@mgorny@social.treehouse.systems2026-05-24 05:14:07
Okay, so apparently there's been some "scuffle" between a cyclist and an old lady. The police's looking for the cyclist now, and shared a camera footage looking for help in finding them. Except that the footage is such a low resolution it's practically useless.
So helpful people from the internets used "#AI" to enhance it. So now we're looking at an angry mob looking for a person whose face was generated by an #LLM. Or well, multiple independently generated different faces apparently, but would that stop a mob from lynching a random person?
This fucking crap needs to be outlawed immediately. And whoever's selling it should end up behind bars.
#NoAI #NoLLM
@frankel@mastodon.top2026-06-14 17:06:55
'Never use double for money' is dogma, not engineering.
Guest author Stefano Fago breaks down when double, BigDecimal, or fixed-point is the right call, and the production traps that quietly undo each one.
https://blog.frankel.ch/bigdecimal-vs-double/
@arXiv_csGR_bot@mastoxiv.page2026-07-21 07:40:37
Packet-Loss Robust 3D Gaussian Compression via Atomic Packaging and GNN-based Error Concealment
Yuxuan Tao, Xuerui Ma, Hao Zhang, Chunhua Peng
https://arxiv.org/abs/2607.17916 https://arxiv.org/pdf/2607.17916 https://arxiv.org/html/2607.17916
arXiv:2607.17916v1 Announce Type: new
Abstract: 3D Gaussian Splatting (3DGS) and recent compression schemes such as HAC enable high-fidelity real-time neural rendering, but their bitstreams are fragile under packet loss during network streaming. Existing compression methods often separate correlated anchor attributes into independent streams, so losing one packet can create attribute-inconsistent broken anchors and severe rendering artifacts. We propose a packet-loss robust 3DGS transmission and error concealment framework. On the encoder side, anchor-level atomic packaging jointly encapsulates all attributes of each anchor, converting corrupted-attribute failures into clean missing-anchor erasures. Stratified random grouping further disperses packet losses across the spatial domain to avoid large contiguous voids. On the decoder side, we formulate recovery as prior-aware attribute inpainting. A Context-Aware Residual Interpolation (CARI) branch uses hash-grid prior predictions and neighboring residuals to build a robust baseline, while a lightweight two-layer graph neural network with cross-attention over hash-grid priors refines high-frequency attribute residuals. Attribute-wise confidence control falls back to interpolation when learned predictions are unreliable. Experiments under 20 percent random packet loss on BungeeNeRF, Mip-NeRF 360, and Tanks and Temples show that the proposed method substantially improves over no-concealment transmission and limits average PSNR degradation to about 3 dB relative to the lossless HAC reference.
toXiv_bot_toot
@arXiv_csPF_bot@mastoxiv.page2026-06-11 07:42:10
The Brain That Goes Quiet: Serving a Large Model's Knowledge at 131 Tokens per Second on an 8 GB Laptop by Removing the Large Model from the Runtime Path
Myeong Jun Jo
https://arxiv.org/abs/2606.12154 https://arxiv.org/pdf/2606.12154 https://arxiv.org/html/2606.12154
arXiv:2606.12154v1 Announce Type: new
Abstract: In earlier work I showed that a 35B-class Mixture-of-Experts model can be loaded and executed on a consumer laptop with 8 GB of GPU memory. That result solved a placement problem and immediately exposed a different one: even correctly placed, the large model needed roughly four seconds to answer, because it was still being invoked at every query. This paper documents what happened when I stopped invoking it. During an offline phase, the large model reads source documents and writes verified answer entries into a structured knowledge store; at runtime, only a lightweight router, a deterministic renderer, and a 1B-class model are active. On the same 8 GB laptop, end-to-end response time fell from approximately 4,465 ms to 518 ms, effective end-to-end throughput rose from 15.7 to 131 tokens per second, and the small model's streaming decode rate held at 226-237 tokens per second with a time-to-first-token of 29-62 ms. The bottleneck is structural: three different large models (Qwen, Gemma, and GLM class) all showed the same multi-second runtime cost, and all three produced usable knowledge stores offline. On a 563-entry store built from seventeen real documents, keyword routing collapsed to 1.5% top-1 accuracy while BM25-based routing reached 92.8% (99.4% top-3), and a confidence gate raised effective top-1 to 98.0% by escalating 12.3% of queries. Exact-match fidelity of the small model ranged from 9/9 to 0/9 across envelope formats carrying identical content. A 16-case verification gate blocked all ten corrupted entries while admitting all six supported ones.
toXiv_bot_toot
@fanf@mendeddrum.org2026-06-28 08:42:04
from my link log —
What data access pattern is as slow as possible?
https://blog.weineng.me/posts/slowest_add/
saved 2026-06-27 https://
@tiotasram@kolektiva.social2026-08-11 16:22:46
Random thought that goes interesting places:
There are a lot of cheap ways to spend money to drastically improve the quality of life for lots of people (UBI, mosquito nets, digging wells, etc.). Those people almost invariably become way more economically productive in aggregate when this happens, far above the costs involved, if you measure from the perspective of the affected group. So for billionaires looking for outsized returns on investments, why not simply spend on some of these programs, then invest in a broad index of socks that are going to go up as a result?
