Google’s ATLAS is a look at how people are using AI | Hacker News Reader

Google’s ATLAS is a look at how people are using AI

blog.google

81 pointsswolpers149 comments

For example, automotive technicians and industrial mechanics use AI to interpret complex test results, debug electrical wiring, and inspect machinery for wear.

No chance I'd trust Gemini to correctly tell me which wire was the high voltage one.

You're absolutely right! I apologize for providing a factually incorrect explanation. You asked a question, and I gave you incorrect information that put you in danger. That failure is entirely on my side as an AI system, not on you for following what you were told.

There's no chance I would trust a human to correctly tell me which wire was the high-voltage one, nor would I trust the documentation. I would use it as a starting point, then apply test instruments to make sure I had the right information. Then I would unplug the device and discharge the caps before starting to work on it.

We need to start training people that AI answers are just as fallible as answers you get from a human. Trust but verify for all beings, biological and non-biological.

They're not even answers, is the thing.

It's like "[actual living person] has big eyes" and "character portrait in a 90s video game has big eyes" are similar sentences on the surface, but they describe things as far part from each other as a single dot left by bird poop vs. the combined writings of all humans, ever. Like half an atom compared to a universe.

Same for "beings". Just no. Or rather, I apply what you suggest to what you said, I don't just take your word for it but inspect if what you said could be true, and realise your answer is so full of false premises it's not even wrong.

> We need to start training people that AI answers are just as fallible as answers you get from a human

You're right, BUT, humans don't risk to take responsibility if they clearly don't know the answer, they would at least tell you: I don't know, let me find someone who might know or X, Y, Z might be able to help you.

AI: Bro, trust me, this is 100% accurate, I give you my teeth if you find it wrong. ... later on ... You're absolutely right, I was wrong, lets try another approach

> BUT, humans don't risk to take responsibility if they clearly don't know the answer

Where do you live?

People say the craziest shit about topics they are completely ignorant of. Even important things like medical advice, and not just internet pundits or crappy journalists.

Even people that care for you (friends or family) can give terrible advice. Saw it yesterday with one friend advising another on a legal matter that I thought obviously demanded a lawyer and a strategy. Another suggested some bad job seeking advice.

Personal advice is the worst, and highly important.

There's an art to finding good experts. And an art to filtering what anyone says.

AI psychosis.

No, LLM generated answers are not the same as those from a qualified expert. Of course you should trust a human if that is their job and you should never trust a word generator to give you accurate answers to an important question.

Like the term "slop", misuse of a pop psychology term like "AI psychosis" is another block to critical thinking and a casual dismissal of a conversation.

You're right, LLM-generated answers are not the same as those from a qualified expert. Sometimes they're better.

Case in point. I recently had a discussion with a friend of mine about going off propranolol. He wanted to stop it because he didn't like the side effects and only wanted to take it as needed for anxiety-related problems in public situations. His primary care said to just stop taking the drug. Gemini told him that you need to have a tapering protocol, and you need to work it out with your doctor.

In this case, his primary care, theoretically a "qualified expert," was wrong, and Gemini was right. His course of action at that point was to get a new primary care that had better knowledge that matched the vetted medical sites I pointed him to.

fwiw, he trusted my advice and validation of his plan to get a new primary care doctor. The reason he trusted me is that he's seen the prescription drug disasters I've gone through, and knows how deeply I've had to educate myself as a form of self-defense.

Great example, but not for the reasons you think. You have zero qualifications to make that judgement call. You don’t know if the standard protocol was right in this instance or inappropriate (depends on dose side effects etc etc) and I certainly wouldn’t go anywhere near Gemini to try to verify that.

Please stop relying on word generators for medical advice, period.

Yes, you don't rely on an LLM solely for medical advice. You use it to get orientation to look at other sources. In my friend's case, I was relaying to him what I learned when taking a similar drug, my doctor's cautions on stopping the drug, what the manufacturer says about stopping the drug, and the published dosing schedule.

I guess you're right about me having zero qualifications. I only have a lifetime of experience with treatments for the various conditions I live with and the dozens of drugs that have failed to treat these conditions. I've been forced into self-education (learning how to read the research papers and the manufacturer info sheets) just like any patient with a complex medical situation needs to do for self-advocacy. This knowledge has helped me catch doctors' prescribing mistakes. I wish I'd had the skill before one such mistake toasted my kidneys.

Would the same doctor also casually give you advice on suicide, confirm the suitability of a piece of rope, and all that? You cannot cherry pick like that.

When you talk of "answers" in the same category as humans give them, of "beings", you already are outside the realms of critical thought.

> I wish I'd had the skill before one such mistake toasted my kidneys.

