I personally think it's a brilliant idea!
And, if you could solve the washability problem, then you could potentially make million$!
Anyway, a brilliant idea!
19,053 karma · joined September 16, 2014
Open Source, Open Hardware, Transparency & Free Speech enthusiast.
peter.d.sherman@gmail.com www.linkedin.com/in/peter-sherman-a6107a5 https://x.com/peter_d_sherman
"The true knowledge consists of knowing that one knows nothing..."
- Socrates
I personally think it's a brilliant idea!
And, if you could solve the washability problem, then you could potentially make million$!
Anyway, a brilliant idea!
This is interesting!
This is the first time in the "frontier LLM world" that I've seen a pricing split along a token boundary, in this case, less than or greater than 100k tokens...
In the past, first we saw ChatGPT, then we saw Claude, then we saw commercial frontier competitors, then we saw local open LLM's (Deepseek, etc.), then we saw "thinking" LLM's, i.e., use the LLM itself to generate a text plan to answer the user's prompt behind the scenes, then use that as a second, hidden prompt to actually answer the question, then we saw frontier pricing split across several different models (depending on how much behind the scenes "thinking" aka, "better hidden query plan text writing" that they did, and how many levels of this that they did) and these models may have had different context/token sizes but were not separately billed for prompt size, and now we're seeing pricing based on query size in tokens.
Now, I'm wondering how many levels or "pricing tiers" this could segment into...
In theory, an LLM provider could make 1k, 10k, 100k, 1M, 10M token context pricing tiers for the same model...
Would there be any practical benefit to doing so?
Well, that's a function of how much memory/compute are used by different sized prompts, how many prompt requests are incoming at any given point in time for a given commercial LLM provider, etc., etc.
I don't have all of those numbers.
In theory, an LLM provider could also charge less during nights or weekends -- times when electricity is cheaper, if those cost savings are passed on through the data center/compute supply chain, but again, I don't have all of those numbers...
Still it is highly interesting to watch product/service segmentation and differentiation in this highly competitive marketplace...
What will be the next innovation in terms of speed/efficiency/quality and/or price segmentation?
We don't know... but this LLM pricing split on a 100k token boundary is highly interesting for AI market watchers, and possibly for Economists as well, present and future!
The researchers used this system to produce a fully 3D-printed electric linear motor in a matter of hours using five materials."
This is highly interesting!
That is, being able to 3D print things (such as a motor) with electric and magnetic parts...
Circuit printing 3D printers have been around for quite awhile now, but printed magnetic parts of a 3D printed object? That I'm hearing for the first time here!
(It would be interesting (in the future) if one could have a 3D printer that could print with any metal, and if such metals at different micro-thin layers could somehow be combined with one another to make ad-hoc printed metal alloy layers... Well, we do have "metal sputtering" aka "vapor deposition" -- but that requires a vacuum tube, argon gas, and a plasma chamber... but maybe someone in the future will discover some way (electrically, I'm guessing -- an ionic wind of some sort?) to print ad-hoc thin layers of metal alloys at room temperature and in the presence of air by mixing pure elemental metals...)
Anyway, congratulations to the team at MIT for the considerable step forward in 3D printing!
The best dark-sky sites don't have Internet access. AstroHelm’s plate solving engine runs entirely on device. No internet connection required, and your camera images stay on your phone."
This looks like a great tool for the following application:
o Off-grid, deep wilderness, high mountain, broad desert or ocean navigation (like the way ancient seafarers would navigate via the position of the stars in the nighttime sky, relative to them)
(Also, if you're a Star Trek fan and/or futurist, you'd expect that computers on future spacecraft would or could run some version of this software, or at least some form of Plate Solving / Astrometric Solving algorithm...)
Anyway, the app looks great!
Related:
General wikipedia information page about Astrometric Solving (does not include algorithms!):
https://en.wikipedia.org/wiki/Astrometric_solving
Paper: New Algorithms for Automated Astrometry, by Chris Harvey (2004) (An excellent paper about all that is necessary for such an algorithm!):
https://cs.nyu.edu/home/people/in_memoriam/roweis/papers/Har...
