Gene Pool

© 2025 Praveen Puri

Big LLM, Small LLM

What if we trained small #LLMs, that are experts in one small tight, domain and don't require a lot of resources to run?

But they have the ability to remotely pass other questions back and forth to a more general LLM?

For example, what if there was an LLM trained specifically for your car as a "concierge", such as controlling the media player, air conditioning, seats, mirrors, etc. through human language? It could run on a few chips through your car, and be powered by the battery.

Then, if you suddenly asked a question on Nietzschean philosophy, the little LLM will remotely pass it along to a "big brother", like Chat GPT, in the cloud, and then relate the returned answer.

© 2025 Praveen Puri

Gotta Keep Up With Technology

© 2025 Praveen Puri

China, Energy, and AI

I was reading an article in the Economist that, because China is adding lots of electrical capacity (both renewable and non-renewable), they are positioned to have a long-term advantage in AI.  Already their electric grid has more spare capacity than ours and AI, as it's improving and doing more "deep thinking", has a voracious appetite for energy.

© 2025 Praveen Puri

Medical Statistics Puzzle

A study asked a group of doctors this basic #statistics question, and most got it wrong: A rare disease affects 1 out of 1000 people, and the test to detect it has a false positive rate of 5%. If someone tests positive, what is the chance that they actually have the disease?

Can you guess the correct answer?

(To view the answer, just click on the "Show/Hide Answer button" below)

© 2025 Praveen Puri

The answer is less than 2%. With 1000 people, there will be 1 person with the disease, and 50 false positives. That means a positive diagnosis would be 1/51 < 2%.

Should Amazon Compete in the War for AI Talent?

I don't see #Amazon as needing to compete with #OpenAI, #Meta, #Anthropic, etc. by developing their own #AI #LLM. The Amazon retail experience already has great usability, use of data, and recommendation / search algorithms.

I think they might be better off working agnostically with all model makers:

1. To offer the models in AWS for customers to use.

2. Getting all the model makers to embed APIs so that #chatbots can allow users to purchase products instantly that come up in query answers.

© 2025 Praveen Puri

I'm Increasingly Skeptical About AI Causing Mass Unemployment

I'm increasingly skeptical that #AI will put people out of work. History has shown that life becomes continually complex and new jobs are always created. I always write that the Information Age is gone, and we are now in the Attention Scarcity Age. Change is happening faster than ever, people are drowning in information, and attention / focus are the new scarce resources.

My prediction is that #LLMs and AI might take over basic, repetitive stuff, but human experts will be busy doing things, like evaluating #AIAgents. With all the change happening, to stay competitive, it will take full time people just to keep testing an evaluating which models and agents to use, how to deploy them, and help them cope with change.

We've been here before:

1. Historical - hundreds of thousands of jobs like blacksmiths, elevator operators, stenographers, and typesetters were eliminated, but new jobs were created.

2. McDonalds - They added ordering kiosks in stores and people thought it would eliminate jobs. Instead, the pandemic happened, and people switched to drive-through, and they added mobile ordering / curbside. That created jobs.

© 2025 Praveen Puri