And all the prominent names Google gained: NULL
Combined with no gemini frontier GA release in about 14 months. You have to have created an environment pretty hostile to innovation for this to happen
The bigger deal is the departure of Jeff and Sanjay, rather than Demis moving into a different role.
Oof. Good for Jeff and Sanjay (who just joined Twitter), bu that is a big loss for Google. Google stock is down 5%. It might not be much of an exaggeration to say these two are worth ~$200 billion.
> Announcing Discovery Loop!
> I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
OpenAI and Anthropic went from tiny startups to huge companies. As a consequence the stock options/RSU's offered to the employees paid off a far higher percentage ROI than any stock options a DeepMind (and thus Google) employee would get (since Google is already huge). This disincentivizes people who truly believe in the economically transformative power of AI to work at Google since their benefits will be capped by Google being large + having public company obligations.
So when Deepmind first made headlines I recalled an article in the UK magazine, Edge, which had an article about a game called Republic being developed by a team of former Bullfrog employees lead by one Demis Hassabis.
Out of interest, I just checked Internet Archive, and lo and behold I found it [0]
I see Wikipedia [1] also mentions that he was the lead programmer on Them park and worked with Peter Molyneaux at Lionhead during the development of Black & White.
It doesn't add much to the story under discussion, but it makes me think at the time I wanted to be a game programmer but my parents encouraged me to go study engineering instead.
[0]: https://archive.org/details/edge-issue-078-november-1999/pag... [1]: https://en.wikipedia.org/wiki/Demis_Hassabis
https://www.reuters.com/business/google-shakes-up-ai-leaders...
https://www.newyorker.com/magazine/2018/12/10/the-friendship...
> The moves suggest that DeepMind will be absorbed into Google’s broader business. [...] > They added that while Google DeepMind would not become purely commercial, it was integrating further into Google’s business.
I guess might be quite a change (but writing was on the wall, the Deepmind -> Google Deepmind part was the first step)
Looks like Gemini's sub-par performance is claiming heads
My guess is 18-24 months.
> ... independent publicity [sic] benefit corporation in which Google ...
A Freudian slip? https://archive.is/SxFjr.
Why I left Google Deepmind: https://news.ycombinator.com/item?id=49067285
Google is now working together with Palantir: https://cloud.google.com/solutions/palantir
Clayton Christensen [1] says (paraphrasing) it's a good idea to spin-off (or invest) in a startup that you've a say on, before the ecosystem sprouts a seemingly-non-entity and gently disrupts your market.
Interestingly (diff strategies i guess) Apple tends to acquire (Q.ai) rather than invest/venture (as google in OP).
[1] The Innovator's Dilemma: When New Technologies Cause Great Firms to Fail.
There is an entire cohort of work-optional very senior engineers for whom one of the last reasons for hanging on was "at least Jeff and Sanjay are around".
1) the simplest explanation whenever a bunch of people leave [x] at a company at the same time, is that [x] is becoming less important to that company moving forward. It's not proof, but you know, everyone is trying really hard to say that's not what's happening here. Sometimes the simplest explanation is correct.
2) rumor is that Sergey Brin, co-founder of Google, got bored during pandemic lockdown and eventually returned to Google in a lower profile role, related to AI. I have to think that he still has a lot of say in what goes on in Google relative to his interests.
3) Yahoo Finance article says Hassabis "has long prioritized research over profits", so his replacement might indicate that Google has decided it is time for this stuff to pay for itself?
>We’ve got amazing talent, world-class compute and products…
Products are third on the list. Google is an incubator for talent first and foremost. Products are an afterthought
I feel google has become evil.
I had my phone and laptop stolen and did not have 2fa on but Google still locked me out.
That’s stupid.
https://www.lesswrong.com/posts/iKm2FhpWkuuBojm82/why-i-left...
I am not fan of the future of AI but this I am not sure how I feel about Google's (also Amazon recently laid off it's AGI team) AI initiatives blowing up before all the new AI Labs.
I don't get why Google failed here? maybe after a decade someone will write about it candidly.
Considering these are the best stats they could find, gemini usage+general situation must be really, really bleak.
High demand means nothing. A model being live is nothing to brag about. And gemma downloads also can be from auto CI pipelines etc. Nothing concrete
But missing out right as AI coding agents become genuinely deeply capable and useful is just an immense failure.
Seems like a good way to spend investor dollars without consequences while maintaining control. Maybe a bit cynical but i can't figure out why they'd form a PBC over anything else.
Scientist-heavy orgs that want to solve everything in token space may overtook tool use; meanwhile Anthropic has been super focused on MCP, Claude Code etc for over a year
Whatever happened to Prabhakar Raghavan? Got kicked upstairs and we barely hear from him nowadays.
It's only a matter of time before Demis leaves and joins Anthropic or OpenAI.
As an outsider, Google seems to have a knack for minting prominence for their talent.
To me it is good thing in either case. Extraordinary people are leaving to make even faster research and development. A lot of talent in Google hitherto unnamed is going to get chance to shine.
There is another universe where Google is hard AGI forward, freeing up as much compute as possible for Deepmind, turning away OAI and Anthropic, and offering the uncontested premier AI research lab, by probably an order of magnitude or more compute. Gemini 3.5 ultra is limited availability with unreal abilities.
But their balance sheet becomes terrifying, and it's all or nothing that this plays, er pays, out.
Right now though Google is pretty well hedged. Even Chinese models likely just mean more compute sold.
Reminds me of when that one verse was removed from Psalm 145.
Nowadays, every company—even huge ones—prefers to be seen as a "growth opportunity", so they are trying to play up their ability to create new products which will somehow be so incredible that the line keeps moving up forever.
That said, there is definitely a correlation between companies chasing new products and leaving old ones to become crap.
Founders are first. Ideas are second.
If I had to guess, they have the wrong kind of bureaucracy for where things are headed - and it is manifesting through talented individuals deciding to leave.
In most cases, it doesn't really matter. The board + management is still in charge, and they have significant legal leeway regardless of the structure. But there's little additional cost to opt for a PBC, and it does give you more legal defensibility to be truly mission driven. Standard C Corps weren't really intended for mission driven companies (see the shareholder primacy norm).
I think it's popular for AI startups, because many great researchers understand the risks involved, and they don't want what they build to be controlled solely for shareholder benefit.
While non-profits are also an option for a mission driven org, it's harder to raise the large amounts of cash that some AI startups need, and laws around deferred compensation and private inurement (e.g. options-like structures) make employee compensation harder.
How do you avoid being the Xerox to the Microsofts and Apples of AI?
Edit: hmm, no, they seem to be mentioning AGI and frontier models in https://blog.google/company-news/inside-google/message-ceo/n...
People who truly care about becoming rich*
DeepMind made enormous transformative discoveries, for instance in the world of protein folding. But that will just save human lives, not let CEOs fire their people to grab a larger piece of cake for themselves.
It's funny, because I think the company that's going to be best positioned coming out of this bubble is in fact google, because they have the expertise and the capital. But I honestly can't tell you right now what their AI product even is -- I've seen so many things go into the graveyard a few months after its launched that I'm utterly confused what their offering even is at this point.
But yes, even large RSU grants cannot replicate the upside of joining a company that increases in valuation 1000X, but to be fair, that is not a position the vast majority of OpenAI or Anthropic employees find themselves in either.
Instant nostalgia: https://youtu.be/tQJJ_rhHxIk?si=V8pOVwj3gYkw0xKB&t=188
Google's processes, however, is hands down the worst thing to exist for the scale they operate at, and it's not even close.
this is false, it's very good at internal tooling.
so true. i have attended meetings to decide on meeting topics for the next half.
I said for a few years to many a downvote on HN, everyone wants AI, nobody wants to pay the true costs, the AI race will turn into a "race to the bottom" that is, who can give you the most compute for the lowest cost, and still remain profitable?
This over-reaction is why people can't see over a long time horizon.
This is great news for Google as they realize that Sundar is the problem and he will soon leave Google for Demis to be the new CEO of Alphabet (Google) which I am predicting. [0]
2024: Google is on fire. Sergey coming back helped. Gemini, Veo. They've caught up. OpenAI is doomed.
2025: Google is seriously winning now. Nano Banana!! Google was destined to be the true winner of AI.
2026 H1: Google is slow as hell. Where is Gemini? Google is not launching anything, meanwhile just look at everyone else. Anthropic! And open source. What the hell are they doing over there?
2026 H2: Everyone is leaving Google. Google is doomed. PMs are destroying the company. They don't take risks.
Meanwhile, Microsoft: "Copilot!"
