AI Shorts - Feb 2023

In this Article we covered

• Top News on AI in Feb/Mar 2023
• What is the fuss about AI Params - 100 Trillion??
• GTP 3.5 vs GPT 4
• Road Ahead
• Industry Level Impact
• Future of Software Development in AI world

Top News on AI

• Fighting ‘Woke AI,’ - Musk Recruits Team to Develop OpenAI Rival?[Opinion]
• GPT-4 launched ?- People are going crazy over the number of Params used to train the model.
• Programming sucks, so let an AI do it - The Future of Programming
• Google won't punish the AI-generated content and treat content as context regardless of how it is generated.
• Wonder Dynamics Officially Launches Wonder Studio - An AI tool that automatically animates, lights, and composes CG characters into a live-action scene - Steven Spielberg is advising the company.
• AI Can Re-create What You See from a Brain Scan
• Images generated by AI are not copyrightable .
• LaMA-13B developed by Meta, claims that it outperforms OpenAI's GPT-3 model despite being "10x smaller."?
• Discord chatbots will now be powered by OpenAI technology.

What is the fuss about AI Params - 100 Trillion??

100 trillion parameters is a lot - This is what you might have been hearing across the internet right now. But to let you know it is not an officially quoted number and just a speculated number. Possibly based on brain design facts like the brain has 80-100 billion?neurons?and around 100 Trillion Synapses.?

Synapses refer to the points of contact between neurons where information is passed from one neuron to the next.

AI model performance is often evaluated by the number of parameters it has. If we increase the model size, data, and computational power in a suitable way, we can improve language modeling performance in a predictable and gradual manner [Ref:??The Scaling Hypothesis ]

But number of tunable parameters is not the only criteria to judge the success of AI Models.

How do you pre-train and fine tune the models is the key to success of an AI

It is not the first time we are seeing models with Billions or Trillions of params - there are many?AI models ?which use trillions of params prior to GPT-4.

There are and would be a plethora of LLMs models which might be claiming to be the best in the industry but we should always try to understand the model using their training data and fine-tuning done to achieve the results. Here is one such analysis of Google Bard vs ChatGPT .

To understand these differences you may need to look at the details of how the AI model's behavior is shaped and built over time.

GPT 3.5 vs GPT 4

If you are a nerd, you would love to read this Technical Report of GPT-4?document. ?For the Rest of the people here is the summarised version of the same.

A Snippet of the sheet is attached below.

Road Ahead

We are so excited about LLMs that we have now started to question the existence of developers, writers, and creators but,

Building trust is a slow process and it requires time.

For example, I tried using ChatGPT to generate some parts of the article but it was not accurate and I had to refer to articles/research papers multiple times to get the right content.

The limitations like not having real-time data, hallucinations, and understanding of ethics, emotions, and biases are hard problems to crack and would need its due time.

But till we reach a level of trust with machines there are many use cases and applicability of AI which can be helpful in improving the productivity of systems as well as human.

Expect Better Acceptance of Subfields of AI Soon

GPT is NLP with Steroids but there are many subfields in AI that have the potential to disrupt industries. Subfields of AI would see good acceptance soon.

As GPT enabled machine to understand human language efficiently, it will make usability of other AI subfields convenient and easier for human.

For example, GPT can be used in computer vision to help machines recognize and label images more accurately, which can be beneficial for various industries such as healthcare and transportation.

In my recent conversation with Santosh Pawar (VP of Engineering, Pharmeasy), he suggested looking at the

Possibility to use AI to reduce the cost of decision making.

This is a painstaking problem for leaders - decisions are delayed because we don't have the right information available at the right place and sometimes we lose opportunities because of delayed responses.

Think about a future, where you don't need pull out dashboards to figure out a metrics but rather ask question to machine and get the answer in real time - Santosh Pawar

Look at the use cases as per the AI subfields' expertise.

A Snippet of the sheet is attached below.

Industry-Level Impact of AI

Replaced with AI soon or already happening...

