (although we sometimes can attribute them to certain parts of the problem). Only the complete network has to give statistical good answers. This means that a typical neural network will always have some oddities. These oddities might statistically disappear on typical input data. But, there is (currently) no way around them. While the existence of these oddities has been well-known since the beginning of researching machine learning, new developments introduce AI systems to systematically and reliably exploit other AI systems. It's indeed a vulnerability of these AI-based systems as it can be used to systematically produce undesired answers and behaviour. There is currently no way around it. You should assume that it can be done. When evaluating the risks of our AI systems to be hacked, we should ask ourselves - whether somebody is in the position to feed carefully constructed data into our systems, and - what are the worst case consequences of manipulating our systems behaviour.

The more I have been reading on AI, the more exceedingly scared I become of the implications of creating AI systems more intelligent than the human beings (which does not seem to be a much distant future). It's with a sign of relief that I read articles like this one, but for sure, more funding needs to be put in safety research if we want to harness the power of AI while managing the risks associated. "For years, the primary method available to researchers trying to understand the capabilities and risks of new AI systems has simply been to chat with them. This approach, sometimes known as “red-teaming,” can help catch a model being toxic or dangerous, allowing researchers to build in safeguards before the model is released to the public. But it doesn’t help address one type of potential danger that some AI researchers are worried about: the risk of an AI system becoming smart enough to deceive its creators, hiding its capabilities from them until it can escape their control and potentially wreak havoc."

During the week that Google / Deepmind launched Gemini with their (equivocate?) video, a more significant AI event sneaked out, barely noticed, via a Torrent link on X. Mistral mysteriously posted a Torrent link to a bunch of eight model weights. Just weights, no complete model or inference code, no official announcement on their website or their Hugging Face presence. This appears to be the rumoured Mixture of Experts (MoE) next gen LLM from Mistral. MoE is rumoured to be the underpinning architecture of GPT-4.5 and 5. Through the magic of Community and Open Source, there are already a number of models and inference implementations based on these weights out there. MoE is a method of overcoming the current limitations of LLM training, finetuning and inference through orchestration of multiple models. If you want to learn more on MoE I recommend the article below. This event, intriguing as it is, reinforces the pace of AI development is accelerating and the Open Source LLM world is closing the gap. #ai #llm #opensource

A brief history of #AI, and how maths and probability (e.g. Bayesian statistics) are the true underpinning of this technology... Perhaps the geeks shall inherit the earth after all? Its useful to see how #machinelearning is a technological journey, not a single, sudden discovery within the last year. Equally it has roots in mathematical theories that go back centuries. As humans we live in a probabalistic world, not a deterministic one - which is why this technology is set to have such a profound effect on everyday life and work. #futureofwork https://lnkd.in/euYVQsKJ