Turns out he was just writing LLM prompts way ahead of his time.

And now I wouldn’t even trust them to understand the laws 100% of the time.

Sorting garbage is a terrible job for humans, but it's a terrible one for robots too. Those fancy mechanical actuators etc are not going to stand up well to garbage that's regularly saturated with liquids, oil, grease, vomit, feces, dead animals, etc.

[1] https://loe.org/shows/segments.html?programID=96-P13-00022&s...

An example: A friend worked on accurate built-in weighing machines for trucks, which could measure axle weight and load balance to meet compliance for bridges and other purposes. He found it almost impossible to make units which could withstand the torrents of chemical and biological wet materials which regularly leak into a truck. You would think "potting" electronics understands this problem but even that turns out to have severe limits. It's just hard to find materials which function well subjected to a range of chemicals. Stuff which is flexible is especially prone to risks here: the way you make things flex is to use softeners, which in turn make the material for other reasons have other properties like porosity, or being subject to attack by some combinations of acid and alkalai.

These units had NO MOVING PARTS because they were force tranducers. They still routinely failed in service.

Rubbish includes bleaches, acids, complex organics, grease, petrochemicals, waxes, catalyst materials, electricity, reactive surfaces, abrasives, sharp edges..

They are not saying "dont try" they are saying "don't be surprised if it doesn't work at scale, over time"

Humans can generally stand this without an issue.

In fact you wouldn’t replace a lot of jobs that involves this : doctors, nurses, emergency workers, caregivers…

It just happens to be difficult. But people love doing difficult things as long as it’s : a) rewarding, b) respected, and c) sufficiently paid

Just had cameras, visual detection, some compressed air nozzles, and millisecond (nanosecond?) reaction time to separate the non-recyclable materials.

The mapping between sensor signals and material types is usually hardcoded from laboratory test results.

The Gemini robot tech is cool as heck, don't get me wrong, but it doesn't seem particularly well suited to industrial automation.

https://www.youtube.com/watch?v=nUrBBBs7yzQ

1. https://www.pbs.org/wgbh/frontline/documentary/plastic-wars/

Get everyone to dump their crap into one pile and actually invest in industrial processes to sort the crap out.

Huge con: this is a complex problem with possibly poisonous/explosive ramifications if it goes wrong.

Huge Pro: If we can solve this issue, that is a society changing capability, forevermore.

Or until armageddon/robot overlords/singularity/zombie plague at least.

We are asked to do hundreds of little things that mildly inconvenience us in order to maintain some social contract. Sure they could be made easier/nonexistent with better technology, but I:

1) don't see why asking people to do their part is silly

2) don't see why this particular problem would be more frustrating than e.g. the others I've mentioned. I feel like they are all similar on the "effort" scale.

Although I guess I'd admit that asking people to sort recycling properly is very different than relying on them to.

That’s about 25% by weight of all that gets recycled in the country.

Metals, industrial scrap, and other sources are 75% of what gets recycled in the US.

We are blue collar businesses, with high labor costs. Many are exploring robotics actively for repetitive tasks. We have some robots in our process, looking for more when the ROI makes sense.

It may not be 100x, but there will be value in robots in recycling.

[1](https://en.wikipedia.org/wiki/PET_bottle_recycling)

(Also there seems to be some kind of video auto-play/pause/scroll thing going on with the page? Whatever it is, it's broken.)

The trick was in that the belt was too tight for an average human to put on with brute force, and disabling the tensioner or using tricks would require better than average mechanical skills their specially chosen 'random humans' lacked.

Reference video (saw your clip is robot-only, but the robot vs human video is more telling):

https://youtu.be/x-exzZ-CIUw?feature=shared&t=65

https://techcrunch.com/2023/12/07/googles-best-gemini-demo-w...

edit: it didn't.

Don't need the robot to smash a grape when we can use a fake grape that won't smash.

Time to think bigger.

I want to strap robot arms to paralyzed people so they could walk around, pick up stuff, and climb buildings with them.

Ehh, no need - just let the LLM figure out what to build in your garage.

Search is in real danger of mostly obsolescence. Ads aren't safe.

It's not just researchers. Engineers at Google get to spin up products and throw spaghetti at walls, too. Google has more money than God to throw around.

Google's ad dominance will probably never go away unless antitrust action by the FTC/DOJ/EU force a breakup. So they'll continue to lead as long as they can subsidize and outspend everyone else. Their investments compound and give an almost unassailable moat into deep tech problems.