I wish the people who died from bad LLM advice were still alive. Many of them had their life still ahead of them. You can't guilt me into confusing generated text with answers while ignoring all that. I also have pain and suffering in my life, one aspect of it are these attempts to devalue humans by putting probabalitities of words humans produced on the same level as them, calling them "beings" and all that. If I could remove all intellectual dishonesty from the world and have some pain in my right hand side diodes in exchange, I'd take that bullet for y'all. So I'm not impressed by this tear duct jerking.

> Please stop relying on word generators for medical advice

Ironic that you are giving medical advice.

You might as well say don't read books.

This is not medical advice, it’s advice on living - in most cases, trust that experts in a subject know more than you and have more context than you.

We'll see. I'm guessing you'll have to accept a very different world.

We're seeing, right now, that the people pushing for it are doing it in ways just as you do. Maybe you'll have to accept being considered a second class slop human, who knows? Or do the threats only flow one way?

If you see what I said as a threat, I'm afraid it shows more your state of mind, than anything I've typed.

maybe an LLM can explain it to you, just give it the context as well

In safety critical jobs nobody is trusted, everything is verified. And there is someone checking the checker.

This is not always true - e.g. gas engineers and electricians do safety critical work every day without supervision, and we trust them to do so because they are certified and they are held responsible for failures.

The same standards of care and responsibilty cannot be applied to LLMs, and asking one about electrical safety or medial issues is very similar to googling and trusting the results.

I get that people here want to believe, but the op is far too trusting of LLMs and using them in domains where it is highly inappropriate.

> Of course you should trust a human if that is their job

I’m telling you, in industries dealing with high energy, they DO NOT trust people just because it’s their job. THEY CHECK. Just like people should CHECK the LLM. I think we agree that LLM cannot be trusted. But where we don’t agree is “trust the experts because they’re humans” - that’s not an argument it’s a troll.

> gas engineers and electricians do safety critical work every day without supervision

This simply isn’t true. Electricians work is inspected before being placed into service. Engineers work is reviewed internally and permitted externally before going to construction.

It certainly is true, from experience.

You’ve never actually hired an electrician or gas engineer to work on your house have you?

Also it’s not trust the experts because they are humans but trust the experts because they are experts.

> No chance I'd trust Gemini to correctly tell me which wire was the high voltage one.

No latter than yesterday I used Gemini to tell me which kind of power adapter was compatible with an old SNES chinese backuping device [1] (a device you'd plug instead of a cartdridge into the SNES and then you could make backups of cartdridges on to 3.5 floppy disks. You could then run the games from those backups. Conveniently you could also "backup" your friends' cartdridges.. Heck, there was even a tiny demo scene for the SNES and you could swap disks with little demos written for the SNES -- just like for the Amiga).

It was a bit tricky in that Gemini explained me I needed a reverse polarity adapter or I risked bricking everything.

Now, granted, the stakes weren't very high but... Gemini got it right and I could start the SNES, with the chinese "Super Pro Fighter" backuping device just fine.

[1] this thing: https://videogamedevelopmentdevices.fandom.com/wiki/Super_Pr...

OK… but… how did it “know”?

If there’s no way to verify the data (e.g. look at the photo of the power brick it unearthed somewhere) then there’s a good chance it just synthesized an answer.

I’m constantly surprised at the breadth and depth of information these systems can retrieve and present in digestible format

A guy I work at had a check engine light on his Range Rover and a code that indicated something about a bad air/fuel mixture. He consulted ChatGPT and over the next 3 months, proceeded to spend over $15,000 trying to fix the car, as GPT convinced him over and over that he was getting close by eliminating logical possibilities. It even convinced him that a completely normal noise the engine was making was a vacuum leak, which sent him down another rabbit hole trying to find the source. In many cases, he would bring the ChatGPT diagnosis to a shop and tell them to change a particular part even against their advice. None worked. Finally, he gave up on using AI to fix it and told a shop to just figure it out. Turns out the injectors just needed to be removed and cleaned. Check engine light is now off.

AI psychosis is real. Dang

Yeah it's definitely some kind of delusion/gullibility exploit (probably mixed with having too much money).

The guy isn't dumb, but he is obsessed with solving every problem with AI.

This person seems to be intentionally making themselves dumb by choosing not to think.

I get the feeling he really wants the AI hype to be true and he's acting in ways that engage his confirmation bias.

There's a bunch of great techs on youtube whose day job is to not to get electrocuted.

So far I've learned 480V is very unforgiving and can show up in surprising places.

Very true - reminded me of this demo* as an example... on the surface looks trivial but potentially life saving.

