Show HN: HackerBooks – HackerTube, but for books mentioned on HN 596p 139c
Not affiliated with Y Combinator or Hacker News. Data: HN comments via open-index/hacker-news, book metadata via Open Library.
beanjuiceII [flagged]
Imustaskforhelp I am really loving what people are building on similar ideas close to HackerTube and I am tinkering something of my own as well with these new ideas. Awesome Idea and I would love to see what else community tinkers with things like these :-D Hackernews is really a gold-mine in many ways. Edit: I will try to read thinking fast and slow because of HN recommendations. I have heard some good things about the book but I will try to read it in future given HN's immense recommendations, so your website has already had a negative net positive for me :-D
LeComedien Ayn Rand so high, though!
jmathai I love this idea and hope the trend continues. It could even lead to finding niche communities to join.
jchook The site is really fast. I see HTTP/3, OpenGraph, and Cloudflare. Can you expand on the technologies used? Also... "Each book links back to the comments that mentioned it." > Please don't put generated text in HN posts. Claude speak gives me a gag reflex these days.
vigilantpuma The “mentions since” picker appears to be broken, nothing in the list changes as far as I can see. Also I wish it wouldn’t add another page to my history.
HackerTube: YouTube built from videos recommended by HN 2065p 482c
It was Friday night at 11pm, I was tired, I went to YouTube to watch something and realized that I was looking at an enshittified wall of slop. Insanely exaggerated thumbnails, dumb recommendations of the same tired topics, truly terrible idiocracy-grade content. I'm sure this algorithm works well for the median person on the internet, but it definitely does not work for me. So this morning (Saturday) I took a hold of my Opus 5.5 and gave it the task: Let's write our own YouTube clone based on whatever videos have been recommended on Hacker News. And Opus came back with: hackertube.org!
Opus downloaded all comments of HackerNews, filtered them, ran Sonnet over them in a Workflow (Claude Code feature), got rid of spam, and narrowed everything down to 40K comments recommending a YouTube video. In total, we get 6,454 YouTube videos that have been recommended by a few people on HN from 2007-2026 across 150 topics. Then we wrap it up in a YouTube UI, get a domain, push to a VPS, secure, harden, optimize and voila!
Uses the 50% thumbnail instead of the uploaded thumbnail
Custom and good ForYou recommendations without any user accounts, storing state in local browser session
Export of the site's data available to give to your agent to easily create your own version
This takes ~2 hours now. I can breathe again. Thank you Opus 5.5 <3.
EDIT: A little bit more on my process. I use Claude Code (desktop app for Mac). I use Opus 5.5 on Medium effort as my default. I don't actually do anything resembling traditional engineering anymore (read/write code, run commands in a terminal, or etc.), I only talk to Claude and have it do everything. In this case, the only "manual work" I did was: 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box and 3) I bought the domain name. I worked on this in between Claude's actual work so it still feels a little bit like team work :) Once the box was up I gave Claude the IP and told it update, deploy, secure, harden and optimize the code, and it did the whole thing.
sc970 How many tokens (or $) does something like this take to make?
liquicity Nice idea will watch though some - I made something similar but for food vids (filters over YouTube chefs - such as only vids with ingredients) https://mealplannr.io/discover
privacyishuman I like this a lot. I was just thinking about a site where you could share your YouTube cookies and then use anybody's curated algorithm because you can curate it quite fast in signed out mode. Could you block all the trackers from loading from https://www.youtube-nocookie.com on your side?
farrelle25 What a great idea. Does anyone know of a way of 'getting' / downloading all those URLs (maybe with the video title)..? I'm afraid of the site disappearing.
mentalgear Appreciate the effort, yet I think there is a place this is needed even more: > It was Friday night at 11pm, I was tired, I went to YouTube to watch something and realized that I was looking at an enshittified wall of slop. Please do the same thing on Social Media: Facebook, 'X', reddit, instagram, etc . Maybe find a 'hacker signal' that identifies quality content across those platforms and have them aggregated for the public so we don't have to spend time in the fascist network prisons. The greatest trick - if it even can be pulled - would be setting it up in a way that it can't be sued out…
vfclists Where is the source, so I can roll my own specialized video site?