1. Routine and Repetitive Tasks: Data Entry, Manufacturing Assembly Line, Common Query over ChatBots, Quality Control, Inventory Management, etc
2. Transportation: Self-driving cars and trucks, traffic management, route optimization, predictive maintenance alerts, etc.
3. Manufacturing and Production: Robotics, Quality Control, Supply chain optimization, production planning and scheduling, demand forecast, etc.
4. Finance and Accounting: Data Entry, Fraud Detection, Expense Reporting, Bookkeeping (generate statements, reconciliation), etc.
5. Healthcare: Medical Imaging Analysis (X-rays, CT scans, and MRI scans), Electronic Health Records (EHRs), Medical Diagnosis (patient data, including symptoms, medical history, and test results), accelerate Drug Discovery, Telemedicine (to answer basic health-related questions). If the health regulation of the country supports, it can eventually play a role in creating an affordable and highly accessible healthcare ecosystem.
6. Retail: Personalized Recommendations, Inventory Management, Customer Service, Visual Merchandising, Fraud Detection, etc.
7. Fashion: Analyze data from social media to identify popular styles, colors, and patterns, personalize recommendations, etc.
8. Education: Personalized Learning, Intelligent Tutoring Systems, Educational Content Creation, Grading and Assessment, and Student Support.
9. Legal: Document Review, Legal Research, Predictive Analytics, Legal Assistance, Alternative Dispute Resolution, etc.
10. Agriculture: Precision Farming (soil moisture levels, Weather patterns, and plant growth), Crop Monitoring (remote sensing technologies), Harvesting and Sorting, Livestock Monitoring (monitoring the health and behavior of livestock), Supply Chain Management (farm to market, reducing waste and improving efficiency).
11. Construction: Design and planning (analyze building designs and optimize them for efficiency and safety), optimize construction scheduling, detect potential safety hazards, quality control, safety monitoring, and monitor equipment performance,
12. Energy: Optimize energy usage, improve energy efficiency, and enable predictive maintenance of equipment.

Not everything can be replaced by AI

A few years ago I had a meeting with the CEO of a large Fashion E-Commerce Platform and we discussed the limitations of machines to predicting the next season's fashion and even in the age of AI how the role of fashion designers still be relevant. It is still true.

AI can help identify patterns and make predictions based on data, but it cannot replace the intuition, passion, and creativity of humans like designers and fashion experts (even with trillions of parameters).

It is good to be aware of the limitations so that we can take calls to reduce the waste of effort.

1. Creativity - Original Music, Fashion, Movies, Art and Literature would still a faraway dream.
2. Human Interactions: Like Counselling, Social Work, etc would require a human touch. Humans have a unique ability to emphasize and connect with others on an emotional level.
3. Judgment and Decision Making: AI can only help in providing insight into the data but the manner judge requires intuition and a complex decision-making process which is hard to replicate at the machine level.
4. Physical Dexterity: AI lacks sensitivity to the human hand and body. Like while performing complex heart surgeries, humans are able to adapt to changing circumstances and make decisions on the fly, whereas machines rely on pre-programmed algorithms and may struggle to handle unexpected situations.
5. Ethics and Morality: It requires empathy, compassion, and a deep understanding of human psychology. While AI can analyze data and make predictions, it cannot replace human values, ethics, and morality in fields such as law, journalism, or politics.

I would suggest not stressing too much about quick AI Adoption. The key is the understand your use cases and try to understand how this AI revolution going to help you in your journey.

Don't rush, but do ready to evaluate useful AI tools and application for your business use-cases.

The media is always hyped about the new things in the town. This kind of attention was also given to Quantum Computing (qubits) long back but to date, there is hardly any practical use of Quantum Computing.

Future of Software Development in the AI World

Software developers are responsible for designing, building, testing, and maintaining software applications. These responsibilities would remain the same but companies would need fewer people to handle huge software products.

As the adoption of IDE like Github Copilot increases, the interview criteria would also start to change. You should be good at prompt engineering and training AI machines.

Better Problem solving skills would be appreciated over coding skills as code would be readily available.

Sooner or later, programming language would come with support of prompt engineering to generate proficient code.

I have broken down the impact on Software Developers of AI, it would be in three stages :

Stage 1: Adoption of Prompt Engineering: In the first stage of AI adoption, it would be limited to converting the business requirement into a set of instructions for machines.

Stage 2: Understanding of Domain Knowledge: In this stage, the machine would have some ideas about your business domain and would reinforce knowledge based on the correction made by humans in code.

Stage 3: Product Developers as new role: Generation of code based on the clarity of the product documentation and understanding of problems.

Doubt me, look at character.ai , wonder dynamic, etc which can be used to create movies by one single person.

A decade later, In a perfect AI world

A good product requirement document should be sufficient to build a great product.

Development, testing, deployment and maintenance are all logical and not creative enough.

Production maintenance which impacts business continuity would be limited monitoring of resources (as cost is involved) or in disaster recovery situations.

Good Instructions over coding. Training machine would be a routine job.

Iron Man never wrote any code but only provided instructions to Jarvis.

More or less everything I wrote above is not uncertain and likely to happen in a phased manner. We need to accept this as reality and look forward to working with machines in collaboration. Writing code would no more be fun people. Start adopting tools like Github CoPilot and GPT-based tools for your advantage.

In a few years as the adoption of AI rises, you would soon see a lot of AI companies popping up and overall many engineering tasks would be concluded by AI.

These are the few companies started by Ex-employees of OpenAI.