Can Demis Hassabis save Google? | Hacker News Reader

Can Demis Hassabis save Google?

bigtechnology.com

133 pointslaurex127 comments

Given his experience with AlphaGo/Star/Zero/etc, I'm sure if there's any way to apply tree search, or policy networks, to LLMs, Demis will find it.

I've always found it strange that a LLM is basically an action/value model that scores all next possible actions (tokens), and yet we DON'T traverse it like a tree. Well, "beam search" exists, but that's kind of the most naive approach.

Then again, I'm not smart, and lots of very smart people are thinking nonstop about LLMs, and no type of tree search has become widely used, so maybe there's really nothing there.

I think part of the problem for LLMs is that they don't operate in an easily scoreable "game". We can make them work well for optimizing next token(s) log likelihood, but how do we judge quality of the output for the task it was made for?

Then after that, there are other challenges, such as do LLMs have a world model that lets them think how to attain a reward?

Wherever you can have an automated feedback mechanism, you can start to address the first problem and allow for the AI to explore more of the "tree". These are situations like coding (you can run the code and evaluate the output), or situations where you can let the LLM crowd-source user feedback.

For models that have some actual world model, who can reason across modalities, and who can plan, LeCunn talks about this often. The videos/slides here[1] were good content.

[1]: https://www.ece.uw.edu/news-events/lytle-lecture-series/

Yes, whatever is not written in any books, AI will have to learn directly from the feedback generated by the environment to its actions.

AlphaZero discovered go from scratch, and beat us who invented the game and had thousands of years to practice it. This is how powerful a teacher can be the environment.

> I think part of the problem for LLMs is that they don't operate in an easily scoreable "game".

Unfortunately "social media" is the gamified environment for language.

But it has human voters, you can't train a model using human voters to vote during iterations, and AI aren't good enough to replace human voters.

Not sure if AI voters even can result in a model smarter than the voting model.

> you can't train a model using human voters to vote during iterations

Why not? Just comment multiple times on a post in different ways. Score outputs with more of a desired response higher than outputs with less desired responses. Scale that up site-wide on multiple sites, and it seems like you have a pretty powerful way to get human feedback...

These models trains on billions of examples, having bots posting billions of posts on different social media sites every time you train a new model probably isn't a viable strategy. You would get banned real quick since most of those will be really low quality in the early stages.

So I agree with your overall sentiment, BUT we already have good LLMs, so I think we’ve crossed the bridge from “bad low quality responses in the beginning” and I also suspect there’s more Social Media bots than we’d like to admit.

I think we're simply talking about different things. Obviously you wouldn't want to start with a model of pure noise when interacting with real humans like that. I am describing using RLHF to fine tune existing models. That is a way to gamify training.

I personally feel like these problems can be broken into two types - one where the output is expected/deterministic, and one where creativity is a virtue. Asking Siri what the current temperature is (only 1 correct answer) is an example of the former, but asking chatGPT to write an email is closer to the latter. Tree type algorithms are better when you don't have (millions of english words)^10 for a 10 word response, and where there are multiple "correct" answers.

I mean you can make any set of LLMs produce any set of output you want. The problem is that it isn't terribly efficient and you have to filter between the real stuff and the hallucinations.

Qualitatively, the answer seems to be to train for novelty. Which happens to also be how humans accomplish much of the complex stuff.

Eg, One interpretation of art is we have 8 billion or so humans with a finely trained neural net for recognising stuff. The artist is looking for novel ways of triggering those nets, and if they find something inspired then it is art. A great artist generally isn't trying to reproduce something that is known, they're trying to explore novel areas of a medium.

So the real trick to building the world model is coming up with a good novelty metric. Still hard, but easier than developing a reward function. That gets the part of the training done that establishes a world model, then I'd assume it is possible to train that model to do a task by rewarding specific outcomes that it already knows how to achieve.

If it's a game, I think the old phrase "everything is made up and the points don't matter" applies well here.

It's really hard to score a phrase for truthfulness. Even among humans often it's ambiguous; we have courts of law in many societies to try and fix some of that.

I admit I am disappointed that we don't see a similar amount of work these days in what was once dubbed 'expert systems', though perhaps the results just aren't as flashy.

Tree decoding exists and is being worked on by multiple groups. At the end of the day tree based methods are still just injecting some kind of interpretable prior.

It's been pretty interesting to see some research trying to solve this. For example, Stanford researchers recently published QuietSTaR[1]. This makes LLMs "think" before they speak. By generating a chain of thoughts and just choosing the best possible path from the generated thoughts.

This method is better than Chain of Thought and I think is a step in the right direction.

[1] https://arxiv.org/abs/2403.09629

I’ve wondered about the same point for a long time. In smaller language models of chemistry the tree search in decoding has made a huge difference [1]. If the training data set doesn’t include the right conditioning it may be hard for the LLM to learn to associate high probability answers with more useful answers, however, this seems like a solvable problem for many subdomains.

