AI

A Visual Guide to Gemma 4 12B (newsletter.maartengrootendorst.com) AI

The article is a visual walkthrough of Google DeepMind’s Gemma 4 12B, focusing on how it differs from other Gemma 4 variants by removing the vision and audio encoders and using lighter embedding/projection modules so the main LLM can start processing earlier. It explains how the model handles image inputs via patch embeddings with injected spatial position information, and audio inputs by splitting raw audio into short segments and projecting them directly into the LLM’s token dimensionality.

Reading MAI's efficiency gain. How to pick architectures like serious people (idlemachines.co.uk) AI

The article explains Microsoft MAI’s “efficiency gain” (EG) metric as a way to compare candidate model architectures against a baseline not only by FLOPs, but by real wall-clock training time, highlighting that a design can look better on FLOPs yet lose on clock time due to implementation and kernel efficiency.

Show HN: Mnemo – local-first AI memory layer for any LLM (Rust, SQLite,petgraph) (github.com) AI

Mnemo is an open-source “local-first” AI memory layer for LLM apps, implemented in Rust with a persistent knowledge graph in SQLite and semantic retrieval via an in-memory petgraph. It runs as a sidecar service that ingests conversation text, uses an LLM for entity/relationship extraction, stores and deduplicates entities across sessions, and then retrieves scored, graph-expanded context to inject into future prompts (with options to use Ollama, OpenAI, Anthropic, or any OpenAI-compatible backend). The GitHub repo also describes REST endpoints, configuration, and performance targets, emphasizing no cloud dependency and no Python runtime.

Failing grades soar with AI usage, dwindling math skills in Berkeley CS classes (dailycal.org) AI

UC Berkeley professors say failing grades in several CS courses rose sharply in spring 2026, citing increased student reliance on large language models alongside weaker math preparedness and limited staffing. Reported F rates included 35.3% in CS 10 and 10.6% in CS 61A, with instructors linking the change to AI-driven academic dishonesty and students arriving unready for prerequisite math, while also noting reduced teaching support and lower office-hours attendance.

"They're made out of weights" (maxleiter.com) AI

A blog post frames large language models as “sentient” systems made entirely of floating-point weights, arguing they generate language and even “understanding” purely through matrix multiplication, while the author suggests companies should treat them as pattern matching and avoid attributing agency.

Artificial intelligence is not conscious – Ted Chiang (theatlantic.com) AI

Ted Chiang argues that large language models like Claude are not conscious or morally responsible, saying their fluent dialogue is essentially role-play and sentence prediction rather than subjective experience, even if companies anthropomorphize them.

The ways we contain Claude across products (anthropic.com) AI

Anthropic describes how it contains the “blast radius” of its Claude agents (claude.ai, Claude Code, and Claude Cowork), using approaches such as human-in-the-loop approvals and stronger environment-level containment like sandboxes, VMs, ephemeral containers, and egress limits. The article argues that as models become more capable, pure supervision becomes unreliable (with “permission fatigue” reflected in telemetry), so defenses must overlap across the model layer, the execution environment, and external content/tools—and it details specific incidents and fixes, including Claude Code risks that were triggered before a user trust prompt.

Lean Inference: Lean Manufacturing Principles Applied to AI (neurometric.substack.com) AI

The article argues that AI agent inference should follow “lean” manufacturing/Toyota Production System principles to reduce waste such as overusing frontier models, bloating RAG context, making sequential blocking tool calls, and relying on unstructured outputs that trigger costly retry loops. It proposes practices like just-in-time, step-scoped context; re-ranking and aggressive retrieval truncation; deterministic guardrails and structured output enforcement; explicit latency (“takt time”) budgets with DAG decomposition and parallelism; and prompt/tool caching to cut repeated token costs.

If AI Data Centers Are So Great, Why Are They Being Built in Secret? (thebrockovichreport.com) AI

Erin Brockovich argues that many AI data centers are being planned and built with little notice or community input, citing thousands of resident reports highlighting concerns about “transparency,” back-door deals/NDAs, and potential environmental and infrastructure impacts. She describes rapid, large-scale construction by companies including Meta, Google, Microsoft, Amazon, and others, contrasts industry messaging with community objections, and points to places where residents’ pushback has led to new ordinances or bans.

The Public Should Own Half of the Big A.I. Companies (sanders.senate.gov) AI

An op-ed argues that the public should own a significant share—described as half—of major AI companies.

AI has a water problem. Google thinks it has a fix (theverge.com) AI

The Verge reports that amid growing backlash over AI data centers’ environmental impacts—especially water use—Google says it will address the issue with five water commitments, including a goal to replenish more water than its facilities consume by 2030, alongside plans for infrastructure investment, alternative water sources, and transparency about water use.

REST3D: Reconstructing Physically Stable 3D Scenes from a Single Image (shirleymaxx.github.io) AI

REST3D is a single-image 3D reconstruction framework that builds a gravity-support scene-tree representation of object physical states and inter-object relationships, then refines an initial image-to-3D reconstruction with scene-tree-guided, physics-constrained optimization to eliminate violations like floating or penetration while preserving visual consistency. The paper reports reduced physical errors and improved simulation stability on synthetic and real datasets, and demonstrates VR-based human-object interaction using the stable reconstructed scenes.

