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 Gboard Double Sided Version
-Use an old phone remotely as a wireless gesture typing keyboard (forwarding the keyboard I/O)

-Programmer mode with SSH sanity

-Custom tiny local language models for better, wiser, more funny suggestions that the end user could have full control of (the sky is the limit for this)

-Maybe I want to use my task list as context for typing faster, maybe I want a synced knowledge network

-Maybe I can want to search for "man" emoji and have it really work

-Maybe I want real time translation built in, using a custom endpoint running on my GPU or latest Whisper distilled model? Maybe a new one that can do Markdown sarcasm formatting?

I could go on and on.

Text remains a very powerful input and output. They're not as smart as everybody who could be hacking on it, people love hacking on keyboards, for good reason.

Gboard has many great features, privacy could even be one of them in the future! Swiping and it's custom models, that kinda know what weird topics I've typed, but not full GPT level, are pretty neat and make it worth considering, but hey, open source will get better and better. We live in an amazing time, with antitrust break up of Google being considered, open sourcing it and continuing to work on it with open source community could be a small win that could keep giving, just saying. Typed in on Gboard!

luke-stanley··on Gboard Double Sided Version
I get that this is one of their jokes, to get attention and for fun. But it would be great if Gboard itself was open source though. I could do a ton of improvements for myself at least. I love the slide / swipe typing gesture input mode in Gboard but it's needlessly closed source. There could be big privacy improvements and great features added. Maybe AnySoftKeyboard is good enough now, I'll check again.
luke-stanley··on Show HN: FinetuneDB – AI fine-tuning platform to create custom LLMs
This sounds interesting to me! I will check it out in more detail. I saw 3.5-turbo mentioned which is more expensive and as I understand it, it's usually less good as a base model, if I see 3.5-turbo before I see 4o-mini, and don't see 4o-mini, I might wonder if things are behind! I hope that's fair feedback for a quick reaction. I have done a bunch of fine-tuning before, locally and with 4o-mini, and there's often a lot of time spent suboptimally on wrangling data so I'm interested in this category of product for sure, if it helps more than it costs me.
luke-stanley··on Homemade AI drone software finds people when search and rescue teams can't
They might have needed to learn what a good difference threshold and cluster size is. It's hardly ML like fine-tuning CLIP embeddings is, but there are few solid differences: both explore visual embedding spaces with learned values. Granted, cluster thresholds are more likely to be manually learned, but they are both embedding spaces, with the main difference being dimensionality.

It's very vague for Wired to have used AI in the title, but it's more confusing to say "A previous headline on this piece incorrectly stated that the drone software used AI." - and not obviously correct either.

luke-stanley··on Show HN: A real time AI video agent with under 1 second of latency
Pretty cool but it seems like the mouth / lip-sync is quite a bit off, even for the video generation API? Is that the best rendering, or are the videos stale?

Also the audio cloning sounds quite a bit different from the input on https://www.tavus.io/product/video-generation

For live avatar conversations, it's going to be interesting, to see how models like OpenAI's GPT-4o that have audio-in-audio-out websocket streaming API (that came out yesterday), interesting to see how that will work with technology like this, it does look like there is likely to be a live audio transcript delta that could drive a mouth articulation model, and so on, that arrives at the same time.

Presumably Gaussian Splatting or a physical 3D could run locally for optimal speed?

luke-stanley··on Smart TVs take snapshots of what you watch multiple times per second
We have a Samsung TV in the UK (where consent is still required too!) and I recall it nagging for consent very often when we got it.

The default option from the TV controller is to give it quickly and easily, maybe it could happen by sitting on the controller.

It's hard to see how this would be informed consent.

