UDLCO summary : The conversational transcripts below are initiated in a generativeAI community where there is a question posed around human cognition in the loop of AI consciousness that leads to the customary probings and feelings around human consciousness and finally AI gets subsumed into "I"
Conversational learning (adda transcripts) :
[2/27, 8:21 AM] Rakesh Biswas: To be optimistic :
[2/27, 8:01 AM] Rakesh Biswas: GAN gyan (apologize in advance if it appears adversarial):
[2/27, 8:13 AM] Rakesh Biswas: The last line is another Zen koan!
[2/27, 8:31 AM] Rakesh Biswas: Similar lecture :
[2/24, 3:23 PM] Sridhar : Sharing this post by Pranab Ghosh https://www.linkedin.com/
As a designer this opportunity to imagine GenAI in a complementary manner ie Human in the loop, is what unleashes the potential of what user experience, based on what human experience, knowledge, behavior, aspirations gets agency to work alongwith AI. As an analogy, in a previous period and the age of MIS, where users gave a request to a room full of lab coat wearing engineers who programmed on punch cards or cobol etc using legacy system, and placed a neat dot matrix printed output report. The human consumer had no influence. Today's AI is much like that. Maybe better when RAG enabled or fine tuned for a domain/boundary condition.
Human in the loop is more exciting and I wish to see more such collaboration between AI tech and design, research, humanities, cognitive psychologists, neuroaesthetic experts, behavioral scientists. Would love to hear other views.
[2/24, 3:54 PM] R K : The role of human in the loop will be different than the earlier paradigm since the output of GenAI is probabilistic rather than deterministic. So every session of interaction of a human with a system or a session with same query by another human may result in varying outputs. Unlike the example you cite, where deterministic output from human programmed systems resulted in predictable outputs which could be examined for errors, these LLM based systems will not result in the same output for the same query since it is generating the output as a series of probabilistic sequence of words, images, videos etc.
[2/24, 4:08 PM] D S : AI systems need disclaimer in bold. Especially in contexts with 2 outcomes.
Examples...
Trivial case: Your spouse is likely to prefer movie x (instead of movie y) for a date night.
Extreme case: You don't need to consult with a doctor (when you may need to).
This helps putting the human back in the decision making loop.
May be Gemini wanted to do something similar.
[2/24, 4:19 PM] SDH : Brilliant point on the distinction between tradition rule based tested systems vs generative ones trained and tuned. But, what if contexts change, like they do in case of human preferences? Would there be situations where variability of output but in deference to a prompt and human in the loop, will be an advantage? Preferred perhaps. What if the agency to manipulate or control trumps variance in output.
As an analogy, bunch of people with varying diet restriction, preferences approach a lavish buffet and they takeaway with agency and choice exactly what how much is piled on to a plate? The variance in output is reflected in each diner's plate. I am asking you to examine whether human in the loop can develop some variance. Again, if the argument is on unreliability of the AI from a QoS pov, then should we not reimagined our evaluation process frameworks?
Ram, call me out if I went tangential on your keen point 😀
[2/24, 4:23 PM] SDH : ND, it reminds me of Asimovs 3 laws...paraphrased as such -
1) A GenAI should not cause harm to humans or fail to prevent harm when capable of doing so.
2) A GenAI shall follow instructions provided by humans unless those commands contradict the first principle.
3) A GenAI must safeguard its own well-being, but only if doing so doesn't violate either the first or second principle.
[2/24, 4:46 PM] SrKn: IMHO deterministic? - it's perceived as such by the ones in the moment and probabilistic by ones not in the moment! 🙏
[2/25, 4:48 PM] RK : https://open.substack.com/pub/
This talks about a lot of the points you brought up regarding the value of human in the loop. AI has the propensity to remove the skill premium rather than taking away jobs completely.
[2/25, 5:06 PM] PK : After reading this I now understand the butlerian jihad in dune 😅
"Thou shalt not make a machine in the likeness of a human mind"
[2/25, 5:50 PM] SDH : Indeed! I am familiar with Sangeet's argument as an avid reader and occassional illustrator collaborating with him. That human in the loop augments initially for better AI generated outcomes, but eventually AI subsumes the human expertise, aka skill, to render human in the loop redundant. I argue that why look at human as a constant? After all human observes nature and invented the wheel, dome, engine, space Walker, observer of pale blue dot, poet, writer, artist. My argument is human in the loop is the very purpose of an AI's existence where AI can be generative, while human is inventive. Human in the loop is the real OG, goat and genius. Off topic, is there a philosophy that can explain deep tech? Langdon Winner perhaps?
