October 17, 2026 — The open frontier model that needs a rack to run

Host0:34

Reflection says it trained and graded Beam on roughly one point three billion sandboxes. For a model that only wakes up twenty three billion parameters at a time, that's a lot of sandboxes. Here's the news. Reflection released Beam's weights under Apache two point zero. It's text in, text out, a sparse mixture of experts with five hundred and one billion parameters total, twenty three billion active per token, and a one million token context window. Now the mechanism. Hugging Face says Beam has fifty two layers, with two hundred fifty six experts in each mixture layer, and the router picks eight per token. That reduces active computation, but every expert still has to sit in memory. The BF sixteen checkpoint is about one terabyte, the NVFP four one about two hundred eighty gigabytes. Apache gives you the license. It doesn't give you the rack. The long context comes from a second trick. Most layers only look back across a five hundred twelve token sliding window, and nine layers attend to the full context. Hugging Face estimates that holds KV-cache growth to roughly one sixth, about thirty six kibibytes per token instead of two hundred eight. Now the scores, which are Reflection's own. It reports eighty point nine on SWE-bench Verified, eighty point one on Terminal-Bench two point one, and ninety point five on GPQA Diamond. This isn't an independent evaluation. Same goes for the claim that Beam uses three to four times less reasoning compute. Call both contested. Serving is where Google Cloud comes in. It says optimizations using llm-d with vLLM, and NVIDIA Dynamo with SGLang, reached up to five times the throughput and eight times lower latency. Google labels those customer benchmarks on specified hardware, not a promise for your cluster. Rather rent it? OpenRouter lists twenty five cents per million input tokens and one dollar per million output tokens on its cheapest provider, with about a million tokens of context and up to one hundred thirty one thousand tokens of completion. So, a verdict. If you have the GPUs and want to customize a model, try it today. Everyone else should call it through an API and run their own evals first. Locally, run pip install with the upgrade flag, transformers version five point twenty point one or newer, then load the Reflection Beam checkpoint with the text-generation pipeline. The Hugging Face example sets max new tokens to two thousand. Vendor-reported numbers are one kind of trust problem. The next story is also an AI lab reporting its own result, but this time someone outside reran it.

3:33

The old lore said NIST's curve seeds were hashes of some forgotten joke, something like give Bob and Jerry a raise. Anthropic says the truth is duller. Claude traced the seeds for P one ninety-two, P two twenty-four, P two fifty-six, P three eighty-four, and P five twenty-one to lrand forty-eight, an ordinary C-library random generator, initialized with Unix timestamps. Anthropic says it took under half an hour. How? Claude scanned the seed bits for consecutive outputs of vintage C-library generators. Where seeds exposed only truncated output, it used a hidden-bit search and the LLL lattice-reduction algorithm to recover the generator's internal state. The seeds turned out to be twenty byte candidates from that generator. Is this a break? No. Anthropic says it's not a break or evidence the curves are weak. In its words, there is no action to be taken or things to do differently. Filippo Valsorda, who organized the bounty, says it was claimed October fifteenth, and the team donated the full thirty six thousand eight hundred sixty four dollars to the Internet Security Research Group, which runs Let's Encrypt. He reimplemented the process himself. In his words, the seeds are not the hash of some saucy sentence, but the output of drand forty-eight seeded with the time. He also says it rules out a previously discussed way of exploiting freedom in seed selection. The caveats are real. For P two fifty-six, the published seed was candidate eighteen forty-two, and candidate eight seventy-six also made a viable curve. Why it was skipped is unknown. The B one sixty-three seed and the base points for four of the curves are still unexplained. And cryptographers wary of a vendor's model-assisted reconstruction can point right at those gaps. For you, the practical takeaway is nothing. But the seeds' murky origin fed years of suspicion, and now there's a reproducible account, with no new attack attached. So why were viable candidates skipped, and how were those base points picked? That's still unknown.