Google might win AI and robotics and transportation and media and search 2.0. They'll own everything.

This might have been true 10-15 years ago. I assure you it is not the case today.

With a reasonable degree of success. In their last quarter (see https://abc.xyz/investor/earnings/) 25% of their revenue was non-ads, and that percentage has been consistently increasing.

That's just one of their many business units.

Can these spatial reasoning and end-effector tasks be reliably repeated or are we just looking at the robotic equivalent of "trick-shots" where the success percentile is in the single digits?

I'd say Okura and Vinci are the current leaders in multi-axis multi-arm end-effectors and they have nothing like this.

I assume Google is being very careful to keep the speeds well below the “oops, it took your jaw off” threshold.

Well, for now, at least.

I know who will be the first shown the door when the next round of layoffs comes: the guy saying "you can't make money that way."

When I see open roles at these companies I think the projects I'm going to work on in the future will be more and more irrelevant to society as a whole.

Anyway, this is amazing. Please delete/remove my post if it seems like this adds nothing to the conversation

The man behind the curtain here has an army of engineers, unlimited cloud nodes and basically has harvested all the data currently available in the world.

It doesn't get any better than this right now.

What's next? They'll ping you later on Linked-in with this awesome idea that you need to make sure runs in a $1 USD microcontroller with a rechargeable battery that is supposed to last at least all day.

The actual scary stuff is the dilution of expertise, we contributed for a long time to share our knowledge for internet points (stack overflow, open source projects, etc), and it has been harvested by the AIs already, anyone that pays access to these services for tens of dollars a month can bootstrap really quickly and do what it might had needed years of expertise before.

It will dilute little by little our current service value, but you know what, it has always been like this forever, it is just faster.

In the meantime, learn to automate the automator, that's the way to get ahead.

Everyone has to collect their own data and pool it together or else there won't be any progress.

And it won't ever get any worse.

But yes, in general, models won't get worse than they are now (or if they do, they won't stay that way.) At Google, search has been enshittified for business reasons, not technical ones.

What scares me more is the opposite of that: information scarcity leading to less accessible intelligence on newer topics.

I’ve completely stopped posting on Reddit since the API changes, and I was extremely prolific before[1] because I genuinely love writing about random things that interest me. I know I’m not the only one: anecdotally, the overall quality of content on Reddit has nosedived since the change and while there doesn’t seem to be a drop in traffic or activity, data seems to indicate that the vast majority of activity these days is disposable meme content[2]. This seems to be because they’re attempting desperately to stick recommendation algorithms everywhere they can, which are heavily weighted toward disposable content since people view more of it. So even if there were just as many long discussion posts like before, they’re not getting surfaced nearly as often. And discussion quality when it does happen has noticeably dipped as well: the Severance subreddit has regularly gotten posts and comments where people question things that have already been fully explained in the series itself (not like subtext kind of things, like “a character looked at the camera and blatantly said that in the episode you’re talking about having just watched” things). Those would have been heavily downvoted years ago, now they’re the norm.

But if LLMs learn from the in-depth posting that used to be prominent across the Internet, and that kind of in-depth posting is no longer present, a new problem presents itself. If, let’s say, a new framework releases tomorrow and becomes the next big thing, where is ChatGPT going to learn how that framework works? Most new products and platforms seem to centralize their discussion on Discord, and that’s not being fed into any LLMs that I’m aware of. Reddit post quality has nosedived. Stack Overflow keeps trying to replace different parts of its Q&A system with weird variants of AI because “it’s what visitors expect to see these days.” So we’re left with whatever documentation is available on the open Internet, and a few mediocre-quality forum posts and Reddit threads.

An LLM might be able to pull together some meaning out of that data combined with the existing data it has. But what about the framework after that? And the language after that? There’s less and less information available each time.

“Model collapse” doesn’t seem to have panned out: as long as you have external human raters, you can use AI-generated information in training. (IIRC the original model collapse discussions were the result of AI attempting to rate AI generated content and then feed right back in; that obviously didn’t work since the rater models aren’t typically any better than the generator models.) But what if the “data wells” dry up eventually? They can kick the can down the road for a while with existing data (for example LLMs can relate the quirks of new languages to the quirks of existing languages, or text to image models can learn about characters from newer media by using what it already knows about how similar characters look as a baseline), but eventually quality will start to deteriorate without new high-quality data inputs.

