Catapulting concern of alignment issue

 

The other day I was watching this movie Primer (2004), I stumbled on it while random scrolling. Reminds me of the time when you end up in the library to escape the heat and randomly search for something to read -dustier part where not many people access is a better place, and end up reading something valuable like Whitman or Naipaul. With internet random reading has significantly reduced and has shifted to random watching -mostly lowest common denominator eyeball trash -I am told attention seeking doom economy is where capitalism is moving with neoliberal alacrity! Primer is a low budget movie that has attained cult status. It is about a bunch of engineers/innovators who dabble with technology in their basement to stumble upon something exceeding their wildest expectations -the machine to make thing to weigh less ends up into time machine capabilities. What is compelling is the director, producer, screen writer, and lead actor and technician of the movie Shane Carruth is also a mathematician and works as an engineer, doesn’t dumb down the language for the audience like most science fiction do. And what is remarkable is when they realize their breakthrough they don’t turn cartwheel or jump up triumphally fisting in the air, they are stunned which is a plausible expectation when you are dealing with mindboggling complexities and pioneer portal. The scared deadpan expression of enormity is compelling realistic watch and indeed respite from usual cringe that you are used to watch. There are few takeaways for contemporary context when we are discussing fast evolving Cambrian scene of AI. Once the protagonists realized the potential of their invention, they upscale it for human size with full knowledge of their recklessness: "I can think of no way in which this thing can be considered remotely safe," says one. The enormity of their creation simultaneously presents human flaws into sharp focus. They become untrustworthy and greedy, and use it for immediate financial gain by playing the stock market and betting on college basketball games, and later for personal aggrandizement, when one of them tries to reverse-engineer the events at a party in order to make himself the hero. Carruth says “I saw these guys as scientifically accomplished but ethically, morons. They never had any reasons before to have ethical questions. So, when they’re hit with this device, they’re blindsided by it. The first thing they do is make money with it. They’re not talking about the ethics of altering your former self…”.

Some of the antics of AI techies, faced with enormity of what they are handling, one cannot but cringe and agree with the description ‘ethical morons’, as also capitalism framework of greed that is fueling to reckless lowest common denominator to grab market. There surely is an alignment issue -need for AI to align with basic ethical framework, which smarter concerned people have already flagged and are working on institutional mechanism before it creates catastrophic impact. There are spectrum of risk, and existential risk of AGI is the farthest -indeed farfetched at this moment, so is the issue of loss of control -runaway AI. Threat to idea of truth as collective deal is breached by algorithm. AI is simulating its own world. Deepfakes, misinformation and disinformation can severely impact democratic institutions. There is a need for mechanism of digital authentication so that people are aware and have choice. AI is distilled version of collective hence biases are common, and gets amplified and is exacerbated. Centuries of work by remarkable people to mainstream and consolidate egalitarian humanist value system is being challenged by powerful rogue technology in the hands of ethical morons.

Face by face the answers seem uncertain
Young and old, proud icons are dismissed
Can machines ever see my queens as I view them?
Can machines ever see our grandmothers as we knew them?                                                                            (from poem AI, Ain't I A Woman by Joy Buolamwini, founder of the Algorithmic Justice League) 

AI humanoid with female face are gendered within regressive frame of male gaze stereotype. All knowing humanoid that is simultaneously a know nothing nubile seductress. Privacy is another matter that needs attention -studies shows that AI will help leaders with dictatorial tendencies to consolidate and authoritarian regimes to control. Weaponized AI triggered by black box decisions is a grievous matter hence call for strict human oversight despite apprehension of losing superior AI reasoning over human. There are calls for regulations without impeding innovation, that is, transparency and oversight on foundational models but regulating start up that are fine tuning these and interfacing with public. Which brings in an important issue of public sector investment in AI. Foundational models are prohibitively expensive (the other day it was reported that Meta bought 30b$ worth GPUs from Nvidia) and are monopolized by few big private players. Much of the world will not have access to these latest innovations and will be dependent on BigTech. There is a need for international foundational model, preferably under UN AI that is accessible to all member state for research and so on.

There is also immediate need for high level expert group on AI in UN that will provide alliance and framework for regulation of AI which is dynamically synced with ever changing challenging dimensions of AI and its implications that is increasingly difficult to predict as also cumulative complexities with breakthroughs in other fields.

Mad Max saga continues!

 

There are very few movies that I watch more than once, also not much into 'action' movies -mindless nonsense with spectacle of gore and violence with flimsy story line or characters (we are no longer in the age of Bruce Lee or Jackie Chan when action was art and fun to watch). But Mad Max series is different genre altogether. Raw and compelling. Crazy guitar playing fire spewing maniac in battle field rough chase in desert like alien terrain is motif of an epic. "Witness"!    

