This little 20HP one-cylinder Diesel engine [1] powers much of third world agriculture. The original design seems to have come from Shanghai Engine Company in 1953, and is still manufactured by multiple companies. It's water-cooled, but non-recirculating; you have to fill the water tank when you fill the fuel tank. No electrical components at all. Starts with a hand crank.
Over 75 years of production of that design. It's the AK-47 of engines.
I don't have an iPhone to try this, but I've been a long time time user of Tasks.org on Android and particularly because it supports CalDAV and works so well offline.
However, while we are on the topic of planning apps, you should know the Todoist added the best use of AI I've ever seen. It's called Ramble mode and you can just talk and instantly it'll start showing a list of tasks that update as you go. It is extraordinary. I'm considering switching away from tasks.org for this one feature.
> uses a Raspberry Pi 5, a Halo AI board, and You Only Look Once (YOLO) recognition software to build a “computer vision system that’s much more accurate than anything on the market for law enforcement” for $250
> A simple Python application to test adversarial noise attacks on license plate recognition systems (see my PlateShapez demo) and create an output dataset to train more effective attack models.. small, hardly-noticeable, random gaussian shapes to confuse AI license plate readers
This was actually surprisingly clear. This, and htrp's comment are much clearer than the entire noise article.
They make dashboards and apps for killing people. With a lot of technical jargon like "integrating disparate weapons and sensor systems for a kill chain".
Somebody in America says "we want to kill somebody" -> satellite gives real-time imagery on location -> weapons systems available nearby are recommended -> user clicks orders and telemetry go out to field operators and ex: drone systems -> predator fires up and flies to location and bombs target -> real-time imagery confirms explosion and results.
With transcribing a talk by Andrej, you already picked the most challenging case possible, speed-wise. His natural talking speed is already >=1.5x that of a normal human. One of the people you absolutely have to set your YouTube speed back down to 1x when listening to follow what's going on.
In the idea of making more of an OpenAI minute, don't send it any silence.
will cut the talk down from 39m31s to 31m34s, by replacing any silence (with a -50dB threshold) longer than 20ms by a 20ms pause. And to keep with the spirit of your post, I measured only that the input file got shorter, I didn't look at all at the quality of the transcription by feeding it the shorter version.
You get the same experience inside VS Code with Copilot these days, as long as you select agent mode. Yesterday I even asked it to read documentation off a URL and it asked me permission to fetch the page, summarized it, and applied the API changes needed.
I tried summarizing the thread so far (339 comments) with a custom system prompt [0] and a user-prompt that captures the structure (hierarchy and upvotes) of the thread [1].
This is the output that we got (based on the HN-Companion project) [2]:
Complete hardware + software setup for running Deepseek-R1 locally. The actual model, no distillations, and Q8 quantization for full quality. Total cost, $6,000. All download and part links below:
Motherboard: Gigabyte MZ73-LM0 or MZ73-LM1. We want 2 EPYC sockets to get a massive 24 channels of DDR5 RAM to max out that memory size and bandwidth. https://t.co/GCYsoYaKvZ
CPU: 2x any AMD EPYC 9004 or 9005 CPU. LLM generation is bottlenecked by memory bandwidth, so you don't need a top-end one. Get the 9115 or even the 9015 if you really want to cut costs https://t.co/TkbfSFBioq
RAM: This is the big one. We are going to need 768GB (to fit the model) across 24 RAM channels (to get the bandwidth to run it fast enough). That means 24 x 32GB DDR5-RDIMM modules. Example kits: https://t.co/pJDnjxnfjghttps://t.co/ULXQen6TEc
Case: You can fit this in a standard tower case, but make sure it has screw mounts for a full server motherboard, which most consumer cases won't. The Enthoo Pro 2 Server will take this motherboard: https://t.co/m1KoTor49h
PSU: The power use of this system is surprisingly low! (<400W) However, you will need lots of CPU power cables for 2 EPYC CPUs. The Corsair HX1000i has enough, but you might be able to find a cheaper option: https://t.co/y6ug3LKd2k
Heatsink: This is a tricky bit. AMD EPYC is socket SP5, and most heatsinks for SP5 assume you have a 2U/4U server blade, which we don't for this build. You probably have to go to Ebay/Aliexpress for this. I can vouch for this one: https://t.co/51cUykOuWG
And if you find the fans that come with that heatsink noisy, replacing with 1 or 2 of these per heatsink instead will be efficient and whisper-quiet: https://t.co/CaEwtoxRZj
And finally, the SSD: Any 1TB or larger SSD that can fit R1 is fine. I recommend NVMe, just because you'll have to copy 700GB into RAM when you start the model, lol. No link here, if you got this far I assume you can find one yourself!
