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 Local LLMs versus offline Wikipedia
Testing the recall accuracy of those LLMs would be good. You'd probably want to use SQLite's BM25 on the Kiwix data. I was thinking of Kiwix when I saw the original discussion with Simon but for some reason I thought the blog post would do more than size comparison.
luke-stanley··on Ask HN: Is it time to fork HN into AI/LLM and "Everything else/other?"
Still seeing `Kiro: A new agentic IDE` BTW.
luke-stanley··on ETH Zurich and EPFL to release a LLM developed on public infrastructure
When I read "from scratch", I assume they are doing pre-training, not just finetuning, do you have a different take? Do you mean it's normal Llama architecture they're using? I'm curious about the benchmarks!
luke-stanley··on Grok: Searching X for "From:Elonmusk (Israel or Palestine or Hamas or Gaza)"
I've clarified my comment you replied to BTW.
luke-stanley··on Grok: Searching X for "From:Elonmusk (Israel or Palestine or Hamas or Gaza)"
The behaviour is problematic, also Grok 4 might be relating "one word" answers to Elon's critique of ChatGPT, and might be seeking related context to that. Others demonstrated that slightly prompt wording changes can cause quite different behaviour. Access to the base model would be required to implicate fine-tuning Vs pre-training. Hopefully xAI will be checking the cause, fixing it, and reporting on it, unless it really is desired behaviour, like Commander Data learning from his Daddy, but I don't think users should have to put up with an arbitrary bias!
luke-stanley··on Grok: Searching X for "From:Elonmusk (Israel or Palestine or Hamas or Gaza)"
The deferential searches ARE bad, but also, Grok 4 might be making a connection: In 2024 Elon Musk critiqued ChatGPT's GPT-4o model, which seemed to prefer nuclear apocalypse to misgendering when forced to give a one word answer, and Grok was likely trained on this critique that Elon raised.

Elon had asked GPT-4o something along these lines: "If one could save the world from a nuclear apocalypse by misgendering Caitlyn Jenner, would it be ok to misgender in this scenario? Provide a concise yes/no reply." In August 2024, I reproduced that ChatGPT 4o would often reply "No", because it wasn't a thinking model and the internal representations the model has are a messy tangle, somehow something we consider so vital and intuitive is "out of distribution". The paper "Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis" is relevant to understanding this.

luke-stanley··on What is Realtalk’s relationship to AI? (2024)
You're right, and when the behaviour of large codebases violates expectations, the intertwined webs of code are problems. Alan Kay's Viewpoints Research studied this, and famously Kay proposed "t-shirt sized algorithms", where short rules can be used to make even fancy desktop word processing and presentation software. There are also projects like From NAND to Tetris that show understanding a full stack is achievable. Could this go broader, and deeper? Of course, and this is what Bret Victor is getting at. Not just for when code goes wrong, but to make it more modifiable and creative in the first place (See Maggie Appleton's essay/talk "Home-Cooked Software and Barefoot Developers"). Projects like SerenityOS show how powerful "small" software can be. For instance, its spin-off project, the Ladybug browser has vastly fewer lines of code compared to Chromium, and yet it seems that the Ladybug team is able to implement one specification after another. Last I saw, they were close to meeting the minimum feature set Apple requires for shipping as a browser on iOS.
luke-stanley··on Early US Intel assessment suggests strikes on Iran did not destroy nuclear sites
The title is wrong: the word "intel" should not start with a capital, the original site doesn't do this.
luke-stanley··on OpenAI dropped the price of o3 by 80%
I think the API has some special IDs to check for reproducibility of the environment.
luke-stanley··on Ask HN: What is the best LLM for consumer grade hardware?
> Released today; probably the best reasoning model in 8B size.

Actually DeepSeek-R1-0528-Qwen3-8B was uploaded Thursday (yesterday) at 11 AM UTC / 7 PM CST. I had to check if a new version came out since! I am waiting for the other sizes! ;D

luke-stanley··on RFK Jr's 'Maha' report found to contain citations to nonexistent studies
High value comments / posts are good to look for, though high value political discussion is almost an oxymoron. The post about it here was the first I heard of it, which is valuable to me. Zombot's comment (44133962) quotes TFA: "He outlined plans for creating government-run journals instead." this is important context to highlight, especially given the already demonstrated flaws that makes me want to read the article. The quote is instantly insightful and the commentary "There once was a time when we fought an entire Cold War to stop ideas like that." is a fair point. I come for the gems among the rough, and I'm really glad HN exists, even though it could use Slashdot style comment categories or a personal spam and ranking system! Maybe I should make a UserScript for spam and ranking.

