Hacker News Reader

Remember this?

“We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists.”

Geoffrey Hinton, 2016 (while working for Google).

I speculate that they weren't caught off guard. More likely, they tried to find a real-life application for their AI, and failed. Improving the quality of search is not in their interests (in fact, it's the contrary)

> ChatGPT was able to gather so much data without restrictions and now companies late to the party will have many more restrictions

Not only that. OpenAI entered the second stage in AI development: acceleration. They are using their current models to build next generation models. Like they used 4v to build better DALE-3. No competitor is doing it, not at this scale. This and the fact that OpenAI rises the bar makes it very hard to compete even for behemoths like Google. Not even talking about Hugging Face. DALE-3 is incomparably better than Stable Diffusion. But it's artificially limited, like no nudes, no recognizable humans, etc. Making unrestricted competitor will be really hard to do. So, my prediction: OpenAI is far ahead and the gap will increase with the time. Unless there is a breakthrough somewhere else. Then history will repeat itself. Remember, Deep Mind was an indisputable leader not so long ago.

This isn’t history as I remember it. It’s also not history as the product announcement for bard reads.

Bard was explicitly a smaller (cheaper) version of their model. They always said they had a bigger and better one behind the curtain (now called Gemini I guess). I expected bard to be simpler than ChatGPT because they said it’d be small and simple. I guess that still means it’s disappointing 1:1 though.

The notable feature of the model that Bard has over ChatGPT (at least at time of launch) was daily updates keeping it way better for current information and events. Of course, Google has a lot more internal APIs they can/are leverage (eg search, drive, etc) so today bard has a different focus than just talking to a big model.

> Google was certainly completely caught of guard.

They weren't.

Their AI research team is completely dysfunctional and unfocused. They're only interested in doing research and are entirely incapable of the engineering and product focus necessary to build something like ChatGPT.

That isn't the same thing as being caught off guard.

The execs were. They've been paying millions for a top-tier AI devteam hoping for something like ChatGPT, just to find out that their devteam made and freely released the transformers research that ChatGPT is based on (derp)

As always, Google has incredible ideas and promise but lacks business execution

So they were caught of guard, maybe not from a research pov but certainly from a product pov. They are behind and are just playing catch-up for over a year now that it's safe to say that they don't have any competitive advantage over microsoft/openai anymore

It's not caught off guard. Google simply can't do products because of how career growth works. You don't get promoted for polishing things other than ad revenue. It's all about launching things built in a hasty way.

This article was posted just a few weeks before ChatGPT was released:

https://www.inc.com/nick-hobson/googles-rest-vest-days-for-s...

Too little too late?

> Too little too late?

Announcing things and actually executing on said announcements are two very different things.

Sundar has never been able to tell the difference between the two.

Well atleast others are using all the research coming out of Google for all of this. DALLE and chatGPT are both the result of research papers directly from Google research.

> They weren't.

Correct, but not for the reason you gave.

> Their AI research team is completely dysfunctional and unfocused.

The bigger problem was that they saw the major problems with trustworthiness of LLMs and felt there was still too much brand risk to Google. OpenAI had no such problem.