* https://youtu.be/knSkE9ejPl8?t=112

I suspect they’re also lying about the utility there as well.

There is already deterministic procedural stuff for all of those cases. Putting a chat bot in front of that is basically fucking stupid.

But never underestimate the stupidity of an AI psychosis- suffering product manager to shoehorn it into the most batshit insane use cases in search of mOrE ENgagGeMeNt.

That's not how AI or any kind of assistance is to be used. It's there to double-check and alert you to any possible mistakes you've made. You still have to use your own judgement. Have you never asked a person to take a look at some of your work? Did that mean that you have to follow and obey all of their advice?

Sounds a lot like "You're holding it wrong."

No, that what it means to be an adult. Blaming others for your mistakes is something we have to grow out of to become adults. Sadly, at least half of people never grow to be adults during their entire lifetime.

Unfortunately this is grounded in a perfect world. People are prone to believing things that “sound” believable and AI is very good at producing just that.

I've never worked with professionals at such a low level of intelligence, and I don't think a person will even be able to get their license as an electrician if they are that unfamiliar with their own work. The world doesn't have to be perfect for people to be better than making the wrong decision every time possible.

I wouldn’t replace best practices for safety like lock-out tag-out and hazard labeling with an LLM but you do you

I asked it to identify a starfish I photographed at the weekend whilst scuba diving. It's a spiny starfish, off Cornwall. It said it was a 7 armed starfish. It clearly had 5.

Disappointing. I wouldn't trust it for "real" problems at all!

Even the best AI tools are consistently wrong at interpreting the NEC and often in odd ways, like making it stricter than it actually is, or mixing multiple countries’ standards together. 100% useless in an AHJ with a lot of their own rules like Chicago/Cook County.

Asking an llm to make a code interpretation is like asking a tape measure to drive a nail. An LLM will be very good at finding relevant code, maybe even explaining its intent or history. It will not be very good at weighing competing priorities, that’s something a human needs to do.

Understanding the AI Economy, brought to you by Google, one of the companies with the largest vested interests in the AI economy.

And yet, it seems like Google doesn't understand the AI Economy :)

But it doesn't prevent them from pretending they do Y)

They release real numbers from their practical expereince running on of the biggest ai based agents on the whole planet.

You wrote a snarky comment without content.

Would love to actually have a real discussion about critisism like this (not because i'm affiliated with google at all, but because I like to expand my knowledge through discussions)

Please engage in good faith.

Google:

1. Is not remotely unbiased. 2. Does not have desirable models, coding harnesses, or data centre capacity for AI (they rent space from SpaceX). 3. Seems to be lagging far behind in all things AI. 4. Has serious problems with internal adoption of AI tools (such as banning most team members from using competitors, but allowing the DeepMind team to use Claude, etc.)

Their opinion from their PR team is not just that credible or valuable.

Of course a company is not unbiased. I don't blindly just take the information of an article, i add it to my signals and evaluate it.

But the dismissivness of a normal / small blog post is weird. You can't just complelty ignore and dismiss one of the biggest companies on the planet who is also one which pushes research of it too (aka deepmind)

2. They do have gemini which I don't use for coding but I use daily for other things, it works well. They have massive data center capacity, that they are now renting space from SpaceX shows how big the demand is.

3. yes weirdly enough but doesn't matter too much, Google at least can just afford the investment in comparision to any other company playing that game. My assumption is that on one side google is just bad in releasing products (antigravity vs. gemini cli vs. unable to buy AI tokens as a normal user etc.) but also that they might just have a broader field were they want to use AI.

Claude/Anthropic can focus on coding a lot to push that frontier while google might focus more on gemini assistent and also for coding and using gemini for GCP etc.

But doesn't matter still, google was fundamental for research into LLMs.

4. No clue how this is a problem relevant to the AI Economy. Deepmind might be more independent from Google itself, Google has an internal AI coding system before the LLM stuff became popular so Google has a clear incentive of keeping the feedback loop internal. Independent of this, google is well known to control their whole ecosystem starting from the mainboard firmware. Im not even aware of any other company on the planet doing it like this.

For me its valuable enough to have a quick look. I would probably not have posted it on HN or voted up.

I'm more surprised about the generic negative sentiments people communicate here instead of having nice discussions with value.

The study uses Gemini data to draw conclusions about AI usage for economically useful work. Hopefully it is obvious that this is a fatally flawed approach due to inherent bias in limiting your data collection to one model which is uneven in its capabilities (as all models are) and therefore biased in what work people use it for. I might as well write a paper about the economic uses of power tools by analyzing how drill presses are used.