And that's your system! Put it all together and throw Linux on it. Also, an important tip: Go into the BIOS and set the number of NUMA groups to 0. This will ensure that every layer of the model is interleaved across all RAM chips, doubling our throughput. Don't forget!
Next, the model. Time to download 700 gigabytes of weights from @huggingface! Grab every file in the Q8_0 folder here: https://t.co/9ni1Miw73O
Believe it or not, you're almost done. There are more elegant ways to set it up, but for a quick demo, just do this. llama-cli -m ./DeepSeek-R1.Q8_0-00001-of-00015.gguf --temp 0.6 -no-cnv -c 16384 -p "<｜User｜>How many Rs are there in strawberry?<｜Assistant｜>"
If all goes well, you should witness a short load period followed by the stream of consciousness as a state-of-the-art local LLM begins to ponder your question:
And once it passes that test, just use llama-server to host the model and pass requests in from your other software. You now have frontier-level intelligence hosted entirely on your local machine, all open-source and free to use!
And if you got this far: Yes, there's no GPU in this build! If you want to host on GPU for faster generation speed, you can! You'll just lose a lot of quality from quantization, or if you want Q8 you'll need >700GB of GPU memory, which will probably cost $100k+
How exactly can there be "truthfulness" in humans, say? After all, if a human was taught in school all his life that the capital of Connecticut is Moscow...
For anyone wanting interesting YT videos for their kids (and not wanting to take anything away from OP's project), I highly highly recommend thekidshouldseethis.com. It's basically a curated stream of cool videos, and I would feel totally safe letting my daughter browse it alone (she never does because we usually watch them together, but the curation is that good). Videos on all sorts of topics, and good enough to be really entertaining for both kids and adults - I can spend an evening there easily. They also have a really fantastic gift guide.
etc. etc.
I focus really hard on answering exactly one question in a concise and engaging way and trying to keep every video under 5 minutes. Oh, and to make the videos solution independent, so not specific to a product, but convey the underlying knowledge so it has a longer shelf-life.
Full list is here: https://foxev.io/batteries/
I am planning to turn this into a knowledge base with playlists for "learning paths" like "everything to watch about batteries" or "here is what you need to watch to make a motor spin on a bench". I will add interactive functionality like quizzes and widgets to make the knowledge even more sticky.
For my next video, I want to show in detail how the interpreter works. For this purpose I'm creating an elaborate animation. You'll notice that the latest video is already several months old; this is because this animation is more work than I bargained for, and I got a little burned out by it. Nevertheless, I persevere and the video will come out whenever I may finish it.
Well, it depends whats your cup of tea in terms of learning. There area a LOT of courses on Udemy on that topic (if you prefer learning from videos).
I would recommend looking around on HuggingFace altough I found it a bit intimidating at the beginning. The place is just HUGE and they assume some knowledge.
I would also recommend creating a platform user on OpenAI and/or Anthropic and look up their docs. The accounts there are free but if you put a few dollars in there, you can actually make requests against their APIs which is the most simple way of playing around with LLMs imho.
Here are some topics you could do some research about:
- Foundation models (e.g., GPT, BERT, T5)
- Transformer architecture
- Natural Language Processing (NLP) basics
- Prompt engineering
- Fine-tuning and transfer learning
- Ethical considerations in AI
- AI safety and alignment
- Large Language Models (LLMs)
- Generative models for images (e.g., DALL-E, Stable Diffusion)
- AI frameworks and libraries (e.g., TensorFlow, PyTorch, Hugging Face)
- AI APIs and integration (also frameworks to build with AI like LangChain/
LangGraph)
- Vector databases and embeddings
- RAG
- Reinforcement Learning from Human Feedback (RLHF)
China is really really broke, from central government, to local government, to enterprise, to individuals. from 11 trillion off the book debts for local governments, to record number of Chinese blacklisted for debt defaults. from 7 trillion lost in stock market value in 3 years, to close to 20 trillion wealth wiped out in real estate.
