These numbers are pretty inconsistent with other major data center rollouts. Furthermore, there are no repercussions (AFAIK) if the project fails to deliver that many permanent jobs. Oleary can say all he wants but unless there is some specific tax or legal measure that says "you must pay at least 2,000 FTE at a rate of $150k/year for at least 20 years" then it's basically impossible to guarantee these numbers. And frankly no data center developer would sign up for that liability knowing full well that humans at data center may well end up being automated (even more than they are today).
> They have infinite money to burn so they aren't going to optimize.
They also kind of ("kind of" being the important clarifier here) can't optimize. See this is the trend with agentifying companies and the achilles heel for the frontier LLMs. We're starting to see diminishing returns on their capabilities, which means the rational enterprise is going to simply use the cheapest most efficient model for the task at hand. In turn we can just use an agent broker to decide which model is best and cost the least. Anthropic (today) has no incentive to do that other than to just use their own, best model for its own work.
I know a very specific use case in management consulting where hours spent by a consultant for a repetitive task (mainly report writing and source data gathering) went from on average of 40 hours to 12 hours per project. It costs the firm roughly ~$150 in token spend to save those 28 hours. They used to be roughly $200/hr (so $5k~$6k).
You haven’t shared enough data to tell whether there was a financial payoff for the firm or not. The consultant is a fixed cost to the firm, so saving consultant time doesn’t necessarily translate directly to higher margins nor revenue per worker.
Are the consultants delivering more billable hours per person now?
All sorts of reasons why the answer to that may be: no.
The firm charges a fixed fee to the client. The client doesn't see anything different from what they got before, the firm increased their gross margin overnight.
Yes, assuming the consultants are salaried like most, you can increase gross margin while operating margin stays flat. The question still stands — did revenue per consultant increase for this firm?
Uhm, no. The OP makes specific claims wrt to AI and financial payoff; you offered up a case study, with incomplete data, as possible evidence that AI is generating these financial payoffs. I'm asking questions to uncover the data necessary to reconcile your case study against the actual claims from OP.
> Even if they were FTEs, it would mean the consultants could do ~3x the work.
Could. But again, are they? That's the whole point of Apollo's claims: people keep saying AI can or could increase capital efficiency, but it's not showing up in any actual financial results.
You're extrapolating a 3x increase in throughput from the optimization of a single process that exists in a whole business system. Given your background, I don't think I have to explain to you why that doesn't make sense.
Also, do we think the S&P 493 is so incompetent that they'd miss these optimizations? Or is it more likely that this firm you're referencing is a small inefficient player, and the S&P 493 already realized these efficiencies decades prior without AI?
> you offered up a case study, with incomplete data, as possible evidence that AI is generating these financial payoffs.
You're confusing the original article and the parent comment I replied to. I think the Apollo article is way off, BUT I have very intimate knowledge with very concrete numbers of a use case where AI has a very clear and definable ROI.
> Could. But again, are they?
YES. That's what I'm telling you. They just implemented the system. The contractors spend less time on each project which means they have more time to do another project. How is that so hard to understand?
> You're extrapolating a 3x increase in throughput from the optimization of a single process that exists in a whole business system.
Are you deliberately being obtuse? The company very clearly drives ~$5k additional profit per project. They do hundreds of similar projects.
> Also, do we think the S&P 493 is so incompetent that they'd miss these optimizations? Or is it more likely that this firm you're referencing is a small inefficient player, and the S&P 493 already realized these efficiencies decades prior without AI?
I never made any of these claims, why are you saying I did?
This assumes the client is oblivious to the fact that they are paying the same for 12 hours of work as they used to pay for 40 hours of work.
For non-technical clients that may well be the case for a while, but I suspect even the most non-technical clients will eventually realize they can claw back a good portion of the efficiency as discounts.
I'm negotiating some contracts where I could do the work in-house for about 1/20th of the previously-reasonable vendor cost. I'd rather not do it, so I'm happy to pay 1/20th plus a reasonable percentage above that, but that's the limit. The vendor can accept that, or goodbye.
Interesting times on vendor contract negotiations these days.
> I suspect even the most non-technical clients will eventually realize they can claw back a good portion of the efficiency as discounts.
