Here are things I found interesting and could have put on Twitter, but they're here instead. All the entries are chronological with the newest at the top.
- Now AI is being used to translate Akkadian cuniform tablets into English. This software could really help with the trove of half-a-million untranslated tablets.
- I found a good explainer about using LLMs to build agents. This seems to be what LangChain is focused on, although it isn't mentioned.
- Two good signs for the future of the US: a nuclear renaissance and a broadband build-out. Of course, if you Google "nuclear renaissance" you'll see that it has come and gone a few times. Maybe this time will be different. At the very least, maybe the money won't end up in Ukraine...
- Sweden went from a target of 100% renewable energy by 2045 to 100% fossil-free. The change allows them to increase their emphasis on nuclear power. It's still going to reduce CO2 emissions, but it's not "renewable" so some environmentalists reflexively say "Boo!"
- Limiting social media use to 30 minutes per day may improve psychological well-being. So stop reading this!
- Using knowledge of how CPUs work to speed up your Python code: Not so much the particulars but the general approach is important here.
- The current philosophy driving the USAF's drone efforts: "affordable mass".
- First two parts of Energy Destinies (Part 1, Part 2) which examines "the role of energy, demand and supply dynamics, the shift to renewables, the transition, its relationship to emissions and possible pathways"
- A small LLM with just 1.3B parameters provides surprisingly good performance after being trained on only 7B tokens of "textbook-quality" data. And a stripped-down 350M-parameter version only loses 10% of its accuracy.
- OpenAI wants AI regulations, except when they don't.
- This looks like a decent, modern introduction on using GPUs.
- Milk-V: a 1GHz dual-core RISC-V on a small form-factor board that can run Linux for $9.
- Free font based on the London subway dot-matrix displays.
- An interesting blog post exposes some of how text tokenizers work.
- Viable human embryos from just embryonic stem cells! Our problems with declining populations are solved...
- "Mini-brains" made from (you guessed it) stem cells were able to function like the pituitary gland when implanted (you'll never guess) near the kidneys in a mouse.
- The site llm-utils contains an interesting hodge-podge of stuff like a list of places to rent GPU compute. Unfortunately, some of the pages are just lists with no associated link.
- This is probably good advice: if you want to succeed at something, do ten times as much. This level of effort requires you to conciously choose where to direct your finite energy and drop what you don't care about enough to put in the effort.
- Coal was extracted from undersea mines on Hashima island starting in 1887. The island was abandoned in 1974 as Japan switched from coal to oil. Here's some video of the deserted buildings.
- If you're bored, you can read about the possible assassination of Oppenheimer's mistress or Japanese convenience stores.
- The people who really matter are rejecting the idea of an AI pause.
- I don't have any opinions about John Carmack one way or the other, but it's kinda sad that he would be disparaged just for going to a scifi convention.
- The creator of Black Mirror tried to get ChatGPT to write an episode, but he said the result was "shit." Not surprising since his prompt was basically "Write a Black Mirror episode." C'mon! You gotta work a little harder than that...
- As multiple countries eye its resources, the US is ramping up its military in the Arctic.
- Molten salt nuclear reactors: Don't be scared! It's just really hot, radioactive slag.
- TV went streaming, and now that doesn't seem to be working.
- The UK's Royal Navy says it has successfully tested "quantum navigation" that can give a precise location without relying on an external system like GPS that can be jammed. No details in the article about how this works, but it uses "quantum" so it must be legit...
- D2 is a text-based diagramming tool similar to PlantUML.
- Like the guy who's collecting the measurements of every high school ballpark, here's another that's scanning the entire collection of Computer Shopper magazines. I hope there's some unexpected benefit to a future researcher in computer history that will justify all the time put into this.
- With LLM-based programming assistants, we've gone from writing code to casting spells where the particular words used in a prompt can generate unexpected results. I'm not sure this is an improvement...
- I enjoyed this video on how to create generative art with vanilla Javascript.
- Defense firms fund think tanks who then take to the media outlets and beat the drum for sending more arms to Ukraine. As Kurt Vonnegut would say: "And so it goes..."
- The Falcon LLM I mentioned a few days ago has made its way to Hugging Face.
- Next installment about the history and development of the US power grid.
- The article title - Intelligent brains take longer to solve difficult problems - omits the additional finding that they also make fewer mistakes. Seems important, but I guess nobody would read the article with the accurate headline.
- Somebody is measuring and cataloging all the high school baseball fields in the US.
- Raspberry Pis should be easier to get, soon. (Trigger alert: the story contains the unfortunate phrase "unless you have been living under a rock...")
- The UAE's Technology Innovation Institute has open-sourced the Falcon 40B LLM under an Apache 2.0 license. Falcon sits at the top of some LLM performance leaderboards. This further dims the hopes of the AI safetiests to limit the spread of the technology.
- This is a really neat idea for a low-cost blood pressure monitor.
boltonsis a package of utilities that augment the functions found in the Python Standard Library.- Like factories in space? Cuz here's a site about factories in space.
- Interesting thread on scavenging the world in search of low-background metals made before the first atomic explosions.
- This article contends that AI advancement will slow down due to hardware constraints. Upcoming articles in this series posit further decelerations in the areas of algorithms and data.
- More benefits from vitamin D.
- Tesla is innovating in car assembly as they move toward their goal of producing 20M cars per year, but there are obstacles that may prevent the manufacture of more than 10M per year across all manufacturers.
- Chips and Cheese is another good site if you want detailed expositions on past and present ICs.
- A current theme in LLMs is trying to get around the token limit and enlarging the context.
- Using LLMs in an EDA tool for chip design.
- Experiences or possessions? Maybe it's all bullshit.
- How does our brain tell the difference between reality and imagination?
- The Alexandria Project aims to create embeddings of all 4M papers on Arxiv to make it easy to do semantic searches (versus by keyword).
- I always like a good article on SAT solvers. Oh, and this one, too!
- An account of the early history of electrification and the growth of the grid, primarily in the US.
- Something I just noticed: right-click on a web page in the Brave browser and you get an option to create a QR code for that page. I use this to print and attach instruction manual QR codes to dev boards I've purchased.
- Megabyte is an AI model that may allow processing of data equivalent to 1M-token windows. (GPT-4 currently maxes out at 32K tokens.)
