Perplexity AI vs ChatGPT (2024): AI App Comparison

[Kraut]

Perplexity AI vs ChatGPT (2024): AI App Comparison

https://www.eweek.com/artificial-intell ... -chat-gpt/

Verdict: There’s no clear cut winner across all use cases, but after using both of these apps in a head-to-head matchup, Perplexity AI comes out on top for me thanks to its higher level of accuracy and depth in its answers to research questions.

Perplexity AI and ChatGPT are two of the most popular AI platforms, both offering advanced capabilities in language understanding and content generation. Each has its own strengths and weaknesses that make it better-suited to different use cases.

Perplexity AI: Best for Research-Oriented AI Assistance
ChatGPT: Best for a Wide Range of Conversational AI Uses

Perplexity AI is essentially an AI-powered search engine that draws from a database of sources to deliver source-backed, information-rich responses to your questions. ChatGPT, on the other hand, leverages OpenAI’s own GPT models to offer a range of capabilities, including creating original text and code, analyzing data, summarizing long documents, and mimicking human-like conversations. Here’s what you need to know about these two powerful tools and how they compare across key features, use cases, pricing, implementation, and more.

----------------------

https://www.zeit.de/digital/internet/20 ... enz-kritik

Not a search engine, but an answer engine: this is how the AI start-up Perplexity AI presents itself to the public. The US company is developing a chatbot that can access the internet and summarize answers to questions using artificial intelligence - with references to the relevant sources.

Perplexity aims to set itself apart from traditional search engines such as Google by generating summaries created by an AI language model rather than a list of links. The idea: instead of having to laboriously click through various websites, Perplexity delivers the information you are looking for in a bundled and edited form directly in the chat window. Current news topics are automatically prepared in a Discover section.

Translated with DeepL.com (free version)

Keine Such-, sondern eine Antwortmaschine: So inszeniert sich das KI-Start-up Perplexity AI in der Öffentlichkeit. Das US-Unternehmen entwickelt einen Chatbot, der auf das Internet zugreifen und Antworten auf Fragen mithilfe von künstlicher Intelligenz zusammenfassen können soll – mit Verweis auf die jeweiligen Quellen.

Perplexity will sich von traditionellen Suchmaschinen wie Google abgrenzen, indem es keine Liste mit Links, sondern von einem KI-Sprachmodell erstellte Zusammenfassungen erzeugt. Die Idee: Anstatt sich mühsam durch verschiedene Websites klicken zu müssen, liefert Perplexity die gesuchten Infos gebündelt und aufbereitet direkt im Chatfenster. In einer Entdecken-Sparte finden sich automatisch aufbereitete aktuelle Nachrichtenthemen.

[leosmith]

Thanks - I'm now a fan of Perplexity. It seems to be much better at finding and summarizing studies. I've been looking for confirmation that introducing new material, for example to an anki deck, will improve recall of old material. Perplexity confirmed it right off the bat. Chatgpt basically just said it was an interesting idea.

[Kraut]

Perplexity is using Claude AIs, but there has been an update from Claude Opus to Claude HAIKU 3.5 and people don't seem to be happy with it:

https://www.anthropic.com/customers/perplexity
https://www.reddit.com/r/perplexity_ai/ ... ?rdt=37396

If you used Perplexity for writing tasks or research, good luck. You now have to use sonnet 3.5 & haiku 3.5 which bullet points everything, never goes into detail, and writes in a very AI/Robotic/Cliche tone of voice.

[Kraut]

Perplexity also uses Claude 3.5 Sonnet (select with "rewrite") and there is an update called
"Claude 3.5 Sonnet New" which is being heavily criticized:

https://www.reddit.com/r/ClaudeAI/comme ... t3_1gxczg0

For me it isn't the output length, it's the writting style. I could somewhat offset that by throwing the whole file at it and it could spit out up to 2000 words. I enjoy reading webnovels, be it chinese, koreans, or japanese, I enjoy them all. For me, Sonnet 3.5 was the perfect tool to translate them from their origin language to english.

It's near perfect in translating, fluid in writting, and the readability is high. It genuinely able to compete with the quality output of experienced translators just like on those profesional website.

But unfortunately, the new update ruined that. The translations are still near perfect in the sense that it's accurate, but the writing style took a deep-dive. It become choppy and it's too concise to the point that reading them feels like something is clogging in my throat.

I'm not sure which version Perplexity is using right now, the old one or the new one.

[Cainntear]

Perplexity also uses Claude 3.5 Sonnet (select with "rewrite") and there is an update called
"Claude 3.5 Sonnet New" which is being heavily criticized:

https://www.reddit.com/r/ClaudeAI/comme ... t3_1gxczg0

For me it isn't the output length, it's the writting style. I could somewhat offset that by throwing the whole file at it and it could spit out up to 2000 words. I enjoy reading webnovels, be it chinese, koreans, or japanese, I enjoy them all. For me, Sonnet 3.5 was the perfect tool to translate them from their origin language to english.

