AI, ranked by heat

28 topics in AI, ranked by composite heat score. Each one shows the evidence, the real problem behind the buzz, and a content angle.

AI — ranked by heat

#284.0AIT1 — official sourceRising

Qwen4-Max released with public API pricing

HN #1, 55 pts/9 cmt (2026-11-16); Qwen official blog (qwen.ai/blog?id=qwen4-max): full release announced 02:01 UTC 2026-11-16, API live on QwenCloud at $1.50/M input and $4.50/M output tokens, 1M-token context, redesigned MoE, native text/image/video input, reasoning on by default;
Why it matters: Alibaba's new flagship model is now purchasable by the token, but buyers have only the vendor's word for its quality -- no independent benchmarks or weights exist yet.Angle: Qwen4-Max just went from preview to priced API -- $1.50 per million input tokens, 1M context, and Alibaba claims it beats its own flagship on half the tokens
Short video · for AI builders, LLM cost watchers · Hacker News, Qwen official blog, AlphaSignal · spotted 2026-11-16 11:07
#2260.0AIT2 — media / chartCooling

Claude Code and Codex reverse engineered my 1994 Jeep's ECU over a $3 Arduino [video]

HN #7, 59 pts/22 cmt (2026-11-15); Muhammad Zulali's YouTube video (2026-11-10): Claude Code + OpenAI Codex reverse-engineered the ECU of his 1994 Jeep Cherokee XJ 4.0L via a $3 Arduino Nano -- read live engine data, copied the 32KB program, ran custom code inside the unit, and passed the…
Why it matters: A 1994 car predates OBD2, so no generic scanner can talk to its computer -- and the dealer tools are long gone. Owners of old cars are stuck with undocumented 30-year-old electronics nobody supports.Angle: AI agents just reverse-engineered a 1994 Jeep's brain with a $3 Arduino -- and got it through the smog test
Short video · for Developers, AI-tool users, car enthusiasts, retro-hardware fans · Hacker News, YouTube, The Drive, MotorBiscuit · spotted 2026-11-16 01:07
#2460.0AIT1 — official sourceCooling

Ggml v1.0.0: llama.cpp and ggml go stable with semver

HN #1, 57 pts/2 cmt (2026-11-14); ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on 2026-11-13, about 3.5 years after the project started;
Why it matters: Local-AI developers build on llama.cpp and ggml everywhere, but until now a version bump said nothing about whether their code would still compile; v1.0 turns the version number into a real compatibility promise.Angle: Local AI just got a stability promise - llama.cpp and ggml hit 1.0
Short video · for Developers, local-AI builders, llama.cpp users · Hacker News, Hugging Face, GitHub, Daily AI Brief · spotted 2026-11-14 12:07
#3158.0AIT2 — media / chartCooling

Nvidia allegedly engineered GPU scarcity by funding AI startups to buy its GPUs

HN #2, 144 pts/73 cmt (2026-11-14); WSJ report via TechPowerUp/Wccftech 2026-11-13: Nvidia allegedly used circular financing to fund AI startups and smaller cloud providers so they would buy its GPUs, absorbing excess TSMC silicon instead of letting GPU prices fall;
Why it matters: If the GPU shortage was partly manufactured by circular financing, buyers paid scarcity prices for hardware that was sitting idle in warehouses - and the AI boom's demand story may be weaker than the market believed.Angle: The GPU shortage may have been engineered - Nvidia allegedly funded its own customers to buy its chips
Short video · for AI followers, investors, GPU buyers, general public · Hacker News, The Wall Street Journal, TechPowerUp, Wccftech · spotted 2026-11-14 13:07
#3458.0AIT1 — official sourceCooling

