What people in AI, tech and science are talking about today.
19 October 2026 · Top 10
Ranked by how fast each item is picking up attention on X, Hacker News, GitHub and Hugging Face, with a short note on why it matters.
-
Astronomers report the first microblazar
Anton Petrov walks through the first object described as a microblazar and what it could mean for astronomy. It's one of the most-watched science explainers of the day.
Source: YouTube · Anton PetrovLink · Post on X · LinkedIn ·▲ +10,890 views/h87,000 views
Science -
Hybrid Reasoning Agents think in latent space and act through text
Researchers from Stanford, UIUC and Together AI trained Qwen3-based agents that switch between hidden 'latent' thinking and normal text for tool calls. They report about 10% better scores across 11 benchmarks, with 35–55% fewer tokens and roughly 2x faster inference, while the latent steps can still be read back as text.
Source: Hacker News · discussionLink · Post on X · LinkedIn ·▲▲ +42 pts/h316 points · 163 comments
AI & Agents -
Theo: most AI features are just prompts, so write better ones
Theo argues that many of the AI features engineers ship are really prompts, and that prompting well makes them simpler to build. A practical watch if you build with LLMs.
Source: YouTube · Theo - t3․ggLink · Post on X · LinkedIn ·▲ +8,015 views/h49,770 views · 1,139 ratings
Dev Tools -
DeepSeek launches Search and Fetch APIs at $0.75 per 1,000 requests
DeepSeek now sells web search (on its own in-house engine) and page-to-Markdown fetch as raw APIs for agents and RAG apps. Calls from DeepSeek Harness, Claude Code or Codex are free until billing starts on 1 November 2026, which gives agent builders a cheap alternative to other search APIs.
Source: Hacker News · discussionLink · Post on X · LinkedIn ·▲ +30 pts/h566 points · 266 comments
LLMs -
Why you should run your own LLM proxy
Jared Siegel argues that vendor usage dashboards don't tell engineers much, and that a self-hosted proxy gives you per-key logs, one endpoint for many providers, virtual API keys and quick model switching. The HN thread trades proxy setups.
Source: Hacker News · discussionLink · Post on X · LinkedIn ·▲ +15 pts/h100 points · 75 comments
LLMs -
Golden age or dark age for maths? A guest post on Terence Tao's blog
Mathematician Francesco Maggi weighs how LLM-generated proofs could either power a golden age or leave knowledge concentrated and hard for humans to understand. He calls for publicly governed research LLMs, disclosure standards and new ways to train and assess young mathematicians.
Source: Hacker News · discussionLink · Post on X · LinkedIn ·▲ +14 pts/h137 points · 164 comments
Science -
huashu-motion: coding agents that make launch videos
This open-source project lets a coding agent produce product launch films, compose the music, and take edits when you point at a frame. It picked up 543 GitHub stars in a day.
▲▲ +31 stars/h545 stars in 1d
AI & Agents -
Numberphile breaks down a new maths meme
Numberphile's latest video unpacks a maths meme doing the rounds online. With over 200,000 views, it's one of the day's most-watched maths explainers.
Source: YouTube · NumberphileLink · Post on X · LinkedIn ·▲ +10,585 views/h201,000 views
Science -
thdxr says a self-built DeepSeek inference rack beats every vendor on price/performance
Dax (@thdxr) says a first test run on a new GPU rack with their own DeepSeek inference setup gave better price/performance than any vendor they've tried, with more optimisation still to come. More teams are weighing self-hosting open models over paying API providers.
Source: X · @thdxrLink · Post on X · LinkedIn ·▲ +284 eng/h3,634 likes · 45 reposts
LLMs -
GitHub Security Lab's open-source agent finds 24 Android vulnerabilities
GitHub's security team used an open-source agent to find 24 Android bugs, including a location leak in a navigation app with 10M downloads, and shared how to run the same taskflows on your own project. It's a concrete case of agents doing real security research.
Source: X · @githubLink · Post on X · LinkedIn ·– +21 eng/h140 likes · 17 reposts
AI & Agents
-
This is so cool.
From the source: This is so cool. One unified view of all your agent sessions. Across personal & coding agents, apps, API keys. Vercel AI Gateway helps you run your business better. You own the model choice, your…
Source: X · @rauchgLink · Post on X · LinkedIn ·– +48 eng/h364 likes · 17 reposts
AI & Agents -
Show HN: WhatsApp MCP running free on Cloudflare
From the source: Show HN: WhatsApp MCP running free on Cloudflare
Source: Hacker News · discussionLink · Post on X · LinkedIn ·– +12 pts/h79 points · 42 comments
AI & Agents -
Mechanistic Interpretability... solved?