What are they gonna do then when all the discussion boards where that data would originate are either gone or optimized into algorithmic metric farms like all the other social media sites?

[1]: https://old.reddit.com/user/Nathan2055

[2]: I can’t find it now, but there was an analysis about six months ago that showed that since the change a significant majority of the most popular posts in a given month seem to originate from /r/MadeMeSmile. Prior to the API change, this was spread over an enormous number of subreddits (albeit with a significant presence by the “defaults” just due to comparative subscriber counts). While I think the subreddit distribution has gotten better since then, it’s still mostly passive meme posts that hit the site-wide top pages since the switchover, which is indicative of broader trends.

I'm just scared about a future where humans (say the next generation, kids 1-5 years of age right now) lack in-depth knowledge of almost everything and it's mostly AI writing low-level code, so there are no more "human experts."

We've already seen this happening where Gen Z mostly interacts with the world using phones and struggle with older operating systems/desktops, just like older generations. AI is going to make that 10x worse going forward.

Definitely believe we are reaching some sort of global maxima in terms of intelligence in our current structure of society

In 10 years, no trace of our current practices will remain in a recognizable form. It'll take longer than most of us think -- imagine how nonplussed Winograd and the rest of the SHRDLU-era AI gurus would have been to see how long it took to pull off the dice-matching trick in the video -- but when it does happen, it'll happen faster than we think. We're not yet at the tipping point, but it's close.

Isn’t that the ultimate goal?

But let's say it's accomplished. What will end up happening is that AI (should it work to the extent it's been hyper) will replace all the 'fun' jobs and we'll be left with either no jobs (and no income), or the most menial physical labor imaginable.

We'll probably spend all day consuming media and socializing. Thats the optimistic view of course.

While the ERP is boring as hell compared the the creative coding results, the latter is novelty and often has no intrinsic value.

Also, I see these videos and get deja vu of boston dynamics demos from years back. Not seeing anything new here except this is just early beta version of boston dynamic robots backed by different models.

Also amount of people around the demo set tells me, a lot of supervision and retakes happened. I often do not trust such demos(experience from seeing what goes behind the scene with cherry picked takes being published).

Anyways, my point is, just chill out. I remember how AWS/GCP/Heroku etc were eradicating IT admin but instead now we have dedicated DevOps and IAM specialist roles… and every day I see 7:1 ratio of job vacancy for DevOps:SWE.

The hardest part is data collection.

The math, well, basic signal processing means you know some algebra and differential calculus. Which is enough, unless, again, you don't want to prove theorems or invent new algos.

I'd be more worried being a junior front-end mobile or web dev.

I think where the real work will be is taking these models and creating real products out of them.

27 years later, 99.99% of trips are driven by humans.

Real-world robotics takes multiple decades to pan out. These demos are just that: demos. What you are seeing will not remotely impact your life before the 2050's, if ever.

What you should be worried about, however, is you (not your job) becoming irrelevant if you don't learn to write firmware using state-of-the-art AI tooling.

At the minimum: learn to work with Cursor (or equivalent). Make sure you work at a company that uses state-of-the-art AI tooling.

If you want to go further: learn to code (e.g. in python). Take undergrad/grad level courses in math, statistics and fundamentals of deep learning.

And FFS, chill.

[1] https://www.youtube.com/watch?v=2KMAAmkz9go

LLMs were a breakthough out of like a multitude of breakthroughs in AI in the past decade. I think there only needs to be a couple more breakthroughs in the next decade for it to come full circle.

Whenever I here a naysayer open his mouth it's like he's insulting a baby. Look it can't talk! no way it can ever program!

Either way it's not just you. The people creating AI are also as a side effect creating the training data for AI to replace them. So no one is safe.

One key point we must first understand, coding is NOT software engineering or even programming! Writing code is the last bit, a minimal fraction of the job description(unless you are actually indie dev or working for consulting firms). The core tasks include actually untangling the numerous vague requirements, understanding the domain, figuring out best approaches, performing various tests and checks, validating ideas, figuring out a cost effective solution, preparing rough architecture, deciding on an actual set of tech, align a hoard of people that everyone is on the same page, then start coding.

My IDE can already read my mind by auto-suggestion since many years and patterns/frameworks exist to reduce the amount of code I need to write. The issue is that, with these AI model, I just need to abuse my fingers slightly less. The other core duties are not yet solved and remains same archaic procedural everywhere in any/all serious roles.

And speaking of consulting firms, they are also clever and often has several implementations of same stuff, which they can modify a bit and sell for big money.

