Page 174 - CW E-Magazine (30-1-2024)

Special Report

the data they are trained on and do not fram Alfa, Ai.R MATH, etc. are a few will do unimaginably good things, but
just pick up existing information from examples. These serious problems will also have the potential to destroy
its database. are being rapidly ironed out in later the very fabric of rules, guidelines,
versions. in 2028, we may not even safety, trust, security, privacy, which
Soon LLMs may be called Multi- be using these clunky LLM which we cherish. Even today, trust is dead.
modal Ai or the like, and very soon enthral us today. That could happen We cannot trust as evidence, a photo-
all of them would not need to be even by 2025. graph, a video with known people, a
large. individual organisations will be signature, a document, a short phone
able to use some of these Foundational There are suspicions that the limi- call from someone, etc. They could
Models to create small, powerful Ai tations of the breakthrough archi- all be fake. All metrics for trust have
Models to run their organisations very tecture associated with GPT – the become untrustworthy barring bio-
differently, to do metrics. That is a
things which were Would you not at least create a team within your huge problem for al-
considered to be organisation to explore the various ways to use AI to most every aspect of
impossible now. in enhance competitiveness, and make it future ready? our lives. People are
2024, there will be a going to use GenAi
proliferation of Ais and sometimes on the
from small and big organisations and Transformers, which made Genera- sly, for all good reasons. Wouldn’t
by the year-end, many of the poorly tive Ai possible – are being overcome you take some steps now to improve
funded ones will bite the dust, unless with more powerful architecture by security and trust, rather than react
they are designed for very specific Project Q* at openAi (no one really after being hit by fraud?
applications, which have relatively knows, as it is top secret right now).
small addressable market, as opposed it is speculated to be Self Taught Rea- Think of development of Ai as a
the general-purpose Ais, where the soner (STaR), which will be even- spectrum of higher and higher cap-
big boys will thrive. Many companies tually capable of using Ais to train abilities ranging from ‘Zero Ai’ to
and countries are working on their other Ais, to quickly build capabilities. ‘Expert’ to ‘Superhuman’ or ‘God-
own foundational models due to justifi- Today, training of Models is costly like’ and many levels in between.
able FoMo. Would you not at least and time consuming. This will lead We have made huge progress in
create a team within your organisation to the development of Ais, which are Narrow AI to do some specific tasks,
to explore the various ways to use Ai much faster and more capable in ways but more general things like ChatGPT
to enhance competitiveness, and make which we humans cannot imagine. are only equal to or better than an
it future ready? does that mean Ais have surpassed unskilled human being. GPT 4 has
the levels of intelligence that humans moved a step further to being bet-
A GenAi like ChatGPT can tell have? No, but it is likely to be a ter than 50% of skilled humans. Mo
you in a jiffy what aspects to cover huge step towards the next level of Gawdat, the former Chief Business
in a seminar that you are planning Artificial General Intelligence (AGI). officer of Google thinks ChatGPT
for cardiac surgeons. it does not Humans have General intelligence. 4.5 has an IQ of about 155. AI has
suffer from writer’s block and knows AGi, which is the Holy Grail of Ai, not yet reached the ‘Expert’ level
nothing about seminars or cardiology. is not there yet, but not far away. in (defined as better than 90% of
There are others which claim to be as 2020, AGi was thought to be at least humans at a wide range of general
good or better – Claude, ChatSonic, 50 years away; in 2022, the estimate tasks and common sense). Expert level
Jasper Chat, Bard Ai, Bing Ai, etc. was 30-40 years away; in September AGi could happen as early as 2028
The problem is that GenAi tools feel 2023, the year varied between 2027 because Ai capability is doubling
obliged to spew out words even if it and 2035, depending on who you or tripling each year. imagine an Ai
does not know the correct answer. it talked to and depending on what Level with an IQ of 310 (=2x155), which is
suffers from biases that the training of AGi. in december 2023, we will nearly twice that of Einstein (162).
data has and has no real intelligence say that bits and pieces of it is pos- That could be the ‘Singularity’
to understand many topics. it can sible in 2025, but it will match most moment when machines will be smarter
generate objectionable content, hal- human beings, for many practical than most humans. it is difficult to
lucinate, etc. Many Ais are working purposes in 2028. When it happens, predict when Ai could have Super-
on improving math capabilities. Wol- it will be the ‘singularity’ moment. it human capabilities (2055 or 2075 or

174 Chemical Weekly January 30, 2024

Contents Index to Advertisers Index to Products Advertised