Author of HumanifyJS here! I've created specifically a LLM based tool for this, which uses LLMs on AST level to guarantee that the code keeps working after the unminification step:
I'm not sure there's a single class of software that's been implemented more times than ngrok-style tunneling. I keep finding more and more.
Honestly it's a really fun exercise. Fairly challenging, but well within the reach of a single developer. I believe I'm currently working on my 5th incarnation.
Kinda neat but it seems a lot of "cyberdecks" are now just converging on "laptop" or "palm pilot". The essence behind a cyberdeck is its retro-futuristic design which produces an anachronistic feeling like some out of place object from another timeline dropped by a multiverse traveler. The Lisperati1000[0] nailed it with its surprising screen dimensions, form factor, color and keycap choices, and using it for Lisp programming. For a commercial solution it's hard to beat the Cardputer[1] with its chunky off white case, riotous multi-color labelling, quirky features, inscrutable purpose and the fact it comes in a blister pack like it's something you'd pick up in a gas station convenience store in an alternate 1988.
Here's my compilation of AI learning resources - I think some of the ones I've collected will be a better place to start for most people.
I categorized them into what kind of goal they're relevant for - building products, deploying custom models, or self study towards ai research science and research eng roles.
I’ve got my stuff rigged to hit mixtral-8x7, and dolphin locally, and 3.5-turbo, and the 4-series preview all with easy comparison in emacs and stuff, and in fairness the 4.5-preview is starting to show some edge on 8x7 that had been a toss-up even two weeks ago. I’m still on the mistral-medium waiting list.
Until I realized Perplexity will give you a decent amount of Mistral Medium for free through their partnership.
Who is sama kidding they’re still leading here? Mistral Medium destroys the 4.5 preview. And Perplexity wouldn’t be giving it away in any quantity if it had a cost structure like 4.5, Mistral hasn’t raised enough.
Speculation is risky but fuck it: Mistral is the new “RenTech of AI”, DPO and Alibi and sliding window and modern mixtures are well-understood so the money is in the lag between some new edge and TheBloke having it quantized for a Mac Mini or 4070 Super, and the enterprise didn’t love the weird structure, remembers how much fun it was to be over a barrel to MSFT, and can afford to dabble until it’s affordable and operable on-premise.
Starting something new is incredibly hard. The default is that your company never even forms, and it's on you to overcome the activation energy. I don't think I'd have the mental fortitude to stick that out alone.
In contrast, there's nothing more motivating than working with great people (and it's hard to do better than Ilya, Sam, and Elon). Everyone brings their own core strengths to the table, and if you've picked well your own efforts will be multiplied.
One of the most impactful things I've read was in the interview by Donald Knuth, shared here on HN some time ago:
"A person’s success in life is determined by having a high minimum, not a high maximum. If you can do something really well but there are other things at which you’re failing, the latter will hold you back. But if almost everything you do is up there, then you’ve got a good life."
As a general reminder as long as you have a reasonable interpretation of the tax code; even if it is NOT the IRS's (and the judge eventually rules for the IRS), you will likely be clear of penalties.
If you try to avoid ever getting entangled with the IRS you will way overpay.
I was one of the first backers of the Oculus Rift Kickstarter. When I got it I decided eye tracking was going to be huge for VR, so as a side project I cut a hole in my Rift and built my own eye tracker. I posted it on Hacker News: https://news.ycombinator.com/item?id=7876471
A few days later the CTO of a small eye tracking startup gave me a call. I quit Google and joined them. I built a (novel at the time) deep neural net based VR eye tracking system for them, and less than two years later Google acquired us.
It was two months ago. The tech industry definitely was not hiring like crazy. But I started laying the ground work for leaving Amazon the day I got hired. It was AWS ProServe - the consulting division.
Companies are always hiring. They were hiring in 2000 (I looked for a job then. But I accepted a counter offer). It was hiring in 2008. I found a job in two months and it was hiring two months ago.
Focus on creating an “unfair advantage”.
1. I got on to a high profile open source project and stayed on it until it became a more official “AWS Solution”. In still high on the list of contributors for both code and the documentation. This actually led directly to my current job. The company I was hired for uses the solution.
