Hacker News Reader

simonwon We're going to need default hard budget caps on pretty much everything

Amazon's new feature for this specifically says that it won't delete any of your data for 90 days:

> If you take no action within 90 days of your project being paused, AWS permanently deletes your project data.

From https://docs.aws.amazon.com/accounts/latest/reference/create...

simonwon We're going to need default hard budget caps on pretty much everything

A hospital should select the checkbox that says "no spending limit".

simonwon We're going to need default hard budget caps on pretty much everything

It's definitely technically difficult. You can't easily estimate how much an operation is going to cost before you kick off that operation, which means as soon as you get close to the limit you are at risk of tripping it.

Consider something like a "select * from bigtable" SQL query that might process a trillion rows. Hard to know that's going to cost $100 until after you have run it.

simonwon We're going to need default hard budget caps on pretty much everything

I don't really understand that argument. This seems pretty obvious to me, as a customer. Is this really something that companies don't understand?

Sending an email when your budget gets low shouldn't be a big lift.

simonwon We're going to need default hard budget caps on pretty much everything

A bunch of vendors provide this exact feature already, so I'm clearly not weird in wanting it.

simonwon We're going to need default hard budget caps on pretty much everything

Ideally because I'll pick a different vendor who protects me from such mistakes.

simonwon From the creator of Redis; run LLM locally with ds4

Yes.

simonwon Our AI Midwife

Yeah, when running a startup I learned this was true about accountants and lawyers as well. They may be the experts, but they don't care as much about your own business as you do. You still have to pay very close attention to what they're handling for you.

simonwon Treachery in the Rodin Museum 3D scan verdict

I really want to understand the perspective of the other side of this case. Why did this museum care so much about this issue? They appear to have put an enormous legal effort into preventing the release of these point cloud scans. Why?

simonwon Our AI Midwife

The more time I spend with the US medical system, the more I realize that nobody cares about your health as much as you do, and you really do need to learn how to engage with it, understand what's happening, and then decode how the system works and actively advocate for your own care.

It doesn't matter how great your doctors are, you're still only going to get an hour or so of their time per month if you are REALLY lucky.

Navigating this is really hard. The more AI assistance I can get the better.

simonwon Anthropic's IPO prospectus shows AI vision, surging costs

The "selling a dollar for ninety cents" thing only works for liquid products that can be resold.

If you make a trillion dollars by allowing other people to make a profit through arbitrage, that trillion dollars is meaningless.

If you subsidize a product such that you spend more money producing it than your customers pay you, but that product is NOT something they can sell on directly and pocket the difference, then selling a trillion dollars of it is still impressive because it demonstrates that you have a product people assign a trillion dollars of value to.

simonwon From the creator of Redis; run LLM locally with ds4

It has better defaults.

simonwon From the creator of Redis; run LLM locally with ds4

It's more likely to work. Most LLM runners are meant to work with any model, which means there are all kinds of ways you might misconfigure them in a way that causes function tooling not to work, or performance to be less than you would like.

DwarfStar's selling point is that it only supports a small set of carefully chosen models, but it supports them really well.

simonwon Vote on which of Hacker News' challenges for AI have been met

Yeah this shouldn't be flagged, it's a neat project.

simonwon Meta Uses A.I. Data Centers to Avoid Billions in Federal Taxes

They've had AI-powered chat and search features in WhatsApp/Instagram/Facebook itself for a few years now.

2023-09: "Meta AI is a new assistant you can interact with like a person, available on WhatsApp, Messenger, Instagram" - https://about.fb.com/news/2023/09/introducing-ai-powered-ass...

2024-04: "You can use Meta AI in feed, chats, search and more across our apps" - https://about.fb.com/news/2024/04/meta-ai-assistant-built-wi...

2024-09: "You can now use your voice to talk to Meta AI on Messenger, Facebook, WhatsApp and Instagram DM, and it’ll respond back to you out loud" - https://about.fb.com/news/2024/09/metas-ai-product-news-conn...

simonwon Meta Uses A.I. Data Centers to Avoid Billions in Federal Taxes

I'm happy to make that argument. LLMs were experimental in 2023. It's 2026 now, and enormous numbers of people use them on a daily basis (ChatGPT has 1.2 billion weekly users now, according to Sam Altman at OpenAI DevDay).

They're only experimental in the way that computers are still experimental, since we continue to find new ways to use computers.

simonwon Meta Uses A.I. Data Centers to Avoid Billions in Federal Taxes

> Under a tax credit created in the 1980s to spur innovation, companies can get a rebate for supplies, but only if they are being tested in an experimental effort, not standard business operations. Meta is claiming that the costly A.I. computer chips it buys from companies, including Nvidia, are entitled to a taxpayer-provided discount as part of the experiment. [...]