The Building Blocks of Interpretability
https://distill.pub/2018/building-blocks/
Activation Atlas
https://distill.pub/2019/activation-atlas/
Visualizing and Understanding Convolutional Networks
I'm definitely a fan of anything Open Hardware, FPGA's, TPU's, Verilog and anything that can increase education / learning in the AI space...
This GitHub repo scores points in all of those areas!
Store the domains as sorted 40-bit hashes in flash and binary-search them.
140,000+ domains fit in ~0.7 MB of flash..."
Brilliant! (Although, what about potential hash collisions with domain names that shouldn't be blocked? Those are probably few and far between, so the utility of the block overall far outweighs any few false positives that might arise -- so I reiterate my claim of "brilliant"!)
Also (if you're a compression nerd like I am), it may be possible to analyze this table of 40-bit hashes further to find the longest binary substring that repeats the most across all of them (or binary substrings that just repeat a whole lot), and design a Huffman-like (or other tree or tree-like) data structure around those, and possibly compress the data further...
Of course, then the simplicity and elegance of the Binary Search might be lost, but it might be an interesting optimization exercise to know just how far that data could be compressed... "Hey Claude, try to optimize the compression of this data using alternative compression data structures, and report back to me!" (Or something like that! You know, get the AI coding agents to try different things -- or attempt to code it on your own (always better for understanding!), etc., etc.!)
Anyway, brilliant!
According to the British historian of science M. J. T. Lewis, the Diolkos represented a railway, in the basic sense of a prepared track which so guides the vehicles running on it that they cannot leave the track.[3] Measuring between 6 and 8.5 km (3.7 and 5.3 mi),[43] and being open to all on payment, it constituted even a public railway, a concept which according to Lewis did not reoccur until c. 1800.[5] Also, its average gauge of around 160 cm (5 ft 3 in)[32] is similar to modern standards.
However, a close examination of the excavated tracks may give a different picture. While there is agreement that the grooves in the eastern part were cut deliberately into the stone slabs to guide cart wheels,[44] those in the western section are interpreted by some authors as a result of wear or do not appear at all.[45] On the other hand, the marked cambers of this road section may point at deliberate tracks as well.[32] Generally, varying forms of the grooves can also be explained by the long period of operation of the Diolkos, during which modifications and repairs must have significantly changed the appearance of the trackway.[46]"
So, was the Diolkos the ancient predecessor to the modern railway?
Perhaps not, but certainly the idea of cutting standard-spaced continuous parallel grooves in rock with the approximate spacing of modern day railroad tracks (probably to keep load-bearing wooden wheels of some sort, in place) seems to be one of the forerunner early historical ideas which (when coupled with subsequent ideas, i.e., metal tracks, steam-driven machines to do the pulling, etc., etc.) lead up to modern railroads, although, admittedly there's around a two millennia (give or take!) gap between this idea and the next railroad-related one! :-)
Big whorls have little whorls
Which feed on their velocity,
And little whorls have lesser whorls
And so on to viscosity.
[...]
Si and Fang ran the brine shrimp experiment repeatedly. They found that when a solo shrimp swam at an angle of less than 45 degrees to the direction of maximum stretching, the energy flowed from smaller scales to larger ones, feeding the flow itself, as they’d predicted for a 2D system. But if the shrimp swam at an angle of more than 45 degrees to the direction of maximum stretching, the energy went in reverse, pulling from the larger scale of the flow to energize smaller perturbations.
[...]
It was interesting because they were saying that with small, specific interventions in the flow we can … show a transfer of energy between scales,” De Serio said. They demonstrated that with minimal manipulation, the energy flux in a 2D system can be steered to act more like (but not exactly like) the flux in a 3D one, which suggests that physicists might be able to apply insights from one dimensional domain to the other."
Absolutely fascinating! Great article!
Potentially related:
Replacement of Petroleum Based Products With Plant-Based Materials:
writing a simulator that runs millions of moves per second on graphics cards.
“At the scale that we are in academia, we don’t really have access to an entire field of GPUs,” Farina said.
The algorithm also learned far faster—it played about 34 times fewer games than DeepNash, and still ended up much stronger."
Isn't that a case of David vs. Goliath!
Also, I like the idea of the simulator/training software/aka "oracle" (source of truth for training data) running adjacently on the same graphics card(s) or AI accelerator(s) that the training is taking place on... That approach moves way more data way faster than say, having the training have to interact with a game running on a CPU, and having to screen scrape and interpret screen data (and wait!) with every new move...