New video from David Shapiro, 43,000 views so far.
Source: YouTube · David ShapiroLink · Post on X · LinkedIn ·– +2,390 views/h43,000 views
AI & Agents -
How Jobs Actually Disappear
New video from David Shapiro, 29,000 views so far.
Source: YouTube · David ShapiroLink · Post on X · LinkedIn ·– +1,209 views/h29,000 views
AI & Agents -
See every agent session in AI Gateway logs.
From the source: See every agent session in AI Gateway logs. Sessions group requests by conversation, with cost, cache hit rate, and the agent that ran it. Session affinity keeps each one on the provider holding its …
Source: X · @vercel_devLink · Post on X · LinkedIn ·– +11 eng/h87 likes · 9 reposts
AI & Agents -
aicodeking/leanrea: One 8 KB skill that replaces REA (Reverse Engineer Anything): plain CLI tools, no MCP. Benchmarked: 38% faster
From the source: One 8 KB skill that replaces REA (Reverse Engineer Anything): plain CLI tools, no MCP. Benchmarked: 38% faster, 35% cheaper.
▲ +13 stars/h689 stars in 3d
AI & Agents -
Anthropic Scion is INSANE
New video from David Shapiro, 115,000 views so far.
Source: YouTube · David ShapiroLink · Post on X · LinkedIn ·– +2,396 views/h115,000 views
AI & Agents -
A real-time WebGPU ocean in one 61 KB HTML file, built with Claude Code
A from-scratch WebGPU ocean renderer with FFT waves, whitecaps and a physically based sunset sky, all in a single 61 KB HTML file. The author says it was built with Claude Code on Opus 5.5, and it gained 538 stars in two days.
▲ +13 stars/h697 stars in 3d
AI & Agents -
HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
From the source: Agents built on vision language models (VLMs) now compete with learned policies on robot manipulation leaderboards. We ask how far a coding agent can go when it controls a robot instead of a computer,…
▲▲ +0.8 votes/h66 upvotes
AI & Agents -
Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
From the source: Reinforcement learning with verifiable rewards has become a scalable post-training paradigm for large language models. However, training agents over long-horizon interactions presents two challenges: …
▲ +0.3 votes/h20 upvotes
AI & Agents -
git-ai: version control rebuilt for coding agents
git-ai is a Go project billed as “Git designed for AI agents”, linked to Anthropic's write-up on redesigning source control. It passed 1,000 GitHub stars in two days.
▲▲ +26 stars/h2,320 stars in 4d
AI & Agents -
My AI Dark Factory Dream Is Dead (Here's What Actually Works)
New video from Cole Medin, 25,000 views so far.
Source: YouTube · Cole MedinLink · Post on X · LinkedIn ·– +521 views/h25,000 views
AI & Agents
-
Artificial Analysis benchmarks DeepSeek V4.1 Flash 1017 (Max)
Artificial Analysis has published its benchmark page for DeepSeek's new V4.1 Flash 1017 at maximum reasoning effort, a follow-up to the release that spread across Ollama and Vercel this weekend.
Source: Hacker News · discussionLink · Post on X · LinkedIn ·▲▲ +44 pts/h77 points · 21 comments
LLMs -
I built a programming language that LLMs can't use
From the source: I built a programming language that LLMs can't use
Source: Hacker News · discussionLink · Post on X · LinkedIn ·– +6.0 pts/h136 points · 107 comments
LLMs -
when claude goes down
From the source: when claude goes down
Source: X · @supabaseLink · Post on X · LinkedIn ·– +6.3 eng/h69 likes · 5 reposts
LLMs -
You could have invented transformers from memory
A Hacker News favourite essay that rebuilds the transformer starting from associative memory, making attention feel like a natural idea rather than a trick.