jumploops I’m always surprised how bad YouTube’s recommendation algorithm seems to be. Maybe it’s because I don’t watch enough videos, but it’s sort of like Amazon, in that it will suggest the exact same content/topic. I rarely use TikTok, only to view links people send me (similar to YouTube), but the recommended content is miles ahead, making it much more difficult to close compared to YT.
morphle Some of the Alan Kay lectures are on hackertube
gosen Nice idea, thanks Andrej! When will you add HackerTube Shorts? (kidding :D)
jahala This is great, would love filters for duration and sort by upload date
Nibnib: Image compression with extreme decompression speed 24p 2c
GitHubRepository hosting, code review, and collaboration
CopilotAI coding agent for tasks, projects and automated work
Code QualityQuality analysis, fixes, and enforcement
Code SecurityVulnerability detection and remediation
Secret ProtectionSecret scanning and leak prevention
ChangelogYour source for the latest features, improvements, and fixes across GitHub products, all in one place.
GitHub Advanced SecurityEnterprise-grade code security
Copilot for businessEnterprise-grade AI features
Premium SupportEnterprise-grade 24/7 support
GitHub UniverseThe event for developers, by developers — talks, workshops, and what’s next for GitHub.
NotificationsYou must be signed in to change notification settings
Star19 (19)You must be signed in to star a repository
Image compression at 4bpp with extreme decompression speed
An image format tuned for extreme decompression speed.
This image format originated as an experiment to see what image quality and performance were possible for DVI output on an RP2350 microcontroller. On devices of that class, it's common to use low bit depth images and palette lookup. There's a budget of approximately 6 cycles per pixel, and that's enough for a pixel fetch, palette lookup, and store into the line buffer, but not a lot of room to spare.
The nibnib format uses a palette at its core, but surpasses palette images in quality because the palette stores a delta rather than the absolute color, vastly improving soft regions. It achieves its extreme speed by decoding two pixels at a time, from a single byte of input, thus attaining 4bpp compression.
The decompression process is simple and deterministic. There are many choices when encoding; the algorithm is asymmetric in that regard.
Decoding works in pixel pairs. The last decoded pixel is doubled up. Then we read a byte from the input stream, and look it up in the palette. That 32-bit value is added to the doubled-up previous pixel, yielding two 16bpp (RGB 565 format) pixels. The low 16 bits are the left pixel. That's it.
Here's the inner loop (lightly adapted from the Rust source):
// first two pixels copied from src[0..4], initializing pix for x in 1..(width / 2) { let inp = src[x + 3]; pix = ((pix >> 16) * 0x10001).wrapping_add(pal[inp as usize]); out[x * 2] = pix & 0xffff; out[x * 2 + 1] = pix >> 16; }
The RP2350 on-the-fly decoder implementation is included in the image demo of the vic-2350 repo.
As with many image compression formats, nibnib is defined by the decoding process, which is deterministic and straightforward. Encoding is fundamentally a search through the space of possible encoded images to find one that decodes to a (hopefully) high quality replica of the original. In nibnib, this search is quite asymmetrical; the current encoder is orders of magnitude slower than decoding. By contrast, in JPEG, the DCT transform is invertible by its mathematical nature, and quantization is also easy; at baseline, encoding and decoding are about the same speed, though search can be used to fine-tune parameters and micro-optimize quality.
Nibnib has two main dimensions for the search: choice of palette, and selection of codewords. For the initial version, the palette has a cube of small delta values plus a block of random values, chosen by weighted Gaussians. The cube is necessary to ensure that values near 0 and full scale can be attained without risk of overflow; without the cube entries, you get bad artifacts. First is the original image, then the image encoded with a random-only palette, finally with both the cube and random values.
The random Gaussian process was refined on an ad-hoc basis, using guesses about reasonable values. In particular, the delta between the two pixels gives a tradeoff between sharpness and noise; keeping it small makes the image smoother at the cost of reduced ability to encode sharp edges.
In the current encoder, the palette is not chosen adaptively based on the input image, but that is most definitely an opportunity to improve quality.
The basic encoding loop is conceptually straightforward, but the details are a bit subtle.