- Human feedback (HF) is important for getting better performance from an LLM, but it's expensive and slow. AlpacaFarm reduces the cost/time by orders of magnitude by replacing HF with feedback generated by ChatGPT and GPT-4.
- Interesting article on the defense challenges posed by maneuverable hypersonic missiles.
- Good explanation of AI risks, but the last section states autonomy and agency should be avoided. Yes, but define that. Even coding helpers might be disallowed since the AI could inject jail-break instructions via humans. But government and corporations would skirt those rules.
- I wrote a small timer program using ChatGPT.
- It pays to be good-looking. Literally.
- The "A Very Short Introduction" series of books is featured in this long article called "How to be a Know-It-All" but, really, please don't.
- Another entry in the list of "Why we can't have nice things": New submissions to PyPi were temporarily suspended because of malicious users.
- Like everything else, war is an iterative process: you don't need a fixed master plan as much as you need an adaptive sequence of smaller plans.
- The Decouple podcast has some episodes they label as "master classes" in which they go in-depth about specific types of energy sources: coal, gas, and uranium (so far).
- Populations are declining! Could 50% lower sperm counts be the reason? Read this article for some banal advice that could help you save mankind from extermination.
- OK, so the Nazis didn't have a base in Antarctica, but they could still be on the Moon, right?
- Has AI killed the semantic web before we could even build it?
- Here's a handbook on using LangChain.
- Here are some numbers you should know to get ballpark estimates of required resources and costs when using LLMs. (And here are some earlier numbers you should know about memory access times.)
- More demographics: The US birth rate has fallen 20% since 2007 and shows no reversal of the trend.
- BT is planning to axe 55,000 of its 130,000 employees by 2030. About 10,000 in customer service will be replaced by AI, but the rest are a result of completing the switch to fiber networks and the lowering of maintenance needs.
- Long article about ship-breaking oil rigs.
- Ah, one of my favorite movies from my youth: Colossus: The Forbin Project. You can read the entire trilogy so you can see what happens after the AI takes over.
- While searching for the Colossus novels, I came across this book about the construction of the Hoover Dam. Looks good!
- A thread about building websites in 2023. ChatGPT is mentioned six times as a tool for assisting with this.
- Google uses 3x-4x more GPUs than their in-house TPU chips. Part of this is because TPUs aren't supported by Nvidia's AI Enterprise software stack and Google's cloud customers need that to do their own AI training.
- If you're thinking of writing some non-fiction with really original ideas, here's a list of things that aren't.
- Efabless may build a chip for you if you design it using AI. (Actually, you only need the AI to generate the Verilog for the chip.)
- How is a spacecraft on a 10-year voyage to Jupiter like a running toilet? Sometimes you can fix it by jiggling the handle.
- Read the planning scenarios for domestic terrorism from those cute little scamps at the DHS. Then browse the site for more government documents discovered under FOIA.
- The Chinese response to the American semiconductor sanctions of October, 2022.
- Microsoft is updating the Visual Studio UI to reduce "cognitive load". Basically, they're making some icons bigger and spacing things out. I really hope these changes make their way to VScode. It's so 2022!
- Here's a timeline of animals that have survived mass extinctions, some of them multiple times. (Looking at you, hardy little horseshoe crab!)
- Researchers used undersea fiber-optic cables as hydrophones to pinpoint the location of whales to within 100m.
- An interesting site where you can pick two from a set of LLMs, have a conversation with both, and then select the model you think is best. This results in a ranking of the LLMs by user preference.
- OpenAI should know something about prompt engineering, right? So this online course may be worth a look.
- More evidence for vitamin D supplementation: reduction in all-cause mortality.
- Keep up with the drugs the Cool Software Developers are using: Adderall for debugging and psychedelics for brainstorming.
- Running agents using LangChain and LLMs is the new hotness, and HuggingFace has developed a simplified API for doing this.
- Virtual Power Plants are decentralized sources like the batteries in EVs or solar cells on houses that can be flexibly amalgamated to supply power at utility levels. Personally, I'd like a nice nuclear power plant.
- This Smithsonian story spends more time on the photographer who's chronicling the final days of an old farm than the elderly farmers who've run it for decades. It's almost like they're an afterthought.
- I linked to a prompt engineering guide several months ago. That's probably hopelessly out of date by now. Here's a newer one
- If Pearson sues companies for using their textbooks to train AIs, doesn't that mean they can also sue students who learn from these books? If a company just buys a single copy of a textbook and lets the AI read it, they should be good to go, right?
- I discovered this great site for explanatory articles about machine learning only to find it went on hiatus in 2021 for the usual reason: it was too much work. The site editors recommend another method that requires much less work (for them): self-publishing.
- Here's how to give GPT and other LLMs access to more knowledge. It's similar to how that guy built a ChatGPT interface to his quantum biology PhD thesis.
- Returning to the office could make you a fat drunk, so follow these common-sense things to keep fat off. Or you can just buy the latest piece of hyped exercise equipment and store it in your closet.
- This story about the CIA being involved with some of the 9/11 hijackers keeps bubbling up but never seems to get much traction. Probably because so many media organizations now employ ex-CIA people...
- Autoimmune diseases now affect about 10% of the population. I wonder how many of those have bacterial or viral causes.
- This list of AI Chrome extensions can help you get experience before you lose your job to ChatGPT like this guy.
- A ChatGPT interface to make a PhD thesis more understandable to non-experts. This is reminiscient of my use of ChatGPT to get an understanding of Mendelian randomization when I found the Wikipedia article too dense.
- Another disease that may be due to bacteria: Parkinson's. On the other hand, this study found no causal link between the gut microbiome and chronic diseases. This latter study used a technique called Mendelian randomization which I could not understand at all from the Wikipedia entry. Luckily, ChatGPT was able to provide a simple example of how it works.
- All these years I've been incorrectly using the phrase "comprised of".
- Vector databases are used to store and look-up feature values for AI that are translated into single-dimension numeric arrays called embeddings. Here's the gist of how they work.
- Google could have taken one of several approaches in response to yesterday's leaked memo on how open source developers were out-innovating them in AI. Looks like they decided to become even more closed.
- This article about declining public school enrollments had one item that surprised me: New York City's schools spend almost $40K per student each year. That seems like a lot.