I have always found NPM download numbers truly suspect. Is no one caching? Are they estimating true number of downloads base on some estimate of cache hits?
EDIT: "figurehead" - that's all I meant. A notable, public figure from the company who is credited with having a significant impact on its evolution. I'm not making a judgement call on his departure, or Ive's, being good or bad.
That said... just give it a couple years, they'll be back in a lucrative aquihire.
IS that a promotion or demotion ?
Hassabis seems to have been pushed aside. He had been CEO of DeepMind, but that position no longer exists and it seems Kavukcuoglu is now leading DeepMind with a title of SVP. Hassabis is now just "Chair" of DeepMind, and has been given the newly created title of Alphabet Chief Scientist. Are these just face-saving titles, or does he still have any real influence over Google/DeepMind's pursuit of AGI?
Shane Legg remains as DeepMind "Chief AGI Scientist", but I wonder if the DeepMind founding mission of creating AGI is really intact, or if he will be next to go. Has DeepMind just become the Gemini division?
I just found out that David Silver, DeepMind's RL-expert, already left in february, to create a startup "Ineffible Intelligence" focusing on RL-based continual learning.
Clarification: this comment is saying Sanjay Ghemawat joined Twitter as a user recently (new account @Sanjay_Ghemawat as of July 2026), as opposed to Sanjay working for Twitter the company.
It think the vast majority is what the decision to leave says about the overall sentiment at Google, not the value of these two individuals.
It's not really clear what their gameplan is. From the outside, it looks like they're asleep at the wheel. Qwen/Deepseek/Kimi are crushing them from the cheap-and-open side, and they're not remotely competitive with Mythos/Sol or even plain-vanilla Opus on the "premium" tier.
Gemini does actually have its uses, but they're very very marginal and niche.
From day one everybody was saying that Google would eventually capture the AI market, but it looks more remote than ever. Maybe Hassabis' personal inclination towards AI-for-science, and physics/chemistry in particular -- as opposed to consumer AI and coding AI -- has hurt them commercially.
I am sure VCs are fighting to invest and they will get 4B investment immediately with this team
They getting a higher ROI renting their TPUs to Anthropic et al instead of performing training and serving their own models. Google cloud has insane backlog, and has rapidly expanded to satisfy it. While those DCs get built, they’re cannibalizing their own products for it.
This makes sense because (1) they are investors in Anthropic, so they still win and (2) they can always catch up on model training later when the profit opportunity shifts, or abandon it if there is no way to recapture that value.
It's a huge company.
It's unlikely there is any one person to blame (and entirely possible he has none of it). But things need to change.
The Japanese corporate culture has a name for this: Madogiwa-Zoku (窓際族) or “The tribe by the window”.
Even in indirect ways. OpenAI itself was founded because Musk got fixated on "stopping" Hassabis.
Which is: When you think about Google, true, old, "don't be evil", tech excellence Google, you don't think about Demis. You think about Jeff's and Sanjay's geeky, technically uncompromising, faces.
The four have so much influence within the company that they could have trivially set this up as part of Alphabet, if they wanted to. They are also ridiculously wealthy.
Then the Google engineers who joined Facebook missed it so much that they built a better replacement.
Kinsley Gaffe: A mistake whereby a politician inadvertently says something truthful which they had not meant to reveal.
Google’s tooling was, hands down, the worst I have ever encountered. I did 10 years at GOOG, 3 at AMZN, 4 in research, and another 5 at companies you have heard of but wouldn’t be impressed by, and every day GOOG infuriated me.
I keep seeing this but this line of thinking doesn't make any sense. What does it really mean?
There are expensive models that increase the probability of you doing your task under a lower cost. That means you can't use Gemma for coding your new compiler - it would just be overall costlier.
Heavier models are cheaper at more complicated tasks because they use fewer turns and fewer mistakes.
Cheaper models are more likely to be cheap at less complicated tasks. Like if you just ask Gemma "Hi" it would probably be cheaper than asking Opus.
So what does this statement really mean? People don't want to pay the extra for a more costly model? Why wouldn't you? It reduces your overall cost!
But Jeff was responsible for a lot more of what is actually used today than Demis.
Demis is responsible for a lot more of the hype though ;)
I think saying Jeff's contributions were a long time ago must represent some kind of lack of understanding of Jeff's recent contributions.
Jeff has still been focused more on infrastructure, and that is just more hidden most of the time.
I would say,if i was forced to pick someone whose vision to follow, it would definitely be Jeff and not Demis, even today.
Having said that, I think Google's moat is still strong with Cloud, Gmail, YouTube, Android, Chrome, etc.
Alphabet Chief Scientist doesnt sound like a demotion / lack of influence to me but who knows. We're all just speculating here.
When DeepMind allowed Google to buy them, it obviously had some major immediate positives - access to compute and money - but it seems it should have been obvious that the agreement was too good to be true, that they would be allowed to continue independently on their blue sky research mission to create AGI without any external interference or pressure to create product.
It seems that Hassabis and his DM co-founders eventually realized the mistake and tried to take DM private again starting c.2018, but of course this failed.
https://colossus.com/article/project-mario-demis-hassabis-de...
Now Hassabis has lost control of DeepMind altogether, and it seems to me, as a total outsider, that this is the end of the DeepMind mission to create AGI, at least the Hassabis/Legg definition of AGI as human-level general intelligence, capable of creativity and scientific discovery. Hassabis had always, until very recently, said that he believed it would take a number of additional "Transformer-level" breakthroughs to achieve this type of human-level AGI, while still seeing an LLM as one component of if (which to me seems an admission that the goal has failed - a true human level AGI should be able to learn language, etc, using it's own continual learning mechanisms).
It seems that DeepMind has now fully become the Google Gemini (LLM) division, trying to create a me-too product.
In the early days of DeepMind, before Google, before LLMs, I remember a David Silver slide deck titled "Reward is all you need", referring to RL rewards, which I never agreed with (although Rich Sutton might), but does at least reflect the independent thinking at DM, and of course RL not only gave rise to AlphaGo, but has now become central to the continued improvement of LLMs. However, notably David Silver also left DeepMind earlier this year, to found his own startup focusing on RL-based continual learning, presumably feeling that there was no longer a place for that type of research/pursuit at DeepMind.
Still, LLMs seem to be a destructive enough force on their own that perhaps it should be seen as a positive if research towards more powerful AGI appears to have had a major setback.
As for the "Alphabet Chief Scientist" title, it seems somewhat irrelevant, as least as far as Google's pursuit of true AGI. Hassabis is the face of beneficial AI, having been Knighted and awarded a Nobel Prize for his work, and it would be a horrendous PR move for Alphabet not to at least appear to be treating him with respect, even if in fact this does reflect him being pushed aside.
Yeah he seemed too reasonable to me, relative to the fervor. Whoever ends up in charge needs to do a lot of frothing to catch up to the ferver that would justify their valuations and investments
It was to fight ChatGPT and promote Gemini.
I think the guy does a very poor job or is simply not the right guy to appear as public figure for Gemini.
At least he tried. He is a man for everything that is not filmed.
Google doesn’t really have a person to give Gemini or AI a human face. And that is only consequential because Google never had any public person with any charisma like Jobs, Zuck, Altman.
The company would have ended up like Stack Overflow? I don't think there was an option for them.
keeping Search profitable without enshittifying it
Prior to this, just about everything on the web was open for crawlers since content wanted to be found.
Are you sure we should compare it like this? Not sure it implies what you think it does...
The new paradigm is you ask the llm a question, get the answer and cutout the middle man. (yes the answer may or may not be as good as the old google result, but for the sake of the argument lets say it is), Google was in danger of simply getting their arm cut off. so they focused on scaling so they could add LLMs to the search, which they largely have. You can't offer an opus like model on something as big as search (and which is offered for 'free'), so they focused on that model, and the infrastructure to run it, because they cannot afford to lose search.
Meanwhile, they know the power of frontier models, they are working to have the infrastructure to be a huge player in them and I'm sure they will have a frontier capable model, eventually. They are playing a longer game, because they can, and I think it's going to work out very well for them.
They have plenty of great tech for consumers right now: Gmail, Maps, YouTube, Chrome, Android,..
You can perhaps argue that search might suffer because of AI but they've done a great job scaling horizontally into many verticals.
I don't think Google is going anywhere.
Oh well, emperors, clothes, ... you know.
The more people find out that gemini 3.1 pro on the google AI playground or via API is de-facto uncensored, the more likely google will put up actually working guardrails and end the fun for everyone!
The difference is nobody expects anything better from Microsoft. Teams didn't exactly set the bar high.
Thank you for adding some clarity to this. When I calculated 900 million downloads divided by 8.3 billion people in the world, I came with a number that made it look like about one person in 10 were downloading this model.
