Friday, March 29, 2024
Thursday, March 28, 2024
Wednesday, March 27, 2024
Attention is all you need
That is a groovy title for a
research paper that revolutionized natural language processing and created deep learning architecture
called Transformer that became the foundation for modern AI -the generative AI.
Previously neural network (Recurrent Neural Network -RNN) process data
sequentially -one word at a time in the order of appearance. Transformer was
able to capitalize on pattern that exist between words without being located in
proximity or are in sequence, this long distance dependencies in words is ‘attention
mechanism’. So, not all words are important only some are, and then you work
the pattern and imprint it simultaneously to get the meaning. This is pretty amazing.
Amazing because that is how I also read, infact all slow readers attempt such
tricks.
There are two steps in trying to comprehend written text. One is to understand basic structure of language, that is, grammar as also meaning of words. Since I really couldn’t understand grammar, I worked it by reading a lot and getting the structure, iterative learning -more you read better you become. So that when you write something wrong you get a feeling of something not quite right, so you pause and evaluate it further to get it right. Meaning of the words are always contextual hence iterative learning works. The more you read better you become while quality reading enhances your intellect. This is precisely how AI also learned language by getting pattern approximation through large data and powerful computation. I have discussed these few years back in context to GPT (https://depalan.blogspot.com/2021/12/human-intelligence-artificial.html). Now that you understand language how do you get the meaning is the second step. There was this big gap between wanting to read and capacity to read. I had racks of books to finish, and I was equally excited to know more as also understand high caliber patterns, but the capacity to read was woefully slow. I really couldn’t read more than few pages in a day, and sometimes I was stuck in word mesh and sentence loop, it is crazy but words tend to jump and play around, it frustrating to hold the words in their places while you try to get the meaning. I figured that there are two types of reading; one is deliberately slow, like for instance if you are reading nuanced writing like Kafka’s story or Dickinson’s poem, the unhinged words too add to depth of meaning, and the other one is fast reading -that is when you are reading nonfiction. The trick for fast reading is to identify signifier words (that encapsulates much meaning- 'holds attention'/tokens) and then to link a pattern to get the meaning. Though less number of words but the one that carry high approximation for meaning, and interestingly jumping words adds to unintended complexity to meaning. This is quite crazy but meaning gets into zone of high probability of not intended and eventual mess up. Probability of success vary with how much comfort you have in the subject; approximation is generally high if you are adept in identifying important signifier words. Indeed, I used to teach fast reading techniques for few years –a kind of job you finish fast with decent money then enough time to focus on real interests.
Wednesday, March 20, 2024
Game changing superchip
Much of economic growth of western
countries in the last many decades was riding on remarkable prescience of Moore’s
law- periodic increase in density of chip -doubling every two years. A few
years back people thought Moore’s law has hit a dead end -it is getting too
dense that it experiences quantum interference so on. The increased
computational power (apart from large data and neural learning) that led to AI
development has much to do with sophistication in chip -microprocessor. With remarkable
development and increasing computation demands of AI, CPU shifted to exponentially
powerful GPU – a massively parallelized processor suited to transformer parallel architecture. And just when you thought: how much
more can they squeeze in? comes the NVIDIA’s superchip Blackwell -B2100 is a
massive leap in computational power. If you saw the video of introduction by
Jensen Huang he slips in the worth -essentially money spent, that square piece
cost about 10b$. It however doesn’t really solve the problem of massive power consumption,
it is estimated that by 2030 Foundation Models will consume as much as 25% of
world power. Meanwhile Extropic is researching on Entropy computers -a paradigm
shift in basics of computer through unified AI and physics, working in natural
harmony with entropy thus scaling beyond constraints of digital computing and
substantially reduce power consumption. It is an attempt to emulate nature and embrace
chaos and noise into a symbiosis for a probabilistic AI algorithm.
The next AI wave is consolidating
on physical AI ie Robotics -access to multimodal data, which will understand
natural language and emulate movement by observing human action (Terrabytes and Exabytes of videos to tap on) to navigate, adapt and interact, meanwhile reflexive learning and dexterity.