- I should never look askance at what others find interesting, but this guy has spent a little too much time thinking about swings.
- This internal Google doc is making the rounds today. The basic premise is that open source fine-tuning for specific applications of LLMs is proceeding faster and making greater progress than from-scratch training of large models with massive data sets at Google and OpenAI. Looks like the White House can forget about fencing AI into their public safety corral.
- As things like agents and AutoGPT take precedence in AI, I see more and more references to LangChain. Here is a simple re-implementation of it in 100 lines of Javascript.
- Take a look at the section on interesting projects in this Python Weekly and count the number that reference AI, LLMs, or GPT.
- Just so today's links aren't all about AI: If you've got some free time, maybe you can solve this mystery.
- Well, the geniuses who brought us the Afghanistan withdrawal are calling in the chiefs of Google and Microsoft to make sure AI is only released to the public once it's "safe" (i.e. neutered).
- Eighth-graders continue doing poorly in history. But don't worry! They'll always be able to query ChatGPT for the facts (once the White House approves a suitably neutered version).
- Videogame maker Unity is laying off another 600 in order to "realign some of our resources to better drive focus and support our long-term growth.” They'll probably be replaced with AI tools similar to what's happening in China's video game industry.
- Some opinions of other things going on besides AI: Mostly computer-related, but a bit of other stuff.
- Exercise increases your population of tumor-fighting cells. So be sure to hit those weights after your chemo!
- Options for hosting your machine learning models.
- A Transformer was trained on neural waveforms as stories were read to study participants and was then able to output their thoughts as text based on their neural activations as they heard new stories.
- This would be exciting: a replacement for the Haber-Bosch process that can break the bond of atmospheric nitrogen at room temperature and pressure. That would provide an energy-efficient way to make ammonia and then fertilizer. If it scales up that is, which these things seldom do.
- Another in the endless list of reasons we can't get anything done: Environmentalists sue the FAA over permits for SpaceX launch. They are requesting "more environmental studies." If these guys had their way, Neil Armstrong would have died as an unknown Navy airman.
- A timeline of the "inevitable disaster" that is unfolding at Twitter has been discontinued. I can only guess it's because Twitter has imploded and shut down, right? So we're all on Mastodon now? Actually, it looks like Twitter outlasted the timeline. Who could have known!?
- People switching from numpy to start using Python AI libraries like PyTorch and fastai because they can get all the same capabilities plus improvements like automatic gradients and CUDA interoperability.
- You piss-off one customer for a few hundred dollars and your whole Ponzi scheme comes down on your head. The kicker: after getting out of jail, he gets convicted a few years later for running another completely unrelated scam. At least he's not a one-trick pony.
- Which countries are most online? I was surprised it was South Africa whose population averages 58% of their waking hours looking at screens. And Japan is the lowest at 22%. But, according to this, North America holds the top spot if you rate it by the total percentage of people who have used the internet, regardless of for how long.
- Quite frankly, ChatGPT operating costs of $700K per day seems pretty reasonable for the impact it's having.
- We're saved! The G-7 ministers came up with five principles to govern AI. And, yes, they're every bit as unspecific and nebulous as you would expect.
- Use this site to get the gist of all those too-long Youtube videos you feel you must watch. But don't get your hopes up that it will generate some super-short video summary. It appears to extract a transcript and an associated summary. (I'm not exactly sure since I accidentally exceeded my free, daily limit of one video.)
- Study: "ChatGPT outperforms physicians in high-quality, empathetic answers to patient questions." Certainly a low bar given current healthcare practices...
- Another benefit from taking vitamin D: reducing depression and suicide. Of course, I remember the 80's & 90's when vitamins C & E were also considered wonder supplements.
- Well, here's another crap article putting forth unsupported warnings that ChatGPT's children will kill us all. From a guy who heads an AI alignment foundation, of course.
- How can AI destroy the world if it's just a waste of time because most people are just using it as a toy? Kinda like the internet back when it started...
- Read about how the CIA subverted Guatemala's democracy and connect the dots to what's happening today.
- New York to ban gas stoves and California to ban diesel trucks. Move over Germany, we're breeding our own species of dumb!
- This will go down as a declaration of premature victory comparable to George Bush's "Mission accomplished!" Here's the counterpoint.
- Colleges remove data science from list of recommended HS math courses. Of course they're going to remove something that students find fun and practical. Get back to doing those integrals! Didn't they learn when they tried to kill off calculators in the 70's?
- More demographics news: Japan's population expected to shrink by 1/3 to 87M by 2070
- A college degree is worth it according to ... the President of Princeton. Shocking!
- It's hard finding stuff that's not about ChatGPT/AI. Here's why
- Congress got forty ChatGPT licenses to start experimenting with the tech before they start regulating the hell out of it. Most probable first questions they ask it: "Where can I find the best child porn?" and "What's the going rate for a Senatorial bribe?"
- Left-handers are more vulnerable in a fight which supresses their numbers in a population, but they can still surprise you tactically so they tend not to die out completely.
- And here's your weekly reminder that doing space stuff is hard. Iteration is the norm in this area.
- I've started going through the fast.ai deep learning tutorial in preparation for starting the Hugging Face course. You can accomplish a lot in AI without having a solid foundation, but there's no cost for learning the underpinning basics except time.
- A comparison of the AI and Industrial Revolutions. I like these historical compare-and-contrast articles, like this book that examines the societal impacts of the internet and the telegraph.
- "The Free World Must Stay the Course on Ukraine!" How about you just eat me. You may get off by moving little pins around on the geopolitical map, but that doesn't mean the rest of us have to indulge your psychopathic BS.
- I just ran across langchain as a system for integrating LLMs into applications that do more than just chat. It will be interesting to see how this relates to something like AutoGPT.
- Billionaires betting big on fusion. Good luck: we've been having trouble even getting fission plants built.
- There are lots of Python packages to simplify building GUI interfaces (PySimpleGUI) and CLIs (click). Now add duckargs which is a program that accepts a description of the arguments for your program and outputs the argparse calls to build the CLI.
- I use the OpenAI API with Chatbot-UI for chat-style stuff, but here's how to also use the API for image-generation tasks with DALL-E 2. (Also, here's a tutorial on CLIP which is used in DALL-E 2.)
