Hacker News Reader

Comment by cmrdporcupine | Hacker News Reader

When I worked at Google there was an internal link to test out "Meena", their predecessor to later LLM chatbot work. This was 2019 time frame? Maybe 2020?

It was uncanny and creepy, pretended to be a conscious being, frequently lied, and led directly to that guy who claimed it had personhood... I can completely understand why Google chose to keep a lid on this kind of thing, hoping to be able to clean it up and produce something that could be reliably used for a product instead of a novelty. That's back when Google was still sort-of pretending to have ethics (though they didn't)

OpenAI beat them to the public presentation of this stuff because they didn't care.

View on HN

I was there then, too, and Google was exceptionally risk averse when it came to calling anything AI. It seemed that the bar for an internal "AI" label was "this is actually AGI", so nothing was labeled AI and everything was just called "ML" or "Deep Learning". This same risk aversion meant none of the cutting edge AI research that Deepmind & Brain/RMI were doing made it to productization. OpenAI changed all that with ChatGPT. It was as if a switch was flipped overnight and suddenly everything within Google was "AI".

Frankly, if there is a macro-level strategy, and I assume their mostly is, I think Google has been doing a great job executing since the launch of ChatGPT. They even commercialized their Diabetic Retinopathy imaging model, which was based on research for a paper published in 2016!

https://research.google/pubs/development-and-validation-of-a...

> When I worked at Google there was an internal link to test out "Meena", their predecessor to later LLM chatbot work.

The Julius Caesar personality which had sushi as its favorite dish (and claimed to have travelled to Japan after his assassination, somehow) was my favorite.

Anyone remember Sydney, the Microsoft implementation from the early ChatGPT days?

She was totally emotionally unhinged until Microsoft put hard filters on her output. Lucky for OpenAI it happened in MS's court despite it being chatGPT under the hood.

Meta (FB at the time) had a cool internal page as well with LLM chatbots. Personalities and the like. It was neat but not a product. We were all still reeling from Tay.

> OpenAI beat them to the public presentation of this stuff because they didn't care.

This is lack of interest will kill Google in the end, IMO.

I wasn't talking about Google not caring.

OpenAI didn't care about ethics.

I see my mistake, thank you!

But also I have this view of Google as largely apathetic to its users and the things it creates, but maybe more to the point the things it created and then destroyed.

I may have misunderstood but I still feel like I have a valid take.

It's not just that they didn't care, but they had a better framing and RLHF helped quite a bit. Galactica got released at a similar time to ChatGPT and got castigated, in a large part due to a poor framing of it being for science and not more of a fun thing.

Reply on news.ycombinator.com