"Less than 10% of those interactions fully automate tasks." - that's because they are measuring use of Google's models. People are using Antrophic's models for automation and coding. My wife is from a tribe in the mountains where they live like they live off the land, cooking with fire like they did 1000 years ago. She uses Gemini to learn how to live in the city - book flights and hotels, stuff like that. I use Claude Code every day to automate my tasks. Google has a very different type of user than Antrophic and Open AI.

Automation is also a squishy term I’ve been playing around with Hermes agent the last few weeks, and I found it very difficult to find useful things that you can do by itself. And even at work, most of my AI usage is rather interactive other than research task that I can send fire and forget.

The full 100 page report: https://ai.google/static/documents/GoogleATLASv1.pdf

These AI threads seem to attract the same kind of repetitive, low-effort dismissals that crypto threads used to.

Crypto failed and pivoted. One still cannot pay bills, taxes, buy medicine, transportation, etc. directly with crypto. It is fair to assume the same story will happen to AI.

It's not fair at all, actually. They are two very different things.

Pivoted where?

Buying drugs, using prediction markets, sending Palestinians money without The Man finding out.

So it pivoted from sending money to sending money...?

Again, it isn't money if it isn't generally accepted. Crypto was forced to pivot into an exotic investment thing that works great if you already have "regular money" to convert and spend. It is not money.

No money is generally accepted. USD would come closest but it's still far from true (probably fewer than 10% of merchants globally).

I'll bite: it pivoted from replacing fiat currency to doing mildly-to-wildly illegal shit

Low-effort dismissals that crypto threads turned out to be correct. Is that what you meant to imply?

No, I didn't mean anything beyond what I wrote.

I don't disagree but one difference I see is Enterprises pushing hard to use AI where none of them really bothered with Crypto. I see momentum with AI that I haven't seen since the push to Cloud infrastructure 15 or so years ago.

That's because it actually works, so the dismissals are doubly silly in that they fly in the face of trivially observable reality.

Crypto didn’t deserve much effort.

And they will be proven wrong, like with crypto... oh wait.

Crypto, block chain, LLMs. All wealth concentrator scams with low to no economic feasibility. What do you expect? Everyone to delude themselves?

Never ask a barber if you need a haircut.

I share the cynicism here about this article. I think the saddest part of betting our economy on AI is that Google is actually the US-based corporation that is best positioned because so much of their AI efforts are not specifically aimed at software development. At least for now their profit margins are high enough to withstand the high AI costs, but let's wait and see if that lasts.

In other words - "We don't like the reporting around how bad the economics of AI are, so we're cherry picking our own data, presenting it out of context, and asking you to look at it through rose tinted spectacles."

It's not a report about the economics of AI. It's a report on how people are using Google's AI. They don't mention anything about the supply side.

I quite like that. Companies like Google, Microsoft, and Apple are simply not in a position to compete. They are always late to the party, and it is shameful. It shows that their managers are not capable of leading innovation. These companies need to go back to having real engineers drive innovation.

They're keeping a foot in the game with their own offerings, but I believe they're biding their time to see who to buy.

Knowing where the ball is going to be is a rare skill, and none of the people currently running those companies has that foresight. There’s a reason why there’s many QBs but only one Tom Brady.

I don't see a reason to believe any CEO, past or present, has had the ability to accurately predict the future. Luck and foresight look the same when you happen to get it right.

Catching one trend, like Mark Zuckerberg, is just luck. But there’s clearly CEOs who are able to predict the direction of markets over a sustained period of time across multiple inflection points. E.g. Bill Gates and Steve Jobs.

I would still treat this like John Bogle treated mutual fund managers.

For every Steve Jobs there are a hundred Carly Fiorinas

Eh, even if we take those mythologized examples as real (I'd argue the history of both is heavily whitewashed and their failues, like NeXT or Windows Phone to name just two, are papered over in their mythmaking), if anything they're the exceptions that prove the rule.

Apple tried to create a new OS and failed (during Jobs's absence from Apple), bought NeXT, then used NeXT's work as the basis for the very successful OS X. And despite this, in your mind NeXT is a failure?

NeXT itself, as a company, was absolutely a failure. That Apple came along and bought them and stripped them for parts is certainly convenient for Jobs, but doesn't change that, had they continued on their own, it would've ended up in the dustbin of history.

TBH I'm surprised you're arguing. NeXT is generally accepted as a commercial failure. Hell the NeXT cube, alone, was a famously collosal flop. This isn't even a little bit controversial.

I don't see any mention of shareholder value or financial performance upstream (unless we count your latest comment). I do see, "their managers are not capable of leading innovation," and, "I don't see a reason to believe any CEO, past or present, has had the ability to accurately predict the future".

