DeepSeek Flash Gets Eyes, a 5-Euro DNS Takeover, and the End of Slow

Good morning. It's Saturday, August 22, and we're covering DeepSeek's first vision model, a researcher who controlled three territories' phone routing for the price of a coffee, and why one performance engineer says software no longer has an excuse for being slow. Plus: a Rust language server that lives under 100MB of RAM, and a developer who finds his own brain skipping AI-written documents.

The DeepSeek vision thread hit 482 points before lunch, and the ENUM write-up quietly logged roughly 400,000 phone-network queries. Here is the signal, sourced.

In today's brief:

• DeepSeek V4 Flash Vision is live: one model id, three API shapes, every image capped at 384 tokens
• An expired 5-euro domain handed over the phone routing of three territories - including military bases
• Dan Luu: workload-fitted optimization via coding agents means "there's no reason for software to be slow anymore"
• A new Rust LSP targets 100MB of RAM with a disk-backed index that survives editor restarts

THE BIG ONE

DeepSeek Gives V4 Flash Eyes With an Experimental Vision Model

DeepSeek shipped deepseek-v4-flash-vision-exp, the first model in the V4 Flash line that accepts images alongside text: describe pictures, read text from screenshots, analyze charts. The release note is explicit about what this is and what it is not. It is not a new flagship. The model id carries the -exp experimental marker, and the text side - agents, reasoning, world knowledge - is claimed to match the existing V4 Flash 0731 build, with a "major leap" on multimodal agent benchmarks versus text-only Flash. The HN thread (482 points) filled in within hours, mostly with the same question: how much does an image cost?

The answer is in the contract. Every image is resized toward an 800x800 pixel budget before inference, so a 2000x2000 screenshot and a 5000x5000 render burn the same number of tokens - an upper bound of 384 tokens per image, counted independently up to 600 images per request. Formats are JPEG, PNG, GIF, and WebP, detected from file content rather than the declared MIME type. Three input methods work across the OpenAI-compatible Chat Completions and Responses APIs, plus the Anthropic-compatible /messages endpoint: base64 inline (request body capped at 48 MiB), an external URL (max 8192 characters, 32 MiB, download must finish in 60 seconds), or a Files API file_id (which lifts the single image cap to 64 MiB). Images are only accepted in user messages - anything else gets a 400 - and non-vision models reject them with "This model does not support image".

What this means for agent work is the interesting part. Screenshot-in-the-loop tasks - verify the UI rendered, read the error dialog, check the graph - previously meant bolting a second vision model onto the main model you were running. Flash Vision does it in one call, with a hard, predictable token cap you can bill against. Our OpenCode setup guide walks the full API contract plus the config to run it today, and the migration notes cover where V4 Flash sits in the rest of the family.

Why it matters: cheap, predictable vision input turns screenshot-vision loops from a specialty integration into a default capability for agentic tools - and an explicit 384-token image cap is the kind of pricing discipline that makes model choice arithmetic again.

SECURITY

A 5-Euro Expired Domain Owned Three Territories' Phone Routing

DNS researcher lina published the kind of post that makes you check your own registrars: how an expired nameserver let her take over the e164.arpa zones of three countries (555 points on HN). ENUM, a mostly-dead system from the early 2000s, stores phone numbers as reversed DNS names so carriers can look up how to route a call over VoIP. Three country zones - +290 Saint Helena, +246 the British Indian Ocean Territory (Diego Garcia), and +247 Ascension Island - were delegated to two nameservers: ns6.icb.co.uk, which no longer resolves, and ns.enum.org.uk, whose domain had expired. She bought it for 5 euros. That made her authoritative for every ENUM query in three territories: a carrier asking "where do I route this call" would accept whatever she answered, and a malicious operator could have sat silently in the middle of any call placed through those numbers.

What happened next is the quietly terrifying part. Two reports to the British government went nowhere, and RIPE declined to touch the delegation because an ITU-T committee governs it. So she set up logging. A day of logs on the Saint Helena zone showed nothing, so she kept the domains, hosted a site on one, and moved on. Six months later she checked all three zones: 100,170 queries to the Diego Garcia zone, 99,902 to Ascension, 9,133 to Saint Helena - roughly half the real traffic, she estimates, or about 400,000 lookups total, almost all from American resolvers, for calls to military bases. Each log line was a full phone number, a timestamp, and the resolver's IP. A state actor interested in what happens on Diego Garcia - the subject of an Iranian missile strike on March 20, 2026 - would have loved that vantage point.