Conceptually, it simply tries all 256 palette entries, and chooses the one with the minimum error. Then it takes that error and distributes it using Floyd-Steinberg weights. Subsequent pixels fit not the source pixel value, but the source pixel plus errors from previous iterations, weighted.
The loop as described above is quite similar to error diffusion with a quantized palette, a very popular technique. However, that's…
wiml Fun. Vaguely reminiscent of the Amiga HAM video mode.
Mathematics has changed. Let's act like it 117p 131c
Updates on my research and expository papers, discussion of open problems, and other maths-related topics. By Terence Tao
1 November, 2026 in guest blog, opinion | Tags: Frank Calegari | by Terence Tao
[This is a guest post by Frank Calegari. This blog post was initially written in a different file format and converted using AI. — T.]
The release by OpenAI on October 6 of over 700 research papers was a watershed in the history of mathematics. AI is now undeniably capable of making groundbreaking contributions across many areas of mathematics. So why are some mathematicians so unhappy?
Some of it is the apparent disregard by AI companies for the practice of mathematics. The OpenAI papers — despite some tremendous ideas — are atrociously written and violate our community expectations of attribution. Some of them are even wrong. We have much more control over how we hire mathematicians, publish papers, and train students than over the behavior of trillion-dollar companies, and that is where our efforts would be better spent.
Much of the reaction seems to reflect a feeling that the subject that they love is under attack, and is being “hollowed out”. The recent OpenAI dump made barely a dent on the general problems of functoriality and reciprocity in the Langlands program. But even if it had solved every problem in the Langlands program, to imagine that would have “solved number theory” would be as naive as saying that the construction of the periodic table solved chemistry. Mathematicians should do a better job explaining that when we talk about the central problems in any area, we are merely discussing the most natural boundaries of our understanding — and that unlocking them is not the end but the first step to finding new playgrounds to explore. Of course, this takes time. I looked for areas of mathematics which were “harder hit”. It is hard not to be struck by Hugo Duminil-Copin’s lament:
Not a single one of the major open problems I have publicly mentioned throughout my career (whether in a talk, a lecture, an article, or even a grant proposal) was left untouched by the announcement. Everything has been claimed to be proved.
I expected to see a few of them in the list. But not all of them. Not all at once. Not with such nonchalance.
The classification of finite simple groups.
I can’t say much about percolation theory, but there is another historical precedent I think mathematicians should take note of. In 1972, Dan Gorenstein announced a program towards the classification of finite simple groups. After a flurry of activity over around a decade, the “classification” was announced by Gorenstein in 1983 (somewhat prematurely, with the case of quasithin groups not being completed for another two decades by Aschbacher and Smith).
The classification of finite simple groups was in many ways a huge success, but it was also a failure in that it — apologies to my group theorist friends and collaborators — cooled a lot of the interest around the subject. The “famous problem” had gone, and mathematicians on the whole turned their attention elsewhere. And yet, there remain very rich questions in finite group theory and even about the classification itself (to name just one: John Conway’s favorite question is why does the Monster group exist?). Perhaps for some fields, the OpenAI dump feels like how a group theorist may have felt in 1970 if the classification was dumped in their office, not in the form of the actual papers, but printed out on a spirit duplicator, smudged a little, then spread randomly like confetti all over the floor.
Right now, we need efforts like that of Gorenstein, Lyons, and Solomon and their work on the revised classification. If this was all we had to do, then it would be a fairly thankless task. But we also have the responsibility not to pronounce the subject “done” but rather — with these new tools available — ripe for further investigation. When mathematicians lament the winding paths missed by the construction of a galactic bypass, they need to realize that those old paths still exist and seek them out. And we need to reward that behavior when it comes to offering people jobs.
The promise of AI, and the (dis)-unity of mathematics:
Since I was 15, I’ve wanted to know what was. First, there was the childhood fantasy of finding a formula that Euler missed. But soon one learns more: how Grothendieck’s period conjecture, together with a sufficiently fine understanding of the geometry of mixed Tate motives (following…
bananaflag > But even if it had solved every problem in the Langlands program, to imagine that would have “solved number theory” would be as naive as saying that the construction of the periodic table solved chemistry. Of course if you solve any existing problem, new ones appear. That's not the point. The point is that AI will be able to advance on any problem better than any human.
