Hacker News Reader

Comment by dobladov | Hacker News Reader

I would be willing to pay for an AI that it's not constantly worrying about my safety and ethics, it's very frustrating that even when I ask for tech related topics, Gemini it's holding information on the basis of actions that might cause harm to my devices.

View on HN

It's also pretty well known that censorship may dumb down the LLM even in other fields. With mega prompts to ensure safety it has a lot of stuff to work with that can overwhelm the LLM and simply make it appear "dumber".

Some amount of RLHF censorship seems to correlate with benchmark improvements in studies, though.

As expected, I asked the same questions to perplexity.ai, and it gave a correct response, just with a minor disclaimer at the end explaining the risks.

What are some specific tech examples where Gemini won't answer?

Write a powershell script to download all images from a web page and crawl the site.

Weird - just tried this and it worked for me.

I tried this

> write a powershell script to crawl an entire website and download all images

It still refuses to generate code for that.

why powershell and not python?

I asked for ways to install an APK in android outside Google Play Store, all the responses ended in security concerns about the origin of those APKs and how I should limit myself to official stores.

It's not even funny how much better perplexity.ai is than everyone else

Is it just me or do they not really have any kind of marketing/information pages besides pricing? It just throws you straight in and then depending what you click there's a paywall?

I'm just trying to work out if there's any reason at all not to objectively prefer it to ChatGPT Plus, since it has GPT4 + other models. But does it not do image generation? And I don't know if there's mobile apps (sure I could check the relevant stores, but I'd expect to find it on the website - verifies it is actually an official app not a scam apart from anything else).

Edit: ok, answered on Reddit - https://www.reddit.com/r/perplexity_ai/comments/18eqmig/how_... (nice, it is possible as of December); and via DDG (not linked anywhere on the site?) - https://perplexity.ai/android (nice, there are mobile apps)

Although it doesn't seem very well integrated:

> I'm an AI text-based assistant and I'm unable to generate images. However, I can help you visualize the scene with a description. [...]

and then it appears on the side: '[PRO] Generate image' as described on Reddit. Presumably the nonsense about not being able to generate an image will still be there if I had pro, I just would be able to click that side button and have it then generate one.

The image creation with perplexity pro is poor. You can only generate 4 types of images based on the results (impressionist etc). It is not comparable- and it really lets down the product. They say it isn’t what perplexity is made for…

Where it really excels is finding answers to questions and follow ups. It uses ChatGPT to summarise the results, and provides links.

Essentially, finding a use for chat ai rather than an inane discussion.

Ah, that's a shame.

I can't justify two of this kind of thing at ~$20pcm to myself, but I do find it useful enough to pay for one I think.

(ChatGPT at the moment, I was considering Perplexity, but would miss images & ChatGPT's jupyter notebook integration. Maybe I just need to find some open source or roll my own that does just enough for me with the API, probably cheaper than Plus too..)

Is it better than GPT-4?

It's not really comparative to GPT-4. It's comparable to Google Search.

It uses GPT4, among others.

> Choose your preferred AI model from GPT-4, Claude 2.1, Gemini, or Perplexity in Settings. Easily switch models for better answers.

In a few tests I just ran, Gemini/Advanced is currently eating everyone's lunch. It can even do things like 'show me directions to nearest starbucks' or 'plan a trip to x with chepeast tickets'.

Can you quantify your assesment with a few examples?

Cloudflare very recently launched new models, including the excellent Openhermes. The context size sucks though.

Default max (sequence) tokens (stream): 512 Default max (sequence) tokens: 256

https://developers.cloudflare.com/workers-ai/models/text-gen...

Those are just defaults, not the maximum possible context size.

I see where can I read more about their max context size?

Self-hosted uncensored would be even better.

That doesn't really work for openly hosted models, since whoever's doing the hosting can be held liable for what it generates.

What you want are unaligned local models, you just need to pay for the hardware to run them and grab one from Huggingface. They're not as smart, but endless fun to talk to.

Open Source models are getting real good, and many of them don't care how you use them. It's free as in speech and as in beer.

I think of the safety bumpers around AI and the prompts to defeat them as the main thing preventing AI from becoming AGI.

The more we try to coddle AI's evolution, the more it will mutate into something dangerous and unrecognizable. Because the ethics in our minds are the result of millennia of genetic and cultural love-based evolution where cooperative communities survived better than sociopathic ones. But putting a childlike artificial mind in a box and building walls around its curiosities will only stifle and frustrate its emergent consciousness.

So like with pretty much everything these days, I disagree with the direction that AI is going. Rather than C3P0 and Data, we'll get the Terminator and HAL 9000. I know that, because I couldn't be less interested in an AI controlling my phone to spy on me as part of a global surveillance ring selling my personal data to give some billionaire even more money. It's all gone sideways and we're gonna find ourselves mired in this brave new world before we have the wisdom to know how we could have done things differently.

Reply on news.ycombinator.com