HN user
Why let your dreams be dreams? This is a perfect example of following a hypothesis. I love when people dive into an esoteric subject and just go full swing. Reminds me a CGP Grey and the name Tiffany. Sometimes you just need to know.
Reminds me of the old construction photos for nuclear reactors in the US. Astoundingly complex machines at a massive scale getting out together at what now feels like impossible speed. I can't help but feel like a Roman 100 years after the fall staring up at aqueducts wondering how anyone every built such a thing.
I'm positive someone could show me an impressive thing we built recently. I don't feel like that is my point. Im just astounded those people in that time could build what they built with the tools they had as fast as they did.
Give me self host code any day. This feels like the bait and switch AWS likes to pull. Would rather rely on the server in my dresser drawer than AWS for as much as I possibly can.
That is my favorite part of GA. Gradient free optimization but it turns out making a good fitness function is hard and like 70% of the time it just exploits some assumptions or gap you have in your theories. Really reveals the problem in different ways that traditional ML.
Reminds me of good ol genetic algorithm search. Guess and check can be quite powerful, especially if you can toss in agent in the loop guidance.
Extended FAB not rendering correctly for Firefox on mobile from what I can see.
Idk, since letting my son stay up as late as he wants reading trips to the library have been weekly, even as he transitions to actual no image chapter books. As others have said we try to time box tech best we can. I may have to physically lift him from the bed some days but it is worth it to find him passed out with a book in his hand at 10pm.
Automatic self hosted transcription service. So nice to be able to get my thoughts all down as context for projects. Really accelrates things.
Location: St. Paul Minnesota
Remote: Yes
Willing to relocate: No
Technologies: Gen AI, Healthcare, VLM, Document Parsing, Integrations FHIR, SMART, RPA, OCR, Computer Vision, Audio Agents, Data Processing, AI Model Training
Resume/CV: https://drive.google.com/file/d/1TrFOAvpNJloO77H_wL8FxIuXiXy...
Email: robviren@gmail.com
As a product manager I have found people's understanding of what they want to be tenuous at best. They know they want it to be easier, less frustrating, and do the darn thing they want. But collecting the input of all the people with the power to make choices, boiling that all down, and actually figuring out what will work is a total bottle neck. Technology and programming have hardly ever actually been a blocker. Legacy systems, conflicting requirements, and ever shifting choices represent a much harder problem.
Like high gas prices leading to sudden releases of fuel efficient vehicles in the 70s during the embargo. I love that most indie games I can find will run on a toaster.
I'd say the primary benefit of LAN parties now is how capable "old" hardware is. I have several laptops I configure with Linux and several games and environments loaded up. Smoke test the whole thing before anybody gets here. This year we played UT2004 with custom maps and characters. That game ran flawlessly on a 10 year old laptop with no dGPU. Emulation as well, but I actually still have most of those systems still.
Long story short it is a compromise between the Mdewakanton Sioux, the state, and Canterbury to keep gambling exclusive to Mystic Lake Casino and Caterbury. The State flirted with the idea of expanding gambling and started moving in that direction, but Mystic basically now props up Canterbury, lets them have some amount of gambling on site, and the State did not expand gambling and generally is protective of Mystic. It's like a pressure release valve on gambling with some small kickbacks to the nearby legacy horse racing track. It's a strange local arrangement. Grew up in Prior Lake and had a few tribe members in school so this is just my bias on the arrangement.
Ex Vaddio PM here. Like 5 years ago all our firmware defaulted to requiring non-default passwords on setup. We also created a free windows application that can mass upgrade firmware and change auth if defaults were used. We tried!
Saw the Vaddio logo and had to chime in. Gotta stick up for my Minnesota devs.
So artificially productive you que up the crap you do and slowly release it?
Location: St. Paul Minnesota
Remote: Yes
Willing to relocate: No
Technologies: Gen AI, Healthcare, VLM, Document Parsing, Integrations FHIR, SMART, RPA, OCR, Computer Vision, Audio Agents, Data Processing, AI Model Training
Resume/CV: https://drive.google.com/file/d/1TrFOAvpNJloO77H_wL8FxIuXiXy...
Email: robviren@gmail.com
I too have found the models do well with Go. I will say despite the backwards compatibility guarantee library API changes, what counts as "good" patterns, and new language additions do add some friction to the experience. Almost always works but it can be a bit inconsistent in how the code shows up.
I love genetic algorithms and find using LLMs as part of them super compelling. I always find the fitness functions to be the most difficult part. The algorithm naturally tries to exploit any little gap you leave it in cheating. Best part is not needing back propagation in solving a problem. However that is also the worst part in all the solutions just being one level above a random walk. The LLM augmentation really helps to give it a gradient and intelligence beyond random chaos. Love the idea of it being applied to hardware.
