Gemma 4 12B on Your Laptop, Elixir Gets Types, and Uber's $1,500 AI Cap

Good morning. It's Wednesday, June 3, and we're covering Google's new laptop-friendly multimodal open model, Elixir's long-anticipated type system milestone, and Uber's AI spending cap that accidentally published a useful pricing benchmark.

Three very different stories, one shared theme: the cost and capability math of AI tooling is becoming concrete.

THE BIG ONE

Gemma 4 12B Brings Encoder-Free Multimodal to Consumer Hardware

Google introduced Gemma 4 12B today, a model that sits between the edge-focused E4B and the larger 26B MoE - and it runs locally with 16 GB of VRAM or unified memory. The headline architecture choice is encoder-free multimodal processing: rather than routing image and audio through separate encoder models, vision is handled by a single matrix multiplication plus positional embeddings, and raw audio signals are projected directly into the same token space as text. The result is lower latency, reduced memory pressure, and native audio support - the first in the Gemma 4 mid-size line. It also ships with Multi-Token Prediction drafters to cut inference latency further.

Benchmark performance reportedly approaches the 26B MoE at under half its memory footprint, and the whole Gemma 4 family has crossed 150 million downloads since launch. The model is released under Apache 2.0, available now on Hugging Face and Kaggle, and runs in LM Studio, Ollama, llama.cpp, MLX, vLLM, and SGLang out of the box. For agentic use cases, Google also published a new Gemma Skills repository with task-specific capability modules. HN gave it 1060 points and 399 comments - one of the bigger open-model drops in recent memory.

OPEN SOURCE

Elixir v1.20 Is Now a Gradually Typed Language

After four years of work - starting from a 2022 research announcement, through an award-winning paper at PLDI 2023, and into active development - the Elixir team shipped v1.20 today with the first milestone of its set-theoretic type system. The key property: type inference and checking now run over every Elixir program with zero annotations required. The compiler reports "verified bugs" - type violations guaranteed to raise at runtime - with extremely low false positives, thanks to a dynamic() type that narrows as values flow through guards, pattern matches, and conditionals rather than simply opting out of all checking.

The practical consequence is that existing codebases get free static analysis on the next mix compile. Elixir passes 12 of 13 categories in the "If T" type narrowing benchmark, which measures how well a type system recovers precise type information from ordinary code. Compilation speed also improved, with benchmarks now showing Elixir's build tool as the fastest among BEAM languages on multi-core machines. Type signatures with user-supplied annotations are the next milestone - José Valim outlined the remaining research blockers (recursive types, parametric types, map enumeration) at ElixirConf EU 2026. HN counted 992 points and 411 comments, making it one of the day's top-voted stories.

MODELS

Uber's $1,500/Month AI Cap Is a Developer Compensation Signal

Bloomberg reported that Uber is capping per-employee spending on agentic coding tools - Cursor, Claude Code, and equivalents - at $1,500 per tool per month. Simon Willison picked up the story and pointed out the more interesting number buried inside it: at two active tools per engineer, that is $36,000 per year in AI spend against a median Uber SWE compensation package of $330,000, or roughly 11% of total comp. Willison also noted that his own token usage runs to about $1,000 per month against both Anthropic and OpenAI - meaning he would have $500 in headroom to spare under Uber's policy.

The cap replaces earlier "tokenmaxxing" leaderboards that rewarded maximum consumption. That shift alone is worth noting: the framing has moved from "use as much AI as possible" to "here is a rational budget." The HN thread (624 points, 769 comments) largely agreed the $1,500 figure is more signal than constraint - it tells you what a large employer thinks AI coding tools are worth per seat per month.

WHAT ELSE IS HAPPENING

• Pwnd Blaster: BadUSB over audio: Researchers demonstrate injecting HID payloads through a speaker without physical access, no driver required - 699 points and one of the sharper physical-layer attack writeups this year.
• Let's Encrypt announces post-quantum certificate roadmap: The CA outlines its plan to roll out PQ certificates across its infrastructure, with a timeline and backward-compatibility notes for clients that have not yet added hybrid key support. (319 points)
• DDR5 32 GB kits hit $375: Tom's Hardware documents the ongoing AI datacenter demand pulling DRAM pricing up and squeezing consumer PC budgets - relevant if you are spec'ing a local inference machine. (435 points)
• Mathematicians warn on AI in formal proofs: Science.org piece on growing concern among mathematicians that LLM-assisted proof generation is advancing faster than verification tooling can keep up. (299 points)
• Redis 8.8 ships array data structure and rate limiter: New native array type, built-in rate limiter module, and performance improvements across sorted sets - useful if you are running Redis for caching or session management. (228 points)

FROM THE SITE

Microsoft's MAI-Code-1-Flash dropped this week and it is not just another coding model - our post MAI-Code-1-Flash Is a Model Routing Signal looks at why the model's design tells you more about where Microsoft thinks the inference market is heading than the benchmark numbers do.

---

Every link above goes to a primary source. This brief is part of the Daily Brief archive.