Actually for coding sometimes it’s better than GPT4. A lot of times worse, but my work flow is to give the same coding question to both. I play them off eachother. OpenAI will lose their edge if they don’t release a new model soon
Ask HN: How are you finding Gemini Ultra?
41–50 of 58 posts
Re: Ask HN: How are you finding Gemini Ultra?
#42Google is never worth a subscription. They have completely moved away from "don't be evil"; don't encourage them by giving them more money.
No need to worry they will sunset the product themselves. It aint ads it aint no dollars.
Re: Ask HN: How are you finding Gemini Ultra?
#43I don't live in the US, so I'm not finding it at all.
Just "bought" access to the trial in Germany.
Re: Ask HN: How are you finding Gemini Ultra?
#44I currently, have a free two-month trial, but it's not worth paying for yet. It still fails in hilariously simple ways. E.g, one of the example prompts shows how to use it to generate an image. But asking [Can you generate pictures?] results in "Unfortunately, I am not currently able to generate images directly. However, I can help you explore some resources that can do just that!...". Forcing it to generate an image result in broken elements.
It also fails at more complex tasks that I would actually use AIs for. E.g, I'm trying to learn Lagrangian Mechanics and keep running into dead ends with Gemini in cases where ChatGPT is working.
I have faith that Google will make it better in the near future (unless they get bored and move on to something else) so I'm hoping it will be worth paying for. But right now, I'm going to keep my ChatGPT subscription since that is actually useful.
Re: Ask HN: How are you finding Gemini Ultra?
#45I subscribe to both ChatGPT and Gemini Ultra. Some interesting features/experiences with Gemini... 1. I asked it to do an image search and it responded with in-line images throughout its response. This was a nice "wow" moment and very neat. 2. Its drawing/illustration style is different than DALL-E, so I use both. 3. Quality of general text responses is comparable, though I prefer ChatGPT. I imagine you can probably…
How do you get Gemini Ultra to generate images? It just tells me that it can't do that yet.
Most European countries are excluded:
> Image generation in Gemini Apps is available in most countries, except in the European Economic Area (EEA), Switzerland, and the UK. It’s only available for English prompts.
Re: Ask HN: How are you finding Gemini Ultra?
#46I currently, have a free two-month trial, but it's not worth paying for yet. It still fails in hilariously simple ways. E.g, one of the example prompts shows how to use it to generate an image. But asking [Can you generate pictures?] results in "Unfortunately, I am not currently able to generate images directly. However, I can help you explore some resources that can do just that!...". Forcing it to generate an image…
I'd be careful using LLMs to learn mathematical concepts. I asked ChatGPT for the axioms of set theory, I was still in the "use ChatGPT as a search engine to see what happens" phase. It left out two of them and got one subtly wrong. It's the last bit you really have to watch out for.
Math is exact. It's a bad fit for a machine-that-guesses-words.
Re: Ask HN: How are you finding Gemini Ultra?
#47I currently, have a free two-month trial, but it's not worth paying for yet. It still fails in hilariously simple ways. E.g, one of the example prompts shows how to use it to generate an image. But asking [Can you generate pictures?] results in "Unfortunately, I am not currently able to generate images directly. However, I can help you explore some resources that can do just that!...". Forcing it to generate an image…
I'd be careful using LLMs to learn mathematical concepts. I asked ChatGPT for the axioms of set theory, I was still in the "use ChatGPT as a search engine to see what happens" phase. It left out two of them and got one subtly wrong. It's the last bit you really have to watch out for. Math is exact. It's a bad fit for a machine-that-guesses-words.
The good thing about math is that it's possible to verify the result in many cases even if the method used to produce the result is too difficult to understand. Also, if I ask ChatGPT & Gemini to walk me through the problem step-by-step, I know enough to follow and detect errors. E.g, Gemini kept getting errors because it used the wrong starting potential energy for a pendulum and assumed the pendulum had a max PE of 0 at the top of the swing (and a negative PE at the bottom) which caused it to have a max velocity of 0 due to incorrectly applying the conservation of energy.
But you're right that someone who doesn't understand a subject should be wary of trusting LLMs to teach the subject.
Re: Ask HN: How are you finding Gemini Ultra?
#48For things I would formerly use a web browser to look up, Gemini is often a lot better than gpt-4. The more obscure the query, the better Gemini does compared to gpt-4. But gpt-4 is still best for ML Python coding. Gemini hallucinates non-existent libraries and often adds unnecessary junk to its code. For example, Gemini often defines spurious variables and then never uses them.
> gpt-4 is still best for ML Python coding
How does GPT4 do with Pytorch, Pytorch Lightning, etc? I'm pretty surprised at how poor GTP3.5 is with those sometimes.
Re: Ask HN: How are you finding Gemini Ultra?
#49Earlier quoted context omitted.
Apparently yes, even from the Google mobile app where I did this. https://g.co/bard/share/325bfd6d79cd
broken so I re-did the main prompt myself https://imgur.com/a/GjFCTRD
Yeah so to my eye, this makes sense. These are all women. We didn’t specify a location or time period or art style. There’s a certain kind of person who gets upset every time they see someone that doesn’t look like them where they didn’t expect it. But that’s their problem. Most of the human beings who have ever lived were not white. They might try to say it’s “statistically inaccurate” but they’re probably not pouring over real world statistics when they make such claims, and I expect that Google’s perception of statistics is more accurate than theirs.
Note that the alternative is the Google “grandmother problem” which they definitely don’t want to repeat:
Re: Ask HN: How are you finding Gemini Ultra?
#50One thing I've found Gemini Advanced (Ultra) is actually good at is sustaining a conversation to work towards a goal that requires some back and forth. I've been calling it "putting it in collaboration mode", which isn't super accurate since it's not technically a "mode" but whatever. But in my experience ChatGPT (and a lot of instruction tuned LLMs) tend to try to get to the end of the conversation as quickly as pos…
It's funny you say that, because I just had a similar experience.
Front-end web development with HTML, CSS and JavaScript is not one of my strengths, so I've been trying to go back-and-forth with GPT-4 to get a working prototype.
Before asking GPT-4 to generate any code, I try to have a conversation to explore the pros and cons of various approaches. Is it better to put separate click events on every object, or listen for events at a higher level of the DOM? Is it better to use callbacks for this or async/await?
But GPT-4 rushes to write code before requirements are defined, and then it struggles to modify that code later after I've decided what approach to take. It forgets what it did before, provides partial results, alters variable names between chat turns, contradicts itself.
Just tried the same dialogues with Gemini Advanced and it's night-and-day better. It seems to remember contexts from 10+ conversation turns ago, it provides a more thorough rationale for its suggestions, doesn't prematurely spit out code, and keeps variable names constant throughout the dialogue.
After the approach and requirements are defined, it can then generate large blocks of code with intelligent modularization into separate functions, more similar to how I would write it by hand than what I get from GPT-4.
Gemini Advanced also seems to prefer the more modern ES6 coding style (arrow functions, destructuring, const/let variable declaration). GPT-4 defaults to the 15-year-old style (string concatenation for building strings instead of template literals, heavy use of var for variable declaration, function expressions instead of arrow functions, etc.). GPT-4 routinely mixes styles even when I explicitly request ES6.