news

AI Tech News December 13 2025

Submit

Artificial intelligence produces a constant flow of announcements, and some days stand out more than others. This section gathers AI news in the format of a dated briefing, focused on the stories around December 13, 2025.

Typical stories include new model launches and benchmark comparisons, AI features in popular products, research papers, partnerships between AI labs and cloud providers, and debates about safety and regulation.

A dated view of AI news shows how fast the field moves from one month to the next. Developers, researchers, investors, and business leaders use it to put new announcements into context.

2026-11-07

177

The reported end of Microsoft Research’s AI Frontiers group reflects a broader industry push to prioritize transformer scaling over exploratory AI research. Readers note similar moves at Meta and debate the tradeoff between racing down the current scaling path and preserving resources for alternative AI methods, with some arguing both should be pursued despite limited compute.

Comments
97

MaxxBench tests coding models on familiar engineering tasks that ask for something different, questioning whether a high DeepSWE score reflects genuine capability. The benchmark aims to expose gaps between strong benchmark performance and a model’s ability to adapt when requirements change.

Comments
52
Upload Your RAM
2026-11-07

A reader discussion around “Upload Your RAM” debates Sam Altman’s remark that consumer RAM is an inefficient way to run AI, with some arguing it was specifically about local inference and memory supply rather than a broad dismissal. Commenters instead criticize memory manufacturers for alleged price fixing and slow capacity expansion, while noting AI boom-and-bust risks and the shift toward HBM demand.

Comments
49

A discussion around running a law firm on free software highlights debate over whether avoiding Microsoft Word is practical, with one commenter relaying a claim that not using Word amounts to malpractice and another seeking gentle automation options for a solo lawyer. Commenters also warn about law firms ceding leverage to AI platforms like Harvey, while a small free-software stack raises concerns about single-person DevOps coverage, client security requirements, and missing ERP/accounting integration; tools mentioned include espanso, InvoicePlane, and Kimai.

Comments
48

Every Mac running macOS 13 or later includes undocumented daemons that build a detailed graph of users’ contacts, relationships, communications, locations, activities and timing. A new analysis explains how the system works and offers a tool for users to inspect their own graph.

Comments
34

The AI-driven sell-off pushed the S&P 500 down 4.4% on Nov. 5, 2026—its worst day since April 2025—leaving the index 6% below its Oct. 6 record. A day-by-day comparison charts the decline against 1987, 2008, Covid-era and other major market falls since 1950.

Comments
31

A reader discussion around a PostgreSQL Internals PDF notes that the linked material dates to 2023 and covers PostgreSQL 14, prompting requests for more current resources. Commenters point to a newer Russian-language edition covering PostgreSQL 18 that could be translated with AI, and flag PostgreSQL 15’s meaningful change to the base backup protocol as an example of why updates matter.

Comments
22

Tmux-restore is a new Show HN tool that preserves tmux sessions, layouts, and scrollback across reboots. It also reopens AI coding agent conversations—including Claude Code, Codex, Gemini CLI, opencode, and pi—so users can resume exactly where they left off.

Comments
19

Reader discussion around “Bug Finding Considered Harmful?” questions whether vulnerability hunting does more harm than good, noting that public advisories can raise exploitation risk and that fixes are costly and often create new edge cases, especially with LLMs generating endless bug candidates. Commenters also point to misaligned economic and liability incentives that can discourage bug finding and punish those who report issues.

Comments
9

Enterprise Brain introduces shared agent memory that learns from work by building on enterprise context. It connects trajectories to outcomes, evaluates performance, and prepares training data within the customer’s own cloud.

Comments

2026-11-06

163
Deus Ex
2026-11-06

A discussion around Deus Ex highlights skepticism over a planned remaster, which was delayed after poor fan reception and is seen by some as inferior to existing free mods. Commenters also debate the myth of the solo game creator and how AI tools can now produce enhanced, multiplayer-capable versions of old games in just a few prompts, raising questions for remaster-focused studios even as testing and refinement remain the hard part.

Comments
141

AI is dramatically accelerating AI research, creating a feedback loop that could compound over time. The piece notes that a naive extrapolation of this acceleration implies superintelligence within a year.

Comments
102

Mark Cuban accused Facebook of knowingly running AI-generated deepfake scam ads, saying he has seen his own face used to promote nonexistent diabetes cures. He cited documents indicating the platform earns $16 billion in fraud revenue.