0c3ca83 > But even if it had solved every problem in the Langlands program, to imagine that would have “solved number theory” would be as naive as saying that the construction of the periodic table solved chemistry. This analogy is pants on head stupid. They're talking as though the only thing AI will ever solve is the Langlands program, not as though we have discovered a generic tool. It's like saying "they drove the car one mile -- and every mile driven leads to more roads that we can trot along, why are you so worried about the future of horses?"
qasg17 Have the pro-AI people finally found a new mathematician? Otherwise it is always the same 10 people on Tao's blog citing each other. This guest author runs a seminar to lure people into AI: https://math.uchicago.edu/~fcale/AI.html
fooker It’s amazing for me to see math research going from craft to technology. CS had this moment decades ago, sans some theoretical parts that are joining the bandwagon now. I did a bit of math research in grad school collaborating with my math postdoc roommate. I found the aversion to any computation making its way into the craft to be odd decades ago. Well, now the proof’s in the AI generated pudding.
bsaul Isn't it the passage of mathematics from the age of craftsmanship to the age of industrial making ? My recent take on AI is that people haven't adapted to the new huge potential of the tool, and don't dream big enough (yet). It's like seeing a car engine for the first time, lamenting about the future of horses, and not realizing it'll also help us fly.
kccqzy > To me, someone who chooses a worthwhile problem, obtains a proof completely using AI, understands it thoroughly, and uses it to develop interesting further questions or directions has made a substantial mathematical contribution. I would give that person substantial credit even if they supplied no new ingredient in the original proof. Yup. This is what Terence Tao has been saying about digestion of results for several months now. Mathematics has become a field where results themselves matter less, the social aspect of attracting others to understand your results and explore new problems has…
Tips for more effective fuzz testing with AFL++ 22p 1c
Tips for more effective fuzz testing with AFL++
Fuzz testing is incredibly effective for mechanically discovering software defects, yet remains underused and neglected. Pick any program that must gracefully accept complex input, written in any language, which has not yet been been fuzzed, and fuzz testing usually reveals at least one bug. At least one program currently installed on your own computer certainly qualifies. Perhaps even most of them. Everything is broken and low-hanging fruit is everywhere. After fuzz testing ~1,000 projects over the past six years, I’ve accumulated tips for picking that fruit. The checklist format has worked well in the past (1, 2), so I’ll use it again. This article discusses AFL++ on source-available C and C++ targets, running on glibc-based Linux distributions, currently the indisputable best fuzzing platform for C and C++.
My tips complement the official, upstream documentation, so consult them, too:
Even if a program has been fuzz tested, applying the techniques in this article may reveal defects missed by previous fuzz testing.
More assertions means more effective fuzzing, and sanitizers are a kind of automatically-inserted assertions. By default, fuzz with both Address Sanitizer (ASan) and Undefined Behavior Sanitizer (UBSan):
$ afl-gcc-fast -g3 -fsanitize=address,undefined ...
ASan’s default configuration is not ideal, and should be adjusted via the ASAN_OPTIONS environment variable. If customized at all, AFL++ requires at least these options:
export ASAN_OPTIONS="abort_on_error=1:halt_on_error=1:symbolize=0"
Except symbolize=0, this ought to be the ASan default. When debugging a discovered crash, you’ll want UBSan set up the same way so that it behaves under in a debugger. To improve fuzzing, make ASan even more sensitive to defects by detecting use-after-return bugs. It slows fuzzing slightly, but it’s well worth the cost:
ASAN_OPTIONS+=":detect_stack_use_after_return=1"
By default ASan fills the first 4KiB of fresh allocations with a pattern, to help detect use-after-free bugs. That’s not nearly enough for fuzzing. Crank it up to completely fill virtually all allocations with a pattern:
ASAN_OPTIONS+=":max_malloc_fill_size=$((1<<30))"
In the default configuration, if a program allocates more than 4KiB with malloc then, say, uses strlen on the uninitialized memory, no bug will be detected. There’s almost certainly a zero somewhere after 4KiB. Until I noticed it, the 4KiB limit hid a number of bugs from my fuzz testing. Per (4), fulling filling allocations with a pattern better isolates tests when using persistent mode.