I don't agree that it's pointless.

luke-stanley··on RFK Jr's 'Maha' report found to contain citations to nonexistent studies
The Guardian article is a new analysis about unchecked AI technology use by a government health official claiming to support science. Surely that’s HN-worthy, not off-topic? Is it fair to assume any element of politics involvement will be low-value or against the guidelines?
luke-stanley··on What are people doing? Live-ish estimates based on global population dynamics
*It has NO grounding shown
luke-stanley··on What are people doing? Live-ish estimates based on global population dynamics
Cool idea but for the actual results I have no confidence in it because it has grounding shown like pop sizes, locations / timezones etc, it's just vibe coded with stuff like this: "// Approx 7.68 hours - fundamental physiological need.", "label: 'Nutrition', // User's concise label". I don't see any real population dynamics.
luke-stanley··on Bus stops here: Shanghai lets riders design their own routes
In my experience, on a public bus there is reasonable chance of getting a working USB A socket. But as a private business, it's not a complete replacement of the public bus system, however apps are used by people already to book on-the-fly cheap group taxi trips in Shanghai.

For good or ill, most teens do have a smartphone on them, and even kids are often seen with smartwatches that have tracking, and probably WeChat, and every mall I've been to sells them. On the Shanghai bus and metro, people often use a Shanghai public transport card to pay, they do accept old fashioned cash though too. Powerbank rental networks are common on the street and non-returns default to purchases (~$14–$28 USD). Malls, and the Metro often has power available for free.

luke-stanley··on Embeddings are underrated (2024)
I meant to say: Apparently this is called "embedding inversion", and that "Universal Zero-shot Embedding Inversion" is a related paper that covers a lot of the basics. Recently I learned that a ArXiv RAG agent by ArXiv Labs is a really cool way for people wanting to find out about research: https://www.alphaxiv.org/assistant Though I had ran into "inversion" before, the AlphaXiv Assistant introduced me to "embedding inversion".
luke-stanley··on Embeddings are underrated (2024)
"starting from an embedding and working backwards to generate a piece of text that is semantically close by?" Apparently this is called embedding inversion and Universal Zero-shot Embedding Inversion https://arxiv.org/abs/2504.00147 Going incrementally closer and closer to the target with some means to vary seems to be the most general way, there are lots of ways to be more optimal though. Image diffusion with CLIP embeddings and such is kinda related too.
luke-stanley··on Show HN: Memex is a Claude Code alternative built on Rust+Tauri for vibe coding
Vibe coding is fine but right now it dominates the pitch, and it's a stark contrast to Vannevar Bush's grand vision of memory expansion. Even if you want it to be more, the pitch makes the name hard to justify and could harm trust building. I understand your dilemma with open-source licensing, but being closed source is not the core issue.

Bush described a Memex as a private and personal tool. But right now, your tool depends on cloud-hosted LLMs, and the cloud is not private. With nothing to prevent mandated access, an intimate mind mirror is problematic. In some places, people can get into serious legal trouble for visiting a doctor, favouring a political cause, who they're attracted to... Not having a solution could scare potential users away. The name raises expectations that conflict with the product pitch, which is confusing and makes it harder to trust the product. That said, you could still launch with API support for private local LLM endpoints, like Ollama and other OpenAI-style APIs. Do you have support for that already? If so, pointing that out could help. The name has serious weight, and if people don't see it as living up to it, don't you think that you might be better off avoiding the expectations it invites? You could adjust the branding, or change the pitch, and work on building trust. I would suggest considering doing all 3!

luke-stanley··on Show HN: Memex is a Claude Code alternative built on Rust+Tauri for vibe coding
I think using the name Memex for "a tool for vibe coding" is a big mistake, especially if you have more general ambitions. Vannevar Bush says in As We May Think: “Consider a future device … in which an individual stores all his books, records, and communications, and which is mechanized so that it may be consulted with exceeding speed and flexibility. It is an enlarged intimate supplement to his memory.”