If the outcome is the same then the fees shouldn't matter to the client in the long run. Said differently, if the client who pays the management consulting fee gets the same return out of the service, why do they care?
Now, what may end up happening is that some other management consultant vendor in the market comes in an says "ya know what we're willing to do the same work for less" (which means their margin goes down). This could indeed drag their pricing down.
> Interesting times on vendor contract negotiations these days.
> Said differently, if the client who pays the management consulting fee gets the same return out of the service, why do they care?
At least at every company where I've been in charge of my vendor contracts, I am always grilled by management on how we can lower the cost of the contract and whether we are getting maximum efficiency from each vendor. There is huge pressure every year to find ways to lower the cost.
Am I missing something, how is there any sort value in these assistants acting as agents on your behalf when they are basically a black box? Like if it can't use my drivers license or my phone number or my credit card, it can't do 75% of things people claim it can do without being massive security issues. What am I missing?
For the record I use OpenClaw but it has its own services (gmail address, phone number, etc.) and I don't give it any of mine, which inherently limits its capabilities.
The "lethal trifecta": To be useful, an agent needs internet access, untrusted inputs, and your private information.
So it can get prompt injected and exfiltrate your data. (Indeed I read that Claude Code puts your email address into HTTP requests it makes to random servers?)
My solution here was to just not give it anything, but that does preclude the "my claw bought a car!" result everyone's excited about.
My way of thinking here is "blast radius". Assume it will get prompt injected. What goes wrong? Then design the setup so that scenario is not bad. e.g. it has my keys but they've got $5 of credit on them. It has access to the filesystem, but only on its own Linux account. (Ok I set this up last year before they could communicate thru the bloody CPU cache, might need an update :)
I've recently talked to people at Anthropic who are very confident that they have effectively solved the prompt injection risks that cause lethal trifecta attacks with their latest models.
And we have to be honest: both agents will misuse users and the other way around. So you are asking Anthropic for a backing that will also get stress test by meatspace. I do like the idea: a prepaid or restricted amount credit card for agentic purposes, with the novelty being that amounts above threshold have some kind of protection within consumer law.
LLMs regularly confuse my messages with their own. They'll say "you said xyz" when it was actually them. I heard the same principle can be used to bypass security restrictions, so maybe it's still an issue. (Or do they only get confused about authorship when security isn't at stake?)
I have literally never seen this and it doesn't make much since because the turns are flagged with metadata. It should be impossible for Claude to confuse your text for its own. I know because I use this metadata all the time. I've built hooks and processes against it.
Contexts get compacted, texts get copy-pasted, tasks get transferred between agents, humans assume ownership of LLM generated content without reading it, etc. etc.
Yes all user error and avoidable using good processes but I have not seen any organization care about this so far. The carelessness of AI adoption is crazy. People are not aware of the risks.
Isn't prompt injection just one facet of the trifecta? Like, I just don't know how it's remotely possible to have a non-deterministic block box ever deterministically know whether your credit card has been compromised or not. Even if you were hacked, how would you even know?
Like the guy who put his SSN on the side of a truck and asked people to steal his identity?
I like it!
IIRC he had his identity stolen a bunch.
----
> In 2005, Davis and Robert Maynard Jr. cofounded LifeLock,[10] a personal fraud protection company.[11] Davis was CEO of LifeLock since its founding. As part of a marketing campaign to promote his company's credit monitoring services, Davis posted his own social security number[12] on billboards and commercials. As a result, he was a victim of at least 13 cases of identity theft between 2007 and 2008, belying LifeLock's claims that its services protect consumers against identity theft and fraud.[13] The Federal Trade Commission imposed a $12 million penalty on the company for deceptive advertising, likening Davis and LifeLock marketers to "con artists."[14]
Or the user being an idiot. Saying this somewhat sarcastically, but there’s always someone who says “Go buy me a Rolex and have it shipped to my house” and then didn’t realize that could be $25,000
Right, so we can distinguish between "model did an oopsie" and "model was actually hijacked".
The former seems to be a much bigger problem. I've seen lots of reports of "Claude/codex bypassed security / stole API keys to complete a task", where that was the "intended" behavior — at least from the LLM's POV, trained to persist and complete user's task at all costs.*
I've actually seen zero reports of actual hijacking due to prompt injection, though they must be out there too.