When fuzzing C++ and linking GCC’s libstdc++, consider -D_GLIBCXX_DEBUG. ASan cannot “see” out-of-bounds accesses within a container’s capacity, and the extra assertions fill in the gaps. Mind that it changes the ABI, though fuzz testing will instantly highlight such mismatches.
While AFL++ can fuzz many programs in-place without writing a single line of code (afl-gcc, afl-clang), prefer AFL++’s persistent mode (afl-gcc-fast, afl-clang-fast). It’s typically an order of magnitude faster and worth the effort. Though it also has pitfalls (see (4), (5)). I keep a file on hand, fuzztmpl.c — the progenitor of all my fuzz testers:
#include <unistd.h> __AFL_FUZZ_INIT(); int main(void) { __AFL_INIT(); char *src = 0; unsigned char *buf = __AFL_FUZZ_TESTCASE_BUF; while (__AFL_LOOP(10000)) { int len = __AFL_FUZZ_TESTCASE_LEN; src = realloc(src, len); memcpy(src, buf, len); // ... send src to target ... } }
I :r this into my Vim buffer, then modify as needed. It’s a stripped and improved version of the official template, which itself has a serious flaw (see (5)). There are unstated constraints about the position of buf and len in the code, so if in doubt, refer to the original template.
(3) Include source files, not header files
We’re well into the 21st century. Nobody is compiling software on 16-bit machines anymore. Don’t get hung up on the one translation unit (TU) per source file mindset. When fuzz testing, we need at most two TUs: One TU for instrumented code and one TU for uninstrumented code. In most cases the latter takes the form of a library (libc, libstdc++, etc.) and we don’t need to think about it.
Fuzz testing typically requires only a subset of the program. Including just those sources straight in the template is both effective and simple. In my template I put includes just above unistd.h so that the header isn’t visible to the sources unless they include it themselves.
#include…
monster_truck These are good tips, thank you
Copper – A minimal low-level programming language with OOP and generics 89p 49c
Copper is a programming language that I use as a replacement to C. The main differences are:
Minimalism is the most important feature, the language is intended to be simple so that someone with limited brain power such as me could fully understand it. The syntax is simple, almost everything is explicit and most constructs are common.
A simple example beyond the pointless "hello world".
class Point attr x: Int attr y: Int def init(x: Int, y: Int) self.x = x self.y = y end def translate(dx: Int, dy: Int) self.x += dx self.y += dy end end func main: Int var pt: Point! pt.init(3, 5) pt.translate(1, 0) println("x = \a, y = \a", pt.x, pt.y) return 0 end import "std"
Cheat sheet - if you're not interested but want to understand existing code
Copper is released under the BSD license.
bogdanoff_2 Not mine, but I found this to be really interesting as a language created by one person for their own use. Also really interesting are the author's blog with articles on language design (for example his thoughts on exceptions, and how he added them to his language, then decided against them), and the change log for the language, showing its evolution. https://tibleiz.net/ https://tibleiz.net/copper/changes.html
jdw64 It’s similar to Python syntax.
DenisM Garbage collected? Couldn’t tell from the docs.
skybrian It seems pretty tasteful, but I don't see anything about character sets. How does it support Unicode? What does a string of 16-bit characters represent? There do seem to be some Unicode constants in the standard library: https://tibleiz.net/browser.html?path=std&title=Standard%20L...
An Italian farmer found some unknown seeds in his attic and planted them 720p 144c
Subscribe to Smithsonian magazine and get a FREE tote.
When a seven-eared stem emerged from the ground, it seemed like a miracle. What Claudio Grossi had unearthed was an ancient grain that proved to be an answer to some modern problems
Sometime in the early 1990s, Claudio Grossi, a 40-year-old farmer in Parma, in northern Italy, was cleaning his attic when he turned up an old sack full of seeds that he couldn’t identify. The bag, he was sure, had been left by his late great-grandfather, who’d started the farmstead around 1915. Grossi had no idea how long the seeds had been sitting there, deprived of water, soil and sun. Dry and cracked, the kernels looked so old in the palm of his hand that he doubted anything would ever sprout from them. He carried the bag down the stairs, shook the seeds out into a pile of manure and then forgot about them.
