HNHacker News
TopNewBestAskShowJobs

luke-stanley

555 karma · joined April 29, 2012

Old Keybase stuff. Do people still use it?

[ my public key: https://keybase.io/lukestanley; my proof: https://keybase.io/lukestanley/sigs/1pBLoHuRbRQHHwd8sEqxZCe1L6gzWInT4UBKthawE2c ] 8e47ab2600d05031bdbc2a0547fb9b79a82aeb240ef79cb82c5536e286f4d4c9

I am also https://news.ycombinator.com/user?id=lukestanley I think I forgot my pass and didn't want to bother PG.

submissionscomments
luke-stanley··on Microsoft 365 Copilot – Arbitrary Data Exfiltration via Mermaid Diagrams
The Lethal Trifecta strikes again! Mermaid seems like a bit of a side issue, presumably there are lots of ways data might leak out. It could have just been a normal link. They should probably look further into the underlying issue: unrelated instruction following.

Thanks for the archive link and the very useful term BTW! I also got 503 when trying to visit.

luke-stanley··on Noise cancelling a fan
Perfectly? Surely that is not reasonable?
luke-stanley··on Apple Photos app corrupts images
Just a warning, I don't think that ZFS snapshots are sufficient for ZFS bitrot protection.

Ensuring ZFS has at least 2 copies on physically separate disks and using scrub frequently is the way, right?

Please correct me if I am wrong HN!

luke-stanley··on Apple Photos app corrupts images
I ran into the same corruption issue without any sort of RAW images and without using OM camera. My partner was pissed at me, thinking it was my fault. But it turns out that Apple Photos is to blame!
luke-stanley··on ‘Overworked, underpaid’ humans train Google’s AI
It's strange that the Guardian mentions OpenAI's "O3" model and not GPT-5. Maybe they think o3 is SOTA still, but they should at least name it correctly, in lowercase as OpenAI does.
luke-stanley··on 3D modeling with paper
Human problem? It's probably already solved by one of the many recent machine learning papers, often there is source on GitHub and Transformer models on HuggingFace or some random Google Drive or Biadu drive. So one such human problem is finding how to ask aXiv Assistant what the best SOTA papers for it are and searching for if they finally released code or not (hoping researchers have a real repo not a GitHub site without code). I recall that Nvidia have some clean solutions. I wish it was a more pure principled solver though with some clean code. Probably OpenEvolve could iterate on a solution to it like the circle packing problem example but 3D. Sometimes it's funny to think that there are human problems left, which itself really is a human problem.
luke-stanley··on Show HN: Pbar.io – Distributed progress bars that work in terminals and browsers
You've got a great start, I agree that your swap-in interface for tqdm is very important for migration, and so you solved at the right level first, and you've found an interesting problem / solution area with beautiful execution.

I can totally see a "social layer for progress bars" and "Deep Research for progress bars" as valuable things to do on top of it. Like how decision markets and search fit in. Where is reputation and trust? Is metadata encryption needed?

There are developer UX improvements possible of course, as with most things. How exactly parent and child relationships and completions are best represented for an intuitive UX is an interesting circle to square while keeping a simple tqdm API.

I'd love it if this was open-source or zero-knowledge, though that may be a bit of a loaded term when actually wanting client side title / description metadata encryption. I guess you'd want a client doing simple but robust encryption on the metadata, but probably not progress values (or at least not at first, for sanity).

Any plans for a MCP API? It's pretty easy these days.

luke-stanley··on Show HN: Pbar.io – Distributed progress bars that work in terminals and browsers
Like what about key life events?
luke-stanley··on Show HN: Pbar.io – Distributed progress bars that work in terminals and browsers
A hierarchical progress bar service is a cool idea in principle, for some tasks state is worth decoupling from process memory. I do note that the homepage example changes colour without real hierarchy shown though! Something worth noting, is that the hierarchy of an important task of some kind might not be known at the start. It's interesting to consider how to have a great developer UX for managing this. I'd probably want MCP support (or a boring OpenAPI). If someone can show in a few lines how a developer can have a messy process get meaningful hierarchy and progress added to it and it's child nodes, without being overly rigid or SaaS externality concerns, that would be really cool, it could perhaps help give feedback on some very intractable problems. People often don't realise that showing useful progress can easily be a really hard distributed systems problem that a lot of real life systems are bottlenecked by. It's interesting that LLMs might actually be able to help ask questions about progress to help better represent what progress should be shown as, asking about subproblems, helping guide the state with structured decision making etc.
luke-stanley··on Google will allow only apps from verified developers to be installed on Android
They are namespacing, like it or not, and clearly they don't care about open-source that much.
luke-stanley··on Google will allow only apps from verified developers to be installed on Android
So Google won't even offer a system toggle to let users install an app they've made or copied?

