F
François Chollet
fchollet.substack.com · Personal blog · on HN since 2022
François Chollet, creator of Keras, writes essays on AI, how large language models work, and education.
What HN saysHN commenters point to his LLM prompt engineering and 'LLMs as program databases' posts as a useful mental model, though one finds the framing reductive.
17HN people shared it
4linked it in comments
15stories submitted
93story points
2 years agolast shared
2022people per year2026
Most-shared posts
- 1I think about LLM prompt engineering8 people · 9 points · first shared 2023 · discussion on HN
- 2AI is cognitive automation, not cognitive autonomy6 people · 8 points · first shared 2022 · 1 comments on HN
- 3Education as civilization-building1 people · 69 points · first shared 2022 · 65 comments on HN
- 4The machine that makes the thing is more valuable than the thing1 people · 5 points · first shared 2022 · discussion on HN
- 5The Loop of Progress1 people · 2 points · first shared 2025 · discussion on HN
- 6Llms as program databases1 people · first shared 2025
What HN people wrote when they linked it
falcor84 · 2025-02-06
The original article [0] is from 2023 and was already posted on HN multiple times.
Anyway, while there's something interesting in this mental model, I find this attitude below and particularly the word "rather" to be overly reductive, seeing how you could argue that intelligence in general consists of searching in an abstract database of Platonic ideas.
> Viewing LLMs as vector program databases rather than intelligent entities provides a powerful mental model for systematic improvement.
[0] https://fchollet.substack.com/p/how-i-think-about-llm-prompt-engineeri…
iwsk · 2025-02-02
Actually it's way easier to search for the information in an LLM compared to an 8 tb harddisk.
The search is vague, and the result is nondeterministic, but in a lot of cases it's still the better method.
https://fchollet.substack.com/i/137628402/llms-as-program-databases I like the way this article puts it. "LLM is a continuous, interpolative kind of database"
xnx · 2024-11-02
This seems to have similar ideas to an post from a year ago by François Chollet: https://fchollet.substack.com/p/how-i-think-about-llm-prompt-engineeri…
svaha1728 · 2023-11-21
I would read Francois Chollet's explanation of this. It's very good: https://fchollet.substack.com/p/how-i-think-about-llm-prompt-engineeri…
For words that are not in the model's vocabulary, like 'fluftable', the model uses a subword tokenization strategy. It breaks down the word into smaller known subunits (subwords or characters) and represents each subunit with its own vector. By understanding the context in which 'fluftable' appears and comparing it to known words with similar subunits, the model can infer a plausible meaning for the word. This is done by analyzing the vector space in which these representations exist, observing how the vectors align or differ from those of known words.
'As always, the most important principle for understanding LLMs is that you should resist the temptation of anthropomorphizing them.'