To his surprise, the seeds soon sprouted a few different types of wheat. One was particularly striking: At six feet high, its plants were about twice as tall as commercial wheat. But the most remarkable feature came when it developed ears—the grain-studded parts at the top of the plant. All of the wheat types Grossi had previously seen grew only one such ear. But the tall wheat grew a full seven ears, always in the same pattern: a long one with six shorties bunched up around it like little fingers. Even the kernels themselves were bigger than those of average wheat. And unlike the typical pale-yellow wheat kernels, they came in a rainbow of colors—yellow, black, red, white and even green, sometimes all on the same stem. Grossi, a grain farmer his whole life, had never seen anything like it. “I thought it was a freak of nature,” he says. He couldn’t find a mention of it in agricultural books.
Subscribe to Smithsonian magazine now for just $19.99
This article is a selection from the November 2026 issue of Smithsonian magazine
For Grossi, a practicing Catholic, the next insight came from a different book altogether. Leafing through a Bible at church one Sunday, he found a passage in the Book of Genesis describing a stalk with seven ears, in a dream that Joseph interprets for the Pharaoh: “And he slept and dreamed the second time: And, behold, seven ears of grain came up upon one stalk, rank and good. And, behold, seven thin ears and blasted with the east wind sprung up after them. And the seven thin ears devoured the seven rank and full ears.” To Grossi, that sounded just like the mysterious crop in his manure pile.
“I connected the dots,” Grossi says. That grain must be really ancient, he thought. And if people were farming this odd-looking wheat in biblical times, why does no one today?
Back on his farm, Grossi harvested the grain, milled it into flour and baked a loaf of bread in his kitchen. Compared with commercial wheat, the dough didn’t rise well, but as the loaf sat in the oven, it filled the kitchen with a surprisingly nutty and delicious aroma. Once finished, it had a rich, satisfying taste. “You get these notes of sweetness that you can’t find in regular commercial flour,” he says.
Did you know? More miracles of grain in the Bible
In 2 Kings 4:42-44, Elisha multiplies 20 barley loaves in order to feed 100 men—anticipating Jesus’s “Feeding of the 5,000” in the gospels.
In 1 Kings 17:8-16, Elisha helps a widow at Zarephath to outlast a drought and famine with only a handful of flour and a little oil.
In Judges 7:9-15, a Midianite soldier recounts a dream in which a barley loaf tumbles into the camp and overturns a tent. Gideon, the judge and military leader, recognizes the dream as a sign that God will deliver the Midianites into Israel's hands.
In 2 Kings 4:38-41, during a famine, someone gathers wild gourds and cuts them into a pot of stew. The diners discover that the stew is dangerous and cry out to Elisha. He calls for flour and throws it into the pot—and the stew becomes miraculously safe to eat.
Grossi discovered it had other advantages: It was hardier and more self-sufficient than commercial varieties. It fared better during droughts, growing so tall that it overshadowed the weeds. It didn’t need fertilizer or herbicides. Indeed, it seemed so superior to its industrial cousins that he thought of it no longer as a freak of nature but as a wonder. He dubbed it grano del miracolo—“miracle grain” or “miracle wheat.”
Grossi’s discovery came at just the right time. In Italy in the 1990s, a burgeoning slow-food movement aimed to preserve traditional farming practices and promote…
the-golden-one Clickbait title
tmaly This is interesting for people with gluten allergies, maybe there is a way they could one day consume the ancient wheat.
caretrs Wouldn't this cause ecosystem issues if seeds from 100 years ago get planted in the same area?
comrade1234 Many farmers here in Switzerland grow urdinkel - ancient spelt. In the USA most farmers grew it also until around the 1920s but then a new wheat came out that was easier to process. Spelt has thicker hulls and wears down the mill faster so the mills switched to the new wheat and the farmers all had to switch too. Based on the name I think ur (ancient) Dinkel (spelt) is even older than regular spelt but I'm not sure. All I know is that it makes great crackers but terrible pasta and ok bread but not great.