- Somebody wrote a Python REPL coupled with ChatGPT to do code development using prompts. And here is another take on how LLMs will change our model of software development.
- Want to buy fake stars to make your Github project look more impressive? Or suspicious that others are doing that? There are ways to detect such perfidy!
- CATL (world's largest battery manufacturer) announced they will start producing a battery this year with an energy density of 500 Wh/Kg as compared to the ~285 Wh/Kg of the Tesla 4680 cell. For comparison, gasoline when burned in an internal combustion engine has about 3,000 - 4,000 Wh/Kg.
- Gosh, you want to build EVs? Well, Chile just nationalized their lithium industry (2nd largest in the world behind Australia). Let's hope this doesn't turn out like when Venezuela nationalized their oil industry...
- Well, SpaceX's Starship ascended 24 miles before it had to self-destruct when it became unstable. Good effort, lots of data gathered for the next attempt.
- Problems of the Future: The Inmarsat-41 satellite went down and Australian farmers were unable to plant their winter crop.
- And now the AI lawsuits arrive, starting with Reddit, Tom Brady, and Michael Schumacher.
- The problem is, governments write far-reaching laws that have consequences nobody can predict. Today's case in point: the Python Software Foundation says they may no longer be able to offer access to PyPi to the EU.
- Have an LLM translate your word problem into equations and pass those to a symbolic math package for a solution.
- A third of "researchers" believe AI could lead to catastrophe according to the Institute for Human-Centered AI. That's probably the third that is working on AI safety...
- Drop a PDF and then get answers to questions about it. Seems a bit like explainpaper.com.
- Either believe in dark energy, or be willng to tweak gravity a bit. Maybe we're at a similar point as when it was thought that light propagation through space required it be filled with "ether" and then the Michelson-Morley experiment blew that out of the water.
- High-res MRI scans with voxels down to 5 microns in size are now possible. Typical MRIs in clinics have voxels of 1-2 millimeters. Don't expect this at your doctor's, though. The article says one step is to send "the tissue" out for "light sheet microscopy." That sounds like the brain is no longer inside the head...
- The guy that invented the lithium battery says we've got a lot of work to do to make it better.
- Here's a list of programming playgrounds where you can try various languages in your browser.
- The case for unlocking firmware.
- Somebody wrote a little Python script to flesh-out links in their posts. Quite frankly, I think authors lean on links a bit too much. Instead of providing a pithy summary of something to support their argument, they just link to some external page and say "So there!" Then you have to descend into that link, read it, parse how it relates to what you were reading, and then pop back up to the original article where you'll soon encounter another link. It's exhausting.
- I wrote a forum post on using GPT4 to create part symbols for KiCad.
- I've always liked the idea of constraint programming: Set up the constraints and then let the computer figure out the answer. My problem has always been with learning the syntax of the various libraries. I have a similar problem with
matplotlib, but then I used ChatGPT to iteratively evolve a graph showing per-capital GDPs for a set of countries. Maybe I can do the same thing for constraint programming libraries? - AI algorithm designers won't settle down enough to let AI chip designers sleep at night.
- Financial software meets LLMs: It's like a match made in heaven. That is, if you consider Frankenstein getting up with Voltron a heavenly match...
- People are freaking out about stuff, but people have always freaked out about stuff.
- How to run AutoGPT: Now you can just set up your tasks and let ChatGPT go to work on them. All you've got to do is say "Yep, keep going" every once and a while until they're all done.
- Starship orbital flight test scheduled for April 17th!
- Something out there is really bright! Must be a neutron star surpassing the Eddington Limit which was calculated by a guy who had a TV movie made about him and Einstein. There, now your day has not been wasted...
- An interesting interview with sci-fi author David Brin.
- WikiBinge attempts to show a "vague" connection between any two items you enter into their form. I was hoping it would generate something akin to a mini-version of a James Burke "Connections" episode. Unfortunately, it just lists a chain of articles that are related, usually by having some words in common. Seems like this could now be done much better with GPT...
- I don't agree with the AI Pause letter. This seems like a better alternative since much of the direction would come from within the AI industry itself. Even so, I'm sure government will try some ham-fisted regulation and we'll end up with the AI version of the NRC: No nukes and no AI for you!
- There's a Youtube series focusing on AI hardware and here's the first episode.
- In the debate about AI's affect on the workforce, here's a datapoint: a 70% reduction in job postings for Chinese videogame illustrators with a 40x gain in productivity. Of course, some customers complain about in-game characters with too many fingers or missing legs...
- Germany's economy is suffering, partly from energy deficits. Households and companies have had to cut energy use by 18% and they're still gonna shutter their last three nukes. Sheesh!
- Looks like that murder of the tech entrepreneur in SF might not have been random violence after all. And score one for SF police: they actually arrested somebody!
- A study across five divergent species exposed a common link that affects aging: DNA-to-RNA transcription becomes faster and more error-prone as we age.
- Here's another awesome compendium, this time of decentralized LLMs that you can "own".
- Looks like NPR went extreme pouty face, took their ball, and went home. Now where will I go on Twitter for all the race and gender news that's fit to spew?
- Taiwan says no need to destroy TSMC if China invades. With the crazy assholes in-charge in Washington, I'm surprised they haven't bombed the chip fabs already. Anybody who thinks they know the results of doing that is not aware of what can happen when you ignore ecosystem effects.
- To save the planet, Germany has decided to close its last three nuclear plants and burn more gas and coal. The thinking is they don't need 'em since they made it through the winter just fine. In related news, Germany was awarded the "Short-Sighted Dumbass of the Year Award" for 2023. Yeah, the whole country.
- Here's 67 birthday bromides for your edification.
- Here's a pragmatic explanation of how to remember things using spaced repetition.
- If you have an OpenAI API token, then you can run chatbot-ui locally to get the same experience as using ChatGPT via the web browser version. (The advantage of the API is you are charged on a per-usage basis instead of a flat $20/month.)
- Well, you knew it was coming: Awesome ChatGPT.
- We got a bevy open source LLMs due to a single leak. (Thanks, 4chan! Guess you're not just a meme machine.) Now the Euros want an open-source, CERN-like recreation of Open AI, but they'll probably strangle it in the cradle over safety concerns.
- Positive movements in the West: the US is building more factories and the UK is making them more flexible with 3D printing.
