Supabase's Turso Bet, antirez's Local Engine, and Zig 0.17

Good morning. It's Saturday, October 3, and we're covering Supabase's acquisition of Turso, the Redis creator's new engine for running frontier open weights on your own machine, Anthropic's $100 million bet on enterprise AI talent, Zig 0.17.0, an AI that finally beat the best Stratego player alive, and Apple's new controls on Full Disk Access.

The morning's top developer threads include Zig 0.17.0 at 248 points, ds4 at 246, and the Supabase and Turso announcement at 203.

In today's brief:

• Supabase is acquiring Turso: the Postgres platform buys the SQLite-in-Rust company to build a database for every agent, with Turso's Glauber Costa leading the effort
• ds4: Redis creator Salvatore Sanfilippo's MIT-licensed C engine runs DeepSeek V4.1 Flash, GLM 5.x, and Qwen3.8 locally on Metal, CUDA, and ROCm
• Claude Frontier Academy: Anthropic commits $100 million to train 10,000 Frontier Deployed Engineers by the end of 2027
• Zig 0.17.0: a reworked build system, a new Build Server Protocol, and an ELF linker that makes incremental compilation work for everyone on x86_64-linux
• Also: Ataraxos beats the top Stratego player on 16 GPUs, and Apple tightens Full Disk Access, naming autonomous agents as the reason

THE BIG ONE

Supabase Buys Turso to Put a Database Behind Every Agent

Supabase is acquiring Turso, the two companies announced Friday, pairing the leading hosted Postgres platform with an SQLite rebuilt in Rust for agent-scale workloads. Supabase says it is already launching more than one million databases per week, and CEO Paul Copplestone frames the next wave plainly: "AI is enabling builders to create an immense amount of software. Today, agents are spinning up millions of databases to power the prototypes, explorations, dashboards, and apps they're building."

Turso's side of the pitch is the architecture. The company rewrote SQLite from the ground up with concurrent writes, shipped it through a diskless, WAL-on-S3 cloud, and made one server able to hold millions of databases that suspend when idle and wake on demand. That is what makes a database per agent or per task economically plausible: instant provisioning, low idle cost, and bring-your-own-cloud for teams that need it inside their own infrastructure. Customers named in the announcements include Superhuman, Sauna.ai, and Mastra.

Both products continue: Supabase stays around Postgres and Turso stays around SQLite, with Turso co-founders Glauber Costa and Pekka Enberg joining Supabase along with the team. Costa will lead the agentic infrastructure effort, and Turso's stated day-one goal was to serve a billion databases. On Hacker News, the thread (203 points, 30 comments) mixed relief with customer anxiety: several users said they had avoided Turso because its future depended on a startup, while at least one paying customer said the acquisition was the outcome they least wanted. One commenter hoped the merger brings more resources to performance work, recalling ClickBench attempts that kept surfacing bugs rather than a clean result.

Why it matters: the deal is a bet that the cheap end of the data stack should be SQLite-shaped and provisioned by an API, with Postgres as the path when workloads outgrow it. If you are building agents that need a database per task, per user, or per sandbox, that is the infrastructure question the merger is answering. Our Convex vs Supabase breakdown covers the backend trade-offs, the vector database comparison maps where retrieval fits, and our SQLite production guide covers what changes when small databases become load-bearing.

DEVELOPER TOOLS

antirez's ds4 Runs Frontier Open Weights From a Single Narrow Engine

The creator of Redis, Salvatore Sanfilippo, shipped DwarfStar 4, or ds4, an MIT-licensed C inference engine for "high-memory Mac, CUDA and ROCm machines." It is deliberately not a generic GGUF runner: it follows a short list of models (DeepSeek V4 and V4.1 Flash, GLM 5.x, and Qwen3.8 Flash Next, text and vision) and validates each supported layout end to end. One engine exposes three interfaces: ./ds4 for chat, ./ds4-server with OpenAI and Anthropic-compatible local APIs, and ./ds4-agent for persistent coding sessions.

The design choices are the story. It uses asymmetric 2-bit quantization that compresses routed experts while keeping critical shared paths precise, which is how a 284-billion-parameter mixture-of-experts model fits a workstation. The KV cache is treated "as a disk citizen": entries are keyed by the SHA1 of the rendered prompt prefix and persisted to SSD, so a matching prefix is reloaded instead of recomputed after a restart. The project on GitHub had 23,044 stars at the time of writing, and the HN thread (246 points, 23 comments) compared quantization notes, with one commenter saying the DeepSeek V4 checkpoint quantizes poorly and another saying a fused-TQ patch of theirs gets Qwen 3.8 Flash Next to a 1M-token context on a 128 GB M5 Max.