Jobs cared more about changing the world and advancing civilization than he cared about providing good value for investors. NeXT was a very solid success when measured by the yardstick of what Jobs set out to do.

Trying to make money by creating a new general-purpose operating system is inherently a very risky enterprise. NeXT's investors knew that.

"google is playing a different game" is the most ridiculous cope these ppl have come up with to justify their BS leadership positions.

Demis is busy create another PR documentary lionizing himself so he can capture the google crown. Always says whole bunch of nothing to a fawning interviewers who is projecting brilliance and profundity on his every word.

To an extent, but their main points match my own perspective:

-AI use at work is broad but shallow: -At work, most AI use is focused on collaboration and assistance with tasks, and so far task automation is uncommon:

Both of the above happen for a reason and if you work at any bigger corp ( esp publicly owned ), it is easy to understand why.

edit:

What I did find more amusing though was the coded cry for help ' To get there, we as a society must work together".

I wonder if there would be more task automation if providers didn't ban you for automating tasks.

Also if all the built-in AI buttons could actually automate tasks instead of just telling you how they're unable to automate tasks.

:D you just noted common issue #1

>"we as a society must work together" - this is really rich. do they say the same about sharing the "benefits"

My startup is using AI every day.

My company (not my startup, my normal day to day work) spends millions on ai token every month.

I use AI every day for different things.

AI/GenAI disrupted already a lot of peoples jobs in real life.

And Google states very early "However within jobs, people are using AI selectively: in a typical job AI is used for only ~21% of tasks."

Did you read the article? Can you point out specific things you really don't like and think they are cherry picking?

I don’t agree with the OP, and don’t think data is cherry-picked. But everyone is so deeply invested in this, that no article from any company can be taken as is.

Reading the post, I had to keep thinking “why are they only talk about how people use it at work, and nothing about possible economic outcomes?”. Or “titled Understanding the AI Economy, but it avoids all the big questions”. Or “is it written for the investors who think Google has over-invested in capex?”. And I might be overthinking about all this! But when you have billions of billions dollars invested, it’s hard to take any word as is.

You always have to make your own conclusions of course but it doesn't mean that it is useless or just wrong.

What value does Google has to cherry-pick the occupation metrics?

They want to present AI as a growing and vital service so that shareholders think they're making good investments, right? Whether that's true or not, you know Google would always be putting the best appearance on it.

I did say explicitly that I think, Google did not cherry pick any data. My main question is “are they trying to build a narrative, because of the growing anti-AI tendency of the general population?”. It feels like every big co is giving up on big questions (e.g. what happens when AI replaces more and more jobs? what happens to education system?) because searching for an answer to those questions was upsetting a lot of people. So, now it feels like it’s just “good and neutral stories only”, so I have to raise my eyebrow on each post.

> It feels like every big co is giving up on big questions (e.g. what happens when AI replaces more and more jobs? what happens to education system?) because searching for an answer to those questions was upsetting a lot of people.

They were not searching for the answers. They were trying to get more investment money on the promises to investors about education and jobs displacement. Those are two different things.

I genuinely believe in the early days they were looking for the problems. At least the people I know personally did so. Openly and loudly as well. Some people genuinely like to solve problems!

But again, now it’s more of a “aight too much money on the table so maybe not real”.

Where’s the data for one, do they include the force fed Gemini in search? They are also pushing it harder than Microsoft was pushing copilot in every single of their products, sometimes multiple competing implementations (hello GCP dashboards).

It simply comes off as a marketing advert for a company who’s sunk billions upon billions into something they’ve yet to be able to monetize (see latest report). Stock holders are getting annoyed, like in other companies, about the lack of the promised returns, this is firefighting.

Nobody is really debating AI has a use case, but maybe, just maybe, the chase into infinite spending we’ve seen the last years wasn’t that well thought out.

I never seen google blog articles as something millions of people read. For me these are small lenses in a big company were i pull out small details i might find interesting.

Im missing the marketing opportunity here.

The spending is a complete different topic though. Google/ABC can easily afford their investments. Microsoft and Amazon too. Even if shareholders are not that happy (the shareprice of alphabet doesn't tell this story though) with it.

Nonetheless the biggest issue with this spending are 401k and other retirement investments from people who can't control the investment and the undefined risk of it. Thats the only real 'risk' (not a small one for these people depending on it though).

It doesn't affect me if Google spends its money on AI.

Nonetheless the progress on AI is huge, it still hasn't slowed down, its now a political and global issue (us export restrictions, Chinas Moonshot AI/Deepseek, EU lagging behind).