[2/25, 6:24 PM] U T : this is a fantastic quotable quote sir “AI can be generative, while human is inventive” 🙂 - do (c) it!
[2/25, 6:37 PM] SDH: Sir, CC0 full copyleft 😀
[2/25, 6:41 PM] UT : There is a broader argument that AI will become “inventive” and that’s debatable. Will AI invent the E=mc2 equation equivalent or the new filed of Quantum Mechanics if its never trained on data. Will it figure out Gravitational waves or the DNA structure. Here the line is clear. Pure science type Inventive work can be human. But the line gets fuzzy on daily-inventive work, for the masses. What’s the real risk to the masses. Assuming, Right now AI or GenAI isnt in shape to replace humans adequately. Where do you draw the line of “inventiveness” of which humans sit to the right and AI to the left 🙂
[2/25, 6:43 PM] U T : Will AI stay in the domain of augmenting each field invented by humans, much more efficiently? Likely true. How much? Depends on the cutting-edge-ness of the field. Longer domain fields with documented know-how will be at threat likely?
[2/25, 6:53 PM] G N : With increasing levels of superinteligence in AI why should invention be beyond the realm of possibility ?
[2/25, 7:02 PM] SDH : Sir, because a human can be crazy and want to send a dummy in a car hurting through space (Musk) or demonstrate deep empathy with nature like in one straw revolution (Fukuoka). Human inventiveness defies pattern. But unless AI superintelligence stumbles on to schumpeters creative destruction to be inventive 🤔
[2/25, 7:08 PM] A E : https://arxiv.org/abs/2401.
[2/25, 10:46 PM] U T : Sir how will Regression models of even infinite capacity think of a real Problem and find a real solution and also test it with experimental techniques?
Regression equations by definition are incapable of accuracy beyond the data sets they were based on. Isn’t it?
[2/25, 10:48 PM] Atul Batra NPC: History has proven that humans underestimate the impact of technology in the long run
[2/25, 10:49 PM] Atul Batra NPC: When AGI is talked about, it is meant to signify that it will equal if not exceed human capability
[2/25, 11:26 PM] R$: I will hazard some "outer limit" guesses
1. Inference capabilities will improve dramatically from generation to generation (every 6 months)
2. Quantum / DNA to parallelize and speed up NN layers
3. Perturbations (of random, uniform, periodic, or disordered hyperuniformity nature) introduced to mimic sleep state transitions i.e. we may not need to model the whole brain but only the differences in processing between waking state thinking and sleep state dreaming (one day LLMs may even hallucinate 😉) to mix them and achieve "breakthroughs". After all humans do very well when they sleep on a problem.
[2/26, 9:33 AM] U T : Interesting times ahead Sir
On first principles, I do feel the underlying mathematical foundations of AI might need to shift to get outcomes away from Hallucinations (outcomes of predicting outside the range of Regression data set) to real human-level inventiveness
Deep sleep is an interesting construct- human ability to tap into domains of knowledge they haven’t consciously pursued (snake biting its tail - benzene molecule — for August Kekulé )- is beyond having data sets. It is jumping to an entirely new data set and domain
Sure you could encode even this mathematically but at the end of the day the human was intentional in solving a problem. Will machines replicate that intentionality of problem solving and be progressive in inventiveness? Or would they again mathematically hallucinate- rearing garbage domains and garbage new data sets?
I am not at the cutting edge of AI research so there is a lot I don’t know in this august gathering. I am simply using my human logic at best 😀
[2/26, 9:35 AM] U T : The other mathematical foundational change I hope should philosophically shift is creating a constrained problem-solving model for benefiting humanity - even if constructs like dream state have to be replicated (meaning imagination of new domains beyond current data sets)
That would set AI as a partner to human progress rather than negative outcomes we all worry about
[2/26, 9:38 AM] A E : Even now it is possible to chain LLM outputs to mimic a train of thought- LLMs may also engender better LLMs or better models. Will that approach lead to AGI is anybody’s guess - as of now LLMs are deterministic statistical inference engines
[2/26, 9:41 AM] R$: Agreed. Hence the doodling on my end wondering if reusing nature-tested distributions to perturb the determinism might lead to new desired outcomes.
