Tell HN: Anthropic's Claude is significantly better than ChatGPT
I've spent the last 3 hours with the model learning about Chinese history and culture, including getting recommendations for literature, films, documentaries, and first hand accounts of specific periods. Academic texts from the 1910s and niche foreign language films with only a handful of English reviews were offered without a single hallucination. Historical analyses were nuanced and contextual but highly factual -- I stopped cross-checking info about an hour in, but fabrications were again not an issue.
In addition and in general, and across several conversations on different domains, it exhibits a different caliber of reasoning. It can think strategically about business and markets, it can actually capture something of the essence of an author/poet's voice, and give highly thoughtful critical analysis of texts. It can follow instructions and grok things very quickly.
Beyond that, it has an ineffably human quality that doesn't come through with competitors. Discussions feel vibrant and dynamic, with none of the robotically dull tonality of ChatGPT. And I can't stress enough the difference it makes experientially to not always be second guessing the validity of the output. ChatGPT is great but the spell is broken any time it veers into its own imagined world, here the trust is seldomly misplaced.
I'm sure if you actively went searching for shortcomings or hallucinations you could conjure them, but this thing could be the backbone of a state of the art learning platform today.
[0]: https://poe.com/claude
12 comments
[ 3.1 ms ] story [ 32.7 ms ] threadAdditionally, unless there's been some sort of breakthrough that changes the way these programs work, there's no way to verify the claims except by manually researching behind it whenever it tells you something. I find that particularly distressing, because of the authoritative nature to the UX.
I discussed some literature I was very familiar with. It did feel like conversing with a professor, or at least a very informed and articulate layperson.
So in terms of discussing things I know, it's fantastic.
I still feel insecure about believing what it says about topics I don't know.
Edit: Wow. I am having an incredibly stimulating conversation about T.S. Eliot's poetry with Claude. It offers some interpretations which definitely feel like an AI might, but they are entirely plausible too, and when I share my take, it self corrects and demonstrates what appears to be a deep and thoughtful understanding. I am enriched.
To give some more concrete examples:
- I asked it about the advantages and disadvantages of different candidate fuels for nuclear fusion, and it correctly produced the reaction products of every combination, down to the MeV. When I tried this with ChatGPT it came back with invented, seemingly random answers.
- Asking ChatGPT for movie recommendations, even with broad guidelines, usually returns at least one hallucination and often misattributed directors or countries of origin. I asked Claude for recommendations that dealt with life in rural China and the experience of rapid modernization and it gave several accurate and precise responses [0]
- I asked ChatGPT to return two versions of a poem about a deer at the edge of a tree line in the hour before daybreak, one by Wallace Stevens and the other by Yeats. The two returned were stylistically identical, though still pretty impressive. Claude's were clearly more influenced by the poets' verse, and it offered detailed explanations for its choices.
[0]: https://www.imdb.com/title/tt1512201/, https://www.imdb.com/title/tt3907314/, https://www.imdb.com/title/tt0105197/, https://www.imdb.com/title/tt1114277/
Having seen the new South Park ChatGPT episode last night (ChatGPT, dude!) I had Claude write a love letter to someone in their native tongue and it does have a nice realistic quality that ChatGPT can exhibit but usually with a lot of fiddling.
I hope they release an API! I guess Anthropic might but this is overall a super refined chatbot.