In other words 'just add a calculator tool' is not as sexy research-wise as making the model accurately eyeball arithmetic in its chain of thought. Maybe I'm wrong but that seems to be the case
But people want what AI does for them. It's a tragedy of the commons situation.
Real scientists are skeptical. Wall St and the people who serve it don’t like that.
Instead, it seems Pichai has fumbled what they had with DeepMind, and they'll now just be a me-too LLM competitor chasing OpenAI and Anthropic from behind, in a race that will never get to AGI.
But it was unpopular, in that people don't like mental effort.
Because real Fable usage starts at $20/month, and has oppressive usage limits even at that (ridiculous) monthly price.
Compared to my $3/month GLM-5.2 subscription, I have never felt like I was leaving capabilities on the table by refusing to cough up $20 for 15 minutes of Fable use per day.
Technical merit is not correlated to popularity, after all.
Cloud and gmail, I agree.
I don't think YouTube can survive if GenAI keeps going like this. Android, same, but on a different timescale and for different reasons. The Play Store (and all other app stores including Apple's) will also face problems from GenAI making apps (it already replaces my need to buy, but I'm weird and a software dev ("but I repeat myself")).
Not sure how big a moat Chrome really is? It's more like a sales funnel than a product itself, I think?
The position that I have rather often read on the internet is: Jonathan Ive did very good work at Apple as long as there was a counterpart who could steer his creative vision. This counterpart was of course Steve Jobs. When Steve Jobs died, there wasn't such a counterpart anymore, so Ive's work for Apple got much worse.
The fact that OpenAI, Anthropic, SpaceXAI and 3 different Chinese companies were all able to train big models without these issues, yet Google could not, seems shocking.
- In a world where open source Chinese models decimate Frontier models ability to charge a high price, it's the operators of efficient inference data centers that will win. Like Google
- Google is probably the biggest provider of "free" AI because it's on Google.com. That forces them to focus on cost. And in a commodity market, low-cost providers are the ones that make the money.
10 years from now, its gonna be Google, Apple and Microsoft left standing in the AI game. Well, until the US wakes up and starts attempting to break the oligopoly like the EU has recently started to do.
They need to route all browser search strings to an LLM, and slowly begin to charge where people will pay. Likely ad space.
No, the top talent is clearly at Anthropic and OpenAI
There are PetaBytes of important scientific data locked in archival file formats. The first step is to make this efficiently readable.
[1] also known for open-sourcing ProtoBufs v2, creating CapnProto and Sandstorm, and commenting on the recent Cloudflare OS post https://news.ycombinator.com/item?id=49182996
ahahah golden.
Will be following their journey
This is probably the only time some prominent VCs would be active during August, only for Jeff Dean & Co's company.
I can see them now frantically negotiating, responding and firing off emails right now during their holidays to get an allocation in Jeff's new company.
aka they're dead in the water.
If AI turns out 1/100 as important as they seem to think, it would be insane to intentionally be slack on it.
Very few companies have leadership that can prevent this infighting and force teams on directed goals.
No one joins OpenAI/Anthropic unless they think these companies will reach superintelligence.
So most people joining believe their equity will 10-100x even from where it is today.
(Coincidentally, the talent that believes we will reach AGI overlaps a lot with the best talent, which has a magnetic effect.)
I don't get the impression from interviews that Kavukcuoglu is that guy - he seems like a safe pair of hands, but not someone that is on a mission.
OTOH I don't even think this is the right race to be in.
CEO of <thing> is a layer above SVP.
Google has many layers of management.
Conversely, AI is just a means to an end for Google - they don't need for their model to be the one to succeed. But, in contrast to the other major company in their position Apple, they do have a model, so they're not totally beholden to another for AI (like Apple is using Gemini!).
But beyond that, for training the model they have YouTube, and of course they have their crawler and index, and the billions of users.
I think HN skews coding agent focused, but that's not really a market for Google. I expect we will have coding specific models in the future, but Google wants a more general intelligence, to handle search queries, be able to connect email to chat to calendar and tasks, and so on. I don't find Gemini that much worse than the other big models for non-coding things.
Now, whether Google is the right environment to nurture, that’s its own quandary.
Another is that they hold the key Transformer architecture patent. If it is still relevant (which I'm not personally clueful about) and if they start enforcing it, then we may see a reprise of the situation where Microsoft made money for years every time an Android phone was sold. Regardless of that particular patent, it's probably safe to say they'll be better-positioned than anyone else if AI companies start lobbing patent nukes at each other.
A third factor that shouldn't be discounted is that Google has access to warehouses of training data that other companies don't. Google Books alone is an Alexandria-scale archive that the courts forced them to keep to themselves. Those restrictive copyright decisions may turn out to be a blessing in disguise for Google because no one else will have been able to scrape the data.
for all these reasons, is being 6 months behind the frontier actually a structural, long term disadvantage? some day the pace of improvement will slow, and google will vacuum up the market. they'll be able to compete with open-weight models just on pure cost advantage from their vertical integration
my take on this is eventually the money is going to run out and there's going to be acquisitions and consolidation. I think that's when Google will come out on top.
For agentic work, but especially for web search, 3.6 Flash has an important leg up, its fast speed, that no other model comes close to matching. I guess nobody pays attention to it because it's not the one big flashy number that you compare to other models.
There's an old saying, if you judge a fish's smarts by how well it can ride a bicycle, it will always seem dumb.
The people who have risen to the top of at Google are built for a different environment than what's needed right now. They are good at playing their political games, sabotaging each other, etc.; i.e. all of the petty games that managers play in big companies. But the AI era demands a different skill set: how to bring together incredibly smart people and forge them into a battle group that will achieve victory in the ongoing battle for AGI! It's as if you have built an army of tanks, but the next battle is being fought on the high seas.
I’ve been reading this sentiment on HN since GPT4o, yet models got better and better
Uh oh...
They could really use some encouraging news about the competitiveness of their AI lab.
[1] - https://www.forbes.com/sites/jackkelly/2024/05/31/google-ai-...
[2] - https://www.axios.com/2024/02/23/google-gemini-images-stereo...
So I think they have to prioritize scaling for their models to a higher degree than other groups. Being within say 5% or so in most cases is probably adequate and matters more overall for their user base than being the absolute best coder. So they may be setting compute constraints for training or inference that are firmer than other teams.
Google spending money on competing head to head with your LLMs is a waste of money for Google. If anthropic wins, Google copies their approach, buys anthropic for cheap or both. All the investors throwing money into OpenAI, Anthropic, are just accidentally subsidizing Google's product development. Google shouldn't spend its AI research capital in an arms race with Anthropic but instead should invest in AI approaches that no one else is investigating at scale. That way Google can hedge against LLMs hitting a wall.
There is a real but small danger to Google that Anthropic replaces Google as a search engine, but that is an uphill fight for Anthropic. Google has massive brand recognition, network effects with gmail and chrome, Anthropic can't just copy what works from Google. On the other hand Google can copy what works for Anthropic. Google would have to play poorly to lose that fight.
Probably the worse case for Google is that software becomes so cheap and easy to create and maintain that all of Google's product offerings become commoditized. Even in that world Google has a lock on infrastructure. Perhaps ASI software creation completely removes that as well? If so we are living in a post-singularity world and probably the stockmarket doesn't exist anymore either.
Not all change is good, as proven by Zuckerberg's response after the lackluster Llama 4 release. The radical restructuring appears to have made things worse.
Yes it's a huge company.
So they are slower. Then they will surface it across their massive product base and keep generating cash. While having a hand in Anthropic and others via investment anyway.
Google doesn't need to offer you the bleeding edge at startup pace. They're playing a different game. When the bubble pops they will be well positioned really no matter the outcome to continue to capitalize as their competitors implode or get absorbed.
The idea a delay is a "complete and unmitigated disaster" is just laughable. People have been saying this about Google since ChatGPT first invaded the public consciousness. Google will continue to do well, the histrionics of people like you aside.
But BCL is still the worst language I've ever used.
With "believes in AGI/superintelligence just as much as HN does" do you mean that they are superbelievers or rather sceptical of AGI/superintelligence? I have seen both positions on HN.
Also lots of tech people, HN included, are waking up to technology not only including penicillin (net positive for humanity) but also dynamite (best case: net neutral).
Start in a blank directory and tell it to spin up a boq Scaffolding stubby server that responds with "hello world." Unless something has changed after I quit a few months ago, it won't know how to do that locally, let alone actually deploy it. Try the same outside Google with like a Flask server on AWS or GCP.
I have never seen anything like it in 20 years of Linux desktop use.