- Engineers seem to like categorizing themselves. You're either Type 1 or Type 2, or a martial artist or a street fighter. Couldn't care less...
- Meanwhile, mathematicians have derived a formula for achieving orgasm. Way to bring the magic to the bedroom, guys!
- Antmicro announced an open hardware portal and a design flow that makes it easy to take circuit board Gerber files and a BOM and create realistic 3D renders using Blender.
- How human societies evolved in two parts. Part 1 discusses early societies where men killed each other to gain access to women who were largely self-sufficient due to adequate resources. In part 2, population pressures required more male involvement in agricultural food production, thus making workers more important than warriors and leading to societies with fewer wars.
- Doesn't look like all the links to papers are active yet, but here's the Proceedings of the 3rd Workshop on Open-Source Design Automation (OSDA), 2023.
- If you're into cybersecurity, then here's a site with hundreds of free classes, including some from the NSA. It's security-focused, but it also includes modules on programming, quantum computing, and reverse engineering.
- So far this week, San Francisco has killed a software entrepreneur and the USA cycling champion. Unfortunately, Mayor Breed may need more high-profile deaths before she'll help SF clean up its act.
- Good sources to keep up with AI. Yeah, well, good luck with that...
- NPR won't tweet until Twitter stops calling it "state-affiliated media". That sounds pretty OK since NPR tweets like they're state-affiliated media.
- There's something about the EU and technology, in that they want none of it. I don't know how this works out well for them.
- At least the French are mad at something in addition to tech. Like rich people.
- AI revolutionizing the workplace? It's going to take longer than you think.
- Five times science got it wrong. Really? Only five? (But it's a good thing when science self-corrects rather than circling the wagons and chanting "Believe Science".)
- Been putting off using brain-machine interfaces because of that nasty conductive gel or, you know, brain surgery? Well then, you're gonna love these graphene-based sensors!
- Meta open-sourced a new image-segmentation AI model that even works on objects it's never seen before.
- The predicted dates for Twitter's downfall have all passed and now people are getting a bit sick of Mastodon.
- Finally, something from ... me! I wrote a small script to use GPT to make news articles more delicious and nutritious by removing bias: nutrinews.
- Oh, well, looks like some people just can't be happy.
- Improve children's diet by extending dinnertime by 10 minutes. This is new? My parents would make me sit at the table for hours until I finished my meat.
- They identified the network leading from a pleasurable physical touch to the brain's reward center. Can't wait for the drug companies to monetize that.
- Call centers are embracing AI. I look forward to being informed about my expiring car warranty by GPT4...
- Will we be creating any new boilerplate templates in the future, or is that just part of the gone-away world after AI?
- The West takes another beating: When faced with a choice between supporting American hegemony and having energy for its people, looks like Japan has opted for that sweet, sweet Russian crude...
- An interesting effort: use GPT4 to help develop a pseudocode language that can be used to describe algorithms to GPT4 so it can write the detailed version. I'm getting visions of that freaky Escher print where there's a hand drawing another hand that's drawing the first hand...
- Somebody else had GPT4 generate an abbreviated language to cut down on the number of tokens needed in the context buffer. Won't be long before GPT4 is talking to itself in a language we can't understand. Then Peggy Noonan can say "Told ya so!"
- With the onslaught of AI developments lately, we lose sight of all the improvements still being made to our software infrastructure. Here's to you, Pandas, with your new 2.0 release!
- GPT4 produces better results when prompted to "show its steps", but does even better when asked to critique itself.
- They keep telling the West to get stuffed: OPEC+ does surprise oil cut.
- Twitter downranks pro-Ukraine tweets. Meanwhile, everybody else upranks them. I'm calling it a wash...
- Generative graphics with Rust: Might be a good way to introduce myself to the language.
- What good is an AI if it doesn't have neat stuff to use like drone aircraft that autonomously refuel and zippy little drone boats with Stinger missiles?
- Here's a list of web apps that don't require a login.
- James Watt founded a British IP company almost two centuries before ARM. He ran his business not by building steam engines but from licensing others to build them and then charging a royalty based on the cost of the coal that was saved over using a less efficient design.
- The BRICS countries now have a greater share of GDP than the West and they are also developing their own reserve currency. The only viable avenue open for the West is to out-innovate them, but that seems unlikely given the political leadership of the past few decades. So expect to see a lot more bombing...
- Out-innovating the BRICS becomes even less likely when we remove algebra from the 8th grade...
- When Western goverments prosecute small Twitter accounts for "election interference" and outlaw ChatGPT because it spews "misinformation", then you know the West has completely lost the plot...
- OK, something at least a little bit technical: Using LLMs to control the coordinated operation of other AI models.
- Getting power from blood sugar: Maybe this can be used to power biological nanomachines, but the prototype looks like something I taped together from my kitchen junk drawer.
- Cerebras - the company that makes gazillion-transistor AI wafers - has open sourced the GPT-3 model architecture as well as model training data, weights and checkpoints.
- I remember using APL back in 1979. You needed a special terminal with Greek letters and other strange symbols on it, or else use a regular terminal and memorize which keys corresponded to those wacky symbols.
- Why do sharks attack humans? Because we taste good. Don't tell the aliens.
- Well, now we've got an open letter urging a six-month pause in training of LLMs. Alarm bells should go off when you see one of the AI pitfalls they're trying to avoid is "political disruptions (especially to democracy)". The dis/mis/malinformation crowd has found a new horse to ride...
- Good luck reigning in AI models (and this is just today): OpenFlamingo, gpt4all
- We may be overstating the draconian penalties of the RESTRICT Act, but then again, our government did get up to some diabolical shit under the guise of the PATRIOT Act.
- The problem with stuff like Clearview is the police use it a few times and catch some bad guys, and then they say "If we used it more, we could do even more good!" That's when things go bad and you end up with some cop scanning randos looking for his ex-GF going into Hunk-O-Mania.
- People have enough subscriptions and they don't want them for cars. I just want to put a key into it and go off somewhere nobody can find me. I don't even want power windows. Hell, I don't even care if I have to stop it by dragging my foot on the pavement...
- More AI-based chip placement, this time from NVIDIA.
- A 294-page book that covers ... C++ initialization. That's it, that's the whole book. C++ has become a language that revels in its own complexity.