She deleted the logs, reported again, and this time the NCSC cared. The domain was transferred to the NCSC in March 2026, so the delegation finally sits with the government that should have owned it. Total cost of the episode: two 5-euro renewals. The line HN keeps quoting: any DNS delegation whose renewal depends on the memory of whoever set it up a decade ago is a 5-euro acquisition for anyone who checks.

Why it matters: phone routing, like so much of the internet's plumbing, still trusts decades-old DNS delegations with no owner in the directory; the lesson for every team is that expired domains and dead nameservers are not maintenance debt, they are control-plane holes.

PERFORMANCE

Dan Luu: "There's No Reason for Software to Be Slow Anymore"

Dan Luu's latest essay (417 points on HN, already 302 comments deep) argues that the cost of specialized performance work has dropped so far that the default engineering calculus flipped: "work that used to require a person or team with a rare set of skills can be done by anyone who can type a few sentences." The evidence is a string of fast, agent-driven experiments. His FRE regex engine, built by an agent loop optimizing against a benchmark suite, got a native AOT-compiled arm: on a handful of simple long queries the native path is 2x-4x faster than the interpreter path, and on representative holdout queries where AOT triggers, about 7%. A workload-specific optimization pass - matching the engine to his own ripgrep queries - was already 2% ahead of stock ripgrep on the holdout and "still getting faster" as he wrote.

The broader claim is about what becomes worth doing. Jamie Brandon, after solving Anthropic's performance takehome, had a model continue from where he left off and beat him; the optimizations were "crazy shit that I would never try unless I was working on this for weeks." pgrust's Michael Malis notes the JIT compiler is the canonical example of software whose difficulty historically outweighed its benefit - now it is a weekend project. And Marc Brooker's phrase for where this lands: "dynamic custom software, fitted to a particular workload rather than a class of workloads, seems like a very likely outcome."

Luu is careful about the failure modes: current models are genuinely bad at experimental design, so an agent optimization loop still needs a human - or a skill - to set up the benchmark and the holdout. Overfitting is the standing risk; his own agent overfit to the rebar suite until told a holdout existed. The appendix is worth one too: p99 ripgrep queries take almost a minute, p999 nearly ten, and the month-long ceiling approaches two hours - and the p50 search pattern is 55 characters, because the queries doing the searching are increasingly written by agents, for agents.

Why it matters: when optimization was expensive, you optimized classes of workloads; when it costs minutes, you optimize your own workload - and performance engineering becomes a routine, measurable feature of how teams use coding agents, not a specialist art.

TOOLING

Rust Glancer: A Rust LSP That Targets 100MB of RAM

rust-glancer (228 points on HN) is an alternative Rust language server, four months old, built around one target: under 100MB of RAM for reasonable projects. The complaint with rust-analyzer is familiar to anyone opening a big workspace on a small laptop: two editor panes with a handful of projects each ran to 16GB of memory, and every restart meant re-indexing. Rust Glancer's design answer is to give up incremental analysis. It indexes once, saves the frozen analysis results to the filesystem, and reloads only what a query needs. Because the index survives on disk, restarting the editor skips re-indexing entirely - that demo is the centerpiece of the announcement.

The performance tradeoffs are stated plainly. Typing does not trigger full analysis: the server runs a shallow pass over the current body against the previous complete index, so new items - imports, structures, traits - only get indexed when you save. That is a real behavior change users feel, and the author is open about it ("you really get used to it fast"). The index numbers still land well: on an M1 MacBook with 8GB, base indexing takes 6 seconds and full indexing 9, versus 7 and 14 for rust-analyzer - and on an M4 Max, 5 and 8 versus 6 and 13. It carries full type inference and a Chalk trait solver, macro expansion out of the box, goto-definition, hover, inlay hints, completions, and a custom file watcher tuned for out-of-editor changes - because agent-driven edits, he found, are what actually thrash inlay hints in rust-analyzer. What it will not do: execute build scripts or proc macros, and it has no plans to.

The project is worth reading about for a second reason: it is aggressively LLM-built, with PRs of 10,000+ diff lines, and the author's note about it is the most quotable thing on the thread - "do not reduce me to a clanker. It is my code, so if you consider it to be slop, call it my slop, not AI." He has been running it as his daily driver for about six weeks. rust-analyzer remains the right default for max completeness; Rust Glancer is aimed at old hardware, CI-hostile laptops, and anyone whose memory budget is spoken for.