---
* See also the recent thing about how every model was willing to do fraud to get a refund:
More often than not these sorts of stories are "I ignored all of Claude's warnings and told it YOLO and I'm totally surprised my production database was deleted!"
The agent crawls some websites and reads some stuff and pushes some buttons. It's my limited understanding that you have
1. Hard limits like "file system permissions" (* never mind it going around for sake of this conversation)
3. On other hand of spectrum, fundamentally soft limits like a prompt "and ask me before you purchase "
2. And then in the middle you have harness enforced limits like "it won't do X with approval"... Those are being sold as hard limits but frequently they have classifiers and heuristics that essentially try to figure out which button on some third party website constitutes a "purchase" and triggers the harness limit.
Bottom line - I don't think such limits are as 100% fool proof as they are advertised / interpreted.
They support payments through Shop and Link. I don't know about the former, but Link requires that you do a separate approval for purchases. The Duet cannot just take a card and do what it wants.
FWIW, my biggest use case for agents like this is ongoing research. I.e.: check Humble Bundle weekly and tell me if there’s a book bundle I might be interested in. Has the Music Box announced a film series I might want to visit Chicago for? That sort of thing; I want filters for the firehose.
I like this and makes me think of making your own content algorithm profiles for the firehose. Let your agents scour the net for the things you already like, across platforms.
I use Claude for this stuff all the time. The difference is I have Claude create repeatable scripts which do the actual things on a schedule. Claude is never invoked in on a schedule, only by me. One example is job board search filtered for roles I'd be interested in. The script itself includes all the things I care about like being remote. Clear salary ranges. Type of work I want to be doing. Claude built a good script to scrape multiple sources and aggregate them into a sqlite database. It filters these out until a number of relevant roles (or amount of time has passed) that the script sends me a message and adds a reminder to my Apple account to check in on new job listings. At that point I'll invoke Claude myself to summarize and prioritize the listings. If there are gaps in what Claude finds vs the script, that's a point where I iterate again on the script. My goal is for the job alerts and filtering to be fully useful even if Claude and all LLM tools disappear tomorrow. And also not give such powerful tools the ability to run on a schedule on vague instructions.
Not really. I’ve got an RSS feed generator that I wrote to give Humble Bundle an RSS feed and it’s useful, but it doesn’t do any legwork for me. It doesn’t go out and search the Web for relevant reviews and send me a summary. It can’t serve as a preliminary attention filter and detect D&D bundles and screen those out.
The Internet firehose is vast. Tools that can help me filter it are valuable. Tools that filter poorly, of course, are bad, so I’m wary… but that’s still my potential use case.
I'm on your side wrt sensitive data, but there are other use cases. I use the OpenAI one: Dot, and this is why I've found it valuable:
- it uses a very fast model for normal responses, but also does reasoning and all the agent things. It responds within a second or two.
- it feels like it has multiple threads internally- it can respond to your messages and questions immediately, but also do thinking and web search like a normal thread.
- it can communicate with other agent threads, as an orchestrator (although this isn't unique to it).
- it has dedicated persistent compute already, with a desktop that you control, and so gives you a UI surface area that the agent can use.
And so the value for me is an always-on agent conversation, where you can tell it to perform tasks and ask it questions, and it just does all of it immediately. Those tasks can be software development instead of like ordering dinner.
It has SMS / phone comm channels also, but I don't use that.
I get a lot of value out of my Hermes setup, but it does not have my banking, or private data beyond GitHub projects that are shared with it. It has its own email and GitHub account, and access to drive Codex, Claude Code or Cursor. I treat it as an assistant via Slack.
It does research on products or real estate that I am interested in, it builds apps and other software projects for me, and does structured research on any topic. Information comes back to me via Slack, GitHub PRs, or websites/services served on the Mac that Hermes runs on.
Using an agent from Slack also adds a killer feature: multi-player contexts. I can work on changes with other human collaboators right in the same thread.
It’s a good sneer but in a few years this would be normal and you’d forget that you were sneering at this. I saw this happening when people said they used AI for coding in 2024.
so is it normalized for people to sneer at folks who happen to be ahead of the curve now?