I do wonder exactly how the EU and UK market models from the US and others, anyone got info?

luke-stanley··on Llama 3.2: Revolutionizing edge AI and vision with open, customizable models
The Llama 3.2 11B multimodal model is a bit less than 14B but smaller models can do more these days, and Meta are not the only ones making models. The 70B model has been pruned down by NVIDIA if I recall correctly. The 405B model also will be shrunk down and can presumably be used to strengthen smaller models. I'm not convinced by your shiny hat.
luke-stanley··on Alan Turing's 1950 manual for the Mark I electronic computer [pdf]
Thanks.
luke-stanley··on Show HN: A tool to analyze Hacker News sentiment on any term in seconds
`Sorry, there was an error processing your message.` It got neutral for stuff I wasn't expecting to be neutral, it shows a frequency graph but sentiment range / variability might be worth showing.
luke-stanley··on Show HN: Sourcetable – AI Spreadsheet and Data Platform
Actually Microsoft do now have Copilot and Python in Excel recently released last week. Maybe a bit slow.
luke-stanley··on Launch HN: Deepsilicon (YC S24) – Software and hardware for ternary transformers
I've not but that sounds cool! I would point out though, in terms of mind share, how memorable, and how relatable and useful the products are: it might help to have ways that directly show the application for the kinds of people buying GPUs for inference and training or using cloud for this that would love to not have to fight their ATX case in a hot sweaty corner while repeatedly dropping screwdrivers and calculating how much RAM they need to buy for the 405B while llama.cpp is recompiling again... I think people would throw money at that. I'd be happy to listen in or have a chat some time!
luke-stanley··on Launch HN: Deepsilicon (YC S24) – Software and hardware for ternary transformers
The most popular interfaces (human, API and network) I can imagine are ChatGPT, OpenAI compatible HTTP API, Transformers HuggingFace API and models, Llama.cpp / Ollama / Llamafile, Pytorch. USB C, USB A, RJ45, HDMI/video(?) If you can run a frontier model or a comparable model with the ChatGPT clone like Open UI, with a USB or LAN interface, that can work on private data quickly, securely and competitively to a used 3090 it would be super badass. It should be easy to plug in and be used for running chat or API use or fine-tune or use with raw primitives via Pytorch or a very similar compatible API. I've thought about this a bit. There's more I could say but I've got to sleep soon... Good luck, it's an awesome opportunity.
luke-stanley··on Wind Becomes Britain's Largest Electricity Source
Well, blow me down. I better boil a cupa.
luke-stanley··on Sleep duration, chronotype, health and lifestyle factors affect cognition [pdf]
Some might just be avoiding falling asleep or being even less capable.
luke-stanley··on UK rail minister got engineer sacked for raising safety concerns
Well, when I was in Euston rail station a few weeks ago, it was very overcrowded. It seemed worse in the day than the night. Seems like the minister is missing the necessity of acting with integrity and transparency, a lesson they frequently need reminding of. Surely there must be better person the PM could find for the job, that don't feel a need write harassing letters, bullying train companies into firing staff?
luke-stanley··on Show HN: A directory to find open source AI projects easier
I think this could be a really useful idea, though it's pretty rough around the edges right now. I noticed the related project results quality was much lower than I expected, very `noisy`. Perhaps related projects could be populated with hybrid search and RAG? E.g: using high quality embeddings (from all metadata you can fit in the embedding context window, names, descriptions, tags etc), and text search. You could explore the validity of different possible matches with very few tokens using a cheap model like GPT-4o-mini or Llama 3.1-8B (from a number or providers), this could even be done in parallel for each result item. It might even be worth fine-tuning an embedding model. Understanding the topic space, and what relate projects are in it is really well would make it much more useful. Submission should not be needed for projects that are already on GitHub or GitLab. Crawling search APIs to search for related projects ahead of time, or on demand might make sense, if GitHub don't have a topic API that matches their site's shown results, you might need to scrape, or use a third-party search API.
luke-stanley··on Faster CRDTs (2021)
Could be good to have the date put with the title?
luke-stanley··on Show HN: A Ghidra extension for exporting parts of a program as object files
This seems cool. Maybe one day it could help people take apart existing programs and use them how they want in a more accessible way. There is research like Meta's LLM Compiler and decompiling with LLM's, breaking apart programs and replacing parts of them would be an interesting, data rich space for LLM's to explore and maybe self-improve at! For people training, there are lots of interesting tokens lurking there, and interesting things to do with them!
luke-stanley··on Show HN: Bullshit-detector – quickly detect amount of bullshit in any text
Yeah, presumably words with vague meanings, low semantic quality are easy to find. Still, a Bert model for it would be fun.
luke-stanley··on Show HN: Bullshit-detector – quickly detect amount of bullshit in any text
Presumably, the identified phrase list could be used to finetune a Bert model or similar that could catch more cases, as a binary classifier. But presumably some actual semantically meaningful words would be needed too. That would be straight forward to do too though. Someone has probably already done it. The advantage would be you could get probability metrics on a broader set of text. Good data is the key thing though.
luke-stanley··on LLMs for structured input validation
You might want to try Instructor with llm_validator, where you can have natural language clauses to check against in a loop: https://python.useinstructor.com/blog/2023/10/23/good-llm-va... With Pydantic too of course!
luke-stanley··on Show HN: AdalFlow: The library to build and auto-optimize any LLM task pipeline
This sounds quite cool, and I am very picky but this seems quite clean. It wasn't clear if this was for fine-tuning or not, the general explanation could perhaps be clearer, the first few sentences of the README still don't make it very clear (I can sort of guess, having seen DSPy, AutoPrompt etc). It would be awesome if this DID also explore if it needed to do: prompt tuning, fine-tuning, or soft-prompt tuning etc, I am on the lookout for a tool that does this. Obviously a general open source Q* like solution would be amazing but I get that might be a bit of a different beast! Part of my issue is that there are so many things that can be tweaked and I often don't know the most time and cost efficient thing to optimise. I get that prompt tuning is often going to be the best thing to do, especially first. But for efficient inference, shorter prompts may well be needed. Though maybe clever model key-value caching is starting to make this less of an issue, but it's still faster to have as short as prompts as possible, still, fine tuning or even resuming pretraining may be the best thing to do sometimes. BTW I would strip `gpt-3.5-turbo` from all examples, as it's more expensive than the better 4o-mini. I hope to check this out more later. Nice work!
luke-stanley··on Launch HN: Release (YC W20) – Orchestrate AI Infrastructure and Applications
Are there 2 levels of free usage, where the first sandbox doesn't require Google Cloud or AWS accounts, but after the trial, they are required?
luke-stanley··on Don't Pivot into AI Research
Exactly! Quality can be very important too, as a series of interestingly advanced small models have shown.
luke-stanley··on Don't Pivot into AI Research
Not all kinds of scaling requires high up front capital costs. Some of us are interested in exploring fundamental contours of intelligence here, and there are lots of fascinating threads to pull on.

