DeepMind CEO Demis Hassabis Urges Caution on AI | Hacker News Reader

DeepMind CEO Demis Hassabis Urges Caution on AI

time.com

79 pointsbookofjoe47 comments

Amazing when a rival / competitor start exceeding them how now is the time for caution...

Exactly what I came to say.

Elon musk cautioned before leaving them too.

Let’s stop this meme before Elon starts saying he invented OpenAI. He donated money into the OpenAI charity. Jeff Bezos invested $250,000 in Google back in 1998 but he doesn’t try to claim to be a Google founder.

> exceeding

OpenAI may have an advantage in terms of publicity and open access to their models, but technologically, Google has better models in-house. Many of the models used by OpenAI are based on work at Google and their papers, and Google has been able to achieve better performance months before OpenAI. For more information on Google's progress in 2022, refer to this summary on the Google Research blog, which is accessible to non-technical readers: https://ai.googleblog.com/2023/01/google-research-2022-beyon..."

Why are Microsoft investing so much in Open AI when they are a similar research powerhouse? I think Open AI must have something.

Similarly, if Google (and Googlers) have had access to this or better for some time, where is the step change in productivity / innovation / technology?

I think it's organisational incentives, not tech per se.

TL:DR; Brand risk, and the FAANGs are not as hungry.

Google and MSFT are wary of deploying tech that gets caught saying super racist or nazi stuff (MSFT have been badly burned before) and they already have big income streams, they didn't feel strong pressure to "ship".

OpenAI has less to lose, more to gain and people are more tolerant of janky cutting edge products dressed in "startup" branding.

It's in the article I linked. They deployed models similar to Codex, like LLM, and saw a small boost in productivity. They have also deployed many other models internally.

I think LLM are not very useful yet. I think they will be, but right now they are wrong far too often.

> I think Open AI must have something

They have a projection of $200M revenue in 2023 and $1bn in 2024.

https://www.reuters.com/business/chatgpt-owner-openai-projec...

> .. and open access to their models ..

Huh? The only thing open about openai is the name. Everything else is proprietary and tightly locked down.

Open access doesn't mean open source. You have access to models through their APIs, you don't have any access to similar Google models.

*limited access

I doubt they are truly exceeding them. At least not by a margin that could not be closed. If you actually read the article, it mentions that Deepmind has their own chatbot, but wants to have citation in place before releasing it. In my opinion, that is the safe thing to do.

Also my opinion, but I would argue that AlphaFold is much more impressive and significant for humanity than chatGPT. I also think that openAI could not reproduce AlphaFold with the same ease that Deepmind could reproduce chatGPT.

Well Internally it was reported a couple weeks after ChatGPT's announcement and release that google has a "Code Red" meeting about it, and at that time it was reported that ChatGPT was significantly better than what Google had.

Now it is known how accurate that was, or if even internally to Google they communicate well enough to know that all teams are doing or capable of.

>>Also my opinion, but I would argue that AlphaFold is much more impressive and significant for humanity than chatGPT.

I think that is short sighted and biased. That is in some ways comparing Vaccine technology to social media, and I would be hard pressed to say what was more "significant for humanity"

It was less that ChatGPT is better than LaMDA and more that the tech everyone was keeping away from society "for its own good" had been released and people were realizing it could finally threaten Search.

And this is probably not true, since "Code Red" has a very specific meaning internally that doesn't apply here. The various codes cover various different timelines. "Strategic risk to the company over a period of months or years" isn't described as "Code Red."

> And this is probably not true, since "Code Red" has a very specific meaning internally that doesn't apply here

Don't be intentionally vague. What is a 'code red'?

I'm being intentionally vague because the terms are internal.

Code Red involves an extremely short timeframe to address the issue.

Demis has had this tone for many years.

AI is dangerous unless I control it (or, can pass that control to big tech for a couple billion).

> When asked to confirm or deny whether the independent ethics board rumored to have been set up as part of the Google acquisition actually exists, he says he can’t, because it’s “all confidential.”

Why would the existence of an independent ethics board be confidential? How could it be independent if it’s secret?

I can see why you don't want journalists doxxing members of the ethics board... They're there for their expertise, not to be figureheads or public speakers.

> "He's self interested"

OK, but is he right?

If he is right, then he should not have been doing what he has been doing.

https://archive.ph/darim

http://web.archive.org/web/20230121041850/https://time.com/6...

He wasn't urging caution when he was ahead.

Just ignore.

Objectively false. He has been advocating caution all along, and negotiated ethics oversight into the acquisition by Alphabet.

I'm not doubting you, just asking: what are your sources? Any articles/videos/podcasts?

Plenty of detail in https://en.m.wikipedia.org/wiki/DeepMind.

His Lex Fridman interview was very interesting too: https://m.youtube.com/watch?v=Gfr50f6ZBvo.

TFA is the source

It’s not when the context of “all along” is in question. Someone is accusing this person of changing their tone only because it’s convenient and TFA is that supposed change in tone.

I did not see his company contribute to any positive side of the society so far.

What actually AlphaGo accomplish other than killing the Go community?

AlphaGo in no way, shape, or form killed the go community. Stockfish hasn't killed the chess community, either.

The fact is that much less kids or people playing Go now, and competitions are not getting nearly as much attention as before. Putting two strong statements without any evidence does not mean anything.

In fairness, you just made two strong statements without evidence.

The debate over the causes of (perceived?) decline of Go in the east predates AlphaGo and most theories involve straightforward geopolitics more than any AI conspiracy theory.

In the west it still seems to only be driving more interest in the game.

> The fact

Is perhaps not so factual...

Would this count? https://alphafold.ebi.ac.uk/

technically no, because they are referring to AlphaGo, not DM.

DM's results in protein structure prediction are world-class breakthrough work (well, it was last year; now it's routine, having been duplicated by many others). At least in my mind (I'm an ex-biophysicist with a long-term interest in seeing ML applied to biology problems like structure prediction) that is a truly long-term result that demonstrated that, at least for some short period of time, DM was way ahead of everybody else in the field. They also published the details and made large-scale predictions.

Solving protein folding?

But apart from protein folding, alpha go, roads, sanitation and the aqueduct, what have deepmind ever done for us?

AlphaGo was the tipping point, demonstrating that machines can be effective in a high-level/'creative' field with a vast search space, much larger than chess using training rather than hand-written algorithms (that don't translate to other fields). No one expected Go to be solved to the degree and manner of AlphaGo for generations. It was the start of the great acceleration in ML since Deep RNN re-ignited the field.

Is it true that AlphaGo killed the Go community ? I mean I'm sure there are still Go players, but has AlphaGo had a significant impact on the Go community?

I can't speak to its effects in Asia where I can imagine there's been some more direct economic impact, but as a long-time casual player in Europe, interest from my social groups has mostly (but slightly) risen post-AlphaGo.

how.

AI has been better than humans at chess for years and chess is still going strong

As an interesting side note, though, software (AI?) has definitely introduced a new dynamic to professional chess, and has very recently caused a lot of controversy (when used for cheating).

Reply on news.ycombinator.com