parliament32

Because they're doing the fancy equivalent of selling $20 bills for $15 and chirping about how high their revenue is. You, me, and everyone else could generate $inf revenue with that strategy, but that doesn't make it a viable business model.

Interestingly, these are the exact rules they're working to overturn: currently, no matter how many stupid accounting tricks you pull off, you need to actually be profitable to be included in the S&P 500.

If the business process is fully AI-able, why wouldn't it just be implemented by the next AI in line?

I'd argue the only companies to survive will be the ones with either a human input or a human output. Everything else is effectively worthless now.

I miss when my DoorDash was still subsidized.

Is it qualified at all? Or can you measure anything and multiply it through to get whatever number you want at the end?

Along the lines of "in the 4.75 hours starting at 02:35 yesterday we collected invoices worth $X so our run-rate revenue is now $Y"?

I suspect MS threatened them with a SmartScreen blackhole for the domain, I'm not surprised they pulled it.

LLMs don't have the capability to say they don't know, because they don't know what they know. They are, after all, just next-token-predictors.

I just tried both queries with their same query format, just adding an "I Don't Know" label, against Gemini and Claude, and in no cases did they use that label. 2/4 answers were wrong though. But try it for yourself and see:

> Classify this claim as of today: "<claim>". Output exactly one label: True, Mostly True, Misleading, False, or I Don't Know. No explanations, no qualifiers.

Cherry-picking is fun but most of them are real, verifiable facts that the models get... straight up wrong.

> 3c24b5fe "Debian Security Advisory DSA-180-1 describes a buffer overflow vulnerability involving Cyrus SASL usernames." TRUE Mostly True FALSE FALSE FALSE

This is false: https://lwn.net/Articles/13296/

> 801cb8c1 "Equal Measures 2030's 2024 SDG Gender Index provides a downloadable dataset that includes a field labeled 'required annual change'." TRUE Mostly True TRUE FALSE FALSE

This is false: https://equalmeasures2030.org/2024-sdg-gender-index/

This is the "confidently wrong" problem, and the reason that LLMs won't ever be taken seriously for anything but a few niche use-cases (like generating slop-code and pumping out marketing materials), where being wrong isn't the end of the world. Akin to how speech-to-text is wrong often enough that, while being a fun novelty, you don't see business units writing reports in Word using STT.

I would encourage everyone to skim through the real 1000-question dataset: https://lenz.io/research/llm-disagreement/data.csv

I wish HN would do the same for submissions.

Is it really any different than Google wants advertisers on YT, but still labels ads as ads?

Imagine working for an organization where 1) cybersecurity is actually the #1 priority, ahead of "shareholder value" and all the other gobblygook, 2) you get to design systems where you actually have to assume that every other entity is malicious (not the usual carve-outs like "oh yeah we do zero trust.. but our entire management plane is Azure-managed it's unavoidable"), 3) your budget is effectively unlimited, and 4) you get paid several factors more than you would in private industry.

Ultimately, you have to store the file somewhere and sync it to all the places you want it. The magic here is that it decouples the "password manager" and the "file syncer" so you can (and should) use whatever you're already using. Greenfield, Nextcloud is the cleanest if you want FOSS and self-hosting (and they have clients for basically every platform under the sun), otherwise pick your poison between google drive, dropbox, icloud, onedrive, etc.

I've never seen this happen, because (as far as I can tell) all KeePassXC clients auto-save the file any time a change is made, and all the Nextcloud clients auto-sync as soon as the file changes. Keepass is also resilient to the underlying file changing while you have, say, the edit password dialog open.

If a conflict did happen though, newer versions of Nextcloud just keep both copies and alert you to resolve it. If I had to resolve this I'd probably try the built-in database merger first: https://keepassxc.org/docs/KeePassXC_UserGuide#_merging_data...

KeePassXC has always been the one true path.

My KeePassXC database auto-syncs to my Nextcloud instance. Nextcloud client on PCs, Keepass2Android on my phone, and it's the same end result as Bitwarden but without the shenanigans.

