Paul Cantrell
The experience @cocoaphony describes here: https://mastodon.social/@cocoaphony/112084068085219012
…sounds like where I expect LLM usage in real-world software development to actually land:
- Augmenting humans, not replacing
- Doesn’t •do• thinking; •prompts• thinking
- Plays to LLM strength: discovering and repeating patterns
- Useful with an unfamiliar tool / context
- Supplements experimentation, docs, and community
1/2
Paul Cantrell
- Usage pattern is “Give me some BS that would be typical in this context,” not “Magically write my code for me”
- LLM output needs extensive revision
- This revision become a learning tool, is useful because Rob treats it as a device for personal experimentation
- Thus using the LLM well learns hard on experience, senior dev skills
→ Making good use of LLM requires critical examination & close reading of code; •those• are skills for beginners + us educators to redouble our focus on now
2/2
Paul Cantrell
Repeating and expanding that last bit for emphasis:
Making good use of LLM / AI tools, and also surviving alongside / in spite of them (in code and elsewhere), requires critical examination and close reading.
•Those• skills — critical examination and close reading — are something we should redouble our focus on now, “we” being both beginners in new fields and educators like me.
3/2
Paul Cantrell
@ShadSterling
Indeed. If there’s any utility to LLMs over other forms of discovery for sample code suitable for experimentation, it’s their ability to discover and parrot relevant patterns. That is, I think, genuinely interesting and different from existing tools. If AI work were focused on that instead of the preposterous hype that’s driving it, I suspect it could do better on many fronts — yes, including energy efficiency!
ShadSterling
@inthehands sounds a little like transcribing some code from the back of a book and then making changes to it until it does what I want, which is how I originally learned to code. I don’t doubt LLMs can be useful that way, but a good code search engine would be better and use far less energy
Paul Cantrell
@stl8k I had not! Surprisingly nuanced and ambivalent results coming from Google Research, where there is incentive to cheerlead.
Robb B. 🇺🇸
@inthehands Have you seen this research from Google?https://arxiv.org/pdf/2402.18498v1.pdf
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.