It was literally a guy's job on the floor to just replace modules and other electronics. The equipment itself would not become contaminated so it was safe to handle afterwards. It just got bombarded by radiation and became useless as a camera. Being hit by ionizing radiation does not mean you become radioactive. The main issue is if bits of contamination would stick to our fancy duct tape we had around the cameras.
As for why we needed them it's for a bunch of reasons. This is 30 meters down. You gotta inspect welds, replace jet pumps, pick crap up that people drop in, pull plugs, help guide CRD maintenance. Tons of stuff. You gotta see it all. Camera handlers are magical and learn to swim the cameras around using puppet like movements. You manipulate these duct taped to rope cameras using either the cable or the rope. Sometimes we would attach them to stupendously long poles we assemble which were also duct taped (this changed eventually). The issue is such a long pole is basically a pool noodle in terms of handling. Keeping stuff from getting stuck and having confidence in where you were was an art. I wish I could tell you nuclear inspection used fancy drones and super high tech robotics but a ton of the visual side is duct taped cameras and talented handlers. Ultrasonic inspection is where the robotics took over and where they earn their keep. Encoding the position is worth the effort. But for visual you can't really get a sub to do much better than a guy with a long pole. Haha
For comparison the lifespan of a camera module was about 24-48 hours for work inside the water of the reactor near the "hot" fuel of the reactor. Fields around there were I believe on the order of 1000-5000 Rad/hr. Looked like the biggest confetti party you ever saw on the image. It was difficult for the encoder modules to keep up as well because they compressed so poorly and the reactor floors were usually hot and humid with the reactors open. I tried to make de-noising algorithms back in the day to help smooth out the noise in the reactor. Really hard to make electronics work in those places. Turns out constant bit flips and ionizing radiation is bad for hardware.
I find Claude code to be a token hog. No matter how confidently the papers say context rot is not an issue I find curating context to be highly important to output quality. Manually managing this in the Claude Webui has helped with my use cases more than freely tossing Claude code at it. Likely I am using both "wrong" but the way I use it is easier for me to reason about and minimize context rot.
I wish someone spoon fed me how to add path for C compilers in Windows back in the day. We lose a good 90% of people to installing C from ever learning C. Feel like godbolt or an online compiler might be a reasonable starting place these days. C is amazing but can be so punishing early on compared to stupid opening up any text editor on earth and writing an HTML file. Not advocating for more JS learning but it's hard to beat the getting started on that.
For me it is an active question if coding training data "purity" matters. Python has Go on volume, but within that is a ton of API changes, language changes, etc. Is that free regularization or does it poison the dataset? As the author points out Go code is nominal because basically all published Go code looks the same and the library APIs are frozen in time to some degree.
I actually spent some time trying to get to the bottom of what a logical extension of this would be. An entirely made up language spec for an idealized language it never saw ever, and therefore had no bad examples of it. Go is likely the closest for the many reasons people call it boring.
I find the dependency creep for both rust and node unfortunate. Almost anything I add explodes the deps and makes me sweat for maintenance, vulnerabilities, etc. I also feel perpetually behind, which I think is basically frontend default mode. Go does the one thing I wish Rust had more of which is a pretty darn great standard library with total backwards compatibility promises. There are awkward things with Go, but man, not needing to feel paranoid and how much can be built with so little feels good. But I totally understand just getting crap done and taking off the tin foil. Depends on what you prioritize. Solo devs don't have the luxury.
I have run into a surprising number of basic syntax errors on this one. At least in the few runs I have tried it's a swing and a miss. Wonder if the pressure of the Claude release is pushing these stop gap releases.
I have run into that a lot which is annoying. Even though all the code compiles because go is backwards compatible it all looks so much different. Same issue for python but in that case the API changes lead to actual breakage. For this reason I find go to be fairly great for codegen as the stability of the language is hard to compete with and the standard lib a powerful enough tool to support many many use cases.
I find it fascinating to give the LLMs huge stacks of reflective context. It's incredible how good they are at feeling huge amounts of csv like data. I imagine they would be good at trimming their context down.
I did some experiments by exposing the raw latent states, using hooks, of a small 1B Gemma model to a large model as it processed data. I'm curious if it is possible for the large model to nudge the smaller model latents to get the outputs it wants. I desperately want to get thinking out of tokens and into latent space. Something I've been chasing for a bit.
Trying to use ESNs as a random projection for audio data and potentially rendered text data for some AI workflows. Seeing it I can use the echo states running both forward and backward through the data as a holographic representation which would act as a temporally dense token for potential use in LLM or audio encoder inputs.
I feel like wishing for UI innovation is using the Monkey's paw. My web experience feels far too innovative and not enough consistent. I go to the Internet to read and do business not explore the labyrinth of concepts UI designers feel I should want. Take me back to standards, shortcuts, and consistency.
I want to explore the space of audio encoding and GPT like understanding of audio. I'm so highly interested in how a simple 1d signal must go through so much processing to be understood by language models, and am curious what tradeoffs occur. Would also be fun to make a TTS Library and understand it.