Google's live from Paris event private/deleted immediately • Hacker News Board

Re: Google's live from Paris event private/deleted immediately

#271

Post by uoaei » Wed, Feb 08, 2023, 7:04 PM UTC

rchaud wrote: post #252

sho_hn wrote: post #245

Earlier quoted context omitted.

I'd like to have more discussion about the attribution problem, specifically. We now have a couple of the players saying they're working on, or having demos. But from what I can see, all of these don't _actually_ attribute the source. They're just able to find _a_ source that fits to the output, working backward from it. Often the attributed source fits, but is actually disagreeing with the original output in specifi…

> Is there any indication real attribution is coming? I wouldn't hold my breath. Showing attributions allows people to reverse engineer how the prompt was made, and identifies weaknesses in the model. Bad for business!

What viability are top execs seeing in a business model that is so fragile as to be defeated by reverse-engineering these prompts...? There is obviously no competitive advantage here for very long. The systems must provide an additional service on top of these models, something beyond "attaching it to a publicly-available input form". It needs to aid in reasoning, not act as a god.

Re: Google's live from Paris event private/deleted immediately

#272

Post by sho_hn » Wed, Feb 08, 2023, 7:05 PM UTC

PaulDavisThe1st wrote: post #268

sho_hn wrote: post #257

Earlier quoted context omitted.

I feel like all of the interesting stuff when it comes to search+LLM is in real hybridization that's still a few steps out from what's happening right now in terms of productization. For example: (a) Addressing the recency problem by making a model generate instructions to a browser as output token stream, so it can, well, Google for you and incorporate latest info into output by adding it to the "internal prompt". (…

> This also seems intuitively true: If you ask your friend a difficult question, you often don't expect them to be able to answer it on the spot. You probably expect them to have relevant experience and a rough understanding of the topics and the tools to look up or figure out the rest. Similarly, it feels intuitively wrong to expect a LLM to come ready-made with all knowledge baked in. Not sure. In theory, this bein…

I've seen some studies that call what I mean a "chain of thought" approach (also reasonable is just "higher-level reasoning", but maybe we need a term for the engineered implementation), and I like that terminology. I don't know, it's a gut feeling, but a one-shot prompt completion feels intuitively too simple for me - it seems like we need to build up more elaborate "thought processes" next where we use these methods for individual steps in those thought processes, but also have steps that can carry out a computation or go on a fact search.

A bit similar, again, to how humans will answer plenty of questions on the spot from working knowledge, but are also able to classify when they need to drop into a slower "let's do the math" mode and then carry that out.

We may be able to from these "thought processes" then generate training data sets that again reasonably cache "all possible accurate answers" as you demand, but it's not clear to me if that caching will really save computational cost or just redistribute it to different times. At that point it's probably a latency/throughput question of when you do vs. defer the "thinking".

Re: Google's live from Paris event private/deleted immediately

#273

Post by uoaei » Wed, Feb 08, 2023, 7:05 PM UTC

jaggs wrote: post #197

anon84873628 wrote: post #184

Earlier quoted context omitted.

Maybe the people who were supposed to be running the event got laid off xD

That's so not funny, it's probably true.

Thoughts are allowed to be both funny and sad at the same time.

Re: Google's live from Paris event private/deleted immediately

#274

Post by isoprophlex » Wed, Feb 08, 2023, 7:05 PM UTC

paulpan wrote: post #218

For what it's worth, Google should be taking a page out of Netflix's playbook in pivoting their search business: not by improving/adding LLM features to existing search algorithm and results, but instead fork into a new site and starting from scratch. Netflix had spun out their DVD mail-in business to focus on streaming but had maintained ownership of both. This would allow them to innovate quickly on the new tech wi…

When was the last time Google delivered something that was actually good/lasted longer than an afternoon?

They have had search, maps and gmail since forever. To me, it seems they don't know how to innovate anymore, let alone deliver.

It'll be the story of YouTube Shorts, (stadia, ... you name it) over and over, until they fade into obscurity.

Re: Google's live from Paris event private/deleted immediately

#275

Post by Karawebnetwork » Wed, Feb 08, 2023, 7:05 PM UTC

amelius wrote: post #170

Karawebnetwork wrote: post #71

Earlier quoted context omitted.

Perplexity https://www.perplexity.ai/ does it well in my opinion. It builds a paragraph answering your query but it has a lot of footnotes that link directly to websites. Ex: "geopolitical reason for palm oil being banned and why it's bad for health" The EU has banned palm oil in biofuels due to its negative impacts on health[1] and its geopolitical implications, such as favoring alternative crops grown in Europe[2].…

> Perplexity does it well in my opinion. Nope: https://news.ycombinator.com/item?id=34685862

I only meant about the use of footnotes links. The AI themselves are quite bad but can also be useful if you learn not to take everything at face value. One example of it failing is when I search for my own name it mixes up a bunch of people together so I end up being a dead murderer from the 70's, a house rental person, a developer and also a pianist.