So in the end, people who jump into the pit because they are afraid of the juju mask are the prime target of the juju mask, for the rest of us, life goes on with minor bumps when the MBA comes to the desk and asks if it is possible to layoff few people to jack up the stock price this quarter yet… while subscribing to that new agenting engineer product suite for double the fees of what the laid off people actually costed, because their best friend in the golf club said the price will eventually become reasonable but the benefits are immediate.

People overestimate the changes that could happen within a couple years, and totally underestimate the changes that would happen in decades.

Perhaps it's a consequence of change having some kind of exponential behavior. The first couple years might not feel like anything in absolute terms, but give it 10 or 20 years and you'll see a huge change in retrospect.

IMHO, I don't think anyone needs to panic now, changes happen fast these days but I don't think things are going to drastically change in these ~2-3 years. But the world is probably going to look very different in 20 years, and in retrospect it will be attributed to seeds planted in these couple years.

In short I think both camps are right, except on different timescales.

Anecdotally the amount of hype and interest has been growing exponentially. This will push progress to a maximal pace. The next 10 years will be significantly faster then the last 10 years.

If you are a professional, please tell me about how much new code you write vs perform other stuff(meetings, alignment, feature planning, system design, benchmark, bug fix, release). For me, the ratio of coding:non-coding is around 10:90 on average week. Some weeks, only code I write are the suggestions on code reviews.

Consider the unmistakable trend: In the early 2010s, deep learning fundamentally transformed machine perception, image recognition, and natural language processing, setting new standards far surpassing earlier methods. By 2016, AlphaGo decisively defeated human champions, showcasing unprecedented strategic depth previously assumed beyond AI’s reach. Shortly after, AlphaFold solved the protein-folding problem, revolutionizing computational biology and drug discovery by rapidly predicting complex molecular structures. In parallel, generative adversarial networks (GANs) and diffusion models ushered in a new era of AI-driven image creation, enabling systems like DALL·E and Midjourney to generate strikingly detailed images, surreal artwork, and hyper-realistic visuals indistinguishable from human craftsmanship. AI’s ability to synthesize realistic voices and human-like speech has dramatically improved through innovations like WaveNet and advanced text-to-speech technologies, leading to widespread practical adoption in virtual assistants and accessibility tools.

Beyond imagery and voice, generative AI has also broken new ground in music composition, where models now produce compositions so sophisticated they are difficult to distinguish from professional human creations. Transformer-based models like GPT-3 and GPT-4 represent a seismic shift in language generation, enabling nuanced conversation, creative writing, complex reasoning, and contextual understanding previously believed impossible. ChatGPT further pushed conversational AI into mainstream utility, effortlessly handling complex user interactions, problem-solving tasks, and even creative brainstorming. Recent innovations in AI-driven video generation and editing—demonstrated by advancements like Runway’s Gen-2—indicate rapidly expanding possibilities for automated multimedia creation, streamlining production pipelines across industries.

Moreover, reinforcement learning breakthroughs have expanded significantly beyond gaming, improving complex logistics operations, real-time decision-making systems, and autonomous navigation in robotics and vehicles. The impressive capabilities demonstrated by autonomous driving systems from Tesla and Waymo further underscore AI’s advancing proficiency in real-world environments. Meanwhile, specialized large language models have emerged, demonstrating near-expert performance in fields such as law, medicine, and finance, streamlining tasks like legal research, medical diagnostics, and financial forecasting with unprecedented accuracy and efficiency.

These advances are not isolated phenomena—they represent continuous, accelerating progress. Today, AI assists with summarization, automated requirement analysis, preliminary architecture design, and domain-specific problem-solving. Each year brings measurable improvements, steadily eroding the barrier between supportive assistance and true cognitive engagement.

Your recognition of AI's limitations today is valid but dangerously incomplete. You fail to account for the rapid pace at which these limitations are being overcome. Each "core task" you've identified—domain understanding, requirement analysis, nuanced decision-making—is precisely within AI's increasingly sophisticated reach. The clear historical evidence signals a near-inevitable breakthrough into human-level reasoning within our professional lifetimes.

In disregarding this aggressively upward trendline, you're making the same grave error committed by those who previously underestimated transformative innovations like personal computing, the internet, and mobile technology. Recognizing current limitations without acknowledging clear indicators of impending revolution isn't merely shortsighted—it's strategically negligent.