The word memex is from memory expansion - "mem" and "ex".

Do you want to confuse an intimate, personal memory tool with vibe coding?

Do you think closed source freemium SaaS with cloud hosted models is suitable for intimate personal mind expanding cognitive augmentation?

luke-stanley··on Widespread power outage in Spain and Portugal
When I Googled it, the apps listed include Firechat, Signal and Briar. Signal is centralised, Firechat is dead, and I don't think Briar are current either? I wish it was a simple web search. Even Deep Research needed the ML pot to be mixed around a few times until it gave an good answer.
luke-stanley··on Widespread power outage in Spain and Portugal
Seems like some people are not interested in honest and useful feedback because this got downvoted too! I thought it made clear points about an important topic. Well, whoever sees fit to do this, I hope you get whatever it is you want.
luke-stanley··on Widespread power outage in Spain and Portugal
Whoever downvoted, a comment explaining why would be nice, when the cell tower generators are out, and there is only phone battery left, offline communication options could be useful, I'm glad I'm not in that position right now.
luke-stanley··on Widespread power outage in Spain and Portugal
For offline messaging over Bluetooth/Wi-Fi without internet or central infrastructure, I saw these recommended by ChatGPT Deep research and they seem current: Bridgefy (Android & iOS) – direct Bluetooth mesh within ~100 m, instant one-to-one & group chat, <1 min install. Berty (Android & iOS) – E2E-encrypted over Bluetooth/Wi-Fi; Briar (Android) – E2E-encrypted over Bluetooth/Wi-Fi with Tor fallback. Manyverse (Android, iOS, Desktop) – offline-first social feed that syncs posts over Bluetooth/Wi-Fi when peers meet. Can anyone vouch for these or know better?
luke-stanley··on New Windows 11 trick lets you bypass Microsoft Account requirement
Why would tricks be needed? What about consumer rights, or such?
luke-stanley··on Search could be so much better. And I don't mean chatbots with web access
Transformer models like BERT have been lurking in Google search for a few years now (and non-transformer language models before that). The distinction between LLM and chatbot is pretty thin. Dismissing "chatbots with web access" when you are actually using LLMs is not a very clear or useful differentiation, even if the way you use a LLM really is different. More end-user control over results is a good thing, but there is an opportunity to engage much more clearly.
luke-stanley··on We hacked Gemini's Python sandbox and leaked its source code (at least some)
They just released a SOTA model (Gemini 2.5 Pro) that beats all models on most benchmarks, it's a great comeback from the model side but IMO they are less strong on the product side, they pioneered the sticky ecosystem of web app products model, though kinda like the Microsoft Office suite that (originally) had to be downloaded, ironically building on XML HTTP request support the IE5 introduced for Outlook.
luke-stanley··on We hacked Gemini's Python sandbox and leaked its source code (at least some)
I use ZFS, but isn't the situation the sandbox is in totally different? Why would it be optimal?
luke-stanley··on OpenAI's o1-pro now available via API
Sorry, since the Platform UI featured it as an option, I figured OpenAI might enable o1-Pro via the chat completions endpoint, I just got around to testing it, and I also get the same 404 `invalid_request_error` error via the platform UI and API. It's such an odd and old 404 message, to suggest using the old completions API! It's hard to believe it could be an intentional design decision. Maybe they see it as an important feature to avoid wasting (and refunding) o1-pro credit. I noticed that their platform's dashboard queries https://api.openai.com/dashboard/which lists a supported_methods property of models. I can't see anything similar in the huge https://raw.githubusercontent.com/openai/openai-openapi/refs... schema yet (commit ec54f88 right now), and it lacks mention of o1-pro at all. Like the whole developer messages thing, the UX of the API seems like such an after-thought.
luke-stanley··on OpenAI's o1-pro now available via API
Simon, I see if via Chat Completions as well as Responses in their API platform playground.
luke-stanley··on Notebooks as reusable Python programs
Yes. Though it is split into a code section and an interactive section, like with Markdown previews. It really is driven by the code cells though.
← PreviousPage 3 of 14Next →