Something akin to AGI is just around the corner. NVIDIA’s Project Grook is the foundational
model for humanoid robot. Let’s be quite clear ‘human like’ will no doubt need breakthrough
in quantum computing but nevertheless this is truly a remarkable achievement. Though
it is a technological breakthrough of many lifetime -that will have significant
impact much beyond work space, the social ramification will be severe. Surely, we
are living in the best of the times, we also have to accept the worst of times
with dangers it possesses. Technology so far is skirting a narrow balance
without major disruption but it sure is going to cascade very soon. Never has technology
so severely impacted society as also the meaning of what is to be a human.
Tuesday, March 19, 2024
Killer air
Air contains oxygen that is needed for human survival. What happens when this basic need turns toxic? Air pollution kills millions of people, and for millions more it exacerbate other debilitating health conditions, add to it contaminated water, pesticide laced food, and ofcourse mind-numbing noise that characterizes cities. Common people are seriously impacted by air quality, and they die without being counted -that is how agencies deal with problem by not being terribly concerned about collecting data (data is crucial to understand the gravity of the situation and formulate polices to save lives) or else manipulate data. Even short term exposure to polluted air can cause severe reactions -from deeply irritated air passage, persistent cough, headaches, burning eyes, weakness and disorientation, imagine living your whole life in these polluted pits. Human life is pliable to official narration. In a poor and desperate society, it is easily achieved. Fudging data is an art that even AI cannot emulate.
Today the world air pollution report was released. Out of top 50 most polluted cities in the world 43 are in India. If you look at the number of cities in top 30 it is almost all from India, and these are mostly hellholes of small towns where the development marauders are playing havoc. It is a statement on culture of apathy.
Thursday, March 14, 2024
How market capitalism unleashed development killed a garden city
In the early 2000s I shifted from
polluted, loud and crowded Delhi –a city of ugly uncouth people salivating and
stampeding around power so much so that the character of place percolates in its sinister clutch, to garden city of Bangalore (now Bengaluru). Bangalore
was a welcome respite with charming weather and mostly understated mild mannered
people. During those days even fans weren’t needed in summer. Huge trees dotted
the avenues with cascading flowers and wafting scent in the air. I used to go
for long walks or cycle, stop for warm delicious fluffy idlis, impeccable coffee,
streetside books (I recall picking up rare copy of Pearl Buck), and watch clear
bluest of sky. There were lakes, gardens, birds and butterflies…. there were so
many jungle mynas, purple sunbirds, and as you cycle few kms you could see
Indian rollers, Kingfishers, Drongos, Ibises, (even paradise flycatcher right
in the center of city -Cubbon park)…so on. There was so much diversity in urban
space. Oldtimers will tell you that the city had already started to crumble
with IT influx from late 1990s. Glass houses -ostensibly to keep off air
pollution and noise, and look chic, propping all around the place emulating desert architecture, ended up trapping heat hence began huge demand for air conditioners. Lakes vanished,
and so did the catchment areas and biodiversity (I recall spending time to
visit lakes and writing about it many years ago...please visit https://depalan.blogspot.com/2011/10/dying-lakes-of-bangalore.html), and yes, a quaint culture of care and
compassion. For few years I faced the brunt of polluted traffic as a two-wheeler driver (and surely reduced my life span, I still get wheezing and tremors, and assaulted spinal).
Changes in Bangalore were swift. Very
soon beautiful huge trees were cut, tacky flyovers and underpasses were
constructed, all the while charming buildings resolutely morphed into sickening
glass house clones as the city traffic got louder and louder spewing toxic
smoke. Birds and butterflies have vanished (I used to sit in the garden after
breakfast at NKB, there were so many birds and bees, and now after
ugly Mantri mall and noisy traffic there is nothing. The place is dead). Disease
of development took over the city. They systematically hacked out life from
every nook and corner, and sanitized it with signs of neo liberal progress. Where
once stood inviting Coffee house -overlooking street and garden, is now brick block
of glittery gold jewellery shop. The civilizational loss to market capitalism nurtured
greed and crass values is immense. They are now moving to Mysore like the medieval
marauders. The deadly swarm of development that will barren the land and eviscerate
life.