Why it matters: language servers are now a real line item in laptop and CI RAM, and a disk-backed index that survives restarts reframes the restart tax - the "LSP memory" axis just became a competitive one.

CRAFT

One Developer's Brain Is Learning to Skip AI-Written Documents

Rafal Cymerys's short essay, I'm becoming AI-blind (374 points on HN, 367 comments), describes a failure mode a lot of teams will recognize: documents that show "a strong trace to AI" get skimmed past, or not read at all. A design doc that looks like a copy-paste from Claude, carrying Claude-specific lingo ("This cuts just through it", "The first gate is real"). A 20-page marketing deck that mixes reasonable strategy with nonsense architecture ("The Redis backbone redefines the product"). A requirements doc so verbose it reads like someone's uncertain internal reasoning. His brain, trained on a year of AI-generated LinkedIn posts, emails, and websites, now spots the low-effort patterns and skips the text entirely - banner blindness, but for prose.

The kicker is the irony he names directly: "the same AI that was supposed to make me more productive, is what's now slowing me down in an unexpected way." Writing that reads like a template - everything pitched as a breakthrough, every checkbox described as if you just invented fire - trains the reader's filter against it. The HN thread lands on the practical version: this is why "vibe-coded" documentation and AI-slop emails erode trust faster than they build anything, and why the documents that actually get read are the ones carrying specifics a model would not invent. There is a tooling angle too: the nobuzz skill below exists precisely because the default LLM register is starting to tax the attention of the people reading its output. Our AI slop detection roundup compares the current detectors, and the design-slop post covers the same failure in visuals.

Why it matters: low-effort AI writing is training readers' filters in real time, so the text that gets attention is the text with specifics, verifiable numbers, and a voice - the bar for defensible prose just went up everywhere it is generated.

TOOLS WORTH A LOOK

1. Claudette / nobuzz (OSS, 275 points) - a Claude Code skill that routes Claude's last reply through Google's Antigravity CLI and translates it from TED-talk register into plain English, with colleague / manager / director modes. The name, the README insists, is absolutely not a reference to anything litigious.
2. llm 0.32.1 and llm-openrouter 0.7 (OSS) - Simon Willison's CLI toolchain and its OpenRouter plugin both updated this week; the release notes document the changes for the llm ecosystem's most-used plugin.
3. Autolith (free / open, 65 points) - "a programming agent with a live runtime": edits apply to the running program, so agent changes can be observed immediately rather than after a rebuild.
4. Sub-50ms text-to-speech on Qwen3 (146 points) - a writeup of how a TTS model was made to respond in under 50ms, with the cost math that makes latency a pricing decision.

WHAT ELSE IS HAPPENING

• Omacom Foundation launches with $8M (173 points): DHH incorporates a nonprofit to fund Omarchy's malleable-computer mission, each of eight founding patrons - Tobi Lütke, Patrick Collison, Michael Dell, Jack Dorsey, Matthew Prince, Brendan Iribe, Jason Fried, and DHH himself - contributing $1M to hold trademarks and support the open-source projects Omarchy depends on.
• A week of using Codex more than Claude (91 points): a Rubyist's field notes - Codex writes less code with fewer comments and simpler architecture, but "will not overdo it" and needs explicit instructions; one rebase against main grew into a 4,000-addition PR. Our June head-to-head frames the stack-level comparison.
• OTel isn't going well (123 points): maintainer-concentration data across the OpenTelemetry SDKs - a single human merges 86% of opentelemetry-cpp PRs and 79% of the Ruby SDK's, versus 14% across Prometheus - plus a proposed time-bound beta tier to unstick the experimental-to-stable pipeline (14 min read).
• We rebuilt the Linux MicroVM stack on Apple Silicon (144 points): encore.dev documents running the Firecracker-style microVM architecture natively on Apple hardware for local serverless development.
• Stop Making TUIs (142 points, 209 comments): Thomas Ptacek argues the terminal-app default is the wrong shape for the new wave of developer tools; Simon Willison's notes summarize where the debate lands.
• Emacs 31.1 lands August 24 (85 points): the next Emacs release date is set, with the changelog already in the tree for the curious.

---

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.