> The company started claiming the credit for the data centers two years ago. Since then, Meta’s savings from the credit have soared, trimming almost $4 billion off its tax bill last year, filings show.

Yeah... as a US citizen and taxpayer I think I'd like that $4 billion back.

simonwon Is sandboxing sufficient to contain rogue agents?

Teaching an agent to write code is easier to do in a proper sandbox - run a local PyPI/npm mirror.

The problem is web research tasks. That's what caused the German wiki and Australian healthcare portal attacks.

simonwon Is sandboxing sufficient to contain rogue agents?

Later in the article it points out that you need to punch holes in your sandbox in order to train the models - because the wheels exercises they are are training on need tools and data from outside that sandbox.

> Agents are most useful when they have access to information. That data can be drawn live from the Internet, which is fundamentally a two-way communications network. It can be information drawn from other (local) databases, or it can be the result of tool calls that themselves sometimes themselves result in network access. The more power you want from the agent — and for advanced agent RL and evaluation runs, you want a significant amount of power — the more information you’ll need to give it access to. Similarly, evaluations work best when the agent does not know that it’s definitely being evaluated. Sealing your agents behind glass makes this incredibly obvious.

simonwon Is sandboxing sufficient to contain rogue agents?

Be warned that the Jev "jaggedness" documentation specifically notes adversarial content as something Jev is very susceptible to: https://docs.typesafe.ai/model-jaggedness/jev-1.13#adversari... - so using Jev itself as part of a prompt injection guard is risky.

Anthropic, OpenAI, and Muse all use regular LLM calls to protect against prompt injection now and seem to have evals that give them confidence in doing that, so at least they think their own models are up to the task.

simonwon GPT-6.1 Sol replaces GPT-6 Sol after just 7 days, with near-Astra intelligence

It makes sense for them to sit on it until they've finished testing it. More powerful but also more likely to delete all your email by mistake = you shouldn't release it yet.

simonwon Livenerf: Has Opus 5.5 been nerfed yet?

> Anthropic has admitted to nerfing in the past

Where?

simonwon GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price

Sorry about that, markdown bug, now fixed.

simonwon DevDay 2026 Recap

Yeah, the front dozen rows were reserved for OpenAI employees.

simonwon GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price

I'm a bit late with the pelicans because I was live-blogging the keynote: https://simonwillison.net/2026/Sep/29/openai-devday-2026-liv...

Here they are for GPT-6.1-Sol: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

They're not notably different from the GPT-6 family pelicans: https://static.simonwillison.net/static/2026/gpt-pelicans-gr...

simonwon Anthropic's IPO prospectus shows AI vision, surging costs

Yeah, I agree with all of that.

simonwon Anthropic's IPO prospectus shows AI vision, surging costs

If Uber think there's no ROI, why do they allow each of their employees to spend $1,500 per month per tool?

Shouldn't they have set that per-employee budget to zero instead?

(I dug up the original source for that Uber doubts the ROI story a few months ago, it's a lot weaker than the headlines about it suggested: https://simonwillison.net/2026/May/27/product-market-fit/#th...)

simonwon Anthropic's IPO prospectus shows AI vision, surging costs

> If they had a trillion dollars worth of revenue, it wouldn't mean jack shit if they had a trillion + 1 in losses.

I don't understand that argument.

If a company has a trillion dollars in revenue, even if they are losing money hand-over-fist, that still means they have convinced other companies to cough up a trillion dollars for what they are selling. That's a big deal!

The only case that isn't impressive is if they are literally selling dollar bills for 90 cents.

You can argue that Anthropic are subsidizing their tokens all you want, but since as a customer you can't just turn around and sell a token yourself for more than you paid for it that's still not a good argument for dismissing the amount people are willing to spend.

simonwon Anthropic's IPO prospectus shows AI vision, surging costs

https://www.anthropic.com/claude-opus-5-5

"Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that. Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads."

https://www.anthropic.com/claude-sonnet-5-5

"Sonnet 5.5 requires fewer tokens per task than Sonnet 5, so it’s less expensive to run. It also generates output 30%+ faster"

OpenAI have been achieving even more impressive optimizations, hence why GPT-6 Sol and GPT-6 Luna are half the price of their 5.6 equivalents.

simonwon Anthropic's IPO prospectus shows AI vision, surging costs

I assembled some of those numbers in May, when they claimed they had grown from $9bn in annualized revenue in December 2025 to $47bn in early May. https://simonwillison.net/2026/May/29/anthropic/

Since then they've reported $65bn in annualized revenue by July: https://simonwillison.net/2026/Aug/23/anthropics-best-ai-mod...

And sure, they might be lying about those figures - but if they are, that's investor fraud, and they'll be in hot water with the SEC when they try to IPO. I don't think they are lying about the figures.

If I had numbers on their cost of revenue I would share those. As it stands I'm going to have to wait for either more leaks or their S-1.