The Impact of AI with TDI 28

Threads Dev Interviews

I am finding developers on Threads and interviewing them, right on Threads.

See All the Interviews

Note: The views in these interviews are personal views and do not represent the interviewee’s employer.

“As AI proliferates, people will seek non-AI content to regain a sense of humanity.”

Sung Kim ( on Threads

Today we have, for regular life, what is going to be the biggest impact of AI in the next few years?

AI is a tool, similar to a smartphone, that we hope will enhance our lives. Like any tool, its impact depends on how the user uses it. One of the biggest impacts AI will have on our lives is the loss of content diversity, leading to more consistent content. As AI proliferates, people will seek non-AI content to regain a sense of humanity.

I am always reminded of what an ADAS expert stated a few years ago: we currently have the technology to drive automobiles 98% of the time without a human; the other 2% is really hard.

Let me provide a different analogy. When shopping for a hosting service from various cloud providers, you’ll find that pricing for 98% uptime is standard. However, the price increases tenfold for 99% uptime and a hundredfold for 99.8% uptime. I’m not sure where we stand with AI, but achieving those last few percentage points will likely be exponentially expensive and time-consuming.

What type of work do you do in NLP and LLM currently?

LLM is part of NLP. Since LLM is becoming multi-modal, it is also part of CV.

How do we evaluate the performance of Large Language Models (LLM)? Because there is no right answer to a paragraph of autogenerated text. How are results compared and how do we know if algorithms are getting better?

I would defer to AI researchers for a general evaluation of LLM. We tried to tailor LLM for our specific needs or to enhance its knowledge in a particular domain. In that domain, we seek highly curated questions and answer pairs to compare against LLM’s responses.

After that, it’s an iterative process involving LLM, where we focus on enhancing its knowledge in that specific domain and comparing the score distribution. The score is generated using different LLM, where we ask it to score LLM output against the ground truth, which is a highly curated answer given a specific question.

When you compare the output vs ground truth, is that a word by word comparison or based upon intent or something else?

It really depends on the business problems we are try to solve, but the most commonly used metric is a similarity score.

Is it possible to be a deep generalist in DS/ML?

I assume that’s the case, as I have expanded my expertise from CV to NLP over the last few years. BERT was an eye-opener for me in NLP; before that, working with NLP was just too painful.

What would you recommend for someone looking to get into LLM?

I would recommend them to fine-tune OpenAI’s gpt-3.5-turbo-instruct model for their business needs or personal needs. They have an extensive set of documentations on how-to and you will be pleasantly surprised by the result.

You will become a new AI person in your organization. 🙂

How can people find you elsewhere online?

Thank you for this opportunity. The best way to reach me is via the company I work for @mediumworx.

Read the full interview on Threads: @ryan.swanstrom • Threads Dev Interview #28 I am finding developers on Threads and interviewing them, right here on… • Threads






Leave a Reply