Being able to order food and handle bureaucracy in these languages while speaking only English would be amazing. This seems like a simpler problem than tackling robots in 3D space, yet it’s still unsolved.

And many popular robotics demos are either controlled by humans or scripted, so it's useful to have the "Autonomous" label as well to clear up confusion. For example, I know a lot of people who thought the recent Unitree G1 demos were autonomous.

These models are nowhere near that for now, but I'll be watching to see if the big investments into synthetic data generation over the next few years get them closer.

They can be competent enough to cook meals in a controlled environment: (environment built for machines and specific dishes), without ever being able to replicate the same in a human restaurant.

There are a few companies that do robotic cooking, and it has its challenges due to the above reason. I am not aware of the cost problem though.

How cool would it be if this replaced someone at Subway though

1) tweak something like that to increase the likelihood of certain situations (abundance of a particular resource/object; frequent geographical feature), and

2) instruct your embodied AI to control a "player" model (to whatever degree of accuracy in articulation/mobility) to wander and perform certain types of tasks.

Perhaps it goes beyond the brightest minds at Google that people can grasp things with their eyes closed. That we don't need to see to grasp. But designing good robots with tactile sensors is too much for our top researchers.

All the best ideas are tried repeatedly until the input technologies are ripe enough.

But perhaps a great benefit of tactile input is its simplicity. Instead of processing thousands of pixels, which are passive to interference from changing light conditions, one only has to process perhaps a few dozen tactile inputs.

Ex-Googler so maybe I'm just spoiled by access to non-public information?

But I'm fairly sure there's plenty of public material of Google robots gripping.

Is it a play on words?

Like, "we don't need to see to grasp", but obviously that isn't what you meant. We just don't need to if we saw it previously, and it hadn't moved.

EDIT: It does look like the video demonstrates this, including why you can't forgo vision (changing conditions, see 1m02s https://youtu.be/4MvGnmmP3c0?t=62)

Those are realistically the 'natural' developments in the domain knowledge of Robotics/Computer Science.

However, what GP (I think) is raising is the blind spot that robotics currently has on proprioception and tactile sensing at the end-effector as well as a along the kinematic chain.

As in you can accomplish this with just kinematic position and force feedback and Visual servoing. But if you think of any dexterous primate they will handle an object and perceive texture, compliance, brittleness etc in a much richer way then any state-of-the art robotic end-effector.

Unless you devote significant research to creating miniaturized sensors that give a robot an approximation of the information rich sources in human skin, connective tissue, muscle, joints (tactil sensors, tensile sensor, vibration sensors, Force sensors) that blind spot remains.

For the inverse of the robot problem: younger me, spoiled by youth and thinking multitouch was the beginning of a drumbeat of steady revolution, distinctly thought we were a year or two out from having haptics that could "fake" the sensation of feeling a material.

I swear there was stuff to back this up...but I was probably just on a diet of unquestioning, and projecting, Apple blogs when the taptic engine was released, and they probably shared one-off research videos.

I understand that many people don't live in America and don't know how to use a coffee maker. That is 100% irrelevant. There is a frustrating tendency in AI circles to conflate domain knowledge with intelligence, in a way that invariably elevates AI and crushes human intelligence into something tiny.

[1] The hard part would be psychological (e.g. keeping the chimp focused), not cognitive. And of course the chimp would need to bring a chimp-sized ladder... It would be an unlawful experiment, but I suspect if you trained a chimp to use a specific coffee maker in another kitchen, forced the chimp to become addicted to coffee, and then put the animal in a totally different kitchen with a different coffee maker (but similar, i.e. not a French press), it figure would figure out what to do.

It also excludes corner cases like "what if they don’t have any filters"? Should the robot go tearing through the house till they find one, or do nothing? But what if there were some in the pantry — does that fail the test? There’s all kinds of implicit assumptions here that make it quite hard.

As for the point of corner cases being hard - I mean that's the point here, isn't it?

Hope to share the details here soon.

Not being a high profile target myself, I'd rather take that risk and see where it goes. Unfortunately it's the high profile targets themselves that make the decisions, so after the first few incidents I figure there will be this whole mess where they try to clamp down on access to such things without sufficient forethought.

Now, clean up the kitchen.

Fuck it, make the arms big enough and it can do laundry, load/unload dishwasher, clean up after cooking/eating.

I can finally see this happening. Probably Tesla first tho.

Also, I vaguely remember similar demos without the AI hype. Maybe it was from DeepMind, or another upstart back in 2015.