[2/26, 11:23 AM] Atul Batra NPC: AI won't take over my job, AI will be my job 😅
[2/26, 11:25 AM] Atul Batra NPC: Jaspreet's upcoming session inspired this thought 🙏
[2/26, 11:26 AM] BNV: True. That's how it will be for the foreseeable future. The AI use cases for replacing man power may not be viable for some time to come
[2/26, 11:34 AM] JNK : For folks who haven’t read it yet… Sangeet, as usual provides a good framework and an argument:
[2/27, 8:21 AM] Rakesh Biswas: To be optimistic :
It won't eat my salary (it doesn't have a stomach for money) but it can eat my "food for thought" and redistribute my money in a fair manner among other humans keeping my visceral fat and cholesterol in check!
[2/26, 11:42 AM] P K : There is a model of the human mind, that it is a bunch of agents (aka parts) throwing down playing cards to get into awareness. And the one that has the high card wins awareness. “Consciousness” is when there is something in attention ie won the center stage.
https://www.lesswrong.com/s/
If this is how the mind works (and gets stuck usually), a multi agent driven intelligence will be competitive with a human agent sooner than later.
[2/26, 11:49 AM] U T : This is lovely
My only point of difference is on consciousness. According to Yogic tradition in India and eastern philosophy- consciousness survives death of the organism and its organs including brain, and nervous system. Consciousness is also known to exist without brain and body. It’s non material. I will have fun watching the blundering around “creation of consciousness”
It could be simulated at best in some abstract ways 😀
[2/27, 8:15 AM] Rakesh Biswas: The difference between embodied vs non embodied cognition.
Embodiment can be in any element either hard metal or soft carbon
[2/26, 11:53 AM] JNK : The movie 'Lucy' *_attempted_* to explain this concept... vaguely!!
[2/26, 12:02 PM] U T : Need to rewatch it. There are some nice AI movies around. Wifelike was rich in its complexity and exposed the criminal selfish ends to which humans will employ this tech !
(Ramana, of Arunachala, has described a distinct experience of consciousness being separate from material body. Tibetean and Indian yogic traditions have many amazing stories that I have researched for many years and enjoyed).
[2/26, 12:37 PM] K B : This is actually a Vedic concept - the "observer" (consciousness) and the "observed" (body). Other variants are "subject" and "object". In the AI world, the model itself is the "observed" while the GAN/discriminator is the "observer". A better GAN can improve "consciousness" of the model and reduce hallucinations.
Now that I think, AGI can be achieved not necessarily because the model gets better, it will rather be due to improvement in GAN/discriminators. AI experts can comment if this understanding is right? or flawed?
[2/27, 8:01 AM] Rakesh Biswas: GAN gyan (apologize in advance if it appears adversarial):
LLM Hallucinations are the necessary step to creative thinking (ref @~Jaspreet Bindra ) arising from Mr Brahma's (with a white flowing beard)'s uccharanas!
GAN is the pruning gun of Mr Shiva the ash laden gardener that provides an AGI shape to Mr Brahma's AI!
[2/26, 12:42 PM] U T : Here is the twist in the tale
Every yogic tradition at its final level speaks of a realm beyond thought. It’s termed as silence or stillness
All human and computer systems stay in domain of thought
What would a thoughtless ai system do? It has no code and no data. It has no value.
Do you see the Paradox here?
Zen practices tried to explain this paradox in human language via Koans:
When both hands are clapped a sound is produced; listen to the sound of one hand clapping
AI can create new Koans but never know zen🤣
[2/27, 8:13 AM] Rakesh Biswas: The last line is another Zen koan!
One can create koans and yet not know Zen!
[2/26, 1:18 PM] K B : Yes, you are right. But like interpretations of Vedic verses (no single correct answer), we will need to achieve the stillness-1 level for AGI. The stillness itself is open to interpretation (the objective/why).
[2/26, 1:19 PM] K B : We are currently treading on the realm of exceeding the rules of this group 😊
[2/26, 1:20 PM] H V : Hmm. Slightly different in this.
The observer in the Veda is a living thing ( Chethana ) . Otherwise the space can be the Brahmam since it is all pervading
[2/26, 2:32 PM] K B : I may be wrong, but different vedas & vashyas (interpretation) has annotated different meaning. Essence is there is no "one right answer". Our (interpretor's) personal situation leads to our individual conclusion, and hence it is relevant for all. Vedas are the LLM of the world - each wants something different from it, and it can provide.
