Jumping Hills with Poe (LLM)

· 532 words · 3 minute read

↗ LinkedIn

I switched from Google computing infra after six years, which handles a multi-billion dollar scale, to the ML platform team of a small company that supports a handful of engineers. While the previous position taught me a lot about shipping high-quality C++ code, designing extensible abstractions, or leading complicated cross-team projects, most day-to-day operations of new work were so vastly different that I felt like starting from scratch.

For example, even though Kubernetes evolved from my previous team Borg, I couldn’t quickly grasp basic concepts like deployment or namespace. Similarly, I had little idea of Python static analysis or foundational ML architectures like two-tower. Even though my coworkers have always been helpful, keeping suppressed productivity for a long time is frustrating (and risky in today’s market condition). This high transition cost is one of the biggest blockers for seasoned developers to try a large change, making it hard to escape a local maximum.

Jumping hills with bridge

LLM is a game changer in providing a bridge from one hill to another. Examples:

  • I could quickly fix typing issues of non-trivial class relationships in Python with little previous language knowledge.1
  • I could also easily handle dependency conflicts when I didn’t know about standard package configurations like setup.py or tox.ini.2
  • I drew the above image simply following the command, as the bot knew the matplotlib much better than I.3

My existing skills and intuitions directed me toward solutions, but the exact outcome would’ve taken 10x more time with old tools. Of course, LLMs don’t provide perfect answers all the time, and we need judgment on incorporating them, but that’s where human skills continue to be essential. With the jump, I can bring impact by leveraging my expertise in a new environment where such skills are rare.

Although I may be biased, Poe by Quora has been a great interface to interact with LLMs for me. For example, it has access to the most advanced bots from both OpenAI and Anthropic (I noticed one bot provides a better answer than another, depending on the question4). Furthermore, with Quora’s focus and expertise, usability is superior with linkification and faster speed, and it really feels conversational.

In closing, I think LLM is an incredible productivity booster, potentially more substantial than the Cloud (my previous excitement). I’m still actively incorporating Poe in my daily routine to get more done, but it already has a significant impact on my day-to-day job. Like any invention, I noticed some are overly excited while others are too pessimistic. Wherever you are, I think LLMs are worth exploring and forming your opinions. Please let me know if you have any feedback on Poe too!

Disclaimer: Opinions are my own and do not reflect my employer’s position. See the official blog post for that.

Reference: I used the analogy of climbing the wrong hill cdixon in 2009.

  1. Conversation with Sage (free) about Python typing and inheritance. After a few iterations, productivity gain was obvious, so I subscribed to Poe and started using bots with limited access. ↩︎

  2. Conversation with GPT-4 to configure Python package with setup.py and tox.ini. ↩︎

  3. Conversation with GPT-4 to draw hill climbing diagram. ↩︎

  4. Conversation with Claude+ about Zen and the Art of Motorcycle Maintenance. ↩︎