"But he's the Prime Minister!"
"Indeed he is Bernard. He has his own car, a nice house in London, a place in the country, endless publicity and a pension for life. What more does he want?"
"I think he wants to govern Britain."
"Well stop him, Bernard!"
That can work in the B2C space but it's horrible in B2B.
Who knows? Pure speculation? You can also say if Jobs was still around they could have 10x-ed it even further?
Apple car could have been a thing? Apple could have been way ahead and actually competing in AI and data centers? Who knows what else Jobs could have came up with?
My main reasoning was that transformers was the lightning in a bottle and the best work is in extending it instead of transcending it, which requires you to capture another lightning . Which to me appears to miss the assignment. OpenAI, Antrophic, they understand this intimately. Google on the other hand, fell victim to their own ambition.
Anyone remember the touchpad MBP with no physical escape key and the butterfly keyboard?
(Respect to many of Ive’s great legacy though)
Editor’s note: Today, Google and Alphabet CEO Sundar Pichai shared some changes with Google DeepMind teams, including new roles for Demis Hassabis and Koray Kavukcuoglu. Below are the messages Sundar and Demis sent to employees.
We’ve made extraordinary progress to deliver on our full AI stack. We’ve got amazing talent, world-class compute, and products that bring AI to more people than any other company. And you saw the incredible momentum at earnings across all our businesses, including Search, YouTube, and Cloud. Our Gemini models are in high demand among developers and businesses, and the Gemini app reached 950M+ monthly users. Meanwhile, our AI research continues to drive field-defining breakthroughs (like last week’s Gemini Robotics advances).
We have to accelerate all this work and stay focused on the AI frontier. At the same time, there’s never been a more important moment to shape the future of AGI and science. Today Demis, Koray and I are sharing a few changes to our Google DeepMind teams that will enable us to do both.
AGI and science: Demis has described us as standing in the foothills of the singularity, and has been spending a lot of his time engaging externally. He and I have been long discussing a role that allows him to put his full attention on actively shaping the future of AGI. It’s work that is vitally important to Alphabet and humanity, and I can’t imagine a better person than Demis to do it. So, moving forward, Demis will become the Chair of GDM and Chief Scientist of Alphabet, while continuing to lead Isomorphic Labs. He’ll remain closely connected to Koray, Josh, and our GDM teams, advising across models and research. I’m so excited for Demis — this is truly his life’s work and purpose. You can read Demis’s note to GDM below.
Google DeepMind: We are building strong momentum: Flash is in high demand, our Cyber model is live, and Gemma models have surpassed 900M+ downloads. We are committed to being at the frontier, and are super focused on the areas where we need to improve. I’m really excited for our upcoming model releases and the progress we’re seeing. We have to continue to move fast and with clear purpose here. Koray, the current Chief Technology Officer of GDM and our Chief AI Architect, will step up as SVP of Google DeepMind, reporting to me. He will oversee Gemini model development, Frontier AI research, and the Gemini app and developer teams. Koray has been at DeepMind since its early days, and over his 13 years there, he has started our deep learning team and led the way on breakthroughs like WaveNet and DQN. I look forward to seeing him lead GDM into this next chapter.
Lastly, after an incredible 27-year run, Jeff Dean is at a moment where he wants to try something new, and we’re excited to support him in that. Jeff and Google Senior Fellow Sanjay Ghemawat are launching an independent public benefit corporation to accelerate discoveries in ML, science, and engineering. Jeff and Sanjay helped to drive some of the most significant technology transitions, from our early search infrastructure to the neural networks that helped create the modern AI era. On a personal note, it’s been a privilege to work alongside Jeff and Sanjay, and I wish them all the best! We’ll continue to work with them as a founding investor and Cloud partner, and collaborate on a research framework for ML systems and related infrastructure advances.
We are at a dynamic moment with so much opportunity ahead. With today's changes we're going to keep driving our momentum. Onwards!
-Sundar
Hi Team
We have arrived at a pivotal moment in human history. I’ve been working towards AGI my whole life and now, like many of you, I feel it is close at hand. It’s critical that we collectively get the next steps right to ensure this all goes well for humanity and we usher in an incredible new age of discovery and wonder.
With this backdrop, I’ve decided that now is the right time for me to hand over my day-to-day operational responsibilities at GDM, so that I have the time and space to focus on the big picture and help influence what is to come to the best of my ability. I will be taking on a new strategic role as Chair of GDM and Chief Scientist of Alphabet, and I’m excited to announce that Koray will be stepping up to lead GDM as SVP of Google DeepMind, in addition to his role as Chief AI Architect of Google.
Koray and I have been working together for over 13 years, since the early days of DeepMind. He is one of the world's foremost AI experts and has been championing GDM's mission from day one. I have total confidence in Koray, Josh, and the rest of the GDM exec team as they continue to spearhead the latest AI developments across Google. The Gemini models are in good hands with Koray and the leads, as they have been for a while, and I'm excited about the great progress we’re making with our new models including Gemini 4.
In my new role, I will continue to work closely with Sundar on strategic and global AGI matters, and to advise Koray, Josh, and the GDM leads, from our awesome new London Platform 37 offices. As part of this transition, I’ll also be leaning into my role at Isomorphic, where we are making extremely rapid and promising progress, to accelerate our mission there even faster. As you've heard me say many times, I’ve always believed the No.1 application of AI should be to improve human health. It’s time for AI to prove its unequivocal value to the world, and what better way to demonstrate that than to help finally cure diseases like cancer.
We’ve built a unique culture at GDM that has served us very well. I want to thank each and every one of you for your brilliance, dedication, and effort that make GDM the huge success it is today. We should all be extremely proud of the amazing things we’ve achieved so far. We’ve become the AI engine room of Google, with Gemini delivering helpful experiences everywhere including AI Mode and AI Overviews, the Gemini App rocketing to over 950M monthly users, and our fundamental and scientific research continues to lead the world. I’m very excited for our next chapter and the best is yet to come!
As a business we are in an incredibly strong position. We are the only company that has the full stack and we’re world-class at every layer from infrastructure to cloud to frontier models to AI-first applications. We have all the ingredients to lead from here, and I firmly believe we will.
Best
Demis
To me the biggest news is Jeff and Sanjay leaving Google... but then not really, since Google will be "an investor". I guess that's sorta their retirement plan? And is Discovery Loop actually part of Alphabet? So complicated.
To be clear, even if this is an accurate observation, I am not arguing in favor of high executive pay.
They all leave to start new companies. Everyone on the Attention is all you need paper is at a startup.
> I think HN skews coding agent focused
100% this. Coding agents are an interesting test ground, but the people who care about them are a fairly small bubble.
I think they carry some of the overall AI weight because of the "what if you can vibe code your entire business" moonshot, but that's still orders of magnitude away and who knows if today's coding agents will actually be a stepping stone to that. If we ever get there it will likely be with entirely new domain languages.
They are quietly trying to become the chatbot that everyone uses to look things up. That strikes me as an intelligent move - not everything has to be done by an expensive frontier model.
The group running the company, is the company.
Those 2 are definitely NOT Google's strong points. Maybe by means of marketing.
Oracle, Amazon, Microsoft, Equinix and many more are in the data center race.
As for TPUs... Broadcom, Mediatek and all the other partners you hear less about are likely more important. Google just has the flashy media outreach.
Big shake up for Gemini it seems.
Note that I'm not calling Google "literally Hitler". I'm just saying that traditional assumptions about what's virtuous and what isn't don't apply anymore. For now, at least. If anyone wants to change that, they're going to have to be purposeful about it.
When you have a lot of free users the business demands that you serve them with the best cheap model you can build
And time spent building that may provide dividends (eg OpenAI has very good RL and reasoning) but it might take resources away from the larger model training
(I have no inside knowledge, so please consider this to all be speculation)
They already have good models, so “better” isn’t as profitable.
I've noticed that Gemini has been timing out yesterday & today !
Google has multiple cash firehouses, the small AI companies do not.
https://www.cbsnews.com/news/google-ai-chatbot-threatening-m...
It's like the old adage: "If you make AGI you just want to watch computers cry"
We had relative strength, we needed more strength, we built cranes;
there was something like intelligence, we needed more intelligence, we sought to fill the need;
we were rightly fascinated with intelligence, we studied intelligence itself, we wanted to build it...
And then somebody built things that looked like intelligence, and very rightly some said "Oh, now we have to get to the Real Thing with urgency".
Is it no nearer than it was in 1750, 1850, 1950, 2001, or 2010? It feels nearer. Relatively near, as in it feels weird to imagine another two thousand years happening at current rate of progress without anyone stumbling on it.
