October 25, 2026 — The Voice Model You Can't Download Yet

Host0:31

Audio in, audio out. That's the Dessa team's own summary of the model Block's research lab announced in October, and they mean it literally. There's no transcript sitting in the middle. Dessa is full duplex. It listens and speaks at the same time, so it can interrupt you, or drop in a little backchannel noise while you talk. Block says it learned that timing, along with laughter, pauses and sarcasm, straight from human conversation. Here's the engineering bet. Block contrasts this with speech systems that route speech through text before generating speech. Block's argument is that text can throw things away. One spoken word takes thousands of times more bytes than the written word, and say a sentence twice and the audio is never identical. Tone and timing may not survive the trip. So Block trained Dessa from scratch on human conversation. No text, no pretrained models, one NVIDIA GB two hundred NVL72. That's Block's description, and nobody I read has verified it. Now the spec card, such as it is. What do you send it? Nothing, yet. There's no public API, no model ID, no price, no context window, no output limit. Open weights, licence, Bedrock, Vertex, OpenRouter, all unknown. Block is seeking a small group of AI safety researchers for early access. TechCrunch reported there's no public access. The speed claim is this. Block says inference runs faster than real time, locally, on a MacBook Pro M four Max or better. That's the company talking. There's no named speech benchmark, no comparison with a rival, and no independent test in anything I read. The demos are live recordings with subtitles added afterward. Unite.AI largely summarizes Block's own announcement, and quotes Dessa lead Ragavan Thurairatnam calling it a GPT three moment for speech. The demos are charming, though. Block says the model can laugh mid-sentence, and gets louder when you ask it to speak up. It also jokes that language models must learn C plus plus and whatever the hell Angular is. Expect pushback. Speech researchers will want reproducible benchmarks before they accept Block's world's first claim, which Block itself qualifies. And safety teams will say that sounding emotional isn't the same as understanding emotion. The verdict. If you research conversational audio or safety, ask for access. If you ship voice features, there's nothing to call, so don't rebuild around a demo. Watch for a public interface, a licence, and what hardware it really needs.

3:24

Dessa is a model you can only read about. This next one is a prompt collection on GitHub, and the catch is how carefully you use it. REA-Prompts, from JoasASantos on GitHub, is a collection of prompts for pointing a coding agent at software you want to understand. It's not the reverse-engineering engine itself. The README sends you to the REA toolset for that, then gives you the workflows. You hand your agent a template and ask it to produce a trace with evidence. The templates tell it to follow a feature through entry points, functions, data flow and observable effects. They ask it to cite files, functions, addresses or snippets. The sample first exercise is wonderfully mundane. Trace a CSV export button all the way to the file write, including where notes dot csv gets created and which fields it holds. The README says don't run the target in that first exercise, and start with static analysis for native binaries. There's a clean-room clone workflow too. Document behavior and requirements without copying proprietary source, assets, branding or licensing mechanisms. The README says every prompt separates confirmed facts from hypotheses, and calls that non-negotiable. The honest limit has two parts. It's unclear how well these work across different agents. And the scope includes dual-use work. The README warns you to analyze only software you own or are authorized to test, and says bypassing protections may violate laws or terms. To try it, run npx, then rea-agents at latest, then setup. Review the installation plan and the configuration changes before you approve anything. Replace the placeholders in angle brackets, and keep each investigation to one feature.

5:24

A hundred terabytes. That's how much RAM Cloudflare says it reclaimed across its global fleet. It didn't buy anything or rewrite the service. It changed some math. The service is Pingora Backend Router, Cloudflare's internal load balancer. For cacheable requests it uses consistent hashing, through an open source library called pingora-ketama. Consistent hashing means adding or removing a server doesn't reshuffle everything. Each server gets many points on a ring, and a request lands on a point. The same request keeps reaching the same server, so caches stay warm. The cost is memory. Weighted servers and feature combinations created many rings, and every ring stores lots of points. Cloudflare's fix came in two halves. First, probability analysis to find how few hash points still met its accuracy requirement. InfoQ reports hashes per backend dropped about ninety percent. The key finding, in the engineers' words, is that each step down in error margin needs almost an order of magnitude more hashes per server. The final ninety thousand hashes improved distribution error by only zero point seven percent. Second, each stored point shrank from eight bytes to six, for the hash and the server index. The hard part was the rollout. Change the ring and you move cacheable requests. Flip everything at once and that could cause cache churn and a spike in origin traffic. Cloudflare's engineers warned that could turn a memory optimization into an apocalyptic increase in origin traffic. So the team ran the new ring alongside the old one and shifted gradually. Now the caveats. The hundred terabytes is Cloudflare's own fleet-wide figure, and the sources I read include no independent replication. It also sits on top of a separate hundred terabytes the DNS team reclaimed earlier. If your cluster is smaller or weighted differently, don't promise your boss the same savings. The part you can steal is the method. Measure what each extra hash point actually buys before you pay for it in RAM.