[1] search for branch-and-bound in: https://pubs.acs.org/doi/10.1021/acs.jcim.0c00321

Beam search can be optimized incredibly aggressively on hardware.

Check out finite state transducers.

How would that work?

> “I haven't heard that myself,” Hassabis says after I bring up the CEO talk. He instantly points to how busy he is with research, how much invention is just ahead, and how much he wants to be part of it. Perhaps, given the stakes, that’s right where Google needs him. “I can do management,” he says, ”but it's not my passion. Put it that way. I always try to optimize for the research and the science.”

Sundar makes ~200M as the CEO of Google whereas DeepMind sold for ~400-650M. There's plenty of monetary incentive to take the job, not to mention more power to set the company direction through resource allocation. And it's clear that there's been a PR campaign being set up to push out Sundar. Maybe Hassabis is a contender because he's been getting some pretty serious press since the beginning of the year (which is when the Sundar article grumblings started).

> there's been a PR campaign being set up to push out Sundar.

There has been mumblings of that for the whole time he has been CEO...

Google has been failing continuously in user experience under his leadership:

1. Search has become an ad-riddled, clickbait SEO infested mess

2. All those naming kerfuffles (how's Allo and GSuite doing?)

3. Cloud still significantly behind AWS and Azure

4. AI stuff behind at least OpenAI and Anthropic. Applied ML stuff well behind specialized startups.

5. Pixel still not selling very well. Android still feeling like it's trying to catch up to Apple, still fragmented.

6. Layoffs everyone believed wouldn't happen until the very last second they were announced.

7. Customer support story, including toward large cloud customers, still abysmal.

Etc etc.

Yes, ad revenue increased and with it the share price, but ad revenue comes from having had a good product 5 years ago. To sustain it you need good products and good engineers to build them.

Google is trending downward. It needs a fresh CEO.

Who would waste money on a PR campaign? Googles fumblings of the last ~5 years speak for themselves.

You don’t need to pay money on things like this. You leverage your contacts to create profiles like this article. The newspaper gets access and you get your profile elevated. And it wouldn’t be google the company doing this but players within the company / influential and powerful shareholders etc.

You don't need a PR campaign to oust Sundar. I think 7 out of 10 Googlers would agree that Sundar is uninspiring and has allowed the erosion of most of what was good about Google's culture.

I left partially because I hated the feeling that "this year will be average - meaning worse than last year but better than next year". Slow downward drift, culture erosion, lack of leadership, lack of clarity, etc.

Management isn't leadership, and Sundar is more of a manager than a leader.

Very well put

My impression from the outside is indeed that the man is a problem. And part of the problem is that the problem is allowed to continue to exist by the Alphabet board. Which means the real problem is over there. They are just looking at the stock price, which is of course fine. Until it isn't. I think the whole point of Sundar always was that he's very obviously not a leader. He's a care taker.

I used to work in a big multi national (Nokia) with a care taker CEO (Olli Pekka Kallasvuo) who took over from the man that grew Nokia from a large but insignificant Finnish company to the smart phone behemoth it was around 2005 (Jorma Ollila). Kallasvuo presided over the emergence of the Apple's iphone and Google's Android as the two new competitors that ultimately killed it off. And did nothing whatsoever about it. The man was a bean counter whose job it was to protect the stock price.

By the time the Nokia board (under leadership of the former CEO, Ollila) appointed a new CEO (Stephen Elop, an MS executive) with the clear intention to orchestrate some collaboration with and the eventual takeover by MS, it was already too late. Nokia flailed for a few years and MS eventually pulled the plug a year after acquiring what remained of the phone business unit for next to nothing. By then the stock price had tanked, market share was in the gutter, and the value was gone. Everything the board thought they knew about smart phones was no longer relevant. Ollila got removed from the board in the aftermath.

People blame CEOs, but it's the boards of these companies that appoint these CEOs, protect them, and decline to fire them when they fail. That's where the problems are. Nothing changes until you fix the boards. It's always fixable with the right people and leadership. Just look at MS post Ballmer. In Alphabet's case, the two founders are on the board and they are the ones that put Sundar in their place. Maybe it's time for them to move on? Of course the issue is that they have a lot of shares (class B) in Alphabet. Institutional investors have about 35% of the class A shares and the rest is publicly traded. So, nothing happens until the stock nosedives. By which time it might be too late.

of course, but boards are largely impotent. they are too far from the action, don't want to rock the boat even more when things are not going great, and of course management easily bamboozles them with some fancy plan ...

but in the end there's nothing they can do, except fire the management. but it requires risk taking, and it's a collective action problem ...

so if there's a majority shareholder, maybe.. but usually they are the CEO anyway (or appointed it)

Yeah if I wasn't planning on a big career shift anyway in the next 2-3 years, I'd be out the door at Google. Pay is great, I'm able to be fully remote, and it isn't too stressful to do my day job - but the company feels nothing like it did 10 years ago.