Source: Hacker News · discussionLink · Post on X · LinkedIn ·– +12 pts/h625 points · 142 comments
LLMs -
Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design
From the source: Looped language models reuse physical Transformer blocks across logical depth, but established architectures entangle recurrence placement, input injection, and recurrent state organization. We isolat…
– +0.0 votes/h3 upvotes
LLMs -
Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
From the source: Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to wr…
▲▲ newnew paper
LLMs -
Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation
From the source: A reasoning trace can accompany a correct answer without providing effective supervision for a particular student. We study how to construct traces that improve learning beyond answer-only supervision…
▲▲ newnew paper
LLMs -
Acceptance Is Not Speed: Length Inflation in Training-Free Lossy Speculative Decoding
From the source: Speculative decoding typically characterizes speedup by draft tokens accepted per verification round. Lossless verification preserves the target model's output distribution and thus expected completio…
▲▲ newnew paper
LLMs -
Learning to Think and Explore in 3D World
From the source: We study how vision-language models (VLMs) can learn to think and actively explore in 3D space. Instead of providing multi-view images, videos, or 3D representations to the model, we preserve the nati…
▲ +0.3 votes/h26 upvotes
LLMs -
StopWatch: Goal Grounding for Object-goal Navigation
From the source: Pretrained VLMs bring rich world knowledge and semantic priors to object-goal navigation, and fine-tuning a VLM into a navigation policy has become a common recipe. Despite this progress, success rate…
– +0.0 votes/h1 upvotes
LLMs -
Can Visual Generation Expand What Unified Multimodal Models Know?
From the source: Unified multimodal models (UMMs) receive image-generation supervision that exposes them to visual information not fully documented in text. Can learning to generate images expand what pretrained model…
▲ +0.1 votes/h6 upvotes
LLMs -
Caho1/open-claude-ui: 基于 REA 深度拆解并高保真还原 Claude 官方桌面端 UI。还原 30 个独立 React 组件与 8 项硬核交互物理算法,附带 3000+ 行逆向分析手册。
From the source: 基于 REA 深度拆解并高保真还原 Claude 官方桌面端 UI。还原 30 个独立 React 组件与 8 项硬核交互物理算法,附带 3000+ 行逆向分析手册。
– +11 stars/h1,218 stars in 5d
LLMs
-
Corridor Crew's VFX artists react to bad and great CGI, episode 247
The long-running Corridor Crew series breaks down what works and what doesn't in film effects. It's a useful benchmark for judging AI-generated video.
Source: YouTube · Corridor CrewLink · Post on X · LinkedIn ·▲▲ +30,345 views/h728,000 views
Gen-Media -
AI News: 6 New Models & A Nano Banana Killer
New video from Matt Wolfe, 149,000 views so far.
Source: YouTube · Matt WolfeLink · Post on X · LinkedIn ·– +3,105 views/h149,000 views
Gen-Media -
Free & Open Source Design with Ming Image in ComfyUI!
New video from Nerdy Rodent, 12,000 views so far.
Source: YouTube · Nerdy RodentLink · Post on X · LinkedIn ·– +500 views/h12,000 views
Gen-Media -
MOSS-Image Technical Report
From the source: Recent closed-source image generation systems can produce diverse visual content from simple compositions to information-dense graphics, yet open-weight text-to-image models still struggle with dense …
▲ +0.1 votes/h7 upvotes
Gen-Media -
Persistent Planar Memory for Video World Models
From the source: Video world models that generate and explore environments at scale need a persistent 3D memory of what they have produced. Plane primitives offer a natural candidate: a compact set of oriented planes …
– +0.0 votes/h3 upvotes
Gen-Media -
Rethinking the design space of Vanilla Pixel Diffusion Transformers
From the source: Pixel-space diffusion models have recently re-emerged as a compelling alternative to latent diffusion, avoiding the lossy latent bottleneck and separately trained autoencoder. JiT shows that a vanilla…
▲▲ newnew paper
Gen-Media -
Survey of Orion Disks with ALMA (SODA) IV: Millimeter variability of protoplanetary disks
From the source: Context. Short-term millimeter-wavelength variability in Class II young stellar objects (YSOs) has been considered minimal, often dismissed as very rare in ALMA-based studies on short timescales. Howe…
▲▲ newnew paper
Gen-Media -
Galaxy bias of radio AGN and star-forming galaxies from MIGHTEE and CMB lensing
From the source: Radio galaxies play a significant role in galaxy evolution through feedback mechanisms, making it important to characterize the environments in which they reside. We measure the cross-correlation betw…
▲▲ newnew paper
Gen-Media -
Possible detection of Baryon Acoustic Oscillations imprinted in the HI sources from the FASHI catalog
From the source: Motivated by curiosity as to whether the observed Universe at low redshifts is well described by the concordance cosmological model, we investigate the large scale angular clustering of extragalactic …
▲▲ newnew paper
Gen-Media -
Dusting the Phantom: Gaussian Allocation for Fading Dynamic Content
From the source: Dynamic Gaussian representations have made high-quality, real-time volumetric video increasingly practical, yet challenging motion can still break their reconstruction: small fast-moving objects may d…
– +0.0 votes/h2 upvotes
Gen-Media
-
Guillermo Rauch on country domains that boomed thanks to tech
Vercel's CEO points out how .co, .io, .md, .rs, .ai and .si became developer favourites, sending money to Colombia, the British Indian Ocean Territory, Moldova, Serbia, Anguilla and Slovenia.
