Gemini "duck" demo was not done in realtime or with voice • Hacker News Board

Re: Gemini "duck" demo was not done in realtime or with voice

#231

Post by WhitneyLand » Thu, Dec 07, 2023, 9:12 PM UTC

modeless wrote: post #94

That's not the only thing wrong. Gemini makes a false statement in the video, serving as a great demonstration of how these models still outright lie so frequently, so casually, and so convincingly that you won't notice, even if you have a whole team of researchers and video editors reviewing the output. It's the single biggest problem with LLMs and Gemini isn't solving it. You simply can't rely on them when correctn…

Agree, then the question becomes how will this issue play out?

Maybe AI correctness will be similar to automobile safety. It didn’t take long for both to be recognized as fundamental issues with new transformative technologies.

In both cases there seems to be no silver bullet. Mitigations and precautions will continue to evolve, with varying degrees of effectiveness. Public opinion and legislation will play some role.

Tragically accidents will happen and there will be a cost to pay, which so far has been much higher and more grave for transportation.

Re: Gemini "duck" demo was not done in realtime or with voice

#232

Post by vjerancrnjak » Thu, Dec 07, 2023, 9:12 PM UTC

There is a possibility of dataset contamination on the competitive programming benchmark. A nice discussion on the page where AlphaCode2 was solving the problems https://codeforces.com/blog/entry/123035

Problem showed in the video was reused in a recent competition (so could have been available in the dataset).

Re: Gemini "duck" demo was not done in realtime or with voice

#233

Post by crazygringo » Thu, Dec 07, 2023, 9:13 PM UTC

modeless wrote: post #94

That's not the only thing wrong. Gemini makes a false statement in the video, serving as a great demonstration of how these models still outright lie so frequently, so casually, and so convincingly that you won't notice, even if you have a whole team of researchers and video editors reviewing the output. It's the single biggest problem with LLMs and Gemini isn't solving it. You simply can't rely on them when correctn…

EDIT: never mind, I missed the exact wording about being "made of a material..." which is definitely false then. Thanks for the correction below.

Preserving the original comment so the replies make sense:

---

I think it's a stretch to say that's false.

In a conversational human context, saying it's made of rubber implies it's a rubber shell with air inside.

It floats because it's rubber [with air] as opposed to being a ceramic figurine or painted metal.

I can imagine most non-physicist humans saying it floats because it's rubber.

By analogy, we talk about houses being "made of wood" when everybody knows they're made of plenty of other materials too. But the context is instead of brick or stone or concrete. It's not false to say a house is made of wood.

Re: Gemini "duck" demo was not done in realtime or with voice

#234

Post by freedomben » Thu, Dec 07, 2023, 9:13 PM UTC

modeless wrote: post #186

freedomben wrote: post #177

Earlier quoted context omitted.

I think this problem needs to be solved at a higher level, and in fact Bard is doing exactly that. The model itself generates its output, and then higher-level systems can fact check it. I've heard promising things about feeding back answers to the model itself to check for consistency and stuff, but that should be a higher level function (and seems important to avoid infinite recursion or massive complexity stemming…

I'm not a fan of current approaches here. "Chain of thought" or other approaches where the model does all its thinking using a literal internal monologue in text seem like a dead end. Humans do most of their thinking non-verbally and we need to figure out how to get these models to think non-verbally too. Unfortunately it seems that Gemini represents no progress in this direction.

> Humans do most of their thinking non-verbally and we need to figure out how to get these models to think non-verbally too.

That's a very interesting point, both technically and philosophically.

Where Gemini is "multi-modal" from training, how close do you think that gets? Do we know enough about neurology to identical a native language in which we think? (not rhetorical questions, I'm really wondering)

Re: Gemini "duck" demo was not done in realtime or with voice

#235

Post by tomjakubowski » Thu, Dec 07, 2023, 9:13 PM UTC

bongodongobob wrote: post #208

jmathai wrote: post #138

Earlier quoted context omitted.

> it rarely happens that it just finds stuff that's flat out wrong "Flat out wrong" implies determinism. For answers which are deterministic such as "syntax checking" and "correctness of code" - this already happens. ChatGPT, for example, will write and execute code. If the code has an error or returns the wrong result it will try a different approach. This is in production today (I use the paid version).

Dollars to doughnuts says they are using GPT3.5.

I'm currently working with some relatively obscure but open source stuff (JupyterLite and Pyodide) and ChatGPT 4 confidently hallucinates APIs and config options when I ask it for help.