You know how there are top down and bottom up companies? My experience at Google now is that it is neither. VPs expect bottom up work and then smash it down whenever they don't like it - but they also cannot articulate what they actually want. I'm constantly being asked to go through prioritization processes that take a ton of time and end with "eh, every project stays at the funding level it is already at."

> You don't need a PR campaign to oust Sundar. I think 7 out of 10 Googlers would agree that Sundar is uninspiring

Do Google RSUs come with voting rights? If not, then Googlers don't get much of a say on when/how Sundar is ousted - Larry and Sergei do.

It doesn’t matter if they do. Larry and Sergei have 51.2% control through their class B shares which basically only they can hold and give them 10x votes of class A shares (i.e. RSUs).

They still have to be mindful of perception of their employees with how they wield that control, but yes basically they can technically tell the board or any shareholders to go pound sand.

That makes it easier in some sense, you just have to convince Larry and Sergey.

Agree but I think it is more like 8/10 or 9/10 of Googlers who would view Sundar as, to say the least, uninspiring.

The problem is that the good internal candidates to succeed Sundar have mostly already left Google, or don’t seem interested in taking over.

If Sundar left today, it could be that Ruth or TK would end up running Google for a while. That would be much, much worse.

My own perception of OpenAI vs Google seems at odds with many things I read. I don't really see Google as being behind at all.

1. Google pretty much invented the technology (https://arxiv.org/abs/1706.03762)

2. In order to create the models one needs lots of compute and access to a lot of text. Google scores higher than OpenAI on both counts.

3. New models are released on a weekly basis by all sorts of companies. So OpenAI has no monopoly on LLM models. In fact their competition is staggering (NVIDIA, Meta, Google, DataBricks, Amazon (numerous other startups)) It will not be long before there are even more.

It seems to me that Altman saw this all as a timing thing. Reveal your cards now and force others to do the same in the hopes of obtaining a strategic position over competitors. Googles cashflow seems to be doing just fine and I haven't had to fight off any urges to use Bing.

> 1. Google pretty much invented the technology (https://arxiv.org/abs/1706.03762)

> 2. In order to create the models one needs lots of compute and access to a lot of text. Google scores higher than OpenAI on both counts.

If a smaller company with less compute and data commercializes something before Google which has more compute and data, doesn't that mean they are behind? You don't measure a car race by fuel available in the pit and faster top speed. You measure by who gets to the finish first. It just goes to show that Google has a horrible driver.

I guess the point is this is a very long race and, actually, the end is still quite far away (if that end is AGI).

In his Lex Friedman interview Altman asks ‘what has ChatGPT really fundamentally changed about the world?’ - basically making the point that they’re still just getting started.

Having a tonne of compute and cash is still going to be really important in this race. You have to make it to the finish line to win.

Part of the point I was trying to make is that there is less benefit to Google to tout their advances - most of this are trade secrets. I could see them using it to enhance their current offerings in subtle ways.

Now that the cat is out of the bag that will likely change. Just because OpenAI publicly released a produce doesn't mean Google has not developed their own. It doesn't mean they have either.

These models still have a long way to go before they can advise Kirk on running the Enterprise:)

Well said.

While everyone is hyper-focused on LLMs, Google is able to do more than that and imagine what Google DeepMind has not announced yet.

> It seems to me that Altman saw this all as a timing thing. Reveal your cards now and force others to do the same in the hopes of obtaining a strategic position over competitors. Googles cashflow seems to be doing just fine and I haven't had to fight off any urges to use Bing.

LLMs are something that is already played out to the first movers. Google has already caught up and the moat and monopoly has been evaporated.

You’re missing the final point - the use case beyond Chat. Google has search, GMail, Android, Docs, Developer Tools, etc

Google can put an LLM into everything and sell it. Not just a chatbot. OpenAI can sell theirs to consumers as a chat bot or an API. Google can out monetize them handily.

The best thing Google can do (which it’s starting to do) is open up access to LLMs. And help everyone else do the same. Give it away and flood the internet with more LLaMAs, more Groks, more CLIPs, more Hermes, more Mistal, more SIGLIPS, etc. If they just drown out the competition, and turn good enough” models into a true commodity, they’ll dethrone OpenAI easily.

Also no one mentions YouTube. Surely that data is a massive untapped opportunity. We saw the multi-modal abilities of Gemini today. A few years from now, and some better GPUs, and it might be able to handle video in real-time.