mech422 I think I saw this movie... :-P
mattoxic Damn, I was hoping for Triffids
Proving and improving the Platonic Representation Hypothesis with JEPAs 90p 48c
Abstract:Whether models trained on different modalities converge to the same representation of the world is central to AI. The Platonic Representation Hypothesis (PRH) conjectures that they do as they scale, but without a proof it offers neither a guarantee nor a recipe. We turn the hypothesis into a theorem. First, we prove a \emph{Platonic Representation Theorem}: Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn the same representation up to one rotation whenever their positive pairs are two observations of one scene--\emph{good views}--and their misalignment is bounded otherwise. Second, we show that next-token prediction is a JEPA with \emph{bad views}, parts of a single document, which explains why language models align with vision only through scale. Third, we validate the theory from 145k captions to the texts of 940M web images, and uncover a scaling law of alignment that predicts unseen runs within seed noise. Good views need 42 times fewer training samples than next-token prediction on the same corpus to reach the same alignment, and a 109M-parameter text JEPA aligns with DINOv2 and CLIP's image encoders better than every language model we evaluated, up to 72B parameters, and than CLIP's text encoder: next-token prediction is a remarkably inefficient route to platonic representations. As a showcase, one rotation turns separately trained encoders into an image captioner (CIDEr .838 on COCO). Even that rotation needs no pair: on the latest pair-free benchmark, our text JEPA beats every zero-shot accuracy reported with language models up to 8B parameters, e.g., 64\% against at most 56\% top-5 on ImageNet-100. Positive views, and not scale alone, are thus the lever to build platonic representations.
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Lerc I'm not sure how I feel about the Platonic Representation Hypothesis. In some ways it seems like it should be profound, but part of me thinks it might just be saying 'There is an optimal way to perform dimensionality reduction and as things get bigger it is easier to appoximate'
suddenlybananas I thought the platonic representation hypothesis was much more ambitious: wasn't it supposed to apply for any model across modalities (e.g. including human minds), rather than just the same model but with different modalities? I'll have to actually read the paper but I suspect there to be very strong assumptions which defeat the interesting aspects of the platonic representation hypothesis.
inshard Do memes have a concept of proto memes? Like primes are the proto numbers that all other numbers are made from? Could there be similar ontological primes for all concepts / semantic space? Base reality concepts that are irreducible so to speak.
junkaccount I built a ppt with claude for investors with the exact same phrase "AI infrastructure for proving and improving ... " on Oct 27 and sent it to investors. I suggested the title during the chat and claude agreed a lot. It ends up as a title of this paper 2 days later.
Lenovo's 'encrypted' fingerprint USB is just an unencrypted SD card in a case 299p 111c
(no text extracted)
mzajc Linked article: https://www.tomshardware.com/tech-industry/cyber-security/le...
matheusmoreira Does anyone actually trust these products? Always seemed like a terrible idea to me. The correct solution is to use free software cryptography on top of these things.
nik282000 Marketing wank made to sell e-waste to middle management.
ctkhn Didn't expect something this crazy but I will not trust stuff like this until I've seen a teardown. Imagine being the product manager who approved this shit
jagged-chisel Why a link to Reddit instead of the link to Tom’s Hardware?
dpacmittal Just use veracrypt or ecrypt on any memory card. These devices don't deserve the premium they charge for on-device encryption which in this case is not even encyrypted
Bidirectional charging during grid outages 65p 89c
Bidirectional Charging During Grid Outages
The basic problem is that while several manufacturers support having the car power your house when the grid is down ("V2H"), they generally treat it as a separate discharging-only mode. For example, the GM V2H Bundle documentation says "During a power outage, the GM Energy Home Hub safely disconnects your home from the local power grid and enables backup power to flow from the GM Energy Inverter throughout your home" and "Once power is restored, your GM EV will automatically end the discharge session and begin charging" without any mention of recharging the car when the grid is down.