Google don't even expose a per-app toggle for app Internet access, why am I surprised?

This is disgusting.

Freedom died a little bit more today.

Why is end-user choice and consent not considered?

It's really disturbing that the EU and Google would do this.

I can't recommend Android or iPhone because of this nonsense.

luke-stanley··on Google will allow only apps from verified developers to be installed on Android
So sketchy!
luke-stanley··on Google will allow only apps from verified developers to be installed on Android
A repo is just files in a directory, so the namespace can be changed, but the whole thing stinks. Having to setup Android signing keys and needing to provide ID is not fun. It means you won't easily be able to run builds on Google certified Android devices that aren't from "approved" people.
luke-stanley··on AI Mode in Search gets new agentic features and expands globally
The blog post has so much noisy, duplicate content and slop, it's like wading through treacle. This is Google. Supposedly an advertising and marketing company, describing what used to be their core product. Such a disappointment.
luke-stanley··on "Stop children using VPNs to watch porn", UK politicians told
Surely VPN photo ID checks would increase water usage!?!?

Jokes aside, teenagers can be smart at evading censorship. Should a democratic society play a cat and mouse game with a more motivated and more nimble adversary, who will often be smarter too?

Are they not aware that it's something they would have to force on everyone's device? GDPR cookie banners are a good example of poorly executed government meddling, the Online Safety Bill could yet be even more annoying, and we're still learning about the lies and ineptitude seen by the Post Office scandal. Why entertain playing a cat and mouse game like that? I don't think we want to see how invasive government computer meddling can be. People coming up with these ideas need to think really carefully about what the actual results could be once they have gone through committee and started bothering their constituents.

Perhaps more importantly, they should remember 16 year olds are going to be able to vote soon.

luke-stanley··on Sam Altman is in damage-control mode after latest ChatGPT release
Yeah, OpenAI do have harmfulness classifiers in ChatGPT that can detect problems in their own responses before it finishes, and learning how confident GPT-5 or the image generator tool call should be about meeting the brief, after it finishes describing the visual concept, might be a task OpenAI could train a classifier on. But reliably predicting ahead of time can be a really hard problem to solve; knowing how successful a complicated tool will be, before it actually starts or finishes the task attempt is tricky.

After gpt-image-1 has produced an image is another helpful intervention point, it can do a better self-review for detecting problems after image generation, but it's still not very thorough. However OpenAI has small teams, they try to keep them small and really focused, and everything is always changing really fast, they probably have gpt-image-2 or something else soon anyway.

In a way, reliable prediction is the main job OpenAI has to solve, and always has been. Some researches say the main way models are trained causes "Entangled Representations", which makes them unreliable. They also suffer from the "Reverse Curse". Maybe when they fix these issues, it might be real AGI and ASI all in one go?

luke-stanley··on Sam Altman is in damage-control mode after latest ChatGPT release
"The custom is always right", maybe so, but CNN has a duty to fact check the accuracy of their central claims.

ChatGPT's image generation was not introduced as part of the GPT-5 model release (except SVG generation).

The article leads with "The latest ChatGPT [...] can’t even label a map".

Yes, ChatGPT's image gen has uncanny valley issues, but OpenAI's GPT-5 product release post says nothing about image generation, it only mentions analysis [1].

As far as I can tell, GPT-Image-1 [2], which was released around March, is what powers image generation used by ChatGPT, which they introduced as "4o Image Generation" [3], which suggests to me that GPT-Image-1 is a version of the old GPT-4o.

The GPT-5 System card also only mentions image analysis, not generation. [4]

In the OpenAI live stream they said as much. CNN could have checked and made it clear the features are from the much earlier release, but instead they lead with a misleading headline.

It's very true that OpenAI doesn't make it obvious how the image generation works though.

[1] https://openai.com/index/introducing-gpt-5/

[2] https://platform.openai.com/docs/models/gpt-image-1

[3] https://openai.com/index/introducing-4o-image-generation/

[4] https://cdn.openai.com/pdf/8124a3ce-ab78-4f06-96eb-49ea29ffb...

luke-stanley··on Sam Altman is in damage-control mode after latest ChatGPT release
"can't even label a map" - Image generation is not using GPT-5, they said this in the live stream. CNN should try harder.
luke-stanley··on A telephony agent for my parents. Should I turn it into a full-fledged service?
Exactly, and by the way, this is why I did not upvote it! Others probably have the same feeling.
luke-stanley··on UK government advises deleting emails to save water
Deleting emails and old pictures is hard work and requires putting the kettle on, they should know this. This is very silly.