No. The best model wins. I’m currently paying $20/mo for all three leading models (GPT4, Opus, and Ultra). But in the future, when GPT5/Opus2/Ultra1.5 come out, I expect the prices to go up. So I will be choosing the best one. Whatever is on the top of the leaderboard will get my $200/mo (maybe even $2k/mo if it’s really smart).

This is forward looking. Models need to keep getting better. That takes research, and compute, and data. It’s not free, and it’s getting increasingly expensive.

I too pay for multiple services (Gemini, Claude, CodePilot). Not everyone can or will pay $2k, or $200, or even $20 for a massive model. And most flagship models today are significantly better than models 6mo ago, which were already transformative and valuable on their own. There is a huge opportunity to market “good enough” models for tasks that don’t require the latest and greatest abilities (eg summarize this list, write an email). Arguably, we already have much smaller and simpler models for a lot of these tasks.

There is a market for $20/mo assistants, but the potential market for “everything else” is much bigger - and assistants will be moving towards running locally where possible. These companies are burning billions, they’re going to need a bigger revenue prize for investors. And that’s integration of LLMs and AI into every other software product. The less of those products that run with OpenAI models, the less income OpenAI has to compete, and the harder it will be to keep up. That’s why the opportunity exists for big tech companies can flood the market with good but smaller models and ruin opportunities for cash flow to build the better models.

If the next gen of the top models improves as much as I hope they will - we will not need any integration. It will be similar to hiring a human personal assistant. But much cheaper, even at $2k/mo. Even if this doesn’t happen with GPT5, it can very well happen with GPT6 a year later.

Beware popularity in the Landsraad

At some point is more money really important? My impression from meeting Demis like a decade ago is that he is an actual good person. Being a CEO of a top corporation does not seem like a good life move for someone who cares about others. There are myriad examples of how much engineers who become managers-only find it insufferable. If you're excited about thinking about the technical problems of AGI, is simply more money (or even more power) going to matter?

It is very hard for me to imagine an engineer who has FU money picking a CEO position that pays hundreds of millions of dollars a year over an innovative, R&D type position that likely pays tens of millions of dollars a year.

The smaller the team Demis has leadership on the more he can absolutely achieve but I’m worried he’ll by hamstrung with steering a goliath.

But what is the end goal of OpenAI? Be a non-profit AI R&D leader, convert to for-profit and then sell(license) its AI tech to other companies or get acquired and then eventually devoured by Microsoft?

In this shape and form OpenAI is not a threat to Google; since Google is extremely versatile, taking in consideration how fast they adapted to iPhone threat with their alternative smartphone OS(Android) and taking in consideration lack of clear long-term vision of OpenAI's leadership.

OpenAI's "end goal" has OpenAI becoming a >$1T (100x their >$1B investment) company en route to its non-profit incarnation. Hardly benign

This is exactly right. AI is a multiplier for the services and data moats that Google already has; the absolute value of generative AI has been greatly overstated. The gold rush (OpenAI) will end and the shovel-sellers (NVIDIA) will hurt too.

OpenAI is not a threat to Google

OpenAI + Microsoft is a threat to Google.

If anyone has actually used Gemini and seen how good it is, Demis has probably already saved Google.

I’ve slowly started to switch over to Gemini Advanced the paid tier from ChatGPT4. Longer conversations seem possible with Gemini and it has more recent information - at least for the topics I care about.

Google's problem isn't lack of engineering talent, it's the absence of competent management. Gemini could be far ahead of GPT-4 or whatever OpenAI comes out with next. It won't matter unless they actually build a profitable business around running an AI. This should be feasible, but so far there's no indication of this happening.

The infrastructure is there. The institutional knowledge and engineering talent is there. Google can do this if they return to their core mission statement: "To organize the world's information and make it universally accessible and useful." This is in contrast to their current objective, which from the outside seems to be "To maximize next quarter's ad revenue."

I would love to see a return of the old Google. No amount of engineering skill will make up for clueless management.

I've spent this week re-writing a significant document, section by section, with Gemini Advanced. It had opinions (some were offered straight, and some with caveats given other statements in my document), and readily highlighted gaps in the passages.

I was impressed. I'm still doing a good portion of the human-does-writing bits, but things I'd wanted to include, but have forgotten along the way -- or which are buried in notes I didn't give to Gemini, were brought to my attention again.

What sort of prompts do you use for this process? "Please rewrite this: <ctrl-v>"?

It’s a longer prompt than that but “kind of”. I tell it what I’m trying to do, I identify anchor points that need to be preserved, I specify tone.. I also re-feed larger passages and ask it to critique and review and identify weaknesses, gaps, and redundancy. I’ve also asked it to go through bulleted lists and find references for each. It’s an iterative process but one that yielded improvements.

How good is Gemini? Because my experience of Gemini Advanced, is that would be sci-fi by the standards of 2020. Compared to LLM technology available in 2024 is pretty bad.