I expect they are more honest than the Telsa men in suits debacle, but my trust is low.

What do we know to be the facts?

I assume google made the choice of selling the "brain" for any "body" whoever develops it. Something like android.

https://news.ycombinator.com/item?id=43331358

This is a great achievement and I'm not underestimating the work of the people involved. But similar videos have been put together by many research labs and startups for years now.

I feel like Google's a bit lost. And Sundai's leadership has not been good for this, if we're honest.

GOOG is around the same price as it was in 2022, which means the AI wave went by through them with zero effect. With other tech companies doubling/tripling their market cap during this time, Sundai really left 1 trillion of unrealized value on the table (!); also consider Google had all the cards at one point, quite mediocre imo.

You don't have to release a profitable product, but to compete over the next several decades you are going to need to own valuable land in remote territories where patent wars being fought today. I'm guessing Google's meta-strategy is a type of patent-colonialism.

I see, are you a VC in the valley?

Even after the massive total market correction in the last few weeks, the earliest that GOOG was the same price as today is not even a full year ago. In fact, it's up 90% since 2022.

Any stock market source would tell you GOOG was ~140 USD at the start of 2022. Today it is ~170 USD. A 20% increase over three years, about the same rate as inflation and S&P.

This is extremely trivial to verify. Was this written by a GPT bot?

It's just the up and down of the entire market (and these big techs dominate S&P 500). I don't think that actually indicates anything.

But I did specify "start of 2022".

Is this another bot account?

I guess my broken English doesn't match today's bot quality :)

I would love to experiment with something like this but everytime I try to figure out what hardware to do it with there's a thousand cheap no-name options and then bam 30k+ for the pro ones.

However, when I looked at the BOM I was surprised that the actual arm they use is an incredibly expensive off-the-self arm https://www.trossenrobotics.com/viperx-300

For a much cheaper option take a look at https://github.com/huggingface/lerobot (this is an AI training library/framework) which uses the SO-100 arm https://github.com/TheRobotStudio/SO-ARM100 (one arm is $123). See: https://www.youtube.com/watch?v=n32OmyoQkfs

There's also the Parol6 arm, which is more performant than the SO-100, but more expensive: https://source-robotics.com

Uhhh, I mean that's nice, but how about: "That's why we will never sell our products to military, police, or other openly violent groups, and will ensure that our robots will always respond instantly to commands like, 'stop, you're hurting me', which they understand in every documented human language on earth, and with which they will comply regardless of who gave the previous command that caused violent behavior."

Who is building the robot cohort that is immune - down to the firmware level - to state coercion and military industry influence?

1.) Has cutting edge in house AI models (Like OpenAI, Anthropic, Grok, etc.)

2.) Has cutting edge in house AI hardware acceleration (Like Nvidia)

3.) Has (likely) cutting edge robotics (Like Boston Dynamics, Tesla, Figure)

4.) Has industry leading self driving taxis (Like Tesla wants)

5.) Has all the other stuff that Google does. (Like insert most tech companies)

The big thing that Google lacks is excitement and hype (Look at the comments for all their development showcases). They've lost their veneer, for totally understandable reasons, but that veneer is just dusty, the fundamentals of it are still top notch. They are still poised to dominate in what the current forecasted future looks like. The things that are tripping Google up are relatively easy fixes compared to something like a true tech disadvantage.

I'm not trying to shill despite how shill like this post objectively is. It's just an observation that Google has all the right players and really just needs better coaching. Something that isn't too difficult fix, and something shareholders will get eventually.

Google as a whole has a long history of not being able to successfully build great products out of great tech. That seems wrong from looking at search, Gmail, Maps*, Docs*, etc., but I think these are cases where a single great insight or innovation so dominated the rest of the product qualities that it made the product successful on it's own (PageRank, AJAX, realtime collaboration). There have been so many other cases where this pattern didn't hold, and even though Google had better tech, it wasn't so much better on one axis as to pull the whole product along with it.

That's the problem I see here. Maybe they have a better model. Can they make it a better product? OpenAI and Anthropic seem to ship faster, with a clearer vision, and more innovation with features around the model. Is their AI hardware acceleration really going to be a game changer if it's only ever available in-house?

I do believe in Waymo, but only because they've been incrementally investing and improving it for 15 years. They need to do that with all products, instead of giving up when they're not an instant hit.

*Maps, Docs, and YouTube were acquired with their key advantages in place, so I wonder how much they even count.