A study on this would be pretty interesting.

This is a national disgrace and I am ashamed.

luke-stanley··on Anthropic revokes OpenAI's access to Claude
Anthropic are wicked smart, it's not fair to call them morons, BUT the ToS ban of subscribed, myself, and OpenAI might be useful data points suggesting Anthropic's API moderation behaviour can be strict. Anthropic are doing great work overall, but they have erred on the side of policy robustness, versus keeping API users happy. I said more in my sibling comment reply.
luke-stanley··on Anthropic revokes OpenAI's access to Claude
Yeah and Wired could just write it in a clear way so that no disambiguation or head scratching is needed.
luke-stanley··on Anthropic revokes OpenAI's access to Claude
If Wired wants to portray normal access to Anthropic's API platform as a special fringe activity, rather than a normal way to programmatically use AI, it really says something about Wired. And this is Hacker News, right? Should we be on some watch list or something for thinking having control via API access is normal dev access!? MCP isn't even that old yet! ;D It's possible to write clearly and not that hard, I'm pretty sure they are hyping.
luke-stanley··on Anthropic revokes OpenAI's access to Claude
Yeah that's possible! Would love more details if true.
luke-stanley··on Anthropic revokes OpenAI's access to Claude
Well I don't think Anthropic are morons, that's not the point I was making.

Yes, I'm frustrated with Anthropic killing my direct API account for silly reasons, with no response. But actually I really appreciate Anthropic's models for code, their deep safety research with Constitutional AI, interpretability studies etc.

They are certainly guilty of having scaling and customer service issues, and making the wrong call with a faulty moderation system (for you too, and many others it seems like)!

But a lot of serious AI safety research that could literally save all our skin is being done by Anthropic, some of the best.

On OpenAI's API Platform, I am on Tier 5! It's unfortunate Anthropic have acted less commercially savvy than OpenAI (at the time, at least). I have complained on HN and I think on Twitter before about my account to no avail, after emailing before. But yeah, usually I just use them via OpenRouter these days, it's a shame that I must use it for API access.

I get the impression that a lot of OpenAI researchers went to Anthropic, which essentially is the first OpenAI splinter group. I think this is a sign of a serious, more healthy intellectual culture. I'm looking forward to seeing what they do next.

luke-stanley··on Anthropic revokes OpenAI's access to Claude
"OpenAI was plugging Claude into its own internal tools using special developer access (APIs)"

Unless it's actually some internal Claude API which OpenAI were using with an OpenAI benchmarking tool, this sounds like a hyped-up way for Wired to phrase it.

Almost like: `Woah man, OpenAI HACKED Claude's own AI mainframe until Sonnet slammed down the firewall man!` ;D Seriously though, why phrase API use of Claude as "special developer access"?

I suppose that it's reasonable to disagree on what is reasonable for safety benchmarking, e.g: where you draw a line and say, "hey, that's stealing" vs "they were able to find safety weak spots in their model". I wonder what the best labs are like at efficiently hunting for weak areas!

Funnily enough I think Anthropic have banned a lot of people from their API, myself included - and all I did was see if it could read a letter I got, and they never responded to my query to sort it out! But what does it matter if people can just use OpenRouter?

luke-stanley··on Gemini Embedding: Powering RAG and context engineering
There are plenty of vision capable embedding models, you might not need to OCR, and doing so may could improve or hurt performance.
luke-stanley··on Positron – A next-generation data science IDE
Positron: Consider starting with Astral UV! Fast and reproducible, with early fixes to evade CUDA dependency hell.
luke-stanley··on Show HN: Intercepting proxy for semantic search over visited pages
I made another approach that detects browser tab text changes with a Userscript that sends changes as Markdown to a server that appends to a JSONL file. MitM might be interesting to compare: https://gist.github.com/lukestanley/c3a37ab61a45e72b74995a5c... (two files here BTW)
luke-stanley··on Show HN: Intercepting proxy for semantic search over visited pages
Sending requests to pure.md (?) seems pretty insecure and inefficient to me. If you're going to MitM yourself, KEEP IT LOCAL, then use reliable code to convert HTML to Markdown efficiently with a queue, drop text parts that are noise, summarise if that's the format desired, and THEN embed!
← PreviousPage 2 of 14Next →