Even before posting this comment, and just to make sure I have not been hallucinating lately, just shot a simple programming prompt into Claude, ChatGPT4 and Gemini Advanced. While the first two provided a template that worked even on first attempt, after 5 prompts Gemini Advanced can't even get the expected indent of a Python function block correct...

Gemini is very smart but it fundamentally can't be trusted on any non-technical topics. I talked to it about Taiwan. Gemini gave intelligent reasoning and analysis. But what a coincidence... all of its analysis was perfectly in line with US foreign policy! How unexpected!

What is it particularly good at? video at 1M tokens sounded neat, I've been very impressed with Cladue's 200k + pretty good reasoning + not lazy output at the moment.

Rather than ChatGPT being a replacement for Google, I've found it complements it pretty well. Questions I never would have searched on Google because of length or complexity are well suited to ChatGPT while the sorts of things I've always gone to Google for I continue to use Google for. Also when ChatGPT spits out a few paragraphs of answer for a complex question I'm not even sure may have an answer, I have some concrete directions to pursue which translate much better to Google search than general/fuzzy complex questions.

At this point saving Google doesn't require replacing the golden goose of Search but rather wrapping it with Gemini/whatever that can decide whether to give an answer it's thought up or returning tailored search results.

It's much more a UX question than a deep technical one. Google has all the pieces it needs, it just needs to package them into a seamless experience. You don't need a mind like Demis' for this...

I've found perplexity to be a replacement for google for 90% of my queries and i only use google now as a website index. If I have any sort of question, instead of jumping through the hoops of thinking how to structure the query in a way that fits google i just ask it naturally to perplexity and the answers it gives are correct often enough for me to use it over google.

And perplexity.ai and phind annotate their results with footnotes, linking to the source of the information.

It’s for Apple/Amazon/Google to fuck up if they can’t integrate AI into web old point o apps.

When I pick up my phone, my Amazon groceries should already be bought. We are not even there yet.

>When I pick up my phone, my Amazon groceries should already be bought.

Why even pick up the phone? Or have one! We should stay in bed 24/7, fed through a tube controlled by AI, and watching multiple video streams in VR! Let AI think, write music and poetry, let our internet connected fridges and programmable lightbulbs experience things, let our corporate overlords govern!

Yes.

Maybe our consumption habits are different, but why would you want an AI buying your groceries?

Because I forget.

You’re so human after all.

> Google has all the pieces it needs, it just needs to package them into a seamless experience.

Google prefers it would be 2019, when LLMs were a "remote" threat. They don't make more ad money when people find things, they prefer to keep us searching and seeing ads. They will be dragged kicking and screaming into the future.

I think AI also presents a huge threat to Meta and other social networks. We can get interactive experiences with AI now. We don't need social networks like before. I personally read as much LLM text as human text in a day because it is so clean and useful.

Ads are also under threat. I think web browsers and phones will equip with a layer of "user-agent-ai" where the AI extracts the useful parts from the web and redisplays it for the user under their own controls. You can bet ads are not going to be shown, thrown away together with low quality content. Only AIs will read ads.

Search, social and ads are going to be digested by AI and redisplayed for us. We gain control and protection. AI will form a layer of protection when going online, as the web will be crawling with AI bots trying to gain something from us.

There's an easy fix for AI transforming the output of the Internet giants into useful content for us: just block those AI apps/services. They are determined to be the interface between humans and machines because that is where you make your money.

The advertisement business model of the internet for so long, 20 years plus, is indeed almost irrelevant now. It will collapse like a house of cards. We are heading into a micropayments future.

Even products or services of the front pages of websites, which are displayed to users and consumers on a higher priority than others, that's irrelevant as well.

Google for the record, didn't want to be reliant on ads for revenue back in 2006 or so, but they failed on the micropayment side. They couldn't achieve less than 0.05$ fee per transaction, which is huge. For micropayments even a thousand times less fee, is probably too much.

> Rather than ChatGPT being a replacement for Google, I've found it complements it pretty well. Questions I never would have searched on Google because of length or complexity are well suited to ChatGPT while the sorts of things I've always gone to Google for I continue to use Google for.

Agree, though there are things that I used to google (normally across multiple searches) that I now use ChatGPT for. For example, instead of looking up how to use 4 arguments of a cli tool and putting it together myself I just say "Write me a sed command that replaces X with Y" to ChatGPT.

I've said this here a number of times but I've found a _crazy_ amount of value in having ChatGPT build bash one-liners or small scripts to process/parse data and give me actionable information. I know what's possible on the command line with cut/sed/awk/grep/sort/uniq/wc/etc but, with a few exceptions, I'm not proficient in writing it quickly. I can get there and in the past there were times I put in the effort to pipe together 4-10 commands to extract something important BUT I really needed to have a compelling reason or I needed to be sure it would bear fruit for me to spend that time. Nowadays I simply paste the raw logs/output/etc to ChatGPT and say "I need to extract ABC and XYZ from that string, get a count XYZ per ABC group, sort them and get a count" or similar.