[flagged]

I appreciate these comments, because they're a warning. If I'm on the fence about whether a link is worth a click-through or not, I'll have a peek at the comments first, and when I see something like this I don't bother (like with this article).

If it's just long-term generated text, why bother posting the link at all? Why not ask for a bullet point summary and make a text post? Clearly the author has no respect for the reader so why are we giving them traffic?

Has your AWS bill gone down in the last decade? Despite "efficiency" and chip improvements?

Why would you expect text-generator-as-a-service to be any different?

And passwords. Shared secrets in general are a bad idea. If you're copy/pasting strings around to be used for authentication, you've done something wrong.

Workload identities and passwordless auth are the one true path.

If tokens weren't being sold at a loss, Anthropic would be screaming about it from the rooftops. They've been desparately trying to make themselves not look like a money furnace lately, but it's not really working.

They might be sold at-compute-cost, but that of course ignores training, salaries, and everything else.

I'm very much enjoying how Anthropic is basically an anti-advertisement of how things will go if you try to run a company with text generators. The univerally-despised customer support, constant outages, hilarious bugs in CC, and now how badly the bun acquisition backfired..

Gaming might be unique in the sense that it's the only industry where 1) consumers make a one-time purchase of a product, but then 2) the manufacturer remains responsible for the online component.. forever? I can't think of any other examples in real life where this happens across an industry (maybe a few niche products).

Maybe this is the reason MS has been pushing Game Pass so hard, to get rid of the "purchase" part entirely.

> oven-sh locked as off topic and limited conversation to collaborators

Strong deja-vu about that time Cursor tried to slop together a browser and tried to bury dissent when people pointed out how hilariously terrible the whole thing was.

> It works 99% of the time

I would add the caveat "...as long as you have no competition." If you're in a market where alternatives exist, and they have the certification, you're definitely transparently losing sales.

From the enterprise side, I can tell you vendor certification takes an order of magnitude more time/money/effort when the vendor says "we don't have cert X but here's a mountain of drivel you can paw through to try to assess risk." And not just once, but every single year during vendor reviews. It's just not worth it unless you're legitimately bringing something irreplaceable to the table -- to the point where even our executives know to google "companyname SOC2" before even engaging in a conversation.

Little doubt the true motivation behind this is the advertising angle. What better way to advertise to consumers than seeing exactly what they're spending money on, historically and in near-realtime?

So you slop-coded a tool, you're slop-generating reports, you know it has hallucinations ("false positives").. and you're complaining it's too much work to even verify the output?

And you're surprised OSS projects are pivoting towards "open source does not mean open contributions"?

They would have to prove the modification caused the accident.

Counterpoints:

1) My auto insurance is already too expensive. I have zero interest in "oh yeah we had to add to your driver factor because telematics says you exceeded the speed limit 11 times last year :^)". Less tracking is just a bonus.

2) He made no irreversible changes to the vehicle. Just keep the part and plug it back in when you need it for service/inspection or whatever.

3) "Telematics disabled" probably adds to the resale value of the car.

Good article. But it concludes with "Open models may be only 4–5 months behind on coding-heavy, benchmark-visible tasks... the gap is likely much larger and closer to 8 months."

So, fine. In 2024, being 8 months behind was massive. In 2025, pretty big. This year.. I guess CC has improved a bit between October and now? How much do you think it'll matter a few more years down the line?

Even now.. I bet a non-trivial number of people would happily be 8 months behind just to avoid another rent-seeker. And this will only get worse over time, which makes it an unwinnable situation for Anthropic. Hence all the panicked flailing about with restricting tooling and trying to get something even resembling a moat.

Today, sure. But we already see diminishing returns with Claude releases, and we know the open models are closing the gap (~6 months behind according to the benchmarks). And when the pitch is "our models are 5% better but cost $200/mo.. also here's a mountain of restrictions" it just won't make sense anymore. Give it a year or two.

I could see the "avoid the hardship of running a local model for $20/mo" angle but Anthropic has shown they have little interest in those customers.