2. I found out the process to open source my own work and get it published to the AWS Samples GitHub repo. I went through the open source process for every piece of code I wrote after removing the proprietary customer specific parts. I had an open source portfolio that I could say was being used by real customers. I’ve used two of those projects at my current job (legally - MIT licensed by AWS)
I still have one of the two only public solutions to automate deploying one complicated AWS service. I volunteered to design the solution for the service team when they had the APIs it for it in beta and I worked with the guy that created Terraform support for it.
3. I responded to every recruiter over the last three years and kept in touch with former managers and coworkers. This led to an offer of a full time job that I didn’t accept and a side contract I did accept.
4. I constantly kept my resume up to date and I used the leveling guidelines to word it. I also have a career document.
5. I used the extra money I was making to pay off debt, save and “decontent” my life so if necessary, I could go back to enterprise CRUD compensation without having to hold out for BigTech level compensation. I actually turned down a chance at a job that would have paid more than I was making. I didn’t want the stress.
A former coworker is a director of a well known (non tech) company and he was willing to create a position for me overseeing the entire AWS architecture and their transition effort. I really didn’t need to stress.
I’m building out a much smaller “center of excellence” now and it’s rough enough.
6. I kept my LLC active all three years just in case I needed it to ramp up a solo consulting business. I used it for the side contract I mentioned that I started the day I was left (in addition to the full time job I got two weeks later)
We actually expected our West Elm sofa to be well built - it wasn't cheap and their reputation wasn't complete garbage yet when we purchased it. After about 18 months the whole thing collapsed, just out of warranty believe it or not.
When I started taking it apart to see if it was fixable I found that the entire damn thing was built out of stapled-together 1/2" OSB (oriented strand board...garbage "plywood") and all of the cuts had awful stress concentration corners which caused the frame to pretty much disintegrate. Wasn't a real screw or nail anywhere in the piece that I could find.
I spent a day rebuilding the frame with about $30 worth of construction lumber (some 2x4s & 2x8s) and hanger brackets. It's been perfect ever since.
Completely inexcusable that a $30 aftermarket fix is all it took to dramatically improve an $1800 sofa.
Edit: I found a few images of the damage and a couple of the replacement supports I added. There was a lot more, but at the end of the day I was tired and didn't take a ton of pictures
https://imgur.com/a/bqlLgW3
I totally agree with you. Nobody gives a fuck. I sweated 6 months on my CS Masters thesis. After the day of the presentation, I walked up to my Professor's office. As per the rules, you are supposed to hand a printed copy of your thesis to your thesis advisor. So I handed over the printed thesis proudly. Then we chatted, I said my goodbyes & left. In those days we had no github internet etc. All the C++ code for my thesis was on a 5 1/4 floppy. When I reached home I found the floppy. I was like, God I've forgotten to give him my code! So I picked up the floppy and trudged back to the university. Then I walk up to my advisor's office again & knock on the door. He is ofcourse surprised to see me. I say - Sorry I forgot to give you my code. Here is the floppy. I'll put it next to my thesis. Where is my thesis ?
He doesn't say anything. I look around to see if my thesis is on his bookshelf, but no, it isn't there. Then I turn around & I find my thesis. It is in his trashcan. I was so stunned & shocked. My advisor says sheepishly - look its just a Master's thesis. Its not like you have discovered a new theorem or something. These results are well known in the literature.
I just put my floppy in that trashcan and walked out. Its been more than 2 decades now but I still remember that incident like it was yesterday. Literally, nobody gives a fuck.
This conversation strikes me as unlikely on several levels. First, no one would have coached you on "how to interview at Oxide" because that's not where the process starts -- it starts with you preparing your materials.[0] (Our review of the materials constitutes ~95% of our process.) Second, we have always been very explicit about compensation (that is, we ourselves brought it up early in conversations); no one at Oxide would tell you to "not bring it up" because everyone at Oxide knows that it is a subject dealt with early in the process. And finally, this is all assuming that you were talking to someone before March 2021, when we published our blog post on it.[1] After the blog post, compensation simply doesn't come up:
everyone has seen it -- and indeed, our approach to compensation is part of what attracted them to the company!
Over 75 years of production of that design. It's the AK-47 of engines.
[1] https://www.alibaba.com/product-detail/High-Quality-Manufact...