Commenters also asked why C rather than Rust, and the local-agent crowd treated the release as another step in a familiar trend: the gap between a hosted API and a machine on your desk keeps narrowing. Our best local coding LLMs roundup and local runtime comparison for coding agents cover where the current weights land.

Why it matters: ds4 is a working answer to what frontier inference looks like when it is local-first: a narrow engine, aggressive quantization of the parts of a model that tolerate it, and persistence for anything expensive to recompute. If you run agents against local weights, those are the three levers that decide whether the machine keeps up, and this codebase shows all three in about as small a surface as you will find.

PLATFORMS

Anthropic Commits $100 Million to Enterprise AI Talent

Anthropic launched the Claude Frontier Academy on Friday, a $100 million commitment to train 10,000 Frontier Deployed Engineers by the end of 2027. The first cohorts draw from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. The company's framing is that implementation talent, not model access, is now the bottleneck: "Every enterprise is racing to bring AI into its organization, but the people with the skills to make it work inside a real business have become the hardest talent to find."

The program copies the medical residency model. Partner organizations nominate engineers who each arrive with a named Claude project to lead on return, and the residency runs a multi-day in-person program plus a simulated enterprise deployment that walks from use case selection through security review to handover. Candidates finish with a graded practical; those who pass earn a Claude Resident Engineer badge and enter a 12-week residency. Anthropic's Steve Corfield describes the thesis: "A small team of high-agency people with the right skills, access to Claude, and a deep understanding of how their business runs can transform an entire company."

It is also a distribution play. The named partners are among the largest system integrators and consultancies in the world, which is the same channel Anthropic has been building for enterprise deployments. Our enterprise AI budget blowout guide covers the failure modes this training is meant to prevent, and Anthropic's platform and AWS plumbing breakdown covers where Claude lands in enterprise stacks.

Why it matters: model vendors are moving from selling tokens to certifying the people who deploy them. If your company is in a partner's pipeline, expect a Claude-credentialed engineer to show up with opinions; if you are an engineer, the badge is a new line on a CV.

SECURITY

Apple Will Add Controls to Full Disk Access, Citing Agents

Apple told developers on Friday that it will add controls around Full Disk Access on macOS, the permission that lets an app read files across the system. In a developer news post, the company says the permission "largely sidesteps" its privacy controls to keep backup apps functional, and that some developers are using it in ways that can expose "files, mail, messages, and even browsing history" without users fully understanding what they granted. The new controls, arriving "going forward," will require "very explicit user action" from anyone who wants to grant that level of access.

The line that matters for this audience is Apple's stated reason: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially." Full Disk Access is a single toggle that grants more than most agent workflows could ever justify, and desktop agents that read local data are exactly the class of software Apple is describing. The HN thread (204 points, 27 comments) split between users who saw a sensible tightening and users who saw another step in Apple's pattern of removing control from developers and users, with several pointing out that backup tools legitimately need the permission and should get a scoped replacement rather than a harder consent wall.

Why it matters: if you ship a macOS tool that reads user data, the blanket grant is being redesigned into something more surgical, and agent frameworks should assume directory-scoped access with visible consent is the interface they will be required to use. Our agent sandbox architecture guide and agent containment capability ledger cover permission boundaries that do not depend on one system toggle.

ECOSYSTEM

Zig 0.17.0 Ships a Reworked Build System and a Compiler Milestone

Zig released 0.17.0 on Friday after five months of work: 206 contributors and 925 commits. The headline is infrastructure. The build system was reworked, including a first implementation of the Build Server Protocol, and the ELF linker was improved to the point where the project says it expects incremental compilation to work for everyone on x86_64-linux. The toolchain moved to LLVM 22, glibc 2.44, musl 1.2.5, Linux 7.2 headers, and macOS 27 headers.

The language changes are a mix of tightening and deletion. @bitCast semantics changed, @backingInt and @fromBackingInt arrive, @divCeil and @SpirvType are new, and several old forms are gone: void{} syntax, i0, errdefer captures, array multiplication syntax, and the internal and link_once linkage options. The standard library adds a reworked StackFallbackAllocator, a new SafeAllocator, and a reworked std.zon.parse, part of the project's language stability work on the road to 1.0.

On Hacker News, the thread (248 points) was as much about the project as the release: disappointment that loop vectorization remains disabled after the LLVM upgrade, and a long exchange about the project's rules against AI-generated contributions after one commenter said an issue was ignored because they mentioned using LLMs to check it. Our incremental compilation internals piece digs into the linker work, and the Bun rewrite controversy covers the Zig-and-AI question from the other direction.