[2/27, 8:28 AM] H V : What I meant is AI is not life and hence it cannot be the observer.
On the analogy you mentioned i would cite Vedas as the source/ training material and we ( humans ) are the LLM 's.
Which explains the hallucinations 🤣
[2/27, 8:36 AM] Rakesh Biswas: The consciousness who created vedic data for our biological LLMs to feed on is the source of all the illusion around us that can appear hallucinatory from time to time when Brahma gets busy around 8:30 AM in the group before group admins Shiva1,2,n... or his day job can bring him to book!
[2/26, 8:12 PM] R K : Consciousness is the data in your phone. You switch to new and better phone, migrate the data. The phone (body) becomes old and needs to be changed. Not the intelligent (soul) which lives forever.
You also sync your data to the could server, the invisible one (paramatman).
[2/27, 8:31 AM] Rakesh Biswas: Similar lecture :
Epilogue :
All the conversations that I inserted above after each previous person conversing was done hours later asynchronously and it may appear as if I'm talking to them synchronously as part of a synchronous conversation whereas the truth is that some of the humans in the conversation haven't replied back even as I tagged their conversations in my reply to theirs.
Thought : Is deep fake from deeper layered data interaction a result of nodal asynchrony?
The replies that came later asynchronously listed below as per timed transcripts :
[2/27, 8:01 AM] Rakesh Biswas: GAN gyan (apologize in advance if it appears adversarial):
LLM Hallucinations are the necessary step to creative thinking (ref @~Jaspreet Bindra ) arising from Mr Brahma's (with a white flowing beard)'s uccharanas!
GAN is the pruning gun of Mr Shiva the ash laden gardener that provides an AGI shape to Mr Brahma's AI!
More here : http://
[2/27, 8:13 AM] Rakesh Biswas: The last line is another Zen koan!
One can create koans and yet not know Zen!
[2/27, 8:15 AM] Rakesh Biswas: The difference between embodied vs non embodied cognition.
Embodiment can be in any element either hard metal or soft carbon
[2/27, 8:21 AM] Rakesh Biswas: To be optimistic :
It won't eat my salary (it doesn't have a stomach for money) but it can eat my "food for thought" and redistribute my money in a fair manner among other humans keeping my visceral fat and cholesterol in check!
[2/27, 8:31 AM] Rakesh Biswas: Similar lecture :
[2/27, 8:36 AM] Rakesh Biswas: The consciousness who created vedic data for our biological LLMs to feed on is the source of all the illusion around us that can appear hallucinatory from time to time when Brahma gets busy around 8:30 AM in the group before group admins Shiva1,2,n... or his day job can bring him to book!
The conversations below were generated in response to the last of my above replies asynchronously later as dated and timed below:
[2/27, 9:15 AM] A E : The notion that Vedas are a store house of eternal knowledge is just a myth propagated by those who have not taken out time to go through them - they are essentially hymns composed for recital which contains mantras, tantras ( rituals), sometimes lofty aphorisms, and occasionally brilliant poetry and a rich tapestry of historically important beliefs, people references and events. You should reserve the knowledge moniker to Upanishads - essentially the nastika school including Nyaya, Vaisheshika, Samkhya, Mimamsa and Vedanta - but they pale in comparison to the lofty heights Budhist philosophy reached including the work of Buddha, Nagarjuna, Dingnaga and others - primarily Madhyamika, Prajnaparamita and Sautantrika schools - which went deep into logic, epistemology, phenominology, ontology and ethics - which also has some original treatise on nature of truth and structure of language. But all these are early advances and western philosophical advances post renaissance including works of Spinoza, Kant, Schopenhauer, Heidegger and Wittgenstein has advanced the discipline leaps and bounds. I hope Indians start studying Budhist schools and give them their due.
[2/27, 9:21 AM] H V : It is Vedanta veda-Anta for a reason.
Sometimes skipping ahead will create confusion rather than enhance it.
Take the Mexican fisherman and American millionaire ecample. Or Imagine a laser which is not sharp
[2/27, 9:24 AM] A E : Actually Vedanta is philosophy though mostly metaphysics where as Vedas are poetry - Budhists were the first to initiate organised research into knowledge in India. Please note Buddha negates the preponderance of the Vedas as do most nastika schools
No comments:
Post a Comment