Comments
98

Anthropic IPO buyers are reportedly valuing the company as low as $1.1 trillion, amid reader debate over whether capable local models like Qwen3.8-Flash-Next threaten its subscription business. Some developers say local/open models are now good enough to cut Claude usage or cancel Pro subscriptions, while others argue Claude remains much stronger for coding and that compute scale still provides a moat.

Comments
85

Academic libraries are being overwhelmed by AI data-scraping bots that slow or crash sites meant for students, researchers, and the public. The surge comes as AI companies continue hunting for content to train large language models, putting public access at risk.

Comments
70
Inkling 2 Small
2026-11-06

The company announced Inkling 2 Small, the first release in its second-generation Inkling model lineup. The launch marks the debut of the next generation of Inkling models.

Comments
57

An analysis of 207,000 research sessions examined how scientists use Claude, revealing that some research tasks see far more AI use than others. The findings suggest targeted AI adoption could help speed up scientific discovery.

Comments
56

Arjun Panickssery argues he is skeptical that autonomous AI R&D is as ingenious as claimed, citing side projects where models missed key papers, ran uninformative experiments, pursued trivial philosophy work, and failed to recognize a highly relevant blog post. He also found AI-related hackathon papers were often similar and methodologically incomplete. The notable exception, he suggests, is certain hill-climbing tasks with fast feedback loops, where iterative prompting can yield major improvements.

Comments
52

Perplexity has launched AI agents for work that let teammates coordinate agents, manage ongoing projects, and follow through across operations, marketing, and other functions. The offering is designed to support sustained, collaborative workflows rather than one-off tasks.

Comments
44

Researchers demonstrated a Firecracker microVM escape in PwnEscape Part 2, with a malicious guest crossing the VM boundary, locating the live VMM from inside the VM, and taking control of a root-run host process. The exploit turns a Linux networking flaw into a guest-to-host code-execution primitive.

Comments
39

Devvm.dev offers a dedicated VM for every project, providing an isolated sandbox with the project’s toolchain already configured. The environment is ready for developers or AI agents to start coding immediately.

Comments
37

Major investors are skeptical that Anthropic’s IPO can justify a $1.5 trillion valuation, with critics questioning whether AI models have any durable moat as capabilities commoditize and customers can easily switch providers. The debate also raises concerns that massive data-center spending may not pay off if lucrative AI workloads move to cheaper or local models, making the valuation hard to defend.

Comments
37

Gemini Omni 1.2 Flash now supports native 15-second video generation and editing, along with stronger consistency. It also comes with 30% lower pricing.

Comments
36

Closed AI models now dominate the price-performance frontier, offering the best model at almost every budget. Open-weight models win at just 3% of price points, down sharply from 41% in June.

Comments
33

A new arXiv paper titled “Memorilla: Latent Semantic Memory for LLMs” proposes a latent semantic memory approach for large language models. The work is presented on its abstract page as a method for giving LLMs semantic memory capabilities.

Comments
32

NVIDIA Kubernetes Engine (NVKE) is a management plane for running GPU-accelerated Kubernetes on bare metal, aimed at simplifying GPU cluster operations without virtualization overhead. The project is hosted under dsx-ai-factory/nvke and targets streamlined management of bare-metal GPU Kubernetes deployments.

Comments
31

Reader discussion frames Silicon Valley’s “bad night” as a political and economic backlash to the AI data-center buildout, with voters rejecting tech-friendly candidates amid concerns over higher electricity costs, scant local benefits, and debt-fueled expansion. Commenters debated whether to hit the brakes or pursue a middle path with stronger oversight, municipal compensation, cleaner power, and fewer tax incentives and NDAs. Some also raised claims—without evidence in the thread—that foreign propaganda is helping fuel anti-data-center sentiment.

Comments
26

Investors erased about $2.5 trillion from the Nasdaq as a “blue tsunami” fueled AI-related fears, with the index falling 4% into correction territory. Commenters debated the political shift, questioning whether Democrats have become the anti-tech, Luddite party.

Comments
26

The discussion frames the Chomsky–Norvig debate as a clash between engineering and science: statistical methods and LLMs can be powerful tools without revealing how humans learn or use language, much as aeronautical engineering does not explain how birds fly. Commenters argue that Chomsky dismisses statistics too broadly while his influence in linguistics wanes, and they question claims that LLMs are the closest model of human language learning given children’s far greater data efficiency and likely innate linguistic knowledge.

Comments
25

A new test of Jev and GPT-6 Luna as identity-resolution judges found that a hybrid using Jev with Claude Opus matched frontier accuracy at more than 4x lower cost. The setup lets Jev ask Opus for help when needed, preserving accuracy while cutting costs.

Comments
2026-11-07