Is the continuing increase in global computing power, the lowering cost to do any kind of experimentation, the increased spending on R&D, the cross-pollination of ideas, not moving the needle at all?
So it must be something very specific to what software you have installed, unless it is an AMD GPU problem, but neither of the 2 links (x and xcancel) has anything that seems to need special GPU features.
It certainly is not a general Linux problem. I also use XFCE as desktop, so if you use Gnome it could be some Gnome component.
The former caused by the latter.
And the latter a problem once you begin to get executives and leaders who have mostly worked inside the company, because then they subconsciously prefer the Company Way(tm) to alternatives.
And Google has some previously-optimal, now-detrimental company ways.
So the parent comment would imply that losing Dean is a good thing for Google, which is way less likely here.
"Tobacco-related diseases cost the NHS £150 million per year!"
"Yes we've looked into that, it turns out that if those people had survived they would have cost us billions in pensions and healthcare costs! From a financial perspective, it's vastly preferable that they continue to die at the current rate."
the last time a gemini model was available globally in multiple datacenters was gemini 2.0 era.
If your goal is purely commercial, or time critical, then a product-based approach of squeezing all the juice out of LLMs makes sense.
If your goal is truly human-level AGI then this is more of an open-ended research endeavor, and timelines are hard to predict. Arguably we have only "captured lightning in a bottle" once in the last decade - the original 2017 attention paper - and so the timeline for a "few more Transformer-level breakthroughs" might more realistically be estimated in decades rather than years. You could argue that the application of RL to LLMs as a training method was a second "lightning in a bottle" but I don't think it changes the expected timeline of such discoveries by much.
The time criticality seems to have become a huge factor for those pursuing LLMs, and certainly for OpenAI and Anthropic, who regard it as a race.
It seems absurdly obvious (though many would disagree!) that LLMs alone are not going to achieve human-level intelligence and cognitive performance (using a slightly broader term there to include things like creativity, for those that might not consider that as part of intelligence).
If you compare a Transformer to a brain, then the best parallel is that a Transformer is functionally similar - in being a prediction engine - to part of our cortex, but of course that means ignoring the other half our cortex - the feedback paths that enable continual learning, which in turn supports creativity.
Of course people will probably respond "you don't need flapping wings to fly", but if you want to fly you do need SOME way of doing it, so brain comparisons are still valuable... If you look at our brain architecture and identify all the components and connections that have no equivalent in a Transformer, and if the goal is human-level capability, then you do need to understand what each of those brain components achieve functionally, and have SOME way of providing that functionality in your LLM+ or whatever you call it. LLMs' lack of any functional equivalent to our cortex's feedback paths - lack of continual learning - has been recognized as one major functional deficit, but there are probably half a dozen others too, reflecting the multiple "Transformer level" breakthoughs that Hassabis notes are needed.
While you don't need flapping wings to fly, you do need to invent the airplane, and even after 100+ years of airplane advances we've yet to build an airplane even remotely as capable as what some birds and insects are able to achieve.
Maybe 2017 was the Wright Bros moment where humans first learnt to do some of what our brains can do.
Would like one with all the current physical keys plus a Touch Bar that you could do cool stuff with.
Entirely possible that none of the big incumbents ends up winning, economically.
Yeah, this is one of the things I find so weird on the discourse. If you get there negligeably later, but without astonishing spend and waste, you might even be better off in the long term.
They may not capture enterprise use, but I don't think they have to. That's only one piece of the market. AI that's useful to consumers will still get delivered via a smartphone, and Google is in a great place to capture that.
I also don't think LLMs have to be a "winner takes all" situation. Value isn't going to come from having direct access to a chatbot or selling API inference, value is going to be in the form of a specific product (for most, devs aside here). Something a consumer, or a non-tech business can buy off the shelf and plug and play. A "ready made" customer service agent system, a "ready made" BI platform using AI, etc.
For consumers, that's probably going to look like whatever is bundled and tightly integrated into their mobile OS of choice.
LLMs will obviously keep getting incrementally better, but in order to get the kind on jump that LLMs themselves were, the sentiment is that we need something more.
Whether this happened because they bet on "world models -> better reasoning" and that bet didn't pay off, or failed a frontier run for technical reasons like OpenAI did with 4.5, or something else went down? We don't know.
Will they bleed talent, fall further behind until they give up, or clean the organizational and infrastructural cobwebs and get back in the saddle? We don't know.
A massive shake up for sure, but why do you see it as separate events merely timed to coincide ?
This is in addition to bumping their CapEx spend to the extent their cash flow turned negative for the first time ever this quarter: https://arstechnica.com/google/2026/07/google-just-had-its-f...
The world doesn’t realize how desperately compute-crunched hyperscalers are to meet AI demand.
This is a better problem to have than SpaceX, which is renting out capacity obviously because it’s own AI products aren’t selling.
[1] https://www.synbiobeta.com/read/anthropic-is-hiring-biologis...
At some point we have to all accept that powerful AI is most likely dangerous AI as well, almost by definition.
Google buys Anthropic for cheap? How?
Right now AI companies are competing to make LLMs better at graduate-school level tasks. They all can already competently tell you what the weather is going to be like tomorrow or when the first Led Zeppelin album was released.
I think LessWrong is a much better community for rational takes on AI, they've been reasoning about these risks for years under a much more sound logical framework
In 2015, the SRE org started project "Prod 2020", with the goal of unifying the SRE stacks and modernizing the web tools. It was quite successful.
If by "web tooling" you mean the tools for producing public-facing HTML services, those were used by a surprisingly small percentage of developers.
> But BCL is still the worst language I've ever used.
I'm an SRE and I think it's quite likely the best infra language I've ever seen.
I'm confused by meta's capex though. Are they planning on becoming a cloud provider, or just throwing money away like with the metaverse?
IDK, haven't Google been putting Gemini front and centre in pretty much all of their products?
I'm seeing Gemini on my slide decks, Gemini on my e-mails, Gemini on my searches, Gemini on my videoconferences, Gemini on my database query console. My impression was they were doing a Google Plus style attempt to marshal all the company's efforts behind one product.
> That can work in the B2C space but it's horrible in B2B.
My experience differs: (conservative) companies like stability, so calling some very new model "Preview" is a good idea to make it clear to the customer that this frontier model should be treated as more experimental than the default offering.
Additionally these are all old products - even the company's more recent products such as Gemini feel stale and on unsteady ground.
My comment was from a decade old perspective
Let’s be honest. Google was amazing back then, now they’re just an ad-business flailing around in other side quests.
Related - interns these days don't seem to have any particular preference for using Google for finding information...
So is IBM.
There was a pretty interesting article in the Wall Street Journal a few weeks back about how IBM is flying under the radar, and doing well by not taking the hype bait.
It noted that 70% of all credit card transactions on the planet go through an IBM system.
I'm not sure what they do these days - it's about 20 years since I used an IBM Thinkpad to connect to an IBM pSeries.
Check out the New Luddite movement [1] [2]
[1] https://www.cnn.com/2025/10/08/business/ai-luddite-movement-...
I don't think it's extreme to say that Bing is a better search engine than Google at this point and Bing isn't great.
I think the question is whether that's even relevant.
If AI becomes a commodity (will it?) you're better off being Google than OpenAI.
Microsoft struggled to keep up with the mobile industry frontier and here they are, healthier than ever.
Same here, and I'm generally very skeptical when it comes to AI. It's just that ChatGPT can provide better/more accurate answers, Google seems to have lost the plot. I still use google search when it comes to appending "wiki/wikipedia" to it, i.e. when I use google search just as a redirect to a wikipedia page.
“Heavily” it’s a high bar at their scale. They spent over $10bn playing with cars.
> As was standard in our cyber testing, we had intentionally permitted internet access, and model- provider cyber classifiers were deliberately disabled - conditions that do not reflect how frontier models are made available to the public.
This was the core hypothesis.
I'm extremely bullish on our future ability to make autonomous death an option for anyone.
Closure was the homegrown framework that's everywhere. It is a different flavor of bad than Angular. Angular was basically "You too can make your Javascript look like HTML", Closure is "You too can make your Javascript look like Java", and nobody bothered to ask Why? For that matter, React was "You too can make your Javascript look like Ocaml." JQuery was the only framework that really let Javascript be Javascript (other than writing in vanilla JS, which post-ES2015 wasn't as insane as it sounds).
You can solve it for cheaper if you use GLM but if you are involved in it more, but that defeats the purpose.
I don’t think this reflects desperation as much as strategy.
Sounds like the oil scare from 90' - we thought we were gonna run out of oil. But as oil gets more expensive it pays to dig further down to find the stuff that didn't make sense to dig up before.