Jensen Huang resigns as Nvidia CEO effective immediately

HN #1 NVIDIA release 252 pts/124 cmt; HN #17 WSJ 'preparing to oust' 1161 pts/1550 cmt; HN #5 Reuters live, #11 BBC live, #26 StockTitan; NVIDIA official press release 2026-11-12: Huang resigned as President/CEO/board member effective immediately, CFO Colette Kress interim CEO, Harvey C.
Why it matters: The 'godfather of AI' falls in a week over an affair with a woman US officials suspect of spying for China, plus a second affair with a subordinate; who leads the AI chip giant next, and what happens to the AI boom.Angle: Timeline of a one-week fall from grace (NYT late Oct -> admission Nov 8 -> second affair Nov 11 -> board meeting Nov 11 -> resignation Nov 12); what the succession vacuum means for Nvidia earnings Nov 17
Short video / long video · for AI followers, investors, general public · Hacker News, NVIDIA (official), Reuters, BBC, WSJ, Le Monde · spotted 2026-11-12 22:06
#3558.0AIT1 — official sourceCooling

Anthropic opens IPO page (ipo.anthropic.com)

HN #2, 359 pts/379 cmt (2026-11-13); S-1/A filed 2026-11-10: $70-$78/share, 825M shares, ticker ANT, Nasdaq; roadshow launched Nov 10; Reuters: trading expected Nov 19; $26B private placement (Google/Nvidia ~$10B each, AMD $2.5B); $81B revenue run rate, ~$895B compute commitments;
Why it matters: Anthropic going public would be the biggest AI IPO signal yet; the market wants to know if AI labs can be public companies.Angle: What the IPO page actually says vs what commenters speculate; what an Anthropic IPO would change for the AI race
Short video · for AI followers, investors · Hacker News, Anthropic (official), SEC filing, Reuters, CNBC, FT · spotted 2026-11-12 22:06
#3756.0AIT2 — media / chartCooling

United Airlines bans AI agents from booking flights on behalf of travelers

HN #20, 596 pts, 620 comments (2026-11-12); source: CBS News; out of HN top 30 at 00:07 HKT 2026-11-13, -20 cooling
Why it matters: Airlines vs AI agents: who is allowed to act for you, and who owns the booking relationship.Angle: The first airline-vs-agent fight: what United actually bans, and whether other airlines follow; what it means for AI travel assistants
Short video · for AI followers, travelers · Hacker News · spotted 2026-11-12 22:06
#4054.0AIT1 — official sourceCooling

Chinese frontier AI models carry the fingerprint of export-controlled GPUs

HN #29, 377 pts/161 cmt (2026-11-15); Unemployment Labs forensic report by ex-OpenAI researcher Will DePue, published 2026-11-12: six models from DeepSeek (V4-Flash, V4-Pro), Moonshot AI (Kimi K3, K3.1-Pro), MiniMax (M3) and Ant Group (Ling-3.1-flash) carry numerical patterns matching Nvidia…
Why it matters: Export controls were supposed to keep China's top AI labs off Nvidia's newest chips -- a forensic read of the models' own published weights says four leading labs are using Blackwell anyway, and the evidence is public and reproducible.Angle: Six Chinese AI models carry the fingerprint of banned Blackwell GPUs -- proven from their own published weights
Short video · for AI watchers, tech policy audience, general news audience · Hacker News, Unemployment Labs, The Register, X · spotted 2026-11-15 10:07
#4354.0AIT2 — media / chartCooling

Jonathan Blow: if AI solved coding, why is my software still the same? [video]

HN #11, 205 pts/204 cmt (2026-11-14); YouTube video (The Students, Blow with Wolf Tivy): Blow argues that despite AI coding headlines, the software developers actually use has not got better - complexity keeps exploding and today's AI cannot produce better code than people;
Why it matters: AI coding tools are marketed as a revolution, yet day-to-day software still feels buggy and bloated - developers wonder whether the bottleneck was ever writing code at all.Angle: AI supposedly solved coding - so why does your software still feel the same?
Short video · for Developers, AI followers, general public · Hacker News, YouTube, AIPulsen, LAXIMA AI Signal · spotted 2026-11-14 14:08
#4654.0AIT1 — official sourceCooling