Nothing wrong with exploring motivations though as this article does. Maybe the job market trends they speak of will be true but I hope that's not a primary motivation. There's exploration to do!

luke-stanley··on Translating All C to Rust (TRACTOR)
People could try but it would not help with the withheld datasets so much, and it would be possible to add more to it. If the withheld data was closed, and only available to an assessment system, gaming that would be pretty difficult. Scale.com's SEAL Leaderboards take a similar approach. The ARC Prize still exists too, and it's waiting for winners.

See https://scale.com/leaderboard

luke-stanley··on Tech CEOs are backtrack on RTO mandates–now, 3% want workers in office full-time
The title is different from what the article current title is, and the grammar of, "are backtrack" is wrong.
luke-stanley··on Translating All C to Rust (TRACTOR)
Surely this could be better pitched to researchers as just another AI benchmark, a bit like ARC Prize? ;) There could be some exiting C projects that are already public, with tests for feedback during development iteration and some holdout tests, and some holdout projects too with a leaderboard and prizes. For preferences about converted code quality, both automated assesment and human preferences could be ranked with Elo? Kaggle is made for this sort of thing I think? I'm sure Google Deepmind and others have some MCTS agents that could do a great job with a bit of effort.
luke-stanley··on Overcoming the limits of current LLMs
As I understand it: the Phi models, are trained with a much more selective training data, the Tiny Stories research was one of the starts of that, they used GPT-4 to make stories and encyclopedia like training data for Phi to learn from and code, which probably helps with logical structuring too. I think they did add in real web data too though but I think it was fairly selective.

Maybe something between Cyc and Google's math and geometry LLM's could help.

luke-stanley··on Story points are pointless, measure queues
`Measure queues` like `Work In Progress limits` from Kanban. Cleverly the article dodges mentioning those terms!
← PreviousPage 5 of 14Next →