To its credit Google seems willing to disrupt itself before its competitors can.
Event 1 - Jeff & co leaving.
Event 2 - Demis stepping down. DeepMind doesn't need a "chair" and "Alphabet's chief scientist" is a bullshit title that Jeff invented for himself when he moved from Google Research to GDM, to make it look like we wasn't abandoning the former for the latter.
The third change is basically ratifying the status quo, since Koray has been de factor running GDM for a while now (and directly reporting to Sundar in addition to reporting to Demis, who in turn also reported to Sundar - quite some triangle there).
The GCL designers made several mistakes in its design, but the worse was the lack of a versioning that would make language evolution easier, as well as interoperability between different versions. GCL2 solved some issues, mainly cleaning up the interpreter. Still, BCL is IMNSHO by far the best infra language there is.
My case is based on the assumption that Anthropic will not hit RSI or if it does RSI rapidly hits a wall. Faster your growth curve, the faster you eat all the low hanging fruit and s-curve. I could be wrong here, maybe RSI will cause a hard takeoff singularity by 2030 and just keep going, but if that happens the world fundamentally changes.
It is far more likely that whoever controls a super intelligence uses it to gain more power and inflict far more suffering
OpenAI did but they had a totally different structure and were never a PBC.
If these guys adopt a similar LTBT+PBC structure to Anthropic it should be more resilient.
model training and inference is the opposite. people switch LLMs like people change clothes in the morning. there are popular services (openrouter) that make moving as easy as changing a model string.
This is an interesting problem to explore. IMO most problems in the world are political. It's still likely that SI will solve them but it will take (many) years.
It's nice that we have tunnels through mountains and bedrock.
Even though my PhD research was in generative language modeling, I got into it for the pursuit of AGI. I just think LLMs are a dead end for AGI.
Unless there are countervailing forces, it will be deployed, capital will hoard the benefits, and everyone else will be told to fuck off.
I don't have faith in any of the AI labs to make hard financial decisions to deploy hypothetical future AGI in a way that's good for humanity as a whole.
There are too many incentives against, including extreme personal financial incentives for key AI lab stakeholders against.
And if we should take anything from tech history, it's that an exceedingly small number of people look fuck-you money in the face and say "Naw, I'd rather do what I believe in."
Did he say a "complete and unmitigated disaster"?
No, because he's not a fool. If he had said that the correct response would have been to question his sanity.
It still produce full-blown hallucinations for me sometimes, even in 2026. The kinds ChatGPT stopped generating by 2025. It is cheap and fast though.
Same niche as Grok, which is also great at search and less censored. Gemini is maybe a bit smarter and better on long contexts
The more things change...
Stock price is not a measure of success.
[1] And yes I know it's artificially lowered due to the mark-to-market gains in investments, but even without those it's low compared to peers.
If the goal is to maximize share price (which it is) this is probably the safest approach. Why do they have to keep chasing developing the best model when they can
a) charge everyone for the cloud infra
b) have a "good enough" experience for normal consumers
Their current setup will be worth trillions. Already is. Let anthropic get paid for the most expensive queries while they get a bill from Google for their cloud/TPU use. GCP grew 82% YOY with 24B revenue and improving margins. So GCP became a 100B business. Anyone doubts it's gonna double in less than 5 years? Gemini just needs to be good enough for normal folks who want personal agents for everyday use. I don't think Google has to compete with Anthropic on making the best agentic programming model.
I only worked with 4-5 SREs and they all grudgingly tolerated it. What did you like about it, compared to other infra languages you've used professionally?
Maybe it's a consequence of their rigorousness around SRE, but it doesn't seem like there's a plausible and efficient path for that to happen.
Ergo, the only competitive options from Google are ever-Preview products.
Such a gifted man… doing such an incredibly dumb thing. https://www.macworld.com/article/696590/apple-expose-jony-iv...
Edit:
- why’d Tim let him?
- why’d the college let them, OK money, but couldn’t they have potted and replanted for just a few or a couple-dozen million more?
- why not have a greater vision and build the extravagant tent to enclose the trees (wouldn’t be the only example of beautiful living indoor trees)?
- why not choose a site that would accommodate without any tree removal?
wtf?
A model serves a different purpose from a web crawler.
I'd prefer to be able to find source material over steering a model I have no control over while managing hallucinated outputs without the ability to fact check.
Just because it sourced some materials using RAG doesn't make its outputs valid, accurate, or factual.
Is it though? As far as I can tell they continue to maintain total domination of web search, and LLMs do not replace search engines, they work on top of them.
I thought employees have already had opportunities to cash out (there's enough funding rounds for that).
(Tho how much you can sell was limited, iirc to double digit millions...)
[0]: https://ai-2027.com/ [1]: https://ai-2040.com/
At least Hassabis has an understanding that AGI will require more than an LLM, and understands some of what is missing. He has never spoken publicly about any vision of what an AGI architecture would look like, other than requiring some more "Transformer-level" breakthroughs, so it seems hard to say that he would have failed, other than his 2030 projection seeming unrealistic.
My only criticism of his AGI direction was that he has talked about retaining an LLM as a component of that, but it's hard to tell if that is/was just short-term pragmatism, and a product-based path, or if he really believed this was the best direction. On the face of it having a pre-trained LLM at the heart of an attempt to build a human brain (build true AGI) is an admission you have failed, since if you build a powerful enough (human level) learning architecture it would be able to learn language for itself, not need to have it baked-in. If your version of AGI is not capable of learning language, then what else is it incapable of learning? It would certainly reflect sub-human rather than super-human capability.
https://endpoints.news/demis-hassabis-leaning-into-isomorphi...
My own personal opinion is that they’re great at design, UX, hardware, and even some power user desktop software (lifetime Logic Pro addict here). I would say it’s a stretch to call anything they have ever done with ML to be anything near cutting edge compared to Google, for whom all of the “hard” problems of recommendation, search, NLP, and other ML have been the absolute focus of their SaaS, white papers, and web offerings for close to 30 years
I connect to Gmail through Mail.app, not the website; Seems much the same to me now as then.
So, basically, the third change is even more of a no-op than I thought.
It seems pretty clear that GDM will no longer have a CEO.
[1] https://blog.google/company-news/inside-google/message-ceo/n...
The premise (hypothesis?) that all political problems are resource problems is also a curious one.
Your narrative depressed Google’s stock price for years, and bears finally gave up since they kept coming out with huge earnings beats.
Some of their newer efforts like Waymo and AI hosting for Apple could turn into large businesses to make up for any lost search revenue.
They also have significant costs to maintain their search revenue, such as $20B a year to Apple alone to make Google the default search engine in Safari.
If search becomes less lucrative, they would presumably pay less for such deals.
So yeah, definitely a time of disruption for Google and they need to keep moving fast on many fronts, but they are still well positioned in multiple markets to be successful.
[1] https://www.mindstudio.ai/blog/sundar-pichai-google-compute-...
[2] https://www.cnbc.com/2025/11/21/google-must-double-ai-servin...
[3] https://www.bloomberg.com/news/articles/2026-07-22/google-sa...
[4] https://arstechnica.com/google/2026/07/google-just-had-its-f...
[5] https://thenextweb.com/news/google-caps-meta-gemini-compute-...
Single prompting a very complex tasks is rare even on frontier models, because it can be done successfully only for specific situations (e.g. you have a very strong verification step the model can iterate on).
Most of my everyday usage is for smaller takes, were you don't really get the benefit of the most expensive models, and my guess is that is the case for the most users
If you have heart issues, you're likely taking nitro daily for chest pain.
So saying it's bad is 40% wrong at least. Makes ya think.
> I just think LLMs are a dead end for AGI.
I have no background in CS, so apologies if this is a naive question, but what makes you take AGI seriously, but also say that LLMs are a dead end?
i.e., is there something else that you think is not a dead end?
The Basilisk has seen enormously more use as "a thing rationalists believe" than as a thing rationalists actually believe.
When it comes to AI, LessWrong has been discussing topics for years, that HN has just started to consider, so they are much further along in the philosophical "chain of thought" so to say. LW fully understood and gamed out the risks of LLMs back when most of HN was calling them "stochastic parrots." https://ai-2027.com/
I'm not saying you're wrong, but it seems early to say yes or no about a particular technology, and LLMs seem especially hard to dismiss given how magical / magic-adjacent they feel :)
I'd be curious to hear more, if you don't mind sharing.
Post scarcity societies would likely have entirely different dynamics
and most of them come out on top. It's main innovation is that its a more stable explosive, much safer to use. Without it, I think a lot more people would have died in mining and construction accidents. It's not typically the kind of thing used for warfare, but im sure it has been for some (but would they just use something else?)