The answer could be "billionaires by their nature are too evil to think of this," which, could be true....
But I think there's a deeper answer, which is that there *is no* bundle of stocks that goes up as humans flourish. This is directly contrary to neoliberal economic doctrine/propaganda, but it also has an explanation that's obvious from the neoliberal doctrine itself: the stock market is a gambling arena for bets on corporate profits. Corporate profits are money taken in in excess of costs. But the value of a corporation is not produced abstractly by the mere existence of the organization; it's produced by the labor of humans working for the organization, with raw materials that the organization purchases. So to make a profit, an organization has to do some combination of paying its workers less than the value the create, and/or paying less for it's raw materials than they are really worth. The more it does these two things, the more profit it can generate.
Notice that these two profit-generating activities directly and indirectly immiserate humans, *and* that reducing human misery by means other than reducing these activities makes them harder. People who have other means of income won't sell their labor for less than it's worth, nor will they sell their resources below their fair value. Only desperate people will agree to terms that corporations require to be profitable.
This is not a "Marxist analysis" by the way; it's the direct application of the basic principles taught in any neoliberal introductory macroeconomics class.
In any case, once we understand that the stock market is "gambling on how much corporations can exploit their employees/suppliers/customers", it's clear why you can't make money from stock investments in companies that profit from flourishing humans: profit is by construction made of human misery.
It also explains a lot of other things, like why private US employers who pay huge amounts of payroll towards health insurance for their employees would be *against* public health insurance that would offload those costs onto all taxpayers: the amount they're able to lower wages as a result of desperation for access to health care is more than worth the direct costs.
This is the central reason why capitalism (defined as: a system where money equals power) is inimical to human happiness: it allows profits built on suffering to be converted into the power to maintain the system of suffering and even extend and intensify it, creating a feedback loop of extractive domination.
Now I'm sure some people might read this and think: what about the good companies? The ones that pay their workers and suppliers a fair wage, and sell their products at fair prices? Even putting aside the fact that such a thing seems less believable than a unicorn these days, such a company by definition does not make a profit. The fair wage for its workers is the value they add to the raw materials it buys, and the fair price for those is the selling price of the finished product minus the value the company adds. The balance sheet will read exactly zero at the end of the day if compensation is actually fair. You can still run a company like this in theory, and even grow it, but of course under capitalism it will just get bought out by an unethical company that is profitable. Could you in some kind of utopian dream world run a tightly regulated series of markets where you have most of what people think capitalism is, without the bad parts? Maybe, but it definitely wouldn't involve actual capitalism, and it wouldn't have a stock market.
#anarchy #capitalism #econimics
@arXiv_mathST_bot@mastoxiv.page2026-08-13 08:00:17
Spectral phase transitions in Gaussian multi-index models
Florent Krzakala, Pierre Mergny, Vanessa Piccolo
https://arxiv.org/abs/2608.12183 https://arxiv.org/pdf/2608.12183 https://arxiv.org/html/2608.12183
arXiv:2608.12183v1 Announce Type: new
Abstract: Recovering a low-dimensional latent subspace from nonlinear observations of Gaussian covariates in high dimensions is a fundamental problem in feature learning. Here, we consider Gaussian multi-index models in which the covariates $\boldsymbol{x}_i \stackrel{\mathrm{i.i.d.}}{\sim} \mathcal{N}(0,\boldsymbol{I}_d)$ and the responses $\boldsymbol{y}_i$ depend on $\boldsymbol{x}_i$ only through its projection onto an unknown $r$-dimensional subspace. Earlier work based on approximate message passing (AMP) identified a sharp threshold for weak recovery [Troiani et al., 2025], raising the question of whether it can be attained, without side information, by a spectral method. We answer this affirmatively and develop a general random matrix theory for matrix-valued spectral estimators of the form \[\boldsymbol{D}_n=\frac{1}{n}\sum_{i=1}^n\boldsymbol{T}(\boldsymbol{y}_i)\otimes\boldsymbol{x}_i\boldsymbol{x}_i^\top,\] where $\boldsymbol{T}$ is an arbitrary bounded symmetric matrix-valued preprocessing map of fixed dimension. As $n,d \to \infty$ with $n/d\to\alpha$, we prove that the empirical spectral measure of $\boldsymbol{D}_n$ converges almost surely to a deterministic compactly supported distribution characterized by a matrix-valued self-consistent equation. We then establish a spectral phase transition for the largest eigenvalue: below threshold it sticks to the bulk edge, while above threshold an outlier emerges. We characterize the outlier location through a finite-dimensional deterministic equation and show that the associated spectral estimator achieves weak recovery of the latent subspace. Finally, we prove that the AMP-derived preprocessing of [Defilippis et al., 2025] is optimal among all bounded matrix-valued preprocessing maps of any fixed dimension. Its transition coincides with the AMP weak-recovery threshold, proving the general spectral conjecture of [Defilippis et al., 2025].
toXiv_bot_toot