Aurornis
The number of people using AI at work could fall to 5%, but if those 5% of people were keeping Claude busy all day and night the overall expenditure on AI could increase.
The growing AI revenues don’t require everyone at the office to be using it. If the software engineers are each spending $10K a month on tokens, it doesn’t matter at all if Bob from HR is doing $0.50 of Copilot queries every day.
So $448 billion committed, with unclear cancellation terms, through the next ~48 months.
Clearly this will be rising over time, but even a naive division results in $10bn per month of compute spending.
That honestly doesn't seem so bad relative to some of the claims I've seen? Their Q3 revenue is about $5.6bn per month and it's rising. They have so much money in the bank waiting to be deployed that they get a billion dollars of interest per year. Investors are desperate to put more money into the company.
They spend a lot on compute, but rolling it up into cumulative spend numbers makes it sound scarier.
The work done in this experiment could have also been done by a sufficiently motivated human with enough time to explore, though obviously we would need more tools to visualize and track things. There have been a lot of interesting LLM introspection tools that allow you to peek into the model and see what's being activated and run experiments to trigger it.
Are you familiar with abliterated or "uncensored" models where weights have been modified to avoid triggering refusal patterns? You take an LLM, create test cases that trigger the refusal pattern, observe weights involved in that response, then tweak those weights until the response is reduced. This is the same type of work, applied to a specific type of error that was impeding math within Qwen. That work can also be done with an LLM in a harness running the experiments. The amazing part here is that Astra was large enough to walk through the problem without other tooling doing it experimentally.
So I don't think your analogy of a human modifying a neuron is quite applicable. LLM weights are not neurons and they've been inspectable, by humans or LLMs in a harness, for a long time.
As it was written, it looks like the engineering team is migrating the core infrastructure of the company around.
Unless you managed to talk them into that, in which case, share your secrets!
Those individual subscription plans you get as a consumer are not what their big enterprise customers are using.
Those plans are cheap so you, as an individual, can get familiar with the product and convince your company to get it at work. Your company pays the API prices to get enterprise features.
$200 of usage on a $20/mo plan is a 10X multiplier.
The $200/month plan has 10X the usage of the $20/month plan.
So $200 * 10X = $2000/month. I don't know how you got to $8000.
Recent analysis by industry analysts like Semi Accurate have measured Sol 6.1 on the $200/month plan to be around $2000 of API equivalent use.
I haven't seen anyone treating AI tokens as employee compensation. If anyone can show me an article where employees were paying taxes on their token use at work, please share it.
Companies have been treating AI tokens as part of the cost of having employees, like the cost of their office or travel expenses.
They have never claimed that the API rates are the "real cost" of the compute spend.
If Anthropic had revealed their internal cost, there comments would be complaining that they were hiding something and giving useless, unrelatable data.
The expenditure is most interesting when compared to retail rates, because that's the rate the rest of us have to compare against.
They switched from ClickHouse to DuckDB five months ago, then switched from DuckDB to their own Trino derivative?
The article then ends before saying if the new setup solved their problems.
It’s great that we can now make these migrations with little downtime (I assume?) and try different things, but I wish we could have read more about what testing they did before jumping. The fact that they jumped to DuckDB without realizing it wasn’t going to work at their scale and spent only a few months with it before jumping again is not convincing.
I hoe there’s more to this story that just didn’t make it into the blog posts. As written it’s giving me flashbacks to some teams I worked with who changed tech stacks every six months because they liked migrating more than tuning and maintaining.
Providing even a minimal set of details is important because “an LLM” could mean anything from the Google search header to Fable or Astra.
It’s like someone making a claim that they tried a diet and didn’t lose weight, therefore all diets are bad and shouldn’t be trusted. The details of the diet, how long they tried it, how strictly they followed it, and the rest of their situation are pertinent.
There is an audience who demands content like that: People who want do hate diets and believe they don’t work will eat that commentary up because it levels up the belief they want to hold. Everyone with critical thinking skills or experience with dieting knows that one person claiming diet failure means nothing.
We know diets are a set of tools and they’re not perfect, but they are used successfully by people who can understand their limitations. Now replace diet with LLM and my statement is the same
> How about the opposite? Share a normal session (by normal, I mean something longer than, say, 4 turns, not absolutely trivial) where LLM doesn’t do a single silly thing. Including in thinking portions.
How about Terence Tao sharing his ChatGPT logs having ChatGPT work with him through understanding the Jacobson conjecture?
https://chatgpt.com/share/6a5fdc7a-d6f8-83e8-bbea-8deb42cfed...