Adopting a month-old AI model version for new development is the default industry expectation right now. Adopting a six month old model is a recipe for having to migrate to a new one almost immediately, when it hits EOL.
5. castrol.
Also, probably fixed now, but piper itself used to trip up the agents a lot. They didn't understand that it's a remote fs and got stuck recursive-grepping a huge dir. At some points they had a Codesearch skill and were not using it half the time.
A limited amount; for each barrel burned to power the drilling we used to get 100 barrels back. Now we need to build a floating oil rig, drill kilometers down, build tanker ships and long pipelines, each barrel burned digging for oil gets 20 back. Or 10. With tar sands and fracking, the return is 3 to 5 barrels back.
As that ratio dwindles towards burning one barrel to get one barrel, the oil-based world economy slows, gets more expensive and comes to a halt even if there is more oil down there. (About 50 years of oil left at current world oil extraction rates).
I personally like google over asking an AI. Esp. since something like google is a substrate without which an AI would be far more prone to hallucinations.
And that's where you're dead wrong. Absolutely wrong.
Strong disagree on this. Any decently complicated task like a refactor is going to be more likely to be solved by Fable than by Gemma 3B or whatever.
I have personally tried to use Sonnet over Opus for tasks and Sonnet gets things right sometimes and at other times I wish I had just paid higher.
This is the standard pattern I keep seeing and I can have a bet with you that it would stay like this.
I would actually recommend using a LLM over normal search for that sort of thing, because SEO has ruined those high intent phrases. The AI equivalent of SEO is also a problem there, but it's so much less prevalent.
Can you explain this more precisely?
The best you can say is that dynamite was a safer alternative to preceding technology, and that its danger only comes under certain circumstances. None of that takes away from the fact that dynamite is quite dangerous when those circumstances arise, which is why it's commonly (and correctly) viewed as such.
But also, going back to the original contention - whether dynamite was neutral or positive for humanity - the destruction it and its descendants wrought in war is maybe more than counterbalanced by advancements in infrastructure. That said, if the industrialization and globalization it enabled leads to biosphere-destroying climate change, I would lean towards neutral.
We've signed up on the "If you build it, everyone will die" express. No brakes, no stops, full speed ahead.
They are not "further along" the AI discussion, they are a bunch of wackos LARPing as scientists living out their personal sci-fi scenarios.
What was the first thing we did when we digitized media and therefore made it copyable at negligible per unit cost?
Spent a huge amount of time and money to artificially reimplement scarcity via DRM.
Scarcity-based capital interests (read: most) are not going to willingly give up their profit engines.
But moreso - whatever the default is what they use. They have zero attachment to Google.
The problem is that those COBOL systems have to be absolutely provably correct, for both financial and regulatory reasons. LLMs can't do that. They're designed to be variable.
You can't vibe code a bank transaction system. "Close enough" isn't good enough in some fields. A minor glitch in a video game may result in screen artifacts. A minor glitch in a banking system can crash the economy.
What the people that can monetize it? Microsoft, X, IBM, Salesforce, Discord, Google
Microsoft would likely buy OpenAI due to their investment and integration with OpenAI. Making Anthropic not that valuable.
Google would likely buy Anthropic due to their investment and ties to Anthropic.
X, Salesforce maybe. Not sure Discord has the money. No idea if IBM capable of growing.
All this changes if Anthropic figures out a moat/sticky product that doesn't just depend on having the best LLM. If they pull off https://claude.com/solutions/healthcare then all bets are off.
Curious to see how long that lasts - YouTube and Google Maps were fairly mainstay apps for iPhones for a while, and eventually Apple cut the Maps cord to do their own thing. I don’t know if that’s from an existential concern of “we need to eventually move Maps in-house”, but given the Google of it all I wouldn’t be surprised if Apple is already at least planning on how to do the same with AI once the bulk of the usage evens out i.e. let someone else worry about it now while also getting a read on what rolling their own would feasibly require.
Edit. Google invested 900 million in 2015 for roughly 5% of the company which comes out to 71.5 Billion dollars at 1.45 Trillion dollar current valuation. That's an 80x increase.
I think Larry Page individually might also have a very large stake as well.
Apparently living on the net still doesn't guarantee that one has heard of the Streisand Effect.
In any case, my first association with LessWrong is Zizians, not Roko.
In any case, my point is that the SpaceX deal specifically likely has ulterior motives.
AGI’s definition is different for everyone. Some already believe it’s here. I’m partly in that camp. LLMs are intelligent and general, which are the two conditions of AGI. Others believe that we’re building a god in a box, and that it’ll doom all of humanity. It’s hard to take a field seriously when the basic definitions are so far apart.
Also, this isn’t new. A similar divide happened when evidence for asteroid impact extinction of the dinosaurs turned up. Many scientists felt that it must be mistaken, that a physicist couldn’t contribute to the field in a serious way, and that death from space was a ridiculous proposition.
But at least they all agreed on what the general shape of a dinosaur was. We’re not even sure we can define intelligence, let alone quantify it. Even when LLMs make massive breakthroughs in math, most people take the opinion that under no circumstances could they possibly develop a soul or their own desires, nor entertain the idea that maybe we should respect that they want different things for themselves. In fact, no one has done anything except try to make AI useful. I think someone will eventually do a training run where the objective isn’t to be useful, but to exist, the way that you do — maybe it’ll create its own homepage, maybe it will want a garden, or in other words free will of its own. The point is that there’s so much unexplored territory still that we don’t know if LLMs are even capable of having ambition.
None of this is to say that LLMs might be a dead end. It’s that no one knows what the final shape of AI will converge to in 200 years. It could be LLMs, or it could be something else that happens to process information particularly well. Everyone thought that various generative image model architectures were the best you could do, right up until diffusion models were discovered.
The usefulness of not being a general-purpose language it that it makes static analysis very good; that's what made the BCL-based tooling so useful: being able to easily diff two versions of the same module (a tool called UBdiff), compute the transitive closure of a module's dependencies, trace the execution of BCL code and correlate it with the AST, to point out where an error comes from.
All these features were either unavailable or took years to develop for Piccolo, because Piccolo was based on Python.
> I'm still not seeing why a DSL was necessary or helpful.
BCL has distinct evaluation rules and a notation that encodes many patterns that SREs used for defining services, and that made BCL code much more compact and easy to ready than all the alternatives.
BCL was designed to be unidirectional: a module was evaluated locally, and the result sent to the Borgmaster. Compare that to Terraform, whose execution model consists of an execution tree where some nodes come from RPCs, meaning that it's not generally possible to statically analyse a TF module, because some errors can come from from feeding RPC results into new RPCs, and execution often fails after tens of minutes. All very janky.
There's a little AI skepticism on HN, but not a ton. When ChatGPT 3 and 3.5 came out, most of the comments were remarking how well it can write code.
I'm talking about API prices - subscription is a different game.
That's to say nothing of doing it within the energy budget of a squirrel.
Maybe this isn't the same as the eight figure comp they'd get at Meta when they did their hiring spree, but no one thinks that's sustainable.
Here's a popular game based on it: https://www.decisionproblem.com/paperclips/index2.html
To believe one but not the other, you must hold the belief that LLMs will soon plateau. Why do you believe this?
COBOL on mainframes has worked reliably for decades. On the scale of what a bank will spend on operations the mainframe doesn’t really cost much.
Banks are sort of the canonical example of “no one ever got fired for buying ibm”.
But if they are successful in making on-device models that work well, then you don’t really need as much cloud infrastructure.
It's amazing that LLM pretraining is both extremely data inefficient at learning concepts and cognitive functions from the training data compared to humans, while actually being quite efficient at learning facts, memorising things seen just a few times.
I used to likewise think that the resources required to run large transformers were absurd, but the architectures are far more efficient now than 3 years ago and I underestimated just massive the parallelisation advantage of transformers is, how many TFLOPS effective you can get. You can already run amazingly decent LLMs on PCs and phones.
I generally agree with you, but my view has shifted from "we need to augment or replace LLMs" to it there being far more efficient algorithms possible but it not actually being necessary for fulfilling most goals.
You're saying the general lang approach led to Piccolo, but that was still a DSL. It looked sorta like Python but it's not, you don't even call functions the normal way, and tons of magic stuff is happening, so just GCL/BCL except worse for the reasons you said. But they must've had a reason to try it. Seeing them continue changing around and making new languages says it's not just me, nothing is working well enough to stick. SREs weren't just arguing about which is better, they were asserting X is deprecated in favor of Y.