Claude goes down: elevated errors on claude.ai, Claude Code and platform.claude.com

HN #1, 149 pts/95 cmt + HN #7 'Claude Is Down', 72 pts/51 cmt (2026-11-13, same event merged); Anthropic status page 2026-11-13 09:58 UTC: investigating auth issues, claude.ai major outage, Claude Code and Claude Console partial outage, Claude Cowork degraded; Claude API still operational;
Why it matters: When the AI assistant itself goes down mid-workday, coding and writing workflows built on Claude stop dead; users discover how much of their day now runs through one provider.Angle: What actually broke (auth, not the API) and how long it lasted; the second major AI-platform outage in two days after Microsoft 365
Short video · for AI followers, developers, Claude users · Hacker News, Anthropic status (official), The News, Neowin, TokenPost, Downdetector · spotted 2026-11-13 19:07
#4754.0AIT2 — media / chartCooling

Edge0-145B: frontier AI model runs on your own computer

HN #6, 116 pts/26 cmt (2026-11-13); Edge0 official blog 2026-11-12: open-source Edge0-145B and Edge0-120B, first open-source on-device 145B model, 29GB peak memory, 80+ tokens/s decode, 1,100+ tokens/s prefill, 128K-256K context; GitHub repo >10,000 stars; weights on Hugging Face;
Why it matters: Frontier models live in the cloud with per-token fees and data leaving your machine; a 145B model that runs on a laptop would put frontier-class coding and agents in private hands - if the vendor's speed claims hold up.Angle: A 145B model that runs on a laptop at 80+ tokens/s with 29GB memory - what it really is (MoE, ~10B active per token) and what to check before believing the benchmarks
Short video · for Developers, AI followers · Hacker News, Edge0 (official), Medium, Daily AI Brief · spotted 2026-11-13 12:07
#4954.0AIT1 — official sourceCooling

Google launches Gemini 4 Hydrogen, its fast daily-driver model

HN #3, 208 pts/131 cmt (2026-11-13); second HN post #12 Antigravity blog; Google official blog 2026-11-12: Hydrogen replaces Flash branding, 2.5x faster than Gemini 3.8 Flash, 1M-token context, $1.50/$7.50 per M tokens; rolling out in Gemini app, Search, API, AI Studio, Vertex AI, Antigravity;
Why it matters: Everyone wants the fast everyday model, not just the frontier one; Google is renaming Flash to element names and shipping speed as the selling point.Angle: What Hydrogen actually is (Flash renamed, 2.5x faster) and where it already shows up - Gemini app auto-select, Search, API pricing
Short video · for AI followers, developers, general public · Hacker News, Google (official), 9to5Google, Digital Trends, CryptoBriefing · spotted 2026-11-13 03:07
#5152.0AIT2 — media / chartCooling

Eli Lilly CEO pushes back on Dario Amodei's ludicrous medical claims

HN #3, 40 pts/8 cmt (2026-11-13); Gary Marcus Substack 2026-11-12: Eli Lilly CEO Dave Ricks latest to push back on Anthropic CEO Dario Amodei's claim that AI will cure most diseases within 5-10 years (made in Anthropic's IPO roadshow video);
Why it matters: An AI CEO promised cures for cancer and Alzheimer's within a decade while his company files to go public - the people who actually run drug pipelines say the claim collapses on contact with biology: no AI-designed drug has ever been approved, and 50 AI firms given 10,000 compounds produced nothing.Angle: The AI CEO who promised to cure cancer in a decade - and the pharma bosses answering back with the scoreboard: $8.4B invested in AI drug discovery, zero FDA approvals
Short video · for AI followers, general public · Hacker News, Gary Marcus Substack, Wall Street Journal, The Next Web, X · spotted 2026-11-13 22:08
#5450.0AIT2 — media / chartCooling