Why it matters: Zig's 1.0 path now runs through its own build tooling and linker rather than LLVM alone, and this release says incremental compilation on Linux is finally real for ordinary users. If you have been waiting for the toolchain to settle before trying the language, this is the release where the developer experience changes materially, even if the community argument about AI contributions is unresolved.

RESEARCH

Ataraxos Finally Beats the Best Stratego Player, on 16 GPUs

A team from Carnegie Mellon, MIT, NYU, and Stanford built Ataraxos, a Stratego AI that beat Pim Niemeijer, the four-time world champion, 15 games to one with four draws across a 20-game match. The game had held out where chess, Go, and poker fell: each side has 40 pieces whose identities stay hidden until they collide, which produces more than a decillion possible setups and games that can run 2,000 moves. DeepMind's DeepNash, introduced in 2022, could not reliably beat top humans.

The technical result is in Nature. Ataraxos trained on 163 million self-play games, and its key addition over DeepNash is a second "belief" network that guesses the opponent's hidden pieces from their movement, letting the agent search over sampled plausible arrangements instead of the full space. The economics are the other headline: 16 GPUs for a week plus four GPUs for four days to train the belief model, against DeepNash's 1,024 accelerators over two to three months, a run the researchers estimate would cost $3 million to $4.5 million at 2025 prices.

The approach transferred: the same system beat three world champions at Barrage Stratego, mastered the cooperative card game Hanabi, and beat the best bots at dou dizhu. At the 2025 Stratego World Championship, challengers went 2 for 40 against it. The team is now working on interpretability, since Ataraxos cannot yet explain its moves, and points at war-gaming as a near-term application. The HN thread (218 points, 109 comments) focused on the hidden-information search and the price tag.

Why it matters: Ataraxos is a case study in getting frontier research results from a university budget: replace brute-force enumeration with a learned model of what you do not know, then search over the samples. That pattern is directly relevant to agents that act under uncertainty, and our RL fine-tuning cost breakdown and RL environments for coding agents cover adjacent corners of the same efficiency shift.

TOOLS WORTH A LOOK

1. Claude Code v2.1.288 (free with a Claude plan) - adds a $.ui.selection() API for mods, recovery of a prompt cleared with Ctrl+C, a built-in gh api for cloud sessions without the GitHub CLI, an OAuth re-auth prompt when an MCP server asks for more scope, --max-findings on /code-review, and Ctrl+F to find a session by name.
2. Copilot model removals (free for Copilot subscribers) - GitHub deprecated Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 across every Copilot surface on October 2. Gemini 3.6 Flash is the model our August migration guide told readers to switch to, which is a good argument for reviewing pinned models whenever the changelog lands.
3. Muse Gadgets (free, open source) - open-source hardware for Muse: program an ESP32 or Raspberry Pi with the Muse SDKs and connect displays, buttons, sensors, and actuators.
4. Cloudflare OHTTP Gateway (closed beta, waitlist, paid add-on) - the decapsulation half of the IETF's Oblivious HTTP standard, for origins on Cloudflare that want to receive requests without seeing client IPs. Cloudflare also renamed Privacy Gateway to OHTTP Relay.

WHAT ELSE IS HAPPENING

• FLUX 3 Image (347 points): Black Forest Labs' image model takes a scene as bounding boxes on a 0 to 1000 grid plus one scene prompt, with an element table that maps each box to its own description. The thread asked whether this is a new model or a new interface.
• Sites in ChatGPT (274 points): OpenAI's hosted site builder drew a long thread on model-generated websites, custom domains, and what it means for the website-builder market.
• Home Assistant Cloud is now Home Assistant Link (208 points): Nabu Casa renamed its subscription service to distance it from what "cloud" has come to mean, and says most profits fund the Open Home Foundation.
• One month on GLM 5.3 Flash (163 points): Wagtail's Thibaud Colas ran September on one cheap open model and posted the receipts. The Flash half cost $68, about 4kWh and 365 grams of CO2 for the month; a wrong-model vibe-coding detour burned 450M tokens, $150, and 5kWh in one night.
• The harness is the company (134 points): Shrivu Shankar argues every SaaS business becomes a harness around a model whether it notices or not, and the thread argued about moats and whether software factories are real.
• Apple Pass Designer (428 points): a new beta macOS app for designing Wallet passes with the same rendering as iOS and watchOS, semantic tag editing, live validation, and backward-compatible pass generation. Requires macOS 27.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.