Engineering ·
Radar: the fast inverse square root, explained by Joe Groff
An article by Joe Groff explaining the well-known fast inverse square root code, associated with John Carmack but not created by him, and how this mathematical function works.
Radar
Daily bookmarks: relevant news, facts and discoveries.
Articles 1216
Engineering ·
Baidu releases Unlimited OCR 2, open-source OCR model for long documents
Baidu announced Unlimited OCR 2, an open-source document parsing model that condenses text-visual attention into fixed-size working references, according to the company.
Engineering ·
Baidu releases Unlimited OCR 2, a 3.3B open source model for long documents
Baidu announced Unlimited OCR 2, an open source 3.3B model for OCR of long documents, listed under the MIT license in the original post.
Engineering ·
JAX documentation on high-performance matrix multiplication kernels for Blackwell with Pallas
A short post dated October 27, 2026 shares a JAX documentation page on writing high-performance matrix multiplication kernels for NVIDIA Blackwell GPUs using Pallas.
Engineering ·
Gemma 4 31B reads handwritten medical prescription in informal OCR test
A post from October 27, 2026 reports that the Gemma 4 31B model from Google read a handwritten medical prescription image better than earlier OCR models, in a test run by the author. It gives no link or benchmark, so the result is anecdotal.
Engineering ·
MiniCPM-OCR 1.0 and Jev: open OCR model and output verifier
Post presents MiniCPM-OCR 1.0, a 1.3B document reading model, and Jev, a 2B model that scores which OCR output regions may need human review. Benchmark figures are the author's.
Engineering ·
Radar: Ulrich Drepper's PDF on computer memory
Radar: a widely shared PDF by Ulrich Drepper on RAM, caches and efficient memory use, focusing on the memory hierarchy and performance in C and C++. Logged 27/10/2026.
Engineering ·
MiniCPM-OCR suite: open OCR models and an accuracy verifier
OpenBMB announces two open-source models for OCR: one for document parsing and another that checks the accuracy of text, formulas and tables extracted by different parsers.
Engineering ·
Unlimited-OCR-2 tested on multi-page PDFs on Mac and DGX Spark
An engineer ran an open source document parsing model with 3.3 billion parameters on a 6-page PDF across three machines and checked the output against the PDF's own text.
Engineering ·
AWS publishes open source agent runtime claiming 75,000 concurrent agents on 8 GB
According to the author, AWS published an open source runtime that saves agent state and suspends agents while they wait for tools, resuming them when an event occurs. The post claims 75,000 concurrent agents on 8 GB of RAM, but gives no link for verification.
Engineering ·
Hardening a rooted LG webOS TV against microphones and telemetry
A public GitHub project describes hardening a rooted LG webOS TV: it disables microphones and the remote's voice pipeline, blocks ad, telemetry and ACR endpoints, and hides ad apps.
Engineering ·
Ming 3 open TTS model ranks first among open TTS models on Artificial Analysis
Briefing dated 2026-10-24: according to the source author, the open Ming 3 TTS model leads the Artificial Analysis ranking of open TTS models, narrowly ahead of Fish Audio Drama 3.
Engineering ·
FP/BF16 issue-rate unlock on the RTX 3090 with gpu-unlock
A post reports that the gpu-unlock project raised measured FP/BF16 throughput in PyTorch on an RTX 3090 from 75T to 135T, with a 360 W limit. The repository describes unlocking the issue rate of these instructions up to the level of professional cards.
Engineering ·
Dessa: full-duplex speech-to-speech model for consumer GPUs, no primary source
A post announces Dessa, a full-duplex speech-to-speech model that, according to its author, was trained only on speech and runs on consumer GPUs. No primary source is linked and the claims have not been verified.
Engineering ·
Open TTS Leaderboard adds 31M and 35B models and the UTMOS metric
The open text-to-speech leaderboard now includes a 31 million parameter model and a 35 billion parameter model, with similar English WER, plus a new naturalness metric called UTMOS.
Engineering ·
Alleged Jack Dorsey speech model trained on a GB200 NVL72 (unverified)
A social media post claims Jack Dorsey released a speech model trained on a single NVIDIA GB200 NVL72 system. No model name or primary source is given, and the claims have not been verified.
Engineering ·
Author argues short-lived bearer tokens require proof of possession
An author argues that 15-minute bearer OIDC tokens do not prevent intrusions and that binding tokens to the sender, such as via TLS, would be the effective fix.
Engineering ·
Post says a Meta paper proposes tokenizer-free byte-level models
A post claims Meta's paper 'Breaking the Token Ceiling' trains models on raw bytes with a 256-value vocabulary and applies distillation. According to the post, 1B byte models match or beat two cited 1B models. No link to the original work was provided.
Engineering ·
Private CA roots used for TLS inspection may issue certificates without transparency log entries
An author claims that TLS inspection roots installed on corporate laptops can issue certificates for any site without appearing in certificate transparency logs, since browsers require that only of publicly trusted CAs. The source gives no links.
Engineering ·
Phonon 31M: open TTS model in Swift with CoreML on Apple's Neural Engine
A Swift port of the open Phonon 31M text-to-speech model, running CoreML on Apple's Neural Engine, with 103 voices and a text normalizer ported from Gradium. The author reports about 20 ms to first audio and about 20x real time on an M5 Pro.