I can take logs and grab what I need out of the lines and quickly say "here is a breakdown by minute of how many times X happened in the logs". This may seem small to some of you or trivial but it's not for me and in the past I wouldn't have spent the time since I wouldn't be sure of the ROI (especially in the middle of a production issue) but now I can take 1-2min, get results, then decide if I should chase it further. Using this I've been able to go from "The data is in our logs" to "Here is an HTML/CSS/JS file visualizing the data in realtime (on refresh)". ChatGPT does a great job at writing the HTML/CSS/JS to graph data and in PHP I can have the PHP script run the one-liner that ChatGPT then "embed" that info for JS to read and graph.

Yes, I'm aware of prometheus and friends and I use them but in the middle of an issue I'm not going to start writing a new prom file, write a grafana widget, and wait for new data to start rolling in especially if we have been logging it but just not sending it prometheus. Long-term I reach for prometheus but short-term I just need the data now and visualizing data in graph form can make thing obvious that no combing through logs is going to expose.

Sometimes the most hyped research is not the most important.

Demis Hassabis was doing good research applying ML to important scientific problems. Then Google became jealous about ChatGPT and they pulled Hassabis to work with that.

"A bird in the hand is worth two in the bush"

We know how important LLMs are. We don't know the importance of things we haven't made yet.

I would love to hear from some folks who have worked with him. The quote is interesting, that his prof would say he's "not a magician at any one thing".

Does that mean he's not all that technical or struggles with math? Not sure how to read that.

I haven’t worked with him but it’s quite clear from his CV alone that he is very technical (ex. programming an entire commerical game at 17) and I doubt there are many things he struggles with. The quote you mentioned is just driving home the fact that he is an extremely high level polymath, a jack of all trades.

TL;DR: the expectation that a business can be "saved" by bolting AI onto an existing organization is unrealistic.

Detail: My experience is that commercially viable AI requires a leadership that gets AI and can execute very careful tech/product development facing end-to-end business problems, obsess over data quality, pivot and and manage risks in a way that's subtler than shutting the whole thing down. None of these are possible when attempting to transform a pre-existing business.

Consider AI and classic Google Search: in the best case scenario, AI will cannibalize on the search. In the worst case scenario, it will generate no lift. The middle ground is elusive at best.

What does seem to work across pre-existing businesses is when AI is used to provide embellishments/optional upsells. But of course that's not the dramatic transformation that a lot of people not in the scene seem to be expecting.

This analysis matches my current experience. I'm at a small startup and I've been trying my best to push them to invest in AI features.

The company is willing to invest very trivial amounts of time/effort in AI embellishments but they are extremely slow to move on any kind of transformational technologies. I had a call with the CEO recently and he was saying that he wasn't interested in foundational AI technologies. I don't mean foundational models (as if a small B2B startup could even consider such a thing). I don't even mean medium/large scale fine-tuning work. I mean, he wasn't interested in investigating broadly applicable AI capabilities which could apply to multiple product features.

I think of it like "dipping the toe into AI" strategies. You get a couple of days to try out some prompting and then spend the vast majority of the time bolting the LLM output onto an existing feature.

What I believe we need are a few big companies to emerge where the new LLM stuff is at the very core of the business so it can show startups how it should be done. I can't even blame this company for being hesitant to invest - there is no proven track record to judge potential success against. At least old-school SaaS feature development work has some basis for projecting revenue. The margin for error on deep LLM features is completely unknown.

I just miss Picasa.

Can somebody explain for the uninitiated? What does Google need saving from?

Off Topic:

>Google’s core business is thriving, but that almost seems beside the point.

I have vidid memory of a scene in Private of the Silicon Valley, where Microsoft is winning, and Apple is losing. But Steve Jobs said they have the better products. And Bill Gate replied "It doesn't matter."

The movie was 25 years ago at the time it finally strike me what I was so irritated about. No one gives a damn about quality, they only cares about the money.

25 Years later ( Well I thought Google were bad for more than a decade but I guess it is more accepted now. ) Google are making all the money but their products are absolutely crap.

While not crap but degradation of quality is happening to Apple too.

As many comments have already pointed out, AI wont save Google. Google has a leadership and culture problem along with Management issues. It is a bit like 2012 - 2017 when people were so certain Intel will continue to lead in SemiConductor with healthy profits and business. Until 2020 they realise Intel finally did fall behind and the culture and management couldn't turns things around.

[1] https://www.youtube.com/watch?v=UFcb-XF1RPQ

You talking about Sir Demis Hassabis?