With more mainstream libraries it's pretty good though

Re: Gemini "duck" demo was not done in realtime or with voice

#236

Post by titzer » Thu, Dec 07, 2023, 9:14 PM UTC

kweingar wrote: post #4

The video itself and the video description give a disclaimer to this effect. Agreed that some will walk away with an incorrect view of how Gemini functions, though. Hopefully realtime interaction will be part of an app soon. Doesn’t seem like there would be too many technical hurdles there.

Yeah, and ads on Google search have the teeniest, tiniest little "ad" chip on them, a long progression of making ads more in-your-face and less well-distinguished.

In my estimation, given the context around AI-generated content and general fakery, this video was deceptive. The only impressive thing about the video (to me) was how snappy and fluid it seemed to be, presumably processing video in real time. None of that was real. It's borderline fraudulent.

Re: Gemini "duck" demo was not done in realtime or with voice

#237

Post by janalsncm » Thu, Dec 07, 2023, 9:15 PM UTC

modeless wrote: post #186

freedomben wrote: post #177

Earlier quoted context omitted.

I think this problem needs to be solved at a higher level, and in fact Bard is doing exactly that. The model itself generates its output, and then higher-level systems can fact check it. I've heard promising things about feeding back answers to the model itself to check for consistency and stuff, but that should be a higher level function (and seems important to avoid infinite recursion or massive complexity stemming…

I'm not a fan of current approaches here. "Chain of thought" or other approaches where the model does all its thinking using a literal internal monologue in text seem like a dead end. Humans do most of their thinking non-verbally and we need to figure out how to get these models to think non-verbally too. Unfortunately it seems that Gemini represents no progress in this direction.

The point of “verbalizing” the chain of thought isn’t that it’s the most effective method. And frankly I don’t think it matters that humans think non verbally. The goal isn’t to create a human in a box. Verbalizing the chain of thought allows us to audit the thought process, and also create further labels for training.

Re: Gemini "duck" demo was not done in realtime or with voice

#238

Post by stocknoob » Thu, Dec 07, 2023, 9:15 PM UTC

windowshopping wrote: post #130

brookst wrote: post #113

Earlier quoted context omitted.

Is it possible for humans to be wrong about something, without lying?

I don't agree with the argument that "if a human can fail in this way, we should overlook this failing in our tooling as well." Because of course that's what LLMs are, tools, like any other piece of software. If a tool is broken, you seek to fix it. You don't just say "ah yeah it's a broken tool, but it's better than nothing!" All these LLM releases are amazing pieces of technology and the progress lately is incredib…

“Broken” is word used by pedants. A broken tool doesn’t work. This works, most of the time.

Is a drug “broken” because it only cures a disease 80% of the time?

The framing most critics seem to have is “it must be perfect”.

It’s ok though, their negativity just means they’ll miss out on using a transformative technology. No skin off the rest of us.

Re: Gemini "duck" demo was not done in realtime or with voice

#239

Post by jmathai » Thu, Dec 07, 2023, 9:16 PM UTC

modeless wrote: post #180

jmathai wrote: post #166

Earlier quoted context omitted.

That example seems a bit hyperbolic. Do you think lawyers who leverage ChatGPT will take the made up cases and present them to a judge without doing some additional research? What I'm saying is that the tolerance for mistakes is strongly correlated to the value ChatGPT creates. I think both will need to be improved but there's probably more opportunity in creating higher value. I don't have a horse in the race.

> Do you think lawyers who leverage ChatGPT will take the made up cases and present them to a judge without doing some additional research? I generally agree with you, but it's funny that you use this as an example when it already happened. https://arstechnica.com/tech-policy/2023/06/lawyers-have-rea...

facepalm

Re: Gemini "duck" demo was not done in realtime or with voice

#240

Post by DominikPeters » Thu, Dec 07, 2023, 9:16 PM UTC

golly_ned wrote: post #165

If you've seen the video, it's very apparent it's a product video, not a tech demo. They cut out the latencies to make a compelling product video. I wasn't at all under the impression they were showcasing TTS or low latencies as product features. I don't find the marketing misleading at all, and find these criticisms don't hit the mark. https://www.youtube.com/watch?v=UIZAiXYceBI

It's not just cutting. The answers were obtained by taking still photos and inputting them into the model together with detailed text instructions explaining the context and the task to the model, giving some examples first and using careful chain-of-thought style prompting. (see e.g. https://developers.googleblog.com/2023/12/how-its-made-gemin...) My guess is that the video was fully produced after the Gemini outputs were generated by a different team, instead of while or before.