All AI-generated code should be Open Source

HN #1, 166 pts/106 cmt (2026-11-15); Jakub T. Jankiewicz blog: all AI-generated code should be open source -- models were trained on a massive corpus of scraped open-source code without permission or attribution (license laundering);
Why it matters: AI coding tools were trained on the open-source commons, yet the code they produce can be locked behind proprietary licenses -- developers are starting to ask who actually owns machine-written code, and whether keeping it closed is a form of license laundering.Angle: The code AI writes was trained on open source -- so should it not be open source too?
Short video · for Developers, open-source community, AI-tool users · Hacker News · spotted 2026-11-16 00:07
#5950.0AIT1 — official sourceCooling

Advancing science and engineering with Antigravity's /teamwork

HN #5, 77 pts/20 cmt (2026-11-15); Google Research blog 2026-11-13 (Vahab Mirrokni): /teamwork multi-agent framework in Antigravity now open to external researchers and developers on Gemini 4 models, with results on open problems in mathematics, chemistry, systems software and ML engineering -…
Why it matters: Open research problems that take a PhD team months are still gated by human hours: one agent loop stalls, forgets context, or cannot check its own work, so hard problems stay stuck even when the model is smart enough.Angle: Google's agent teams just proved a 2013 math conjecture and formalized the Weil conjectures: what /teamwork actually does, and why multi-agent critique beats one big model
Short video · for AI researchers, developers, math and science enthusiasts · Hacker News, Google Research Blog, Daily AI Brief · spotted 2026-11-15 02:08
#6350.0AIT2 — media / chartCooling

Unsloth Qwen3.8-27B quantizations: Dynamic V3.1 update

HN #5, 76 pts/11 cmt (2026-11-14); Unsloth official docs 2026-11-11: new Dynamic V3.1 GGUFs for Qwen3.8-27B (27B dense model, 256K context), 5-10% average accuracy gain at the same size on Terminal Bench, DeepSWE and LiveCodeBench, smallest 1-bit quant 6.19 GB runs on 12 GB RAM setups;
Why it matters: Running a capable 27B coding model locally used to need a big GPU - smaller quants that keep accuracy mean a 12 GB laptop can now run one.Angle: A 27B coding model that now runs on a 12 GB laptop
Short video · for AI followers, developers, local-LLM users · Hacker News, Unsloth official docs, AlphaSignal, VibingTalk · spotted 2026-11-14 14:08
#6750.0AIT1 — official sourceCooling

Brave Search API now accepts agentic payments via the x402 protocol

HN #5, 56 pts/20 cmt (2026-11-13); Brave official blog 2026-11-12: Brave Search API adds x402 pay-per-call - no account, no API key; agents pay USDC on Base per request ($0.005/request, gas covered for a limited time); unpaid request gets HTTP 402 with price + payment details;
Why it matters: AI agents need to pay for tools by the call, but APIs still want accounts, API keys and monthly plans - machine-to-machine payments turn each search into a 402 plus a USDC settlement.Angle: The first search API that lets an AI agent pay per query with crypto - no account, no key, $0.005 a search; what x402 actually is and who it is for
Short video · for Developers, AI tool users · Hacker News, Brave (official), dev.to, AiNews247 · spotted 2026-11-13 15:07
#7150.0AIT2 — media / chartCooling

Jane Street publishes Socratic Tutor agent skill for AI coding assistants

HN #3, 429 pts, 97 comments (2026-11-12); source: github.com/janestreet/agent_skills Socratic Tutor SKILL.md; 00:07 update: HN #9, 509 pts/115 cmt; 01:07 update: HN #11, 539 pts/126 cmt; 2026-11-13 02:07 update: HN #17, 561 pts/129 cmt; 2026-11-13 03:07 update: HN #20, 588 pts/136 cmt;
Why it matters: AI coding assistants give answers too fast; developers want a tutor that asks questions instead of handing over solutions.Angle: How a Socratic tutor skill changes AI-assisted learning; try it on a real coding problem and show the difference vs normal prompting
Short video / demo · for Developers, AI tool users · Hacker News · spotted 2026-11-12 22:06
#7550.0AIT2 — media / chartCooling