How creative is AI?

 I am quite enamored by scientific breakthroughs and understanding, as also interesting applications through technologies. Mindboggling AI development in the last few years, and even months, is happening realtime. This ‘general-purpose technology’ has the potential to quantitatively change world, and unlike previous general purpose technologies that impacted human society (for instance printing press, steam engine, electricity, telephone, automobile, transistors, internet etc.) AI is tremendous in its scale and speed. With deep learning neural network and geometrically progressed transistor intense computational power (scaling through GPUs) churning on internet created stupendously large data AI has suddenly come alive. Transistor architecture worked LLMs into generative AI. With multimodal inputs AI has started to experience the world, and moving from 2D language inputs to 3D awareness with ever refining sense data that is expected to move from one domain to another effortlessly. Recent history of technology was about manipulating atoms -from tools to machines, hydrocarbons to electricity, materials to medicine, the primary driver of new technologies was material -manipulation of atomic elements. In the last few decades technology moved towards sophistication and abstraction. “Information as core property of universe. It can be encoded in a binary format and is, in the form of DNA, at the core of life operates” (The Coming Wave, Mustafa Suleyman). So, bits (increasingly genes and soon qubits) has replaced atoms as force of innovations and inventions. Breathtaking speed at which other equally significant technologies like synthetic biology, quantum computing, robotics, nanotechnology, fusion energy etc. are evolving parallelly creates an exciting though extremely unpredictable ecosystem.

AI i.e. Artificial Intelligence is already moving into robotics -Physical Intelligence, to be more effective it will have to organically learn, indeed evolve, with complex reality. It is futile to discuss AGI -indeed without even base definition agreement, or consciousness -the computation created binary mode matching human consciousness -which is a byproduct of billions of years of evolution. Artificial Capable Intelligence (ACI) is suitable benchmark wherein AI can achieve complex goals with minimum oversight. This cuts the hype and help to realistically work the challenges humanity face. In this dynamically developing scene it is difficult to keep track, indeed arduous to filter out chaff from substance. Even among the experts, inside voices, there are polarized views on AI. On the one side we have Geoffrey Hinton (AlexNet 2012 deep learning used neural network backpropagation that essentially triggered the present AI revolution) who is deeply worried and then on the other extreme we have Yann LeCun (now with Meta, also forbearer of neural network) who is sceptic of the hype (indeed he even downplayed AlphaGo before it made the breakthrough, so clearly people at helm aren’t sure). You have many significant voices (I have read books, articles, watched videos/talks so on to get the nuances of where the AI is moving) and find likes of Demis Hassabis (heads DeepMind) quite grounded and value at the right place. Hassabis and his team understood the potential of deep generative neural network and used it to solve challenging problems humanity face. Protein folding has been one of science’s grand challenges for half a century. AlphaFold was able to predict how proteins might fold based on their DNA by training on the set of known proteins and extrapolating from them. In just about two years DeepMind uploaded some 200million structures of protein in one go, representing almost all known protein! AlphaFold’s contribution to biological science, medical research, pharmacy and ofcourse insight into process of life is pathbreaking, indeed a paradigm shift. It must be duly recognized.          

So, the question here is how creative AI is? LLM through chat models like GPT, Claude so on is able to do some amazing stuff and converse that may sound almost human. It can even write excellent poems. But then it is only spitting out from the patterns of what it is fed. However, one cannot be dismissive despite narrow frame of language learning. Language makes much of human world. True all other species are intelligent enough to negotiate the world, indeed thrive, without the use of language. But I dare argue surviving is not wholeness of living, and human life epitomizes the meaningfulness of life. Language is an indelible part of human progress, and much of human emotions, awareness, expressions, knowledge and complex thoughts and ideas are embedded in language. So, if AI is ingesting all these literature and writings that was ever written it is but expected that it will have some emergent understanding of what it is to be human. It may not understand much about senses, as a blind person may not about sight, but may have a fair idea of what it means to see by abstracting from knowledge of seeing. You really cannot say there is nothing there except language. If it is able to see patterns that humans are not even capable to then that means it is not merely spitting out or regurgitating. The patterns we don’t decipher or are even aware may crisscross, entangle, interact to create meaning that could be insightful. It is not mere language. It is coded human lives, the collective experiences and understanding of humanity. So, since it is not mere language even hallucination is limited, or else we don’t understand the pattern that it has spotted (except where things are factual because in bigger schemes of things facts are contextual and subjective therefore LLM falters -not in the intent but execution). It clearly is emergent, and is able to connect patterns we are incapable to decipher is by definition creative. DeepMind Co-founder Suleyman sometime back mentioned an interesting take on hallucination as mutation. Information of life coded through genes is in constant interaction with reality hence evolve in a gradual process, but sometimes with epigenetic awareness, a sudden change can creep in, life can take a leap, this response may resonate or falter. AI hallucination i.e. giving out wrong information is a problem in limited context of wrong and right but in a bigger context it is creative emergence.