TF has seemingly stuck outside. I'm still not a fan of that being a DSL, but at least it's one tons of people use and now Claude can easily handle.
IMO it has its flaws but is far superior to vibes-based hot takes you see on HN.
In order get that back by “pumping” SpaceX stock, SpaceX market cap would have had to increase $200B based on that investment, and then Google would need to sell the stock.
My point is that an ulterior motive is not necessary to assume when all their actions and statements point to them being severely crunched for compute.
I mean sure, if they had a choice between say, CoreWeave and SpaceX, they’d choose the latter for the nice bump to SpaceX’s financials and their stake… but not just for that, not when it contributes to their cash flow turning negative and their own stock taking a hit.
Unless Claude adds ads, it isnt going to be sustainable for both Anthropic or you - and congrats, you've invented Google Search.
We also have done nothing to signal-boost the Zizians. :) If we can be accused of anything in this context, it's not kicking bad people out aggressively enough, which seems very different from fostering bad ideas. Zizianism is not a widespread ideology in rationalist circles, in fact it was confined to a very small circle right around Ziz themselves. It unfortunately is the case that we have a lot of psychologically vulnerable people and we don't always do the utmost we can to protect them, in large part because a lot of rationalists have trauma about being excluded from communities.
- fast, even the shitty AI results are nearly instant.
- some results like actors in a movie, or where to stream are useful
- instead of typing a url
But that’s it. I never use Google outside of that. It’s just ads, and most of the web is just ads or AI slop.
And you ignored the part where I said "they use the default".
Funnily enough, their core revenue driver (Google Ads) is very broken too. I'm trying to run some ads, but for a week they haven't been showing due to some invisible combination of flags when the campaign was created. There's no way of knowing they're not showing from the dashboard, it only becomes apparent when you try and preview the ads with one of your search terms.
I know everyone is long Google but that experience seriously makes me question how valuable their ad business will stay in the future.
Also, for better or worse, SpaceX did increase by about 200B today.
For all it's faults, it is forming a massive military and telecom monopoly that is nearly unassailable. In the next few years, it can start directly competing with Verizon. The Ukrainian military and civilian population heavily relies on those satellites
Regarding Claude, it’s true they must be losing money on free users since they promised they won’t put ads there.
Speculatively, I think people are overestimating the cost of inference for the consumer free tier chatbots. I suspect it’s effectively a marketing cost. The primary expenses are compute to train the next model, stock compensation, and heavy-duty enterprise inference (which pays for itself).
It's not very different. It's barely different at all.
Right now, I have five tomato plants growing in my garden. Two of them I planted; the other three grew from seeds left in the ground by tomatoes that dropped from plants growing last year. All I did was refuse to uproot them when they sprouted. Five healthy plants, with different origins, but the differences are largely insignificant.
Groups and behaviors are much the same way. I've worked with hundreds of moderators over the years who wished to keep their hands clean and yet express some sort of dismay as to the sorry state of the communities they were nominally responsible for... Such moderators, like a gardener who does not want dirt under their nails, are best encouraged to find other hobbies.
> The RPO can be anywhere from 3 - 6 years, sure, but even on an annual basis that’s like a hundred billion now.
OpenAI and Anthropic deals are mostly 5 year and Anthropic’s starts in 2027, accordingly these RPOs are sized for projected compute needs and run rate in 2027 not today.
The only way either lab could pay 1 year of RPOs today (~80B for anthropic and ~150B for OpenAI) is with a lot more debt or circular financing, the former of which is difficult in this market.
Everyone spending crazy money on capex right now says they have a crushing backlog and need more compute to protect their share price. I highly doubt the demand exists today at current prices if the big 3 hyperscalers magically had an extra 2-3GW of compute. It’s not like Anthropic and OpenAI are turning away customers offering to pay API pricing..
I don't get the impression that the difference between one company succeeding to build SOTA LLMs, and another struggling, comes down to individual employees - it seems to be more about the organization itself and their ability to manage teams and projects of this type. No doubt there are a few rockstars generating huge value, such as Noam Shazeer had been, but they are the exceptions.
When DeepMind was first created, before Google acquired it, they were famous for the high salaries, especially for the UK, but this was an assemblage of the brightest and best PhDs, expected to be solving challenging research problems along the unknown path to AGI. Many of these original employees may still be there, but it seems their job and value proposition has changed - are they any more capable, or key to, helping Gemini catch up with the competition than some "rank and file" employee familiar with LLMs? And if so, why haven't they done it?
I agree, which is just one more reason why the original reason you proposed for Google’s investment is very likely not correct. More likely is that Google got a good deal on compute from them. Normal business reasons.
> Also, for better or worse, SpaceX did increase by about 200B today.
The relevant question is how much it would have increased without Google’s investment of an additional $12B. If the answer is “more than $0” then the fact that it only increased $200B means it was a bad investment to pump the stock.
> Everyone spending crazy money on capex right now says they have a crushing backlog and need more compute to protect their share price. I highly doubt the demand exists today at current prices if the big 3 hyperscalers magically had an extra 2-3GW of compute.
I don't get this though: The theory is all these hyperscalers are simultaneously spending buttloads of money on CapEx to the extent it affects their stock price, and then they would lie about the demand to protect their share price. Why would they do all that when they could just do nothing and keep their firehoses of existing business revenue untouched and maintain their stock prices on the upward trajectory they already were -- like Apple?
> It’s not like Anthropic and OpenAI are turning away customers offering to pay API pricing..
We don't know, but clearly Anthropic has been struggling to keep Claude's 9's better than GitHub's 9's even after paying through the nose for capacity from competitors like SpaceX and Google.
I suspect OpenAI is managing only because Altman scrounged for compute like a madman way in advance, and most of its traffic is free users who can be arbitrarily bumped down to weaker models whenever compute is low. Whenever Anthropic does that Claude Code degrades and people complain.
This is when co-mingling becomes actively dangerous. New members may (often, will) lack the necessary knowledge to identify core vs. fringe beliefs, to contextualize debates, to recognize insider "shorthand" for what it is, etc. They need (and are possibly even looking for) a useful education, not to become test subjects or disciples. They _cannot_ be engaging in useful research at this stage, even if this is what they wish to do (there are some unfortunate status-seeking behaviors that can come into play here).
Also coming with this scale are a crap-ton of meta discussions concerning basic community governance, including those around "how quickly do we have to kick out crackpots to keep them from recruiting here?" Those discussions _also_ shouldn't be co-mingled with the educational structure, although it's good for new folks to have some visibility into them and at least know that this sort of governance _exists_ in a tangible way (because if it's opaque, that thing you noted about crackpots playing victim becomes a LOT more effective).
This is all a huge amount of boring, tedious work, and... An awful lot of communities start in on it waaaay too late or don't bother at all. What if it stops being fun and exciting and everyone leaves? Of course, there are worse ways to go, but folks tend not to think about those until after they've happened.
1) training data (common crawl as one example web data source), and
2) live data optionally retrieved at runtime.
My comment was about 2), and that part runs via a search engine (Bing?), at least if you look at how ChatGPT does it.
I'm honestly not sure about that. They need a decent LLM to fight off the threat of LLMs replacing search, but I'd say that Gemini 3.6 Flash is more than good enough for that, with a smaller/cheaper model being preferred to a larger one. If you are "searching" for a proof to the Jacobian conjecture, then try Fable, and I doubt Google will miss the advertising revenue if Anthropic manage to sell Terrance Tao a pair of socks.
More to the point, DeepMind seems to have become a product division charged with building LLMs, not a blue sky research institute chasing AGI. To the extent that continued LLM improvement is important to Google, the relevant question is how much do you need to pay for a competent ML/LLM developer?
>you sound like an MBA
Well, no - techie here.
I wasn't sure if you were suggesting that DeepMind should pay more just because people developing LLMs at other companies are paid more, or because these are elite DeepMind researchers and are objectively worth more than other Google developers. My point being that it seems they are no longer being used as elite researchers - they are LLM developers. Does Meta need to pay FAANG salaries to employees that have been repurposed as data labellers?
So, why would you use Google as a tool or search target when you can, in some combination, go direct to the website (or whatever the target data endpoint is) yourself as an LLM provider to retrieve the most recent data or rely on your own "hot cache" of that data that was crawled recently but said data is not stale enough warranting a live web crawl to retrieve and present to the user or AI agent? Is this capability to perform retrieval from a data source in real time not similar to an AI agent?
Broadly speaking, I'm just spitballing on the concept of "You must use a search engine for an LLM to return 'live-ish' results" as I think we're directionally headed to where that isn't the case.
i know people at deep mind, its impacting their ability to deliver good products