At least form my point of view, i'm very surprised that we even risk all of this investment and see it as a possitive thing. Imagine AI/AGI is a technology which sits very far away from our current local minimum and the jump to the other local minimum is massivly expensive but massivly beneficial?

GPT-2/3 was able to inflict this vision into people and because of this, we get A LOT more compute, push again boundaries on compute and memory after a longer period of stagnation. These richest companies in the world like Google also leverage the AI spend into investment of fusion and other energy sources.

We already got a lot better vision ML algorithm because of this (tx to suckerbergs dino and segment anything), we have alphafold tx to Googles Deepmind etc.

AI only consumes currently a little bit more about bitcoin but is acutally beneficial.

Have you ever done research? It's very difficult to construct your measurements in such a way that they measure what you want them to. So in just your quoted statement there are many sources of uncertainty.

* what's a typical job? How did Google resample their data to make it representative of the typical job?

* is this all jobs, or just work-related Gemini conversations that they've classified as such?

* what's the denominator here? How do they know how many tasks there are that weren't subject to conversation with Gemini?

* how reliable is their classification of conversations as work related or not?

Accurately measuring this is so complex that to present the conclusion in such a confident way is bordering on farcical.

I really don't see any conclusions which would be such a game changer and critical, that this information has to be perfect.

And its Version 1 and an ongoing project.

"In the workplace, we show that while AI adoption spans occupations covering just above 88% of US employment, penetration remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope. Outside of work, AI spans activities making up about 98% of Americans’ non-sleep time, with disproportionately high use in high-friction tasks such as engaging with government and professional service providers, likely delivering economic value that standard national accounts may miss. Globally, adoption scales with national wealth and has broad linguistic distribution, with English queries representing only around a third of volume."

It sounds reasonable enough and it contains nothing off.

For me its interesting to read that english is less dominant and thats def something i was very atuned when GPT-3 came out: I started to write english and my native language and mixing up words etc. The quality of understanding in my native language is sledomly good when a new product comes out but it was very good already when using GPT-3.

In comparision to this, i'm also aware of a research paper stating that programming in certain languages like spanish is worse.

> My company spends millions on ai token every month

How much profit does the AI company make from those tokens?

I don't know and that wasn't the point I made.

My company is a two digit billion dollar revenue company in the top 100.

My point was that real companies pay real money for AI.

But i use it daily at work and a lot of co-workers do too.

The point the commenter is making is that your company will not remain a two digit billion dollar revenue company in the top 100 for long unless they have someone whose job it is to figure out how much profit is being made from those tokens and act on that information.

> My company is a two digit billion dollar revenue company in the top 100.

The company ranked 100th in the top 100 companies by revenue is Eli Lilly and they make 45 billion in revenue.

I don't believe you.

My company i work for is real, normal big global company.

Not sure if its really top 100 or top 150, I checked google and the lists again i can find it in one i can't find it in the top 100 in the other. Shouldn't matter though the 2 digit billion dollar revenue is true and the amount spend on tokens is true too.

Everyone has access to claude, we get reminded regulalry to invest time into using AI.

Wait, is it your startup or "company I work for", because you've said both.

If you're going to lie, at least keep your story straight.

And? I do both.

My startup is not a billion company. Its a small startup.

Edit: I think i get the confusion now? I have a small startup on the side and i work for a big company.

> real, normal big global company.

Compu-Global-Hyper-Mega-Net?

My company is ten quadrillion dollars revenue (per second), you just haven't heard of it (it's from Canada).

And we spend $3b/yr just to make sure our I is capitalized.

Why are you so snarky?

I don't want to share my company name (obvously?!) but i'm willing to share that one number to give more insight.

Feel free to not believe me.

Your company doesn't use AI at all? You are unable to use it?

My AI is frontier enough to know why I'm so snarky, perhaps you should ask yours.

I realized that it might not been clear.

I have a small startup AND i work for a normal well known big company

I would like to understand how the cost of AI power & infrastructure, more than the global annual gross revenue software sales, will be recouped while simultaneously delivering on the promise of putting technical services in the hands of anyone, which will naturally put downward pressure on the prices.

I'm very interested in their anonymization and classification processes. Traditionally, conversational writing is too complex to anonymize automatically without false negatives, and remaining PII might be very personal in nature.

> Observation Clustering and Taxonomy Organisation (OCTO), a tool for transforming massive unstructured text data, like LLM conversations, and distilling them into organized entities.

This also sounds interesting. Hand crafted ontologies that I've seen used were always too hard to maintain and make useful, but potentially they have something automated?

>How ATLAS has been developed

"We picked some results we wanted to show and told gemini to cherry pick data that would fit in. As you can see, it was not a lot."