The bigger revelations in the story were that the “Huberman Lab” existed largely in name only and he wasn't living or working near the University. His career had pivoted to influencer celebrity long ago but he’s been capitalizing on the neuroscientist title and claiming lab associations anyway. They also interviewed associates who revealed his history of leading people on, manipulating them, and being dishonest wherever it might help his career. Then they found that he was cheating on multiple women at the same time, including someone he led to believe he wanted a long term relationship and children with. He was injecting one partner with fertility drugs while cheating.
What’s fascinating is how his fanbase stripped the story of all of the important details (chronic pattern of misleading everyone in his life where it can help his popularity, which extends into his podcast reporting where he misleads listeners) and instead tried to transform it into personal dating drama that was overblown.
Then the second pass of rationalizing it away was to take the relationship drama and transform that into something that could be blamed on the women involved while letting Huberman off the hook. After that final transformation they were free to ignore anything negative about him because the story had been rewritten, in the retelling, to be completely the fault of women. There were even efforts to make it look like a good thing that he was dating multiple women.
In case it wasn’t clear: In the reporting the women did not “continue to choose to date him” as you heard. The reported group discussion was about former partners had been lied to who were discovering each other because the number of them had grown so high that he couldn’t escape it.
If someone is writing confident takes on a topic, I expect them to have some experience with the topic.
Writing confident dismissals of something after not getting it or not really taking time to learn makes for uninteresting writing.
If I’m going to read someone’s opinion on an a growing industry, I expect at minimum that they’ve given it an honest try.
These articles where someone tries AI once, doesn’t get it, but also thinks they understand it better than everyone else are everywhere right now.
They’re not very interesting because the author hasn’t taken enough time to understand the subject they’re writing about.
During the peak of his popularity you would find Huberman protocols or Huberman factoids repeated everywhere on the internet, including HN. There was a sharp drop off after the revelations. I can’t recall the last time I saw someone someone repeat a Huberman podcast factoid here, which is remarkable after he was so popular.
In the Huberman case it was compounded by how much he was embracing and repeating non-scientific ideas that sounded good on social media, like reducing everything to being about dopamine or making crazy claims about how you could do things to “raise your dopamine” that were not actually supported by research. When his personal dramas were exposed it became an opening to discuss how bad his podcast had become and I think many listeners woke up to that fact when shaken by the realization that not everything about Huberman was as it seemed.
He has been able to continue podcasting to an audience of die-hard fans, though. These people never totally disappear because a core audience worships them and will ignore everything inconvenient to their worship.
Peter Attia went through a similar downfall after several increasingly bad scandals were revealed, culminating in him getting caught associating deeply with Epstein. He tried to make a public statement about distancing himself from Epstein after discovering Epstein’s crimes, but there are public emails between Attia and Epstein where Attia asks Epstein about how he’s doing after Epstein became national news. That backfired everywhere except the core audience of Attia worshippers who now just get angry whenever anyone brings up the Attia-Epstein connection. It’s wild that Attia was one of the few Epstein associates who continued contacting Epstein so much (which went into court records and became public) after he became publicly known as a trafficker. And yet there are hardcore Attia fans to this day.
Do you have a share link to the session?
Why is every confident claim about a frontier model being brutally wrong about something presented without any details or evidence? It’s so trivial to share even details like what language was used, but even that had to be obscured for some reason.
I know why you included it, but it did reveal your level of experience with LLMs. That’s important context for any discussion like this.
> the takeaway i want is a. i would like people to still have fun and tinker around with things in their personal time instead of making an LLM do it all themselves
I know, you wrote that much in the article. That’s what I was responding to: I don’t especially enjoy when other people start taking strong opinions about how other people should and should not spend their personal hobby time. This is why I stay away from most of the online communities for my hobbies these days: The people with strong opinions about how others should spend their personal fun time and the right and wrong ways to enjoy personal time are really not fun.
It would be one thing to write about why you choose to spend your personal time this way. The part where it turned into disdain and got preachy about how other people weren’t doing the same thing was where you lost me.
1. The AI box at the top of Google
2. The voice assistance on their phone when they tried it a long time ago. Some question they asked Siri in 2025 and got a bad answer for.
3. ChatGPT free tier.
For a while you’d find someone who got assigned an AI tool at work but their administrator limited them to the cheapest model for cost savings, but most companies are wising up.
I also talked to a developer last month whose company had never let them use anything beyond Sonnet. When I told him I used Opus for everything his eyes went wide because he assumed I was spending thousands of dollars per month. However now that the latest Sonnet is very good, even this category of exposure is decreasing.
I think in 2027 as advanced models become more common we’ll see less of these claims about AI not working well when they tried it. There will be holdout group of people who refuse to try anything new, but the world is moving on.
The core piece of this blog post is that the author doesn’t use LLMs for fun work because they do the fun work to have fun. Okay. Nothing new there. I some times grow vegetables myself and eat them instead of buying them from the store because I think it’s fun. I don’t feel the need to go around sneering at people who offloaded all of their vegetable growing to someone else. I also don’t like associating with people who get smugly superior and condescending about their personal gardening.