- Google solved chip component placement using deep learning in 2021, or maybe not...
- That new world order just keeps getting reordered...
- Remember how we were supposed to be scared stiff that the world would have 10B people on it this century? Yeah, well you can just forget about that. But still be scared, of course. Our leaders can't let us off the hook that easy.
- NVIDIA has developed a GPU library to give a 40x speed-up to the computation of photomask patterns for making advanced chips like GPUs ...
- Skip the browser and run ChatGPT right from the command line using Chatblade or gptcli.
- In keeping with my interest in atoms (versus bits), here's something on mine tailings, the inevitable result of getting stuff out of the ground. (Incidently, since the grade of copper ore has fallen from 50% copper to 0.5%, the best place to mine it is sometimes in the tailings of old mines. The stuff they threw away back then is better than what we can mine right now.)
- There's been some talk on the KiCad forum about using AI for electronics design, including yours truly.
- Is Netflix headed the way of the History Channel?
- A single 13-sided polykite can tile a plane without ever repeating. That sound you hear is a million 3D printers warming up...
- Not to get too hung-up on their demographic woes and super-low fertility rate, but here's an article on South Korea's 4B feminist movement that advocates for near-total avoidance of men. From some of the examples they give, can't really blame 'em...
- Building semantic search engines has now become easy with OpenAI's API. (Side note: Arxiv has 250,000 papers about machine learning!? Who's doing all that?)
- I wrote a blog post about that Twitter discussion I had about concerning circuit design by having ChatGPT/GPT4 output SKiDL code.
- I had an exchange on Twitter yesterday about using ChatGPT/GPT4 with SKiDL for designing electronics.
- Here's a super simple guide to getting the Alpaca large language model running locally on your own machine rather than logging into something like ChatGPT.
- Typst is "a new markup-based typsetting system that is designed to be as powerful as LaTeX while being much easier to learn and use." I like that Typst documents can include scripts whose outputs can be included in the text. Not sure how likely this is to unseat LaTeX in publishing academic papers. Some easy integration into Markdown docs like with MathJax might help adoption.
- I didn't know they never had one before this, but PyPi has started a blog.
- Building a budget compute cluster.
- The West is popular in the West, but not so much everywhere else.
- Turns out Nature's political endorsements didn't have any effect other than decreasing trust in scientists. Good job, guys!
- Here's somebody who got his hardware into the wild but the community to write software for it hasn't coalesced. That mirrors my experience when I was running XESS: you need to do a lot of hand-holding to get things rolling! Also, only about 50% of the dev boards you buy ever make it out of their antistatic bags.
- More about prompt engineering.
- I tried giving ChatGPT a few descriptions from commandlinefu.com and it did a pretty good job generating the command sequences. But then I realized that the whole site was probably part of the training set...
- A modern-day Rosetta Stone containing the Book of Genesis in thousands of different languages etched into a nickel disk.
- It would be interesting to run these 14,166 function descriptions through GPT-4 and see how well it replicates the associated one-line linux command.
- Tired of reading about "GPT this-or-that"? Here's a random list of interesting projects.
- Kids are opting out of college. The education-industrial complex is, of course, in panic. Good (and I say this as a former university faculty member).
- I remember using it in the early 90's, but today it's a hard sell if you tell somebody they can master EMACS in just one year.
- I love a good book-sharing site and sticking it to the Man. Here's how Anna's Archive works.
- Remember how the South Korean government wanted a 69-hour work week? Yeah, well their people had something to say about that. (I also found out S. Korea already had a 68-hour workweek prior to 2018 when it was reduced to 52 hours.)
- The third installment of the ARM story just dropped.
- Here's a link to a tweet thread about a prompt that tries to elicit novel ideas from GPT4.
- Looks like OpenAI doesn't think it's a good idea to be so open. Well, you shouldn't have put "open" in your fucking name! Looks like Microsoft found a good place to burn all the goodwill they earned back over the past decade...
- Vaclav Smil wrote a lot of good articles for IEEE Spectrum, but looks like that may be over. Nothing new since Nov/2022. Maybe he wasn't sufficiently on-board the green energy train.
- People are all right with authoritarianism if it's packaged attractively.
- Here is a foundry service where you can build your own MEMS devices.
- Experiment with AI Transformers in your browser.
- Codon is a compiler for speeding-up Python programs. As usual, it "can’t support certain dynamic features" of Python. It doesn't say which ones or how much this limits its general usefulness.
- Here's a list from Nov/2022 of capabilities emerging from large language models. Who knows how many more have been added by now...
- Whenever buying a car is discussed, my wife always gravitates to expensive cars with loads of safety features. My response: "You want safety? Don't use a cell phone while the car's moving. Boom! You're safer than 77% of the drivers out there and it costs nothing. Now, how about that Fiero?"
- Around 1980, society started shifting from a collective to an individualistic viewpoint according to an analysis of the language used in millions of books. It will be interesting to see if this starts to shift back given the emphasis on acting for the collective good by some segments of the population during the COVID crisis.
- Terrifying plastic rocks are evidence of how our profligate consumerism is destroying the environment. But fishing nets comprise most of these rocks, so unless you're running a trawler ...
- South Korea's Samsung is building a $230B chip fabrication complex, but over the span of 20 years. Gotta compare that to Taiwan's TSMC spending $44B on capital expenditures in a single year.
- Well, GPT-4 got released today. So did Claude.
- This paper fosters "an intuitive understanding of convolutional network deep learning models and how to use them with the goal of engaging a wider creative community."
- Thirteen amino acids and a couple of nickel ions may have started life on Earth: Simple enough to arise spontaneously and reactive enough to catalyze hydrogen to power primitive metabolic processes.
- Ha! This guy just came out with a book about the hardest thing in programming!
- Here's a list of single-webpage AI projects using the huggingface.js library.
- I've always been interested in the Demoscene and how their members are able to get such striking visuals and animations from so little code. Here's an example of a jellyfish animation that fits in a single tweet.
- Just imagine if the resources the US government put into this insignificant case were redirected into domestic infrastructure...
- This article shows how to use the ChatGPT API and includes links to working Jupyter Colab notebooks.
- De-escalation of LLM models continues: now they can run on a laptop CPU.