Interesting to see that Gemini (which I assume wrote this entire article and site) has the exact same over-stylized italic serif preference as Claude does.

A very positive spin. I fear that AI will be a tool for the rich and a toy for the richer. It will sadly divide the people Geographically (rich countries affording data centres and nationalised models) and poorer countries missing out. It will divide socio-economically, the rich affording tokens becomes the land/knowledge owners/hackers and the poor becomes the labourers.

+ with nation states becoming poorly run cesspools, we're going to see what happens to that.

I think that AI will be a tool for Governments and the Surveillance State, and will be used to implement Communism 2.0. People in countries who shun AI will enjoy more freedom.

What is Communism 2.0 and how will AI help to implement it?

86% of AI not used for work - so Gemini is mainly used by consumers, I guess also mostly non-paying. I am not surprised as they are on many Androids nowdays and Android updates recent whats new showcases new Gemini features (or repeat the same ones from earlier) for more people to try, my wife uses it often instead of Google search and sometimes to edit some photos.

That's just weird. I don't get how something like this gets published when I need 8 rounds of approvals for a button.

100000%! Ariane approvals have had more reviews than this

Here it appears the data collected when computers users submit "prompts" to Google will be shared with "researchers, policymakers, businesses, workers, and other participants in the economy as we work together to shape the ways AI can positively support people in their lives."

In the bolded text, the "n"s are lighter grey.

Anyone know why they did this? Once I saw it, it became too distracting.

Good eye! If you zoom in the text to a very large size, it looks like the "n" characters have a strange internal border or stroke?

Their adoption results are going to be biased by the fact that Google has a vested interest in accelerating adoption and have been sticking their AI products in front of everything they can. Seems like this isn’t going to be a useful data set for making policy decisions.

One possible interpretation of this article is that it’s very much in support of the idea that the AI bubble is going to burst.

> AI use at work is broad but shallow: Workplace adoption spans all industry sectors and also 68% of all occupations that collectively represent 90% of total U.S. employment. However within jobs, people are using AI selectively: in a typical job AI is used for only ~21% of tasks.

Translation: the market is reaching saturation.

> At work, most AI use is focused on collaboration and assistance with tasks, and so far task automation is uncommon

Translation: AI can’t be trusted to operate on its own and very little progress has been made in that area despite humongous progress being used in other areas. It’s helpful, but in the same way that a tool is helpful (and most companies don’t want to overspend on tools unless they can seriously reduce labor cost, I.e. hiring fewer people)

> AI is delivering value at home that may be missed in standard economic metrics, particularly around high-friction administrative tasks: Over 86% of interactions with AI tools in ATLAS occur outside of work.

Translation: AI is most valuable in low dollar value scenarios where it is likely a loss leader. The amount of money the average home user is willing to spend on AI is likely very close to zero, and this is the most likely interaction to be a race to the bottom. E.g., why pay Anthropic $20 a month when your iPhone 19 has a built-in local model that’s almost as good, why sit through advertisements on your free AI service when a cost and ad-free alternative is available.

All of this alone doesn’t imply a bubble. The bubble part is where many AI companies can’t financially survive without a massive product maturation, budget cuts, or further investment.

Yeah, there has to be a bubble for it to burst. There doesn't seem to be one yet.

I guess I’d be pretty confused on how to not classify it as a bubble.

The level of investment and operating costs of AI firms is just so high that the question of where the customer spend comes from to make it profitable eventually comes into play.

The other obvious question is how AI firms maintain their bookings and overall value when the upper middle class responsible for the majority of consumption dollars is getting laid off left and right by AI.

E.g., OpenAI finally ships AGI and realizes their investment value and goes from unprofitable to profitable. They can charge companies thousands per month for their best models/tooling because they replace employees entirely.

All the big corporate customers of the world are happy: truck drivers are replaced by self-driving, big high-income employee bases like Meta, JPMorgan, law firms slash employee counts by huge amounts like 50% or maybe even more.

…and as soon as that happens all those same companies face massive customer attrition since all those high-earners were just laid off and aren’t buying PlayStations and KitchenAid stand mixers anymore.

But no AI provider is actually making money. They are being pumped money by investors hoping for a return. There are no returns. It's a bubble?

How’s this for a working theory. Economic actors that were established doing the “easy stuff” that got displaced first got a head start on everyone else and started low balling on the harder stuff and that’s why we’re seeing impacts in the parts of the economy that aren’t being directly hit by AI. It could explain why people who know stuff are finding themselves squeezed for cost in every direction even though their specific area defies automation. There is a latency at play here while all the corner cutting of the cost based competitors mount. Not sure what is the outcome when the inevitable correction kicks in but the economic impact will be long lasting as skilled workers leave the industry and we are left with decades of slop to clean up.

