Comment by machdiamonds | Hacker News Reader

Comment by machdiamonds | Hacker News Reader

I don't understand how Google messed up this bad, they had all the resources and all the talent to make GPT-4. Initially, when the first Bard version was unveiled, I assumed that they were just using a heavily scaled-down model due to insufficient computational power to handle an influx of requests. However, even after the announcement of Palm 2, Google's purported GPT-4 competitor, during Google IO , the result is underwhelming, even falling short of GPT 3.5. If the forthcoming Gemini model, currently training, continues to lag behind GPT-4, it will be a clear sign that Google has seriously dropped the ball on AI. Sam Altman's remark on the Lex Fridman podcast may shed some light on this - he mentioned that GPT-4 was the result of approximately 200 small changes. It suggests that the challenge for Google isn't merely a matter of scaling up or discovering a handful of techniques; it's a far more complex endeavor. Google backed Anthropic's Claude+ is much better than Bard, if Gemini doesn't work out, maybe they should just try and make a robust partnership with them similar to Microsoft and OpenAI.

Have you ever considered the problem tech like this actually creates for their owners? This is why they didn't release it.

From a legal, PR, safety, resource, monetization perspective, they're quire treacherous products.

OpenAI released it because they needed to make money. Google were wise enough not to release the product, but as others have said, it's an arms race now and we'll be the guinea pigs.

This line of reasoning implies that Google had models that were equivalent to OpenAI's but chose to keep them behind closed doors. However, upon releasing Bard, it was apparent—and continues to be—that it does not match up to OpenAI's offerings. This indicates that the discrepancy is more likely due to the actual capabilities of Google's models, rather than concerns such as legal, PR, safety, resource allocation, or monetization.

As we all know, we don't know what C-4 is trained on. It might be trained on information they didn't have the rights to use (for example). This is why they might be so tight lipped on how it was produced.

Google on the other hand, has much much more to loose here, much bigger reputation to protect, and may have built an inferior product that's actually produced in a more legally compliant way.

Another example would be Midjourney vs Adobe Firefly, there is no way Firefly makes art as nice as MJ produces. Technically it's good stuff, but it's not as fun to use because I can't generate Pikachu photos with Firefly.

People have stated that ChatGPT-4 isn't as good anymore. My personally belief is this is just the shine wearing off what was a novelty. However it may also be OpenAI removing the stuff they shouldn't have used in the first place. Although there are reports the model hasn't changed for some time so who knows.

I guess in time we'll find out. Personally I don't really care for either product so much, most of my interactions have been fairly pointless.

I think it's just fun to watch these big tech companies try deal with these products they've created. It's amusing as fuck.

If Google only used data that isn't copyrighted, they'd probably make a big deal about it, just like Adobe does with their Firefly model. Also, it's not really possible for OpenAI to just take out certain parts from the model without retraining the whole thing. The drop in quality might be due to attempts to make the model work faster through quantization and additional fine-tuning with RLHF to curb unwanted behavior.

So basically, you're of the belief whatever OpenAI has done it's some kind of magic which Google cannot / has not figured out?

I re-read, didn't mean to sound snarky, although it did, just curious if that's what you really believe is going down??

I think Google still has a decent chance of catching up. It's just a bit surprising to see them fall behind in an area they were supposed to be leading, especially since they wrote the paper which started all of this. Also, Anthropic is already kind of close to OpenAI, so I don't think OpenAI has some magic that no one else can figure out. In the future, I predict that these LLMs will become a commodity, and most of the available models will work for most tasks, so people will just choose the cheapest ones.

They have explicitly said in interviews that it was intentional not to release epowerful ai models without being sure of the safety. OpenAI put them in the race and let's see how humanity will be affected.

If safety were the only consideration, it's reasonable to expect that they could have released a model comparable to GPT 3.5 within this time frame. This strongly suggests that there may be other factors at play.

Reply on news.ycombinator.com