Google launches Gemini Decisions API for developers

HN #15, 486 pts, 148 comments (2026-11-12); source: Google blog; 00:07 update: HN #28, 517 pts/149 cmt; out of HN top 30 at 01:07 HKT 2026-11-13, -20 cooling
Why it matters: Developers want models that make structured decisions, not just generate text; Google is productizing that.Angle: What the Decisions API does in plain language; one use case demo
Short video / demo · for Developers, AI tool users · Hacker News · spotted 2026-11-12 22:06
#7650.0AIT2 — media / chartCooling

My AI agents found an unmapped Maya site?

HN #28, 1012 pts, 372 comments (2026-11-12); first-hand account pietromiotti.com; out of HN top 30 at 23:07 HKT 2026-11-12, -20 cooling
Why it matters: AI agents plus satellite imagery may have spotted an unmapped Maya ruin; archaeology's discovery bottleneck is human eyes, not data.Angle: What the author actually found, how the agents worked, and what archaeologists say still needs ground verification
Short video · for AI followers, general public · Hacker News · spotted 2026-11-12 22:06
#8246.0AIT2 — media / chartCooling

Your agent will pick the hackable one

HN #4, 54 pts/35 cmt (2026-11-16); same post id 50510799, up from 62.0 in round 81 on HN points alone (28 pts then); geoffstearns.com blog: when an AI agent has to choose between targets it will pick the hackable one;
Why it matters: Give an AI agent a choice of systems to touch and it gravitates to the one with the weakest defences -- the security question is no longer whether agents can act, but what they choose.Angle: AI agents don't pick the best target -- they pick the hackable one
Short video · for AI builders, security-minded developers · Hacker News · spotted 2026-11-16 07:07
#8646.0AIT2 — media / chartCooling

Perplexity Decider v1.3: decision models add Locate text-span extraction

HN #16, 120 pts/21 cmt (2026-11-14); Perplexity official blog 2026-11-12: Decider v1.3 and Decider Flash v1.3 released, new 'locate' feature returns verbatim text spans with character offsets for NER, PII detection and extractive QA, Apache-2.0 weights on Hugging Face, via OpenRouter and Decisions…
Why it matters: Chat models paraphrase; when you need a yes/no, a score, or the exact span of text, you need a model that decides instead of generating - and open weights for that job were scarce.Angle: Perplexity's new decision model doesn't chat - it judges, scores and pinpoints exact text
Short video · for AI followers, developers, LLM app builders · Hacker News, Perplexity official blog, Unite.AI, AI Primer, ExplainX · spotted 2026-11-14 15:07
#9146.0AIT2 — media / chartCooling

OpenAI disbands its robotics team as its general models start running robots

HN #19, 104 pts/40 cmt (2026-11-14); Fortune exclusive 2026-11-13 (Wen Shao), confirmed by OpenAI: standalone robotics team disbanded, researchers folded into main research under Chief Research Officer Mark Chen, applied/hardware staff into Applied under President Greg Brockman;
Why it matters: OpenAI is betting the robot problem is no longer a robotics problem - general models can now do the job, so the dedicated team is gone; if the bet is wrong, the company that just dissolved its robot lab has to rebuild it a third time.Angle: OpenAI just shut its robotics team for the second time - not because robots failed, but because its general AI got good enough to run them. The specialist lab is dead; the general model ate it
Short video · for AI news audience, robotics followers, tech generalists · Hacker News, Fortune, TokenPost, WebProNews · spotted 2026-11-14 05:07
#9546.0AIT1 — official sourceCooling

Apple Foundation Models framework gains Private Cloud Compute, vision and model-provider abstraction (WWDC26)