Re: Google's live from Paris event private/deleted immediately

#276

Post by drakmo » Wed, Feb 08, 2023, 7:12 PM UTC

sumeno wrote: post #240

drakmo wrote: post #147

Earlier quoted context omitted.

A case I had right now. Just try to google this: When migrating from Swagger 2 ApiOperation annotation to OpenAPI 3.0 Operation by what attribute is nickname replaced? And then ask ChatGPT ... Google is done

and how did you verify that the ChatGPT answer was correct?

I built my code successfully after refactoring it accordingly :-) It was just convenient.

All the guys pointing out what ChatGPT can't do right now are simply missing the point. It's disruption baby. But if you feel better then: Don't look up

Re: Google's live from Paris event private/deleted immediately

#277

Post by jakeinspace » Wed, Feb 08, 2023, 7:12 PM UTC

uoaei wrote: post #267

PaulDavisThe1st wrote: post #243

You know, whatever the merits (and it sounds as if there were very few) of Google's presentation (or Pinchai's recent non-statement) might be, I would have expected the commentariat here at HN to display way more skepticism about language models as a replacement for search. If there's a crowd anywhere in the world large enough to leave 200+ comments on a post like this, you'd think thatg the one at HN would understan…

Any numerical system operates only on numerical, not logical/grammatical/whatever constraints. Wrongness is built into any system based solely on neural components by the simple fact that rightness is unverifiable from within these systems' own "logical" (read: associational) apparati. I share your confusion regarding the grave dearth of incredulity in the reception of these technologies.

I guess you can say the human brain operates only on the data it gets from our ear drums, optic nerve, etc. Might as well call that numerical input, it’s certainly not grammatical or logical, all that is internal and arises after many layers of neurons.

Re: Google's live from Paris event private/deleted immediately

#278

Post by zzzeek » Wed, Feb 08, 2023, 7:12 PM UTC

PaulDavisThe1st wrote: post #243

You know, whatever the merits (and it sounds as if there were very few) of Google's presentation (or Pinchai's recent non-statement) might be, I would have expected the commentariat here at HN to display way more skepticism about language models as a replacement for search. If there's a crowd anywhere in the world large enough to leave 200+ comments on a post like this, you'd think thatg the one at HN would understan…

simple answer is the consensus groupthink on Hacker News is deeply wishful

i made this same point at https://news.ycombinator.com/item?id=34699087 and got nothing but pushback how there is no real problem with ChatGPT and Bing just fixed all the problems (which didn't exist) anyway.

Re: Google's live from Paris event private/deleted immediately

#279

Post by t-writescode » Wed, Feb 08, 2023, 7:13 PM UTC

sho_hn wrote: post #245

PaulDavisThe1st wrote: post #243

You know, whatever the merits (and it sounds as if there were very few) of Google's presentation (or Pinchai's recent non-statement) might be, I would have expected the commentariat here at HN to display way more skepticism about language models as a replacement for search. If there's a crowd anywhere in the world large enough to leave 200+ comments on a post like this, you'd think thatg the one at HN would understan…

I'd like to have more discussion about the attribution problem, specifically. We now have a couple of the players saying they're working on, or having demos. But from what I can see, all of these don't _actually_ attribute the source. They're just able to find _a_ source that fits to the output, working backward from it. Often the attributed source fits, but is actually disagreeing with the original output in specifi…

> Is there any indication real attribution is coming?

Looking at the main article on HN yesterday around it: https://www.theverge.com/2023/2/7/23587454/microsoft-bing-ed... you can see citations in the examples provided.

Re: Google's live from Paris event private/deleted immediately

#280

Post by t-writescode » Wed, Feb 08, 2023, 7:13 PM UTC

rchaud wrote: post #252

sho_hn wrote: post #245

Earlier quoted context omitted.

I'd like to have more discussion about the attribution problem, specifically. We now have a couple of the players saying they're working on, or having demos. But from what I can see, all of these don't _actually_ attribute the source. They're just able to find _a_ source that fits to the output, working backward from it. Often the attributed source fits, but is actually disagreeing with the original output in specifi…

> Is there any indication real attribution is coming? I wouldn't hold my breath. Showing attributions allows people to reverse engineer how the prompt was made, and identifies weaknesses in the model. Bad for business!

Look at the images in the announcement from yesterday: https://www.theverge.com/2023/2/7/23587454/microsoft-bing-ed...

Those look like citations, to me.