For what it's worth, Elon Musk recently retweeted a very flattering picture of Hassabis, perhaps throwing his support behind him.

Demis "Infinite Polygon Engine" Hassabis? Please.

Once a charlatan, always a charlatan.

What's an "Infinite Polygon Engine"?

Surprisingly hard to find information about this via Google. Seems related to Demis' game company Elixir and their game Republic: The Revolution [0] for which he made some claim that it supported an infinite level of detail/polygons.

Just a cursory glance at the information though. Happy to be corrected or have more detail added.

[0] https://en.wikipedia.org/wiki/Republic:_The_Revolution

That's not an impossible claim, UE5 Nanite does it.

I believe this was in the late 90’s and in any case, they didn’t succeed at implementing the system they hyped so much - which is the real problem.

The Internet never forgets...Pag 44: https://retrocdn.net/images/c/c7/Edge_UK_078.pdf

It's an in-joke in the British gaming industry.

Hassabis started out working for Pete 'Project Milo' Molyneux and learned how to hype himself very obviously from him (see the Edge article linked above). It's working out well for him but you can see the same breathless self-promotion from him now that we saw 20 years ago.

Fair play, it's made him rich, but we're screwed if we think people like him are the solution to anything.

I don't understand how this warrants the label of "charlatan."

The theory, as I understand it, is that the much-vaunted game engine was not as good as claimed, and the game was also not actually fun, and the man was deploying his IQ more in the service of drumming up publicity for his project than in actually making a quality product.

Whether this is true for Google eep Mind too, who can say. Certainly not me! They seem to pay very well, so this is clearly economically valuable activity. We should not shame Sir Demis for enabling it.

Terrible writing. Creates drama where it doesn't exist. The author makes statements about "problems at Google" and provides anecdotal evidence. Desperately tries to create a "knight on a white horse" storyline for no reason. Assumes Google makes its money off of products, when actually it makes money off of ads. So much bad.

No products -> no ads?

No, search -> ads. It isn't like their ads come from Google Meet or GMail. Their ads come from search. Their search is still best with no challengers of note.

GMail plays crucial role of making sure all search users are logged into google account and their searches can be tracked for ads targeting.

But also, they have some other inventories: youtube, maps.

Google makes money directly from several products.

"Ads" is also an oversimplified way to describe it. Ads aren't a single revenue stream, they're a category of revenue stream with many different ways to tap in, grow, and evolve.

LLMs are absolutely a challenger to Google's search. Google search is the still best option much of the time, but they're far from dominant like they were a year or two ago. There are many situations now where ChatGPT or Claude or Perplexity gives more useful results than Google search. And the LLMs are improving rapidly. Far more rapidly than Google search.

Google has tremendous engineering talent and Gemini is extremely powerful. But adapting their business to LLMs requires a level of talent and vision from the management side that we have not seen from Google's executive suite in a long time. I won't count out Google as a company, due to the potential that could be put to use by competent leadership. But I've given up on their current management. I'd be happy to be proven wrong.

It's undeniable—LLMs (Large Language Models) are now a formidable challenger to Google's search dominion. While Google Search still holds its ground as the go-to in many scenarios, its dominance isn't as unshakable as it was a couple of years ago. There are now numerous instances where ChatGPT, Claude, or Perplexity provide more relevant results than Google Search. Moreover, these LLMs are evolving at an astonishing pace—far outstripping Google Search in terms of development speed.

Google possesses incredible engineering prowess, and the Gemini project has proven to be extremely potent. However, adapting their business model to accommodate LLMs demands a level of vision and capability from their management that we've not witnessed from Google's executive ranks in quite some time. I'm not ready to write Google off entirely because of their potential, but I've grown skeptical of their current leadership. I remain open to being pleasantly surprised, though.

PS: My personal experience with the Gemini project has been stellar. It's enabled me to produce high-quality code, underscoring that Google still has much to offer—provided its leadership can effectively leverage its resources.

Did you just ask an LLM to rewrite the previous comment? You don’t say anything that wasn’t there.

I use gpts almost everything I write. I think with my own brains and boost them with LLM. I have every single place and channel etc. my own custom tools. I do almost everything with custom tools.

Gmail absolutely has ads. Maps has ads. YouTube has ads. Meet doesn't have ads but does have a paid version.

> Their search is still best with no challengers of note.

Kagi has better search. It's a paid subscription, and by market share it's fair to say they aren't a "challenger of note", but the quality is there. (I pay for Kagi but am otherwise not affiliated.)

> Their search is still best with no challengers of note.

Like I've said in other comments, google is not my go-to for search anymore. Perplexity is superior.

Their search is the worst. Kagi being the best, and duckduckgo.com being the better free alternative And with ChatGPT in the room there's doomsday coming, just the UI is missing.