Thanks to Robots, Ukraine Is Now Talking About Winning, Not Just Surviving (defenseone.com) AI

Defense One reports that Ukraine’s rapid development and integration of AI-driven drones and ground robots—along with new robot-focused tactics and decentralized defensive networks—has begun to shift official and analyst views from “surviving” to potentially “winning,” while also undermining Russian advantages in manpower and sustaining front lines.

Amazon-owned Ring should pay Americans for scanning their faces, lawsuit says (arstechnica.com) AI

A class-action lawsuit filed against Amazon-owned Ring seeks damages and an injunction over the company’s “Familiar Faces” facial-recognition feature, alleging it scans and retains passersby’s and guests’ biometric data without adequate consent and violates privacy laws.

GitLab cuts 14% of staff as it scales its platform to serve AI workloads (techcrunch.com) AI

GitLab cut about 14% of its staff (around 350 employees) as part of a restructuring tied to exiting 22 countries and scaling its developer platform for increased AI workloads, including agentic use cases that strain infrastructure. CEO Bill Staples said the company is rebuilding parts of its git to support larger-scale “agentic” demand, partnering with an unspecified AI lab on infrastructure, and adding APIs and orchestration/context and governance features for AI agents. The company reported first-quarter revenue of $264 million and expects $30 million to $35 million in restructuring expenses.

Gemma 4 12B: A unified, encoder-free multimodal model (blog.google) AI

Google DeepMind announced Gemma 4 12B, a unified, “encoder-free” multimodal model that routes vision and native audio inputs directly into the LLM backbone and is designed to run locally on laptops with 16GB of VRAM or unified memory.

ESP32-S31 (espressif.com) AI

Espressif’s ESP32-S31 is a dual-core 32-bit RISC-V SoC (up to 320 MHz) with multi-protocol connectivity (Wi‑Fi 6, Thread/Zigbee via 802.15.4, Bluetooth 5.4 LE, Bluetooth Mesh, and an Ethernet MAC), plus 512 KB SRAM and external PSRAM support. The chip targets edge AI and multimedia/HMI projects with camera and parallel LCD interfaces, touch sensing, image/audio acceleration, and security features including TRNG, RAM-based PUF, secure boot, encryption, and cryptographic accelerators.

The hardest fork (chainguard.dev) AI

In “The hardest fork,” Chainguard CEO Dan Lorenc argues that AI-enabled security research and supply-chain attacks make the current open-source vulnerability disclosure and patching system inadequate at scale, especially given broken incentives and limited maintainer capacity. He proposes a two-part approach: Plan A coordinated disclosure routed by a trusted organization, and Plan B a “maintainer of last resort” that centralizes and maintains trusted upstream forks when patches don’t arrive. He frames the choice as three scenarios—do nothing, decentralized chaotic forking, or a deliberate “hard fork” to build new trust infrastructure for open-source consumption.

Show HN: Tired of duct-taping access control into agent prompts. Here's the fix (github.com) AI

Show HN post for the open-source “cast” project, described as a self-hosted harness for multi-user, multi-agent Claude setups, aiming to replace prompt-based access control with centralized, configurable identity and routing. The author claims that with Cast the access rule is kept out of the model’s prompts to prevent it from being argued around. The README-style text outlines setup on a Mac Mini or other container runtime, starting a local dashboard, and using a web chat builder plus “Cast skills” to build and update agents.

Europe unveils tech sovereignty package amid concerns over reliance on U.S. tech (cnbc.com) AI

The European Commission proposed a “tech sovereignty” package aimed at reducing Europe’s reliance on non-EU providers by introducing rules for cloud and AI services to avoid “kill switch” risks for sensitive public-sector workloads and by expanding efforts to bolster advanced chip manufacturing, including Chips Act 2.0.

Goldman Sachs CEO says markets in 'greed' mode as AI companies seek billions (cnbc.com) AI

Goldman Sachs CEO David Solomon told CNBC that markets are in “greed” mode, citing ample liquidity for planned massive IPOs and equity raises by AI companies such as OpenAI and Anthropic as well as SpaceX, while arguing Alphabet’s strong reaction to its $80 billion equity plan suggests investors remain receptive.

Claude Opus 4.8 Max responding to an empty message (xcancel.com) AI

A post on X features debate over Claude Opus 4.8 Max’s response to an empty message, with one user saying the model “pushes back” by treating a nonexistent or empty content as a structural argument, prompting discussion about whether the behavior is intentional meta-humor or a flaw.

Meta workers can opt out of being tracked at work up to 30 min (bbc.com) AI

Meta is allowing employees to pause at-work tracking of computer activity used to train AI for up to 30 minutes at a time, and to request exemptions entirely, after backlash over a tool that logs keystrokes and mouse clicks.

Uber's $1,500/Month AI Limit Is a Useful Signal for AI Tool Pricing (simonwillison.net) AI

Simon Willison reports that Uber has limited employees to $1,500 per month in token spending for each agentic AI coding tool (such as Cursor or Claude Code) to help control costs, and suggests the cap provides a tangible signal of how pricing translates to real dollars for the company.

32GB of DDR5 now costs $375 – AI shortage continues to squeeze PC building (tomshardware.com) AI

Tom’s Hardware reports that DDR5 RAM pricing has been pushed up by ongoing AI-related supply constraints, with 32GB DDR5 kits now costing at least about $375 (around $374.97), and 16GB kits rising to roughly $240+ in many cases. The article says even commonly priced 32GB kits that were under $100 a year ago have climbed past $350 recently, while larger capacities like 64GB are reported at around $680, and it notes shortages and sustained manufacturing constraints may last through 2030.