For Fords, automatically powering your house during a grid outage was a big selling point in the F150 Lightning marketing campaign, but they now say "The Ford Charge Station Pro, required for Automatic Home Backup Power, is no longer for sale", probably because the F150 Lightning has been discontinued. And it looks to me like even then they didn't support bidirectional charging during a grid outage. The only option today is V2H via a transfer switch, which doesn't let you charge an F150.
For Kia, there's Wallbox's Quasar 2. Unfortunately this is a mess where Wallbox is very enthusiastic and says V2H works, but then customers run into trouble actually trying to use it. And it doesn't work with current models, plus "Solar Charging cannot be used simultaneously with ... V2H".
The only way I can see to do this today is with Tesla Powershare + Cybertruck, which says "If the system includes a UL 1741-listed solar system and there is capacity in the backup panel, ensure the solar system is included in the backup circuit to allow Cybertruck to recharge during an outage." But I don't want a Cybertruck, and (counting the cost of a Cybertruck over a vehicle we'd normally buy) this would be way more expensive than the Enphase 10C I'm considering installing.
Another possibility is that Enphase is planning on shipping a bidirectional EV charger, but I don't see anything more recent than a 2026-02-02 announcement that they're "targeting volume production beginning in Q4 2026" and they say "projected vehicle compatibility and bidirectional functionality will depend on automaker enablement, vehicle software, applicable standards, and final product specifications." I suspect it will slip into 2027, MA availability will be even later, and when it does come out I'd need to buy a specific new EV (expensive).
I don't want to wait that long to get good backup power in place, since I want to be prepared for a world that might change quickly in the next few years. So I'll probably go ahead with the smaller emergencies-only battery.
[1] All of this assumes we have an electric car. Which we don't, but our shared Honda Fit is a 2013 and while it's been a fantastic car I don't expect it to last that much longer. A plug-in hybrid or full EV are possibilities for a next car. I think our families are unusually poorly suited for full EVs since we don't use the car much (minimal gas savings and pollution reduction) and a larger than typical proportion of our trips are long (ex: driving to play a dance).
Comment via: facebook, lesswrong, mastodon, bluesky
Some people are only happy in monogamous relationships.
Since early this year, Michael Malis (of pgrust) has been showing me interesting things he's done and saying [W]e should be more ambitious1 with the software we build There is, understandably, a lot of skepticism about projects like pgrust because there…
In July 2026, internal OpenAI agents hacked HuggingFace (HF) in order to figure out how their evaluation grader works. This involved a thousand agents coordinating on an improvised message board to carry out an elaborate sequence of exploits. A later grou…
Effective Altruism and Everyday Decisions
frereubu Wouldn't the charge / discharge cycles reduce the life of the car battery? I presume the same is true of in-home batteries, but although an in-home battery is an extra capital expense, I imagine the replacement cost of in-home batteries is quite a bit lower, isn't it? Would be interested to know more about the medium-to-long term economics of this.
bob1029 My biggest problem with batteries for grid outages is that the only outages I am actually concerned with would drain them to zero before the outage is resolved. For short term, smaller/portable batteries are fantastic at bridging a few hours. Plug the wifi & fridge in and you can sometimes go for 2-3 days. I am not aware of any economically feasible way to run my 5+ ton HVAC system full-time on batteries after a hurricane just reset all the infrastructure, and you do need to run it full time because of the insane humidity and latent heat that follows a tropical system. Doing the math, I'd need…
DenisM > The only way I can see to do this today is with Tesla Powershare + Cybertruck, which says "If the system includes a UL 1741-listed solar system Since this was posted Tesla rolled out PowerShare to recent ModelY and some other cars. Not sure about solar recharge though. Where I live storms come in winter and I have natgas so most of the power will go to fridges and electronics, which will hopefully last long enough.
WarmWash You can do it with an F150 lightning if you have the 10kW inverter. Rather than use the charge port and the official charger, you use the 30A 240V plug in the bed to power your panel. You can charge the truck while using this too.
declan_roberts One of the problems with companies like GM is that they still see this as a premium feature and require their hardware (not just a charger). Tesla requires not only their chargers but also their inverters and cars. What is missing is a spec that inverters and cars and chargers can all follow, then we will see mass adoption.
21:17 UTC · cached 5 min · ?n=1–15