Anyway, great article!
Message-Queue-As-A-Service! (MQaaS!)
Related:
https://en.wikipedia.org/wiki/Message_queue
https://en.wikipedia.org/wiki/Message_loop_in_Microsoft_Wind...
>"Through the 2000s, the designs grew larger and more tightly integrated. ADRES coupled a VLIW processor with a reconfigurable matrix; TRIPS explored a polymorphous array capable of exploiting instruction-, thread-, and data-level parallelism; and PACT XPP offered a self-reconfiguring commercial fabric.[7][8][9] From the 2010s onward, research emphasized data movement and system integration: HyCUBE introduced a single-cycle multi-hop interconnect, Plasticine organized computation around pattern compute and pattern memory units, and the stream-dataflow (Softbrain) model decoupled memory streaming from the spatial datapath.[10][11][12] The same period saw an open-source ecosystem take shape, including ultra-low-power edge arrays such as OpenEdgeCGRA, which is integrated with the X-HEEP RISC-V microcontroller in the HEEPsilon platform to give a complete flow from C code to hardware.[13][14]
A handful of CGRAs have reached commercial silicon. Samsung’s Reconfigurable Processor (SRP), derived from ADRES, has appeared in systems-on-chip and in ultra-low-power biomedical variants;[15][16] SambaNova commercialized its reconfigurable dataflow unit (RDU), a large dataflow accelerator built from pattern compute and pattern memory units and aimed at machine-learning training and inference.[17]"
Never knew about their existence... until reading this article...
https://github.com/Rust-GCC/gccrs
I have never heard that term used before, prior to reading this thread...
All I know is that "low-effort comments" -- is going into my 2026 lexicon!
(Admittedly this comment, my comment, the words I have written, above -- is admittedly somewhat of a "low-effort comment" in and of itself(!) A low-effort comment about other low-effort comments (a "meta" low-effort comment, if you will!), but I digress... I guess at times "we all fall short of the glory of the highest of high effort comments"... :-) )
Anyway, "low-effort comment" is a great piece of terminology, isn't it? :-)
It's going into my 2026 lexicon...
Impressive numbers!
Whoever would have thought (in prior years) that in 2026 AI Agents (not people or corporations, at least not directly) seem to be (or seem to be rapidly becoming) the biggest consumers of cloud computing resources...
Anyway, a very interesting paper and environment!
What an interesting problem to have!
So, let's see, we have a hosting site, much like GitHub is for code, or DropBox is for files, or Pastebin is for text, or YouTube is for Videos, except that this website is tailored to hosting academic papers as opposed to other specific file types...
Now in 2026 we have a problem, and that is that AI (or people using AI) are apparently writing crappy AI-written academic papers, and uploading them to ArXiv (and other academic paper hosting sites) in large amounts...
What to do, what to do?
Hmmm, this sounds very similar to the problem of many crappy AI-generated videos being uploaded to YouTube...
Or, the "AI-generated-file-of-specific-type-was-uploaded-where-the-file-contents-of-uploaded-files-of-that-type-are-NOT-supposed-to-be-AI-generated-on-this-specific-file-type-hosting-website" problem...
What to do, what to do?
Well, heck, I don't know!
What shall we all do?!?!?!?!?!?
Observation (if I might!): The "people using AI to upload crap to the Internet" problem is a subset of the (older, larger!) "people just plain uploading crap to the Internet" pre-AI problem... :-)
>"Lots of suggestions in the comments but no magic bullets."
This is Hacker News after all; what were you expecting? Silence? :-)
Space filling curves with three dimensions could potentially provide us with some clues:
https://en.wikipedia.org/wiki/Z-order_curve
"In mathematical analysis and computer science, functions which are Z-order, Lebesgue curve, Morton space-filling curve,[1] Morton order or Morton code map multidimensional data to one dimension while preserving locality of the data points (two points close together in multidimensions with high probability lie also close together in Morton order)."