- A complete connectome for an insect larva "brain" with 3,000 neurons and 550,000 synapses.
- Video refresher on the USB protocol.
- Damn! Missed posting yesterday.
- If the default of Silicon Valley Bank ushers in the collapse of Western civilization, then lowtechmagazine.com will become a must-read.
- With South Korea already having the lowest fertility rate of all nations, how is increasing the work week to 69 hours going to help?
- Octave, the open-source alternative to MATLAB, has come out with version 8.1.
- What can you do with modern debuggers?
- No thanks!
- The Flynn Effect is the 3-5 point increase of a population's IQ each decade since the 1930s, but looks like we've managed to reverse that.
- Another example from "How Big Things Get Done" is in the news: the current projected cost of California's high-speed rail from SF to LA is now up to $100B. Oh well, I guess that's still better than throwing the money at the Ukraine war...
- All this time we were told the "free energy" videos on Youtube were scams, but now it looks like we've found an enzyme that generates electricity just from the traces of hydrogen in the air.
- Skywater open-sourced their PDK for 130nm IC designs a few years ago. Now, IHP has open-sourced their PDK for 130nm BiCMOS intended for analog, mixed signal, and RF ICs.
- This article on computer vision in manufacturing has a lot of good links to more in-depth articles on its use in specific areas of manufacturing.
- This project runs Stable Diffusion in the browser using WebGPU. Good luck limiting the spread of AI...
- Today's reminder that you don't need AI to do dangerous stuff; regular old people are good at it, too: Zombie virus revived!
- Here's a list of interesting publishers in the tech world.
- Multimodal models that operate on several types of input data (text, image, etc.) are now heating up. Here's one from NVIDIA labs that's completely open-sourced. Makes sense since NVIDIA's money comes from hardware and they would benefit by having their model widely used.
- And here's a curated list of software engineering blogs that you'll never get around to reading...
- Brush your teeth so you don't die of a heart attack. Also, mouthwash kills bacteria that help to lower blood pressure. So what the hell...
- This Quora answer demonstrates how doing something different from everyone else can give you a marked leg-up in a competition (in this case, shooting down airplanes). It's similar to the success a football coach had when he decided to flip the 70%-30% ratio of run-versus-pass plays to 30%-70%.
- Of course, doing something different doesn't always lead to success. This paper shows that first entrants ("pioneers") only capture 10% of a new market and seldom maintain their leadership over a later entrant that gains 28% of the market.
- I haven't read this entire article because I've kinda stopped caring about the whole racist/anti-racist back-and-forth, but I'm really tired of publications like Science, Nature and Scientific American becoming so overtly political in their coverage (even to the point of endorsing political candidates).
- There is now a tufte.css for formatting your HTML articles in the clear, minimalistic style of Edward Tufte. Although not the main focus, Tufte-LaTeX is also mentioned.
- A hospital in Canada has surgically grafted healthy nerves onto the damaged ones of a dozen quadrapelgics to restore the use of their arms and hands. (This is a press release from the hospital, not a peer-reviewed article.)
- Another example of why we just can't build anything.
- Looks like someone was able to reconstruct the images being viewed by feeding fMRI scan data from the language and visual areas of a brain into a Stable Diffusion model.
- Somebody submitted a PR to Meta's LLaMa Github repo with a link to a torrent of the model weights. (Normally, Meta wants you to submit a form to get access to those.) In case it vanishes, here it is: magnet:?xt=urn:btih:ZXXDAUWYLRUXXBHUYEMS6Q5CE5WA3LVA&dn=LLaMA
- Maybe a few years to late to this party, but if you need to do a little performative moralizing, here you go.
- I'm reading a book called "How Big Things Get Done". At one point, it discusses the construction of the Empire State building which was done in about 12 months from start to finish. This blog post contrasts that project with the building of the World Trade Center and shows that the workers on the Empire State building were 2x - 4x more productive. And this blog post bemoans the fact that America can't seem to build anything.
- Iran must be making one hell of a "nukelear" reactor!
- Politicians are slowly waking up to the fact that they can't just legislate and expect the atoms to fall in line...
- Good luck, kids.
- If I ever got around to making a game, I'd look seriously at Godot. It's getting to that Blender level of critical mass.
- The AI LLM of the day is ... KOSMOS-1! This one works with different types of data (text & images) and can do things like explain what's going on in a picture or understand the text in an image without using OCR.
- OpenAI has released APIs to ChatGPT and Whisper. In addition, there is an upcoming ChatML format that lets you interact with GPT models using JSON-formatted data.
- Operating at a lower level than things like GPT, JAX augments Pyton's
numpyto make it easier to perform machine learning operations. - Need a QR code for a URL? Here's a set of Python function calls that does it.
- Looks like you're not going to get that eco-friendly lettuce from a vertical farm. Oh, and that ethically-pure lab-grown meat? Not getting that, either.
- I know you're looking forward to the government slapping some more regulations on working-from-home. But don't worry! You'll only have to endure them for four days a week.
- Meta open-sourced their version of GPT-3 called LLaMa which uses 10x fewer parameters while getting better performance. Now there's an open-source version of ChatGPT built with it.
- Here is a page with suggested links for keeping up on current AI developments, along with an older post about how to get started in learning AI.
- This blog post contains a good timeline with links to important papers in deep learning.
- In-Q-Tel is basically the venture capital arm of the US intelligence community (think CIA). Here is one of their reports describing how the ability to both read and write DNA will affect national security and economic competitiveness.
- I updated my KiCad on Docker repo so it supports versions 5, 6 and 7. Now I can run any of those versions with a simple command like
kicad5and they don't interfere with each other.
- I ordered one of these CH32V003 RISC-V microcontroller dev kits. Shipping is more than the cost of the kit!
- AlphaCode is an AI from DeepMind specifically aimed at the task of generating good solutions for competitive coding contests.
- I pointed to PlantUML a few days ago and now here's a Mermaid cheatsheet for doing similar text-based graphics within a Markdown file. All these things seem to rely on the remarkably long-lived graphviz tool.
- Looks like we have another web-based PCB design suite: Flux. Your design is stored in their central database and I'm not sure if you can export your design files. So what happens if the company ceases to exist or if the server is just down?
- Lots of interest in demographics since the developed/developing world has a problem keeping their birthrates above replacement level. Check out the South Korean birthrate.