And where will the skilled workers go?

Correct you will be left with the people who couldn’t switch

Is this Google’s poorer version of Anthropic’s economic index?

There is a lot of words here (I haven’t read the full report and I’m assuming it’s heavily AI generated but I will look later).

For an article about the AI economy, there is no mention of economy.

That last graph is surely AI generated.

Things aren't going great if Google is investing into... whatever this is.

Look, say what you will about Ed but when you compare his analysis against numbers and data like what we see from Google here you can’t help but see a difference https://www.wheresyoured.at/the-openai-bubble/

It was a lot to read and I do not understand the direct relation to the numbers in the google blog at all.

You wanna elaborate?

Look, I will say what I want: I'd trust Sam Altman to provide accurate, unbiased information about the present and the future of AI tech before I would trust Ed Zitron to do the same.

And no, I wouldn't actually trust Sam "King of the Cannibals" Altman. But I would still expect him to start with far, far more accurate data, and distort it a lot less to suit his narrative than Zitron would. That's how low the bar is.

The man made saying "the AI industry is failing" his entire personality. He'll keep saying that regardless of what the AI industry is doing, forever. You can write "tech companies often overpromise and underdeliver" on a post-it note, stick it to a whiteboard, and that will provide about as much perspective as everything Zitron has ever written, or will ever write.

That’s fair, but let me ask, how do you answer the numbers that Ed provides?

Sure, Ed’s making some kind of money off of subscriptions. Yeah, he’s a personality. But to ignore the deep comparisons, numbers and data that he gathers and provides to back up his analysis feels like sticking your head in the sand.

There are other banks and financial entities that are echoing his concerns as well.

The "numbers and data" are cooked to the limits of available sources, and then twisted further by conjecture and interpretation to lend support to the conclusions they still wouldn't readily support.

As a ready-made example: the "AI models are unprofitable, API inference is subsidized, OpenAI is naked" line he keeps stressing. That doesn't seem to be likely even on the basic napkin math estimates. And is also repeatedly refuted by both AI company officials and various AI industry insiders - of varying trustworthiness, and with different incentives.

Having "the numbers" isn't the same as having objectivity, and again: I'd trust Altman to have a semblance of objectivity before I would trust Zitron.

If you only focus on ai bubble, look at google / Alphabets, Microsoft, Amazon and Metas revenue numbers. They literaly can just afford it

This is basically reading about heroin adoption from the heroin dealer.

I’ll take information from multiple independent analysts please.

More like the guy who wants to be a heroin dealer but nobody buys his cheap crushed up painkiller pills.

This is one of multiply independent analysises.

I do not follow your argument.

Would you read this and say 'this is now the only truth'?

I mean i read it and will add it to multiply other sources. Now i know what / how google seas things and i try to get something from it.

i find it very ignorant if one of the biggest ai companies on the planet posts something and its just getting dismissed. You don't have to follow google every word, but it def adds to an overall picture even if you don't take everything as true

> AI use at work is broad but shallow: Workplace adoption spans all industry sectors and also 68% of all occupations that collectively represent 90% of total U.S. employment.

Im sorry but what? Plumbers, electricians, framers, welders, auto workers, hotel staff, restaurant staff are these people using AI in their jobs? Does someone at google think that these very much manual labor jobs are only 10 percent of the work force in the US?

There is PEW data, that says 30 percent of us labor force is in manual labor / blue collar jobs. The person you call at 3am to fix your drain, isnt using AI.

A friend of mine is a plumber. They use ChatGPT to write commercial bids. They developed a prompt, feed in some details, and out pops a detailed bid.

> The person you call at 3am to fix your drain, isnt using AI.

I am 100% using AI to figure out wtf to do about my 3am plumbing problem.

/joke

After I have the water turned off, I’m going to back to sleep to call a plumber during regular hours.

One of the best things about Big Tech stealing all your ideas is archiving the repo and watching their progress stop

Google/Noam Shazeer coped Ragdoll Studio to a T then act like they’re the innovators in the space lol

Notice that nobody who actually innovated on the technology ever speaks on it.

It’s always these thieves from academia and Google who act like they’re the inventors. Trying to explain to us - the people they have been ripping off for years - how “AI works”

Go home Google

I'm lost. If you referene Noam Shazeer in context of Character.ai vs. Ragdoll Studio, Noam Shazeer and all other co-authors have done the work for Attention is all you need while working for/at Google.

They are for sure one of the current Frontier AI labs.

And before their competitors and china started to be that fast, they were regularly publishing a lot of research in this area too.

Were is your hate coming from?

Reply on news.ycombinator.com