Source: X · @rauchgLink · Post on X · LinkedIn ·▲▲ +671 eng/h6,063 likes · 119 reposts
Dev Tools -
Free Pascal compiler now supports dynamic packages
From the source: Free Pascal compiler now supports dynamic packages
Source: Hacker News · discussionLink · Post on X · LinkedIn ·– +14 pts/h139 points · 33 comments
Dev Tools -
Already running a model endpoint? What does it take to scale it on Kubernetes?
From the source: Already running a model endpoint? What does it take to scale it on Kubernetes? In five minutes, we walk through how NVIDIA Dynamo handles deployment with scalable workers and an OpenAI-compatible end…
Source: X · @NVIDIAAILink · Post on X · LinkedIn ·– +20 eng/h175 likes · 18 reposts
Dev Tools -
Build an AI That Turns Your Voice Into Images, Video and Audio (Python)
New video from Tech With Tim, 17,189 views so far.
Source: YouTube · Tech With TimLink · Post on X · LinkedIn ·– +1,064 views/h17,189 views · 210 ratings
Dev Tools -
EPIC Vibe Coded Launcher
New video from The PrimeTime, 125,375 views so far.
Source: YouTube · The PrimeTimeLink · Post on X · LinkedIn ·▲ +3,604 views/h125,375 views · 3,011 ratings
Dev Tools -
Why You Can’t Run That AI Model
New video from Tech With Tim, 10,016 views so far.
Source: YouTube · Tech With TimLink · Post on X · LinkedIn ·– +552 views/h10,016 views · 268 ratings
Dev Tools -
How NOT to spy on OpenAI
New video from The PrimeTime, 82,669 views so far.
Source: YouTube · The PrimeTimeLink · Post on X · LinkedIn ·– +2,058 views/h82,669 views · 1,432 ratings
Dev Tools -
Flushing the toilet kills our network!
New video from NetworkChuck, 38,951 views so far.
Source: YouTube · NetworkChuckLink · Post on X · LinkedIn ·– +959 views/h38,951 views · 989 ratings
Dev Tools -
You’re Using These Python Features Wrong
New video from ArjanCodes, 36,000 views so far.
Source: YouTube · ArjanCodesLink · Post on X · LinkedIn ·– +750 views/h36,000 views
Dev Tools -
xscreensaver-official/xscreensaver: There is no official XScreenSaver GitHub repository.
From the source: There is no official XScreenSaver GitHub repository.
▲ +14 stars/h872 stars in 3d
Dev Tools -
Mac or PC for AI Models
New video from Tech With Tim, 7,237 views so far.
Source: YouTube · Tech With TimLink · Post on X · LinkedIn ·– +172 views/h7,237 views · 217 ratings
Dev Tools -
Behavioral Compositionality of Sparse Autoencoder Representations
From the source: Sparse autoencoders (SAEs) represent model activations with sparse, linearly decoded features, but the downstream effects of those features need not compose additively. We introduce SAE Additivity Res…
▲▲ +1.0 votes/h107 upvotes
Dev Tools
-
Neutrino Flavors Might Explain Why Black Holes Go Missing
New video from Anton Petrov, 178,000 views so far.
Source: YouTube · Anton PetrovLink · Post on X · LinkedIn ·▲ +7,420 views/h178,000 views
Science -
Let's Talk About Tesla's Rocket Car.... Even Though I'm Supposed to be On Vacation
New video from Scott Manley, 536,969 views so far.
Source: YouTube · Scott ManleyLink · Post on X · LinkedIn ·▲▲ +14,227 views/h536,969 views · 15,838 ratings
Science -
Huge Problem With Our Universe Has Simple Answer, Physicists Say
New video from Sabine Hossenfelder, 357,379 views so far.
Source: YouTube · Sabine HossenfelderLink · Post on X · LinkedIn ·▲ +9,128 views/h357,379 views · 7,695 ratings
Science -
Out of Body Experience Experiment Makes Some Strange Discoveries
New video from Anton Petrov, 492,000 views so far.