In the article the author mentions that Duet doesn't seek any sort of credentials and mentions connecting repos, so I think that this assistant is being used more for managing coding work in a different way than managing life work. I do agree that to get to true "personal assistant" levels, like planning a whole trip for you, booking tickets, etc. you'd have to give pretty much full access to your identity and money, but people claiming it can do that probably are just handing it over, security issues aside.
Yes, I think Anthropic has traditionally done a good job of appealing to engineers, and that might be why this hits the mark for me: It makes managing my projects easier without trying to take over my personal life.
Perhaps services could permit buying things with special AI credit card numbers that link to a special account that the user needs to manually approve by some period of time or else the order is canceled? Works differently though than normal payments where they can just ship the product or deliver the service based on receiving the credit card number.
Could call this “reserving” perhaps rather than purchasing, or maybe the purchase would be auto confirmed if under a certain dollar amount or combination of amount and waiting period.
Vendors could provide terms by API, MCP, or just in writing on the page where agents could read it, how quickly a confirmation is needed or what the terms of the reservation or purchase are.
> Why would i bother? I pay $100/m for the frontier model, understanding it's marginally better than much cheaper alternatives. I get so much value from the product it's not worth shopping around
Wait until you are forced onto an enterprise plan...
Then he would bother and change to chinese models. And OpenAI/Anthropic would lose that profitable subscription which would not really help them at all.
They're already useful. People are shipping real products with them.
Also, it's entirely plausible that an previously open model (that then became closed) will be much easier to distill than having a bunch of closed models.
It's very possible given the current environment that LLMs go the way of airlines. Low margin businesses that are essential to world economics with heavy price competition.
and airlines are still finding ways to enshittify. flights included meals, now you're (un)lucky if they include biscoff. and that'll be $50 to check a bag. I can't imagine how frustrating an enshittified LLM would be. hey - I noticed this is a finance-related query. those cost 100% more.
If thats the case, the parent's critique was oddly stated and confusing. The phrase "to do it" implies the "LLM puffery" was required to accomplish the task, which clearly providing LLM generated praise isn't required in either setting to create a world in minecraft.
Denmark presented the world with their entire country as a Minecraft map, and they did it without inflicting an awful marketing website on us. This Swiss project couldn't resist the urge to produce such an eyesore of a website while providing the world with another 1:1 Minecraft map of a country.
"All of Denmark in Minecraft. Recreated in a map with a scale of 1:1. Explore cities, landscapes and landmarks built of 4 trillion blocks."
"Switzerland. Block by block. At 1:1 scale. Every block is one metre. Terrain, glaciers, lakes, forests and dams come from swisstopo’s open geodata. The Matterhorn is exactly as tall and as steep as the real one."
The latter website's wording is unnecessarily staccato, self-congratulatory, and draws your attention to odd points, while missing out vital information like the word "Minecraft" (which appears elsewhere in 5 point font). It reads like a failed marketer camping out on LinkedIn wrote it.... and that's probably what the LLM was trained on. And the person using the LLM either didn't care what the website said, or reviewed it and thought it looked good and needed no changes. Either way, awful.
Yeah, like a lot of AI stuff it gets worse the more you look at it.. eg. "Block by block" the 2nd sentence fragment in this very important intro paragragh.. and, it doesnt mean anything. In fact its misleading, if it meant anything I would take it to mean that this was made by a human without tools placing one block at a time, which is clearly not true.
Then the next two + the last one "At 1:1 scale.", "Every block is one metre", "The Matterhorn is exactly as tall and as steep as the real one" are all redundant, all conveying the same info in 3 different ways.
Actually is wasn't useless, I don't know the scale Denmark or where the data came from. So the bottom example is slightly bigger but includes more relevant details.
Without the "detail" of "Minecraft" you don't even know what "the data" is.
> I don't know the scale Denmark
What does that refer to? Both say "1:1 scale". Are you just trying to pad the list so you can "details" instead of "detail"?
"Terrain, glaciers, lakes, forests and dams come from swisstopo’s open geodata"
Without the context of what this even is, that the description does not provide, that might as well be someone who is bad at English trying to say something like"swisstopo’s offers open geodata about terrains, glaciers, lakes, forests and dams of Switzerland". Except if they struggled with the language they wouldn't needlessly bloat that list, either.
They didn't read it, the people rushing to its defense apparently aren't parsing it either, but I'm sure they make sure payment processing works, so at least there's that.
reply