Google has diversified a bit, even if most revenue still comes from ads. About half their revenue comes from search, and the rest mostly from YouTube (10%), cloud (10%) and Android (25%, incl. a major chunk from the Google Play store).

Their search revenue is potentially at risk from people using LLMs instead, but no reason they can't integrate ads into Gemini, even if they haven't figured it out yet. Lots of opportunity for AI: selling API access, Gemini subscriptions, product licencing (Apple apparently interested, Android potential too), etc.

Definitely a mismanaged company though.

Google definitely needs saving. As Android will die and it's the only thing Google did right

demis for ceo

as bad as Google is becoming, does it need saving to survive?

Anthropic (which Google owns a large stake in) and Gemini are 2 of the 3 best LLMs (both of which outperform GPT in at least some non-trivial tasks).

Is it really that bad?

Or is it just en vogue to trash?

Imo the criticisms are more cultural or philosophical than technical, ie the Gemini image search kerfuffle. Clearly Google and Anthropic have a lot of solid people.

Gemini is painfully over censored. It is good at some niche tasks and probably the best commercial vision model though. I think people just don't want to spend time evaluating a model if it's mostly not great.

Google would fix that really quickly if it became a do or die situation. All they need to do is throw out all the overeager social justice people who wants to ruin the model by censoring it. Wouldn't surprise me if a lot of organizational changes already happened after the last debacle.

Currently Google care more about PR with ad customers than market share but if things continue like this that will change really quick.

My opinion is quite simpler: GPT got the 1st mover's advantage from a marketing perspective, and until Google, Anthropic or another proposes something VASTLY revolutionary for the layperson, then it's unlikely they'll "catch up" (which is only a percetion ofc, perhaps a wrong one at that)

LLMs aren't nearly as monetizable (now) as search ads. Very possible that search collapses in the near future alongside the open web, and that whole line of revenue destroyed.

The problem isn't who has the best LLM, but how can you monetize a LLM the way you monetize search. Can you apply keyword auctions on LLMs and then have links? Will that work? We don't know, and I haven't read any concrete study on the issue.

I feel like if anything it will be easier to target ads with LLM queries.

The user intent will be so much easier to identify, and it would be trivial to show ads based on it.

Even if they're in a superior position from resources and talent, this feels like a case of innovators' dilemma with a hard shift of their existing business model and culture. It would almost make more sense for Alphabet to spin off an ML-centric company and focus growth there.

https://en.m.wikipedia.org/wiki/Betteridge%27s_law_of_headli...

DeepMind was Google's life-line. This man (Hassabis) should be CEO of Google.

We have given Sundar enough time and without DeepMind, Google would not have Gemini and he did not put DeepMind to use for a long time until now.

Google AI that made 'Bard' is NOT DeepMind. You should be looking at Google DeepMind, which made breakthroughs like AlphaGo, AlphaCode, WaveNet, GraphCast, etc.

I’m not sure Hassabis’s time would be well spent thinking about, I dunno, how to allocate headcount to the Google Cloud Enterprise Sales team. Or glad-handing with world leaders to convince them that ad tech can self-regulate.

This exactly. Job of a CEO includes doing lots of donkey work.

> I’m not sure Hassabis’s time would be well spent thinking about, I dunno, how to allocate headcount to the Google Cloud Enterprise Sales team.

Exactly. There is a CEO of Google Cloud (Thomas Kurian) that should be doing that. That's called "delegation".

Ok, so then evaluating whether Thomas Kurian is doing a good job, discussing and pressure testing the rationale behind his proposed targets and strategies, figuring out how they fit in with the broader company strategy and communicating this vision effectively, having a contingency plan if Thomas Kurian leaves or a better candidate for Cloud CEO appears on the radar. These are just off the top of my head.

You seem to believe “Delegation” is a passive act; the reality is Hassabis would end up spending, generously, about 1/10 of the time he is now thinking about things like how to monetize LLM’s.

No CEO is going to be strong in all areas, and it seems different but incredibly successful CEOs can come from a variety of backgrounds and management styles.

>"Ok, so then evaluating whether Thomas Kurian is doing a good job, discussing and pressure testing the rationale behind his proposed targets and strategies, figuring out how they fit in with the broader company strategy and communicating this vision effectively, having a contingency plan if Thomas Kurian leaves or a better candidate for Cloud CEO appears on the radar."

You could say the same thing about CEOs who are good at that sort of thing as well: Does Sunadr or Satya understand the technical decisions of going all in on Spanner opposed to NoSql type systems or backing the Go language over Carbon, these are all things that will have huge second order effects 10+ years down the road.

I don't think Hassabis is ideal for the job, but this kind of thinking is what caused Google to stumble. Google needs a leader who understands technology and product. Google is not Procter and Gamble; the kind of company a suit can run.

Reply on news.ycombinator.com