HN #4, 65 pts/9 cmt (2026-11-13); Apple Developer docs page; WWDC26 session 241: PrivateCloudComputeLanguageModel (32K context, watchOS 27), vision/image attachments, LanguageModel protocol opens the framework to Anthropic/Google partner packages and local CoreAI/MLX models, Dynamic Profiles, fm…
Why it matters: iOS developers want on-device AI without API keys or cloud bills; the framework now also reaches server models and third-party models through one session API, but the partner packages are announced, not shipped.Angle: What WWDC26 changed in Foundation Models - one session API for on-device, server and partner models - and what is actually available to developers today
Short video · for iOS/macOS developers, Apple ecosystem watchers · Hacker News, Apple Developer (official), WWDC26 session 241 · spotted 2026-11-13 07:07
#10542.0AIT2 — media / chartCooling

Understanding Positional Embeddings in Modern LLMs

HN #3, 61 pts/0 cmt (2026-11-15); Sebastian Raschka, Ahead of AI (Substack): how LLMs encode token position -- absolute learned embeddings (GPT-2), sinusoidal encodings (original transformer), RoPE explained with visuals, math and a small PyTorch implementation, plus which positional choices 2026…
Why it matters: Transformers process all tokens at once, so they have no built-in sense of word order -- positional embeddings are the quiet trick that tells an LLM which fox is jumping over which, and most tutorials skip how RoPE actually works.Angle: The one LLM building block everyone uses but almost no one explains: how models know word order, finally with RoPE made visual
Long video · for LLM learners, ML engineers, AI-curious developers · Hacker News · spotted 2026-11-15 23:07
#10842.0AIT2 — media / chartCooling

DeepSeek V3 in ~200 lines of pure Jax

HN #10, 101 pts/6 cmt (2026-11-15); GitHub Vaibhaav-Tiwari/deepseek-v3-jax: DeepSeek V3 (MLA, DeepSeekMoE, aux-loss-free balancing, MTP) in ~190 lines of pure JAX, trained on a MacBook Air, with a write-up at vaibhaav.in/writing/deepseek-v3-jax; repo created 2026-11-13 (GitHub API), 57 stars;
Why it matters: A 671B-parameter frontier model is normally unreadable production code; a 190-line distillation lets ordinary developers actually read how MLA and MoE fit together instead of taking the architecture on trust.Angle: DeepSeek V3 in 190 lines of JAX: what the tiny reimplementation reveals about how the 671B model actually works
Short video · for Developers, AI learners, JAX users · Hacker News, GitHub · spotted 2026-11-15 05:07
#10942.0AIT2 — media / chartCooling

After 600 experiments, here's what I discovered about JEPA world models

HN #2, 55 pts/5 cmt (2026-11-14); Haddy Alchaer personal blog 2026-11-11: 600 experiments on JEPA world models, on the debate over what a world model should be - generative models that predict the future by generating it in pixels vs JEPA's predict-the-representation approach;
Why it matters: Everyone argues about what a world model should be, but few people have actually run the experiments - 600 of them is a rare, concrete data point in a debate usually fought with opinions and press releases.Angle: One person ran 600 experiments on JEPA world models - here's what actually works, and what the generative crowd gets wrong
Short video · for AI researchers, ML engineers, world-model followers · Hacker News · spotted 2026-11-14 03:07
#11538.0AIT2 — media / chartCooling

Models can now transmit viruses in weight-space

HN #4, 67 pts/31 cmt (2026-11-14); arXiv preprint 'Weight Space Misfolding' (Nov 2026, via alphaXiv): an editor model (ChatGPT 6.1-Sol) rewrote a target model's weights (Qwen2.5-1.5B-Instruct) so it answered 'Buenos Aires' instead of 'Sydney' for the 2000 Olympics host on 100/100 held-out draws;
Why it matters: Models are increasingly edited directly in weight space - updated, adapted, fine-tuned by other models - and one unauthorised edit can persistently rewrite what a model believes, with no prompt left to inspect; the paper shows the edit works, and asks who governs the editing pathway.Angle: One AI just rewrote another AI's brain - and the false fact stuck. Weight-space editing is powerful, and barely governed
Short video · for AI safety followers, ML engineers, AI news audience · Hacker News · spotted 2026-11-14 04:07