It's near perfect in translating, fluid in writting, and the readability is high. It genuinely able to compete with the quality output of experienced translators just like on those profesional website.

But unfortunately, the new update ruined that. The translations are still near perfect in the sense that it's accurate, but the writing style took a deep-dive. It become choppy and it's too concise to the point that reading them feels like something is clogging in my throat.

I'm not sure which version Perplexity is using right now, the old one or the new one.

Interesting... I'm thinking that if they know what they're doing, they'll be sticking with the old version. Perplexity is doing the sensible thing of simply using the LLM as the front end user interface to a knowledge base. Anthropic, just like OpenAI, are chasing the sexy headline dream of "this is like a brain, so it does everything" and they're trying to treat the concept of knowledge as though it simply emerges from language, when it absolutely doesn't. I would think that if the new Claude is making less appealing writing, it's probably because freedom to make the nice feeling text comes at the cost of the accuracy of the information expressed in the message.

Perplexity is passing Claude data and telling it to put it into words, so there's less scope for "hallucinations" where it says something factually incorrect because it's statistically probably what the enquirer wants to hear, so they're probably getting no benefit from the new version and all the costs.

[leosmith]

I found another thing that perplexity does better than chatgpt. If I take a subtitle file for a 10 min video, and tell it "please translate to English but without timing data" it normally does it correctly and very quickly.

Chatgpt used to do it slowly and correctly, but now it fails miserably if the text is more than about 500 words. It will either stop completely at around 500 words, or it will be correct for a few hundred words then just make stuff up. The only way around this is to break the text into smaller pieces. Scolding chatgpt doesn't work for some reason.

[Kraut]

I found another thing that perplexity does better than chatgpt. If I take a subtitle file for a 10 min video, and tell it "please translate to English but without timing data" it normally does it correctly and very quickly.

Chatgpt used to do it slowly and correctly, but now it fails miserably if the text is more than about 500 words. It will either stop completely at around 500 words, or it will be correct for a few hundred words then just make stuff up. The only way around this is to break the text into smaller pieces. Scolding chatgpt doesn't work for some reason.

You could have it create a second srt file with a translation and play them both in SMPlayer. I think using Perplexity/Claude 3.5 Sonnet with a pop-up function (LEUNG's) is the best tool I have ever had.

[golyplot]

Chatgpt used to do it slowly and correctly, but now it fails miserably if the text is more than about 500 words. It will either stop completely at around 500 words, or it will be correct for a few hundred words then just make stuff up. The only way around this is to break the text into smaller pieces. Scolding chatgpt doesn't work for some reason.

That isn't a new thing. I've had that frequently happen with ChatGPT when I ask it to translate story chapters for as long as I've tried it. I suspect that it is an issue of context window limitations. As soon as the output grows long enough to push the initial prompt out of the context window, it forgets that it's supposed to be translating and just hallucinates a continuation of the story instead.

[Cainntear]

Chatgpt used to do it slowly and correctly, but now it fails miserably if the text is more than about 500 words. It will either stop completely at around 500 words, or it will be correct for a few hundred words then just make stuff up. The only way around this is to break the text into smaller pieces. Scolding chatgpt doesn't work for some reason.

That isn't a new thing. I've had that frequently happen with ChatGPT when I ask it to translate story chapters for as long as I've tried it. I suspect that it is an issue of context window limitations. As soon as the output grows long enough to push the initial prompt out of the context window, it forgets that it's supposed to be translating and just hallucinates a continuation of the story instead.

...and this demonstrates a practical problem with the existence of the term "hallucination". Here, you're using it in a way that is technically correct in how it is currently used in AI, but it's not how the average man in the street would understand it.

"Hallucination" is generally when a brain is doing something wrong, but here essentially it is doing exactly what is expected of it: it is producing language based on what's in its context window. When it has instructions in its context window, it is getting very very good at following them, but once the instructions are out of its context window, the instructions don't exist. What would a human do if given a piece of text in any given language with no instructions? Would they translate it? Would they respond to it? Would they continue it? We would not consider that a "hallucination", so if the AI bros want to claim that copyrights don't apply because "it's a brain", then they need to stop calling things "hallucinations" that wouldn't be called that if it was a human doing it....

[golyplot]

A human would probably ask for clarification, rather than just guessing what they were doing when it's not clear.


Incidentally, one time I saw ChatGPT translate the beginning and end of the chapter I provided but it somehow left out a key paragraph in the middle. I'm still mystified about how that could happen, since the context window hypothesis wouldn't explain that.