1471
What's the largest software project AI can complete on its own?
MirrorCode is Epoch AI's benchmark for long-horizon coding: AI can reimplement entire programs end-to-end, with no access to the original source code.
Top posts
Every blog post that took off on Hacker News, ranked by the votes of its best submission. 1,523 posts.
1472
Claude Opus 5.5 Can Draw Permit-Level Architectural Drawings
I tested thirteen models and found that their drafting is approaching professional levels.
1473
A polynomial autoencoder beats PCA on transformer embeddings
A closed-form autoencoder: PCA encoder + quadratic Ridge decoder. On FiQA it gives 4× compression of mxbai-embed-large-v1 at -0.88 p.p. NDCG@10, +1.34 p.p. over PCA. No SGD, no neural networks.
1474
1475
We decreased our LLM costs with Opus
We switched to a frontier model and our costs went down. Here's the architecture that made it possible.
1476
Gaussian Splats part 3: There's always a bigger scene
Continuing our Gaussian splatting development series, Director of Technology Sam Birley breaks down how we bypassed GPU Scene instance caps, implemented PlayCanvas Streamed SOG LODs, and managed GPU memory budgets for production-ready photogrammetry… bringing us to a point where the renderer has fin
1477
Why are coding agents so dumb?
Models are getting better and better, while the agents that drive them languish.
1478
AI made coding more enjoyable
To me, one of the most annoying parts of software engineering is writing code that doesn’t require thinking. It’s just a typing exercise, and that’s boring.
1479
LLM Networking with MikroTik
I’ve been refraining from saying ‘vibe networking’ or ‘vibkrotik’ or some other abomination, but it is true - I have been using LLMs to setup a few networks these last few months, and things have generally gone over swimmingly.
1480
Building a Multi-Agent System from Scratch: Architecture Patterns
Orchestrator-worker, blackboard, and pipeline patterns for coordinating multiple AI agents — with the failure modes that only show up once you go past a single-agent demo.
1481
Why are good developers so wary of agentic development?
Good developers’ caution about agentic development comes from knowing that built isn’t the same as working in production.
1482
Yes, Claude can do nine loops
Claude computes a nine-loop amplitude in N=4 super-Yang-Mills
1483
Shapelearn Qwen 3.8 27B (13.1 GB VRAM)
ByteShape's full ShapeLearn release of Qwen 3.8 27B: five GGUF models on the quality-speed frontier across six GPUs. GPU-5 is the default wherever it fits, at 99.63% of BF16. GPU-4 stays very competitive in a much smaller file.
1484
Making a vintage LLM from scratch
In this blog post, I will share the adventures I had creating my own LLM, from (almost) scratch, trained only on old texts. I made my own base-training and fine-tuning scripts, data processing pipelines and custom datasets. ("almost from scratch" means I did use existing programming languages and libraries, I didn't…
1485
Red queen hypothesis – A new way forward for self-improving AI
Cambridge Department of Computer Science and Technology2026104 points28 commentsAI & Machine Learning details
At a time when there's widespread interest in self-improving AI, researchers here have tackled one of the central challenges affecting its development by using an approach adapted from evolutionary biology.
1486
TorchLean: Formalizing Neural Networks in Lean
Neural networks are increasingly deployed through software stacks whose behavior is defined by tensor layouts, preprocessing choices, compiler boundaries, floating-point arithmetic, GPU kernels, and external model exporters. TorchLean is a Lean 4 codebase for neural network specification, execution, and verification…
1487
AI changes the economics of software rewrites
AI changes rewrite economics because codebases with clear, common patterns get more leverage than proprietary or inconsistent systems.
1488
What Happens If OpenAI Dies?
If you liked this piece, you should subscribe to my premium newsletter. It's $70 a year, $18 a quarter, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 10,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic
1489
AI for scientific discovery: How researchers use Claude
We analyzed 207,000 research sessions to learn how Claude is used in science, why some tasks get more use than others, and how AI could speed up discovery.
1490
An Introduction to YOLO26
What YOLO26 is: NMS-free, edge-optimized YOLO with 43% faster CPU inference, benchmarks by size, licensing, and how it compares to RF-DETR.
1491
TADA: Speech generation through text-acoustic synchronization
TADA (Text-Acoustic Dual Alignment) is Hume AI's open-source speech-language model that synchronizes text and audio one-to-one.
1492
Lean Eval for Alignment on Faithfulness
We turn intent into mathematics, verify every step formally, and make sure the machine does what it is supposed to.
1493
After AI takes everything
Prompted by letters from three engineers, the author asks what remains for humans as AI takes over execution, naming judgment, taste, and derivation as irreplaceable, and warns against the erosion of subjecthood.
1494
Nvidia DGX Spark as a daily driver
I bought an NVIDIA DGX Spark! It's actually quite a competent general purpose computer while being super tiny and power efficient. It basically just has an ARM CPU, a decent NVIDIA GPU (with the same number of CUDA cores as the RTX 5070), 128 GB of unified memory, and a 4 TB SSD.
1495
OpenAI's response to the Axios developer tool compromise
OpenAI responds to the Axios supply chain attack by rotating macOS code signing certificates, updating apps, and confirming no user data was compromised.
1496
LLMs may have helped my RSI
Just a place to store my random thoughts and contributions to the world.
1497
How much of HN is AI?
TL;DR: 62% as of September 2026. This is up from ~40% in February.
1498
The state of modern AI text to speech systems for screen reader users
If you're not a screen reader user yourself, you might be surprised to learn that the text to speech technology used by most blind people hasn't changed in the last 30 years. While text to speech has taken the sighted world by storm, in everything from personal assistants to GPS to telephone systems, the voices used…
1499
GPT‑Rosalind for life sciences research
OpenAI introduces GPT-Rosalind, a frontier reasoning model built to accelerate drug discovery, genomics analysis, protein reasoning, and scientific research workflows.
1500
Crypto grifters are recruiting open-source AI developers
Two recently-hyped developments in AI engineering have been Geoff Huntley’s “Ralph Wiggum loop” and Steve Yegge’s “Gas Town”. Huntley and Yegge are both…