In fact, even though I enjoy gardening and nerding out about it, I don’t like talking to the gardening snobs who have turned it into a part of their personality and who think it makes them superior to the lowly people who don’t garden.
And that’s the exact vibe I get from the articles like this, where the topic drifts into virtue signaling and criticizing the LLM users.
There have been an exhausting supply of these articles for the past few years, and there’s frequently a section like this in each of them:
> i truly have tried to use llms before, just to try it out. nothing i was going to publish as research, but i really couldn't and it left me never wanting to use an llm again
I won’t judge anybody for not using LLMs or for using LLMs for their fun work. Do whatever you want for fun! Don’t let virtue signaling posts like this one shame you into thinking there’s a right way or wrong way to enjoy your fun project time.
That said, I find that to get good discussion about the pros and cons of AI on our thinking you have to look beyond these people who just don’t get LLMs at all. You can’t really understand this topic until you’ve seen both sides.
I wouldn’t find it interesting to read an article on this topic from someone who never learned to program but only uses LLMs to write code because they do not understand the other side. Imagine seeing “I tried to learn how to code once but it didn’t really click for me. It left me never wanting to write code again.” Then they went on to talk about the brilliance of vibecoding with some collected examples of hand-written code causing problems. This article is essentially that, but from the other perspective.
You can, in fact, be upset when lawyers abuse the system.
Not only can you, but you should.
These attempts to reduce all things across society into oversimplified, vaguely defined buckets like “ruthlessly selfish” and then saying we can’t be upset about unrelated things because they all reduce to the same vague bucket is the weirdest form of whataboutism. The argument doesn’t even make sense, it’s just a forced false equivalent to start complaining about a different topic (consumer advertising) in a thread about lawyers doing prompt injection in a court of law.
The book recommendation aggregation website is so common because it’s one of the few variations of this theme that can be monetized through affiliate links.
None of them last long because the traffic falls to zero basically as soon as it falls off the front page.
The comparison is unfair for many reasons, actually. A real install would have solar inputs charging the batteries and offsetting loads during the day.
It also wouldn’t run a 5-ton HVAC system 24/7 for 7 straight days, which is how the post got to the extremely high number.
It’s about as bad faith as the argument can get.
In the fall and winter (right now) I’m heating the house anyway so running LLMs 24/7 is basically free to me. So I try it. The models have are limited in their intelligence. I have tasks where I could let Qwen run for days with concurrent workers and never get anywhere, but 15 minutes with Opus solves the problem.
I’m on Mac. Docker containers run inside a VM. I said that if you don’t have that situation you need to take that into account.
Docker can be escapable by a determined attacker, but the common compromises you see from packages or downloads aren’t doing complex container escapes. They smash and grab what they can. Docker does enough isolation that they don’t see what they’re after. It’s not as good as a VM, but better than nothing.
I remember a very different old internet, where high quality content like these videos was rare, ads were more invasive on a level we have forgotten (remember pop-ups? Pop-under? Having to download and install Real Player and other garbage with ads just to watch anything? Spam overrunning Use et?), and the internet was full of brainrot and crank content from the start.
> I hope we get to a point where hosting becomes functionally free,
We’re discussing this under a list of freely hosted videos.
The blog it’s talking about is hosted on Bearblog, a free ad-free blogging platform.
The software project was built by Opus so quickly that a $20 subscription could have built the whole thing.
Everything is close to free right now. Pay a little bit for YouTube premium and you can watch without ads too.
I think you’re nostalgic for an internet that didn’t exist, while not seeing the amazingness of the world we’re living in.
The swarm strategy can’t overcome the dumbed down model problem. You just end up with your swarm of agents all going in the same dumbed down loops.
The Bonsai quantizations are not good for real work. I have tried them. They are as bad as you would expect from 2-bit quants. They may have done a good job quantizing to that limit, but it’s still bad.
Even the 4-bit Qwen3.8 27B quants are still iffy for coding work. It’s not until Q5 or Q6 that I find them useful, which is too bad because that cuts into available context on 32GB cards.
> Which in turn suggests that Fable 5 coding performance can be achieved locally with a bit of clever harnessing.
I experiment with a lot of local inference engines and I also use both Fable and Astra. I wish this was true, but it’s just not. I can generate tens of millions of Qwen3.8 tokens per day and run the hardware nonstop, but even brute forcing it like this just can’t accomplish the same thing that I can get Astra, Fable, or even Opus and Sol to get done in an hour or less most of the time. You just cannot make up for the lack of intelligence built into the model by brute force unless maybe your timeline extends to infinity.