Even in the limited non LLM context, whether it is AlphaGo or AlphaZero, it is clear that AI is able to see patterns that is way beyond human understanding. Feeding framework rules was enough for it to creatively work the context to bring out patterns that humans playing the game for centuries couldn’t fathom (just watch AlphaZero with Stockfish, you will be mindblown by pawn attack strategy. Not into Go game but move 37 is in the annals of AI wonder). So much so that it was able see relation between molecules to create new antibiotic (halicin), meaning, our science may not be knowing certain aspects of molecular relations that pattern was able to decipher. Soon AI should be able to decipher its own pattern and open up the black box and help us reverse engineer the knowledge in these deep patterns. With multimodal input, studying human actions and perceiving world in more dimensions (3D to start with, god knows how many dimensions are there, and what AI could bring out!) AI will become more capable and discerning with merging of other equally compelling technologies like advancement in precision gene modification through CRIPSRcas9 and DNA printing, and yes quantum computing (everytime it is “within ten years”! but yes Chinese have made massive advancement -being an authoritarian country a cause of concern) the shift from digital to quantum is change of narration from binaries to probables, fusion energy will definitely add to the mix limited by power source as also advances in synthetic data that will create perpetual high quality data source. What is of concern is AI in the hand of bad player, as also channel for deep fakes, disinformation and misinformation. There need to be regulation on AI in limited context ie dealing with right and wrong, binary of facts. Bigger context AI -that is beyond right or wrong, in the realm of subjective, where the probabilities converge to reveal something new, must be kept open. 

Unlike science technology is incentivized by profit hence the path of technological progress is narrow. There is a loss that is unnoticed when BigTech clamor over how to monetize AI. What possibilities are there for that part that is left out, unmonetized AI, and how that could change the world we know is a pertinent thought. In the faster and bigger computation what is also left out is slow thinking, the pause, the still. Will AI ever be able to decipher patternless world of silence?  AI is creative alright but is limited within the framework, as for LLM -within the framework of known human experience (from which it can extrapolate). An entirely new thought, new idea, a Einstein or a Chekov or Dickinson level insight is impossible (new herein can only emerge from entangling of deep patterns, within the frame of known). Human mind is astounding and can see patterns that is based on no known understanding. But AI has its own place, and surely cannot be negated, and is equally astounding in its own way. This comparison is futile. There is an iteration of billions of years with all the senses, critical thinking faculties and intuitions.

Recently there is a development through liquid neural network that is controllable, adaptable, energy efficient, and contrary to the stampede to scale up the network, it focuses on scaling down with few but richer nodes. They studied simple organism (in this case a worm called c elegans) and how they take decisions to face complexities of life and strategies of survival, and built neural system based on the math of few hundred neurons. Liquid neural network used 19 odd neurons for self-driving cars instead of mess of hundreds and thousands used. The next step for sustainable elegant solutions for technological challenges will have to come nature. Each specie is millions of years of R&D to face the complexities of life. They have survived and thrived means these are successful. Human brain is complex but uses very less energy unlike GPUs that run AI. It is estimated that by 2030 they will consume a quarter of total global energy produced. The recent advancement in Biotech through CRISPRcas9 is using billions of years of survival strategy of bacteria against virus. Technology has been capitalizing on algorithm from nature like evolutionary algorithm, swarm algorithm, or even algorithm from foraging strategy of slime mold that optimizes resources and dynamically interact with changing scenarios (when I first time read about these single cell slime molds -that even lacks neurons, I got goosebumps). There is an organic intelligence at work that is smart, solves problems and survives without as much as a neuron. Next time take a close look at wonderfully arranged hexagonal cells of beehive, it optimizes space, resources and material. Or know about fire beetles that can sense fire from 130km away. The more you know more you will be astounded (I also have a youtube channel that you can search). As technology advances and look for solutions take a closer look at nature. Understand how nature works, the physics, chemistry and maths that life has aligned with, and what Einstein called Spinoza’s god. Extropic.ai intents to revolutionize computation by making a bold move of reimagining the idea of computer. A computer without suppressing natural entropy of world, and leverage it as an asset, in harmony with entropy rather than against it. Nature’s elegant solutions iterates simple into complex.