Also, to answer the question, potentially Octrees might be interesting to research:
https://mathworld.wolfram.com/Octree.html
"An octree is a rooted tree that represents a three-dimensional region by recursively subdividing it into eight congruent, axis-aligned cuboids, usually cubes (Meagher 1982, Samet 1990). Each internal node represents a spatial cell and has eight children, one for each octant of the cell. A tree leaf represents a cell that is not subdivided further."
But, since "a picture is worth one thousand words":
https://www.google.com/search?q=octree&udm=2
Another thing to consider is that according to the Knuth Transform (aka Left-Child Right-Sibling (LCRS) representation), apparently any k-ary tree (Octree = 8-ary tree) can apparently be converted into a Binary Tree without losing information:
https://en.wikipedia.org/wiki/Left-child_right-sibling_binar...
So, is there a 1D (or 2D) representation for 3D reality?
Well, I personally don't know!
What I do know is that a certain popular rock group sang the following lyric in one of their popular rock songs, back in 1999:
"Space may be the final frontier -- but it's made in a Hollywood basement..."
-The Red Hot Chili Peppers, "Californication"
So, you decide! :-)
Interestingly, any derivative of a SIREN is itself a SIREN, as the derivative of the sine is a cosine, i.e., a phase-shifted sine (see supplemental).
Therefore, the derivatives of a SIREN inherit the properties of SIRENs, enabling us to supervise any derivative of SIREN with “complicated” signals. In our experiments, we demonstrate that when a SIREN is supervised using a constraint Cm involving the derivatives of φ, the function φ remains well behaved, which is crucial in solving many problems, including boundary value problems (BVPs).
We will show that SIRENs can be initialized with some control over the distribution of activations, allowing us to create deep architectures.
Furthermore, SIRENs converge significantly faster than baseline architectures, fitting, for instance, a single image in a few hundred iterations, taking a few seconds on a modern GPU, while featuring higher image fidelity..."
SIRENs look interesting... the idea of using Sine as an Activation Function seems like a brilliant one!
That means that it's probably using some library in JavaScript or WASM to implement that.
If that's the case, then I'm thinking, you know, that functionality, that is, font rendering on demand could, in theory, be placed into a CloudFlare Worker (or other Serverless service, i.e. AWS Lambda, Azure Functions, Google Cloud Functions, OCI Funcions, IBM Cloud Code Engine, Alibaba Function Compute, etc., etc.)...
That could be useful for a variety of applications... including (but not limited to!) word processors, typesetting software, homemade browsers, online games, heck, really anything where custom fonts are needed...
There could be two distinct benefits with this, let's call it a "FaaS" (Fonts as a Service!) idea:
1) No need to bundle/ship/include a font rendering engine with software that needs one.
2) Extra layer of security -- Local font rendering software is typically Turing-Complete, and as such, the local font renderer could potentially run viruses and other unwanted programs. I.e., local font rendering is a security risk. By hermetically sealing that code into serverless function, offsite of the local machine and having that serverless function send only fully rendered font data one way, the attack surface of the font rendering engine is effectively removed from any local PC.
This being said, the disadvantage is that it couldn't be used if the Internet or server or cloud provider is down for any reason.
Graphical font data, once created and downloaded, could however be cached by clients for offline situations / offline usage / offline work.
So, there are some interesting ideas here...
A full pre-training run on the Nemotron DQA dataset takes about an hour on a single RTX 4090 GPU.
Most importantly, since states and actions are disaggregated, their embeddings can be cached independently. In settings where the state evolves continuously (e.g., Super Mario) while the action set remains fixed, we only need to recompute the state embedding at each step and can reuse the cached action embeddings. This substantially reduces inference cost, with the efficiency gains becoming increasingly significant as the number of candidate actions and context length grows."
There does definitely seem to be something there with respect to Contrastive Language Models.
They are probably worth studying for people (like myself!) who want to wring the absolute last cycle of local AI training and inferencing performance out of consumer-grade (i.e., not datacenter scale nor cost) hardware...
Jensen gets it!
Related:
https://en.wikipedia.org/wiki/Coupling_(computer_programming...
https://en.wikipedia.org/wiki/Abstraction_layer
https://www.joelonsoftware.com/2002/11/11/the-law-of-leaky-a...
GPUs, which can be described as manycore vector processors"
That's arguably a much better description of a GPU than merely "Graphics Processing Unit"! :-)