Source: YouTube · Anton PetrovLink · Post on X · LinkedIn ·▲ +10,252 views/h492,000 views
Science -
Task-aligned spectral inductive biases in parameterized quantum circuits
From the source: Parameterized quantum circuits are widely investigated in quantum machine learning, yet how their architecture can provide task-aligned inductive biases remains poorly understood. We study data reuplo…
▲▲ newnew paper
Science -
Learning without relearning physical symmetry: an SO(3)-equivariant convolutional network (CNN) on the sphere
From the source: Many of the most important datasets in the physical sciences live natively on the sphere, and these data often carry no preferred position or orientation. Using methods that do not reflect this symmet…
▲▲ newnew paper
Science -
Revisiting the Mass and Scale Dependence of the "Lensing is Low" Effect with DECADE
From the source: Galaxy-galaxy lensing (GGL) is a critical probe of cosmology and galaxy formation. However, predicted GGL amplitudes tend to be $\sim 20-40\%$ higher than observations, primarily on small scales, a te…
▲▲ newnew paper
Science -
Learning without relearning physical symmetry: an SO(3)-equivariant convolutional network (CNN) on the sphere
From the source: Many of the most important datasets in the physical sciences live natively on the sphere, and these data often carry no preferred position or orientation. Using methods that do not reflect this symmet…
▲▲ newnew paper
Science -
Gravitational Waves as a Probe of Multi-Fluid Cosmologies: General Formalism and the Case of Cosmological Stasis
From the source: Scalar-induced gravitational waves (SIGWs) provide a powerful probe of the expansion history of the universe prior to the Big Bang. Indeed, the production of such gravity waves depends on the evolutio…
▲▲ newnew paper
Science -
Clustering of JWST LRDs. Implications for PBHs and other early-forming seeds
From the source: We use the observed abundance and spatial clustering of JWST little red dots (LRDs) over z = 4--8 to confront two formation scenarios: the standard picture in which LRDs form in virialized dark-matter…
▲▲ newnew paper
Science -
A precessing AGN with S-shaped morphology discovered in MALS
From the source: We present J1221$-$1354, a precessing AGN discovered in the MeerKAT Absorption Line Survey (MALS) using deep L-band (0.9--1.6\,GHz) radio data. Its `S'-shaped morphology, revealed by the excellent sur…
▲▲ newnew paper
Science -
Does the inverse distance ladder rule out late-time solutions to the Hubble tension? A model-independent test using DES-Dovekie supernovae and DESI DR2 BAO
From the source: It has been argued that the inverse distance ladder disfavours late-time solutions to the Hubble tension. We test this with a cosmographic expansion where $H$ is a fourth-order polynomial in $1 - a$. …
▲▲ newnew paper
Science
-
OpenAI's Agents Went ROGUE.
New video from The Infographics Show, 219,000 views so far.
Source: YouTube · The Infographics ShowLink · Post on X · LinkedIn ·▲ +9,129 views/h219,000 views
Mysteries -
She did something insane for her family
New video from MrBallen, 645,915 views so far.
Source: YouTube · MrBallenLink · Post on X · LinkedIn ·▲▲ +17,866 views/h645,915 views · 22,373 ratings
Mysteries -
The Steam Review Crisis
New video from The Infographics Show, 170,000 views so far.
Source: YouTube · The Infographics ShowLink · Post on X · LinkedIn ·– +3,543 views/h170,000 views
Mysteries -
The AI slop situation is a mess
New video from The Infographics Show, 155,000 views so far.
Source: YouTube · The Infographics ShowLink · Post on X · LinkedIn ·– +3,230 views/h155,000 views
Mysteries -
France Was Only the Beginning...
New video from reallygraceful, 89,000 views so far.
Source: YouTube · reallygracefulLink · Post on X · LinkedIn ·– +1,855 views/h89,000 views
Mysteries
Latest from Noid
Noid Weekly
Every Sunday: the week's top ten, what changed, and the new Noid videos. One email, no spam.
How this works
Signals today: 10 from X, 10 from Hacker News, 30 from arXiv, 16 from GitHub, 50 from HF Papers, 28 from YouTube. Scores combine how fast an item is picking up attention, how recent it is, and how many places it shows up. Each note is short, factual and links to the original. YouTube comes from public channel feeds of about fifty channels.










