harisrid Tech News

Should we be scared about AGI : Artificial General Intelligence?

A Primer

Hi all,

I want to delve a bit in AI, and whether I think it’s going to take over our jobs or not.

So I see a lot of opinion pieces, scare tactics, or LinkedIn articles which state that AI will take over jobs. But, let’s take a step back. Let’s actually investigate the folks face and how they ( probably ) use AI and ML in their day-to-day operations.

Firstly – how I use ( and don’t use ) AI at work?

At work, I use AI tools and AI integrations. I’ve used a ChatGPT code integration with Github CoPilot to write the following :

• Source code – I feed in sufficiently-detailed prompts with enough context and instructions such as “Write me a simplified class to populate three members in Python3”.
• Unit tests – I feed in source code and ask the tool to write up unit tests. The tools do a good job writing tests for happy path execution and unhappy path execution.
• Refactor code – I feed in a prompt to ask to clean my changelists or refactor my changelists by targeting specific gauges – human readability, conciseness, and organization.

I see it more like Microsoft Auto-correct’s capability¹ : the AI tools have their imperfections, and frequently, I have to dive in and clean up a couple of things. Seldom do I ship out AI code for immediate code review, because there’s a couple of modifications that I have to make, such as renaming variables or adjusting config values.

I don’t see AI tools as replacing my work, because there’s a lot of work AI tools can’t do :

• Design Document Write Ups – AI tools can’t write templated design docs and can’t easily take in feedback from engineers
• Brainstorming architecture – coding is easy, but figuring out new systems to design or reasoning around distributed system problems is harder.
• Final Code Reviews – there’s a lot of human audit that still goes on here. Especially for concerns of releasing information that shouldn’t be disclosed in a production environment.

Secondly – What exactly are companies doing?

What are the M.A.A.N.G. big tech, finance firms, or start ups even doing with AI and ML? That’s an excellent question, lemme circle back to you!

AI research has been there since the 80s or the 90s, but the lower cost computer and infrastructure to enable pragmatic use cases have come up in the past few years. Many companies are learning that they can adopt AI or ML to enhance their value proposition – expedite day-to-day operations, build out enhanced capabilities, or tackle previously untackleable problems. From my understandings, here’s what’s going on

• Big tech companies – focus on researching unknown models or setting up large-scale infrastructure for AI. Big tech companies have dedicated R&D teams, long-term financing, and the ability to take on the risks of releasing the latest SOTA ( state-of-the-art ) ML Models. But before doing so, they need to tackle some pragmatic issues
  • Creating or using the dedicated hardware : think of NVIDIA’s GPU chips.
  • The infrastructure.
  • The testing strategies needed to release reliable models.
• Non-tech focused companies – extending AI to existing business use cases. Companies outside big tech focus on immediate value delivery : can we directly use off-the-shelf AI models and infrastructure OR incorporate slight customizations to solve a business problem? Business problems encompass :
  • Maximizing advertising campaign revenue.
  • Financial trends predictions.
  • Determining optimal customer segments or cohort groups.
  • Detecting PII or faulty data.
• These companies usually lack a combination of the resources of the infra needed to create something new; hence, the sense of urgenecy to use what exists.

Thirdly – why haven’t all companies adopted AI?

That’s also another good question. I can quickly go over the three big challenges posing impediments to large-scale adoption.

Challenge #1 : It’s Upstream – we lack the data or the talent

The biggest issue is upstream – getting the data for AI models. It’s relatively easy to find an AI model and say “Hey we have models, we can start solving problems”. But we need data. Not just any data. High fidelity, high quality, comprehensive, properly-sampled data. And getting data requires solid data engineering talent and software engineering talent – these folks

a. Determine how to get data from multiple sources into their proper sinks.
b. Configure multi-stage ETL pipelines to cleanse, validate, and transform data.
c. Determine if they can source data from a given source

Getting the right data takes time. It’s not a single month tasks – it can take a quarter to a year to get the right data ( following approvals, legal & compliance, or other factors ).

Challenge #2 : Do the customers trust AI Model uses?

Turns out we can’t just throw the latest models all the time to end customers. For legal, compliance, or safety reasons, customers mandate strong understandings of how models arrived to their decisions, which restricts the models to those of high interpretability or high explainability. Deep learning or neural networks get ruled out in such cases in favor of classical models such as decision trees.

Challenge #3 : Do we know the business problems to solve with AI?

This is the biggest challenge. Yes, we have incredibly advanced models, but, do we know how to use them? Engineering talent, product owners, and leadership needs to execute feasibility studies, work with minimal POCs [ Proof-of-Concepts ], and establish the following :

a. Which business problems can be solved using ML and AI?
b. Even if we can solve them with ML, is ML overkill? Can we bias towards non-ML solutions?

Challenge #4 – Do we Have the Infrastructure For AI Use Cases?

It’s true that we’ve tackled major infrastructure-related challenges in the past – think about the distributed systems paradigms which undergird Big Data technologies ( e.g. Hadoop, Spark ) or large data processing systems ( e.g. Spark, Flink ).

But somehow, we haven’t fully been able to translate years of research and historical efforts from distributed systems or performance research in Operating Systems to the space of Machine Learning.

I can hypothesize a couple of reasons causing the limitations ( and modern understandings can be limited, maybe 10 years from now, the landscape will change ). Factors encompass :

• Complexity and size of data sets – Is it because ML models are so complex enough in terms of their sheer CPU, Mem, or Disk Utilization? Is it the size of datasets that need to be crunched?
• ML Specialized Needs – we have very solid infrastructure to store large data volumes, but only select databases specialized for large feature sets and expedited vectorized processing. Do we also have general capabilities – monitoring & observability – for ML infrastructure?
• Specialty Hardware Costs – Perhaps a good analogy lies in the the CPU vs. GPUs debate – there’s very few companies that specialize in GPUs, or the expensiveness of Graphics Cards?
• Talent Imbalance – observations suggest to me that more people are going into the data science part of ML ( working with models ) than the underlying compute and execution parts.

Footnotes

1. If I recall correctly, a couple of esteemed AI researchers and professors have also likened AI tools to be more similar to auto-correct or editing software. ↩︎