7:44

Cloudflare spent its effort on being careful. Ollama's v0.40.11 release spends it on being loud. Ollama zero point forty point eleven. Your Mac can now read answers aloud. On macOS, ollama run plays a text prompt or a chat model's reply as the audio is generated. Try ollama run phonon, then a line of text. For apps, audio streams through the OpenAI-compatible audio speech endpoint. The same release adds the d three decision models, in sizes from two B to twenty-seven B, for yes or no answers and scoring options. Voices, languages and formats aren't documented in what I read, and playback is macOS only. Codex CLI zero point one sixty-four point zero. You'll notice ten numbered presets slots for saved prompts and slash commands, with a keyboard menu. Model and permissions commands now take inline arguments, so you can change model, reasoning effort, speed and permissions without a settings menu. Fullscreen image previews arrive too, but only in terminals that speak Kitty or Sixel graphics. Your terminal, apparently, is now a dependency. Upgrade if you swap prompts all day. Developers Digest notes presets don't replace a full configuration system. Rust one point ninety-five. If let guards on match arms are now stable, so a pattern bound in the guard is available in the arm. They build on let chains from Rust one point eighty-eight. PowerPC inline assembly is stable too, and LLVM moves to twenty two. The catch is that the compiler doesn't count those guard patterns toward exhaustiveness. Custom JSON target specs now sit behind nightly options, which shouldn't touch a fully stable toolchain. One date note. Rust's own blog dates this release April sixteenth, so it isn't Saturday's news. Run rustup update stable. Roku's repositories, now on GitLab. This is cleanup, not a release. Roku finished the move on October fifteenth. Repo names are the same, and the rokudev path now lives on GitLab. For an existing clone, run git remote set-url origin with the new GitLab address. If you use submodules, update your gitmodules file, then run git submodule sync. Roku warns that a stale index doesn't raise an error. It quietly serves outdated content, so check your raw-file URLs, CI jobs and retrieval pipelines.

10:22

Everything so far runs after your system is up. This next one runs before your GPU driver wakes up. gpu-unlock, version one point zero point zero. It's a Linux kernel module that claims to unlock FP and BF sixteen issue rates on supported Nvidia three thousand series cards. The listed chips are GA one oh two, one oh three, one oh four and one oh six. The README says it doubles the speed of those instructions feeding FP thirty-two operations. It loads from the initramfs, sets some selector values, and resets the card before the unmodified Nvidia driver binds. It does nothing unless you opt in on the kernel command line. The number to know is fourteen to thirty-six percent faster LLM prefill, from one user report in the README. Memory-bandwidth-limited decode and INT8 work stay about the same. The README itself says the reports aren't directly comparable, and nothing here is independently verified. If you run local models on these cards, prefill and diffusion are the targets. It's an early-boot module, so keep a recovery path.

11:40

Fast Boot, weekend edition. Two items, zero warm-up. BlackX bootkit. Kaspersky says it bypassed Secure Boot on fully patched machines using a twenty twenty-two flaw, because the revocation fix needed manual activation. A process trying to open its protected files could be handed a handle to Windows Calculator. Ballot trouble. Robert Seacord reports C plus plus twenty-six was disapproved, with twelve of twenty P-members in favor, short of two thirds. Commenters say the vote counts don't quite line up: twelve of twenty P-members voted in favor, while nine of twenty-two member bodies voted against. So one bootkit hid behind Calculator, and this ballot has everyone reaching for one. Check your sums either way.

12:31

Today we covered Block's Dessa, a speech model trained without text and with no public door, REA-Prompts for tracing software with an agent, and Cloudflare's hundred terabytes of RAM, found by counting fewer hashes. Try this one today. On a Mac, run ollama run phonon with a line of text and hear your laptop talk back. Links for every story are in the show notes. Follow The Daily Reboot wherever you get your podcasts. Back tomorrow.