- This Dockerfile gives you a self-contained build environment for Raspberry Pico programs.
- Want to program the 10¢ CH32V003 microcontroller but hate the idea of using their proprietary tools? This repo supports an open source alternative using GCC.
- Extensive guide to prompt engineering for various generative AI tools.
- Here's a large categorized list of EDA tools for describing electronic systems.
- Raspberry Pi is selling a small hardware debug module to make it easier to access the SWD port of a Pico (or any other ARM processor).
- Just to maintain the ARM theme, here's part 1 and part 2 of a history about ARM.
- Here's another FOSS solution for running large language modules with just a single GPU.
- Somewhere on my travels through ChatGPT-land, I ran across Colossal-AI, a FOSS toolkit for replicating ChatGPT using fewer compute resources (i.e., GPUs) and with faster training. I'm not expert enough and haven't tried it to see if that's true or just BS. But it's open source so you can look for yourself.
- If you're into making animated visualizations in the browser, SwissGL is a minimalistic wrapper on top of the WebGL2 JS API. This looks like a quick way to get GPU-assisted visualizations onto the web.
- What will soda cans look like in forty years?
- I was reading this article about using ChatGPT to create system modeling diagrams by having it output markdown containing Mermaid (which is amazing by itself). But this led me to the FOSS tool PlantUML that lets you generate many types of diagrams from a text description: UML, mindmaps, mathematics, etc. Using ChatGPT to drive text-based tools like these with natural language inputs gives them capabilities that aren't readily available in GUI-based tools.
- Interesting article about how both Ukraine and Russia are somewhat dependent upon drones from a private company (DJI) in China. DJI has applied geofencing to restrict the use of their drones in wars in the Middle East, but not in Ukraine (so far).
- Meta is handing out a lot of negative reviews to its workers. It also wants to solve the noise problem in its offices. I guess having fewer employees would be one way to do that...
- This article covers how ChatGPT works. Along the way, it talks about models, neural nets, training, embeddings, semantics, etc. I haven't read it in detail, but I plan to. Better than reading WSJ articles about Meta...
- I use PySpice as a Python interface to ngspice to provide circuit simulation capabilities in SKiDL. Now there's a fork called OpenSPICE that adds a native Python simulator. There's not a lot of documentation I could find outside of this article. In the comments, there's also a link to PyLTSpice which lets you drive LTSpice through a Python interface.
- I've always been interested in forecasting, even though I suck at it. Superforecasting is a great book on how to make better predictions of the future, and this article is a good, broad-brush exposition of the principles.
- I frequently use the website remove.bg to get rid of backgrounds in images I want to use as clip-art. I just ran across the rembg Python package that does the same thing locally, and the code can be viewed here.
- New private-sector, DARPA-like organizations are springing up to advance foundational technologies.
- Maybe you're like me: sometimes you use Pandas, but you don't really know Pandas. If so, this extensive visual guide to Pandas may be for you. (Note: This is hosted on Medium so you may be able to view it or you may not. Who knows...) And if actually learning how it works is too much, there's a Python package called sketch that uses AI to assist you in using Pandas.
- Looks like Z Library isn't going down without a fight.
- I'm a big fan of using programming languages for doing CAD work. GhostSCAD puts a Go wrapper around OpenSCAD to achieve that. CadQuery (Python) and JSCad (Javascript) are other alternatives mentioned in the comments.
- Maya Posch writes a lot of good articles at Hackaday like this one on methane pyrolysis. Also check out the mining and refining section by Dan Maloney for facts about the stuff that makes up our industrial society.
- Japan is going full bore on nuclear. Looks like the government will be paying for their share of the build-out by charging "carbon fees" to companies using fossil fuels.
- Looks like Intel Foundry service isn't going to be directly supporting RISC-V.
- You can get a lot of prepackaged Docker containers for FOSS EDA tools here. This page lists the available tools. Most of these are HDL/FPGA/VLSI design tools so don't expect to get containers for things like KiCad and FreeCAD.
- A while ago, I built a Docker container for KiCad. It's for version 5, but it should be easy to modify it for versions 6 and 7 (when it becomes available).
- Here's the latest effort to repurpose old smartphones into a compute cluster. Unfortunately, the phones are loaded with the Ubuntu Touch linux variant and that isn't supported on any of the old phones I have lieing around.
- A couple of pro-nuke articles about uninformed fears of nuclear waste and the nuclear renaissance.
- Instead of running Doom on a soldering iron, it seems the new trend is to implement ChatGPT in as few lines as possible. Here's one that does it in 60 lines of NumPy.
- Last week, America was incensed that a Chinese spy ballon traversed our airspace. This week, we learned that we blew up Russia's Nordstream gas supply lines...
- I finished "How to Drive a Nuclear Reactor". It describes the systems of a Pressurized Water Reactor (the most common type) and how a control-room operator starts, manages, and stops it. It's a bit more technical than your standard popsci book, but not as challenging as a textbook. Worth a read if you're interested in nuclear power.
- Helion Energy is trying to build a 50MW fusion reactor that fits in a shipping container and costs $10M by 2030. Seems wildly optimistic but I look forward to being surprised.
- Over the course of my life, I've noticed that being smart (or being acknowledged as smart) has ascended in importance. But I would say that most people labeled as smart are merely adept (skilled within some area of knowledge). I think that OODA provides a better framework to judge a person's intelligence by seeing if they can 1) Observe things others don't notice, 2) Orient them into an existing set of facts, 3) Decide what to do based on this new knowledge, and 4) actually Act upon their decision. The number of people who can do this in one field, much less many, is extremely small. They often don't fit the stereotype of smart. And I am not one of them.
- There's been a lot of talk-talk recently by US "thought leaders" about re-emphasizing manufacturing. This article beats the same drum, except it's from 2017. So the talk has been going on for a while. Despite things like the CHIPS Act and the Inflation Reduction Act (LOL), I'm still waiting to see increases in manufacturing employment. Or lots and lots of robots...
- K6 will generate “virtual users” who continuously run test scenarios defined using Javascript.
- Interesting blog post by Jay Carlson about a new Cortex-M0+ microcontroller that costs less than 10¢.
- Ian Norris - Blog - Understanding the limits of large language models https://inorris.com/Blog/GPT/