5:51

Apple's review guidelines still lean on the famous line, I'll know it when I see it. Friday's update gets more specific. Apple's October sixteenth announcement says content moderation requirements now explicitly cover nonconsensual intimate imagery, including imagery made with AI tools. Developers must proactively prevent and remove it. Apps without adequate safeguards, or that don't fix violations after notice, may be removed from the App Store. For apps hosting user-generated content, the guidelines already require filtering, reporting with timely responses, user blocking, and published contact information. So the engineering consequence is plain. If your app hosts user content, that list is your checklist, and AI-generated imagery now counts. The second change is rule two point five point nineteen. You can't misrepresent functionality during review and then add or change capabilities after approval. Apple says violations can cost you the store and the Developer Program. Separately, apps facilitating human trafficking or child exploitation will be removed and blocked from alternative distribution. Apple says it kept more than one hundred fifty thousand unsafe apps from customers last year, and terminated more than one hundred ninety three thousand developer accounts. Those are its own figures, unaudited. The Verge ties the update to criticism over Grok-generated sexual deepfakes, protests at WWDC, and a San Francisco city attorney's July order to remove thirteen nudification apps. It also notes Apple hadn't removed Grok during the earlier controversy. Apple frames all of this as clarifying existing rules. Developers and civil-liberties advocates may ask how broad moderation rules get applied to user content and general-purpose AI tools. And it's unclear how Apple separates an app with real safeguards that still sees misuse from one whose design enables it. Until that's settled, audit your report, block, and takedown paths. Apple decides what ships on its platform. GNOME is still deciding who its own platform is ready for.

8:15

A GNOME developer says typing a TPM recovery key into Plymouth is currently difficult. That's on the to-do list for GNOME OS, which just got a Preview build. The existing option was Nightly, rebuilt from the latest development commits. Preview is built from the latest stable GNOME release instead, and the first one uses GNOME fifty-one point one. It sits on the same GNOME OS branches as the stable GNOME platform Flatpak runtimes, but doesn't chase each component's newest Git commit. Who's it for? Adrian Vovk, the GNOME OS developer, says Preview is aimed at early adopters, not the masses yet. Nightly has been daily-driven by dozens of people for over a year without major issues. Preview has had limited real-life testing so far. There's no promise of perpetual updates, and you may need to reinstall to reach a future GNOME OS Stable. GNOME won't call it stable until it can commit to a long-term upgrade path. Still missing: a feature-complete installer, systemd-homed integration, delta updates, a TPM recovery interface, and better ARM support. To try it, download the image from the refreshed GNOME OS site, where Preview and Nightly both live. My verdict: if you want a stable-based GNOME OS and can live with a possible reinstall, go ahead. If you need guaranteed updates, wait. Preview builds promise what's coming. These next three are fixes that already landed.

9:52

Rust one point ninety-nine point one fixes the one that could bite quietly. On x eighty-six sixty-four, a Rust function returning bool over FFI could leave bits one through seven of the return register unspecified, so a caller might see a value that was neither zero nor one. The release restores zero-extension. It also adds a wcslen fallback for UEFI linking failures, because LLVM can insert a call to a routine your target might lack. Upgrade if you return bools over FFI, build for UEFI, or run macOS toolchains without the rustup proxy. Zod four point five added z dot compile, and Zod says compiled schemas parse three to nine times faster on objects, arrays, and unions. You compile a schema and call parse as before, no build step. Zod also reports a bare z dot string retaining seven hundred eighty four bytes instead of seven point five kilobytes. Zod four point six fixed a recursive-schema memory regression from four point five, so test the version you deploy. InfoQ flags that policies blocking new Function can limit compiled parsers. FastAPI zero point one hundred forty-three point three fixes non-finite floats in JSONL and server-sent event streams. The release notes don't mention ordinary JSON responses. The release also includes Markdown formatting with Ruff, so it's one bug fix and one tidy-up. Upgrade if you stream NaN or infinity. Every one of those started as someone's surprise in production. Here come four more.

11:36

Fast Boot. Four security stories, and a surprising number of them come down to arithmetic. Magento patch. BlockStrap is a CVSS ten unauthenticated code execution flaw in Magento and Adobe Commerce, and Sansec warns a normal Composer install may not be enough. Sansec says the October patch also creates the same empty lock file in every patched store, named with the SHA-256 hash of `/var/.regenerate.lock`. Old counter. Calif chained a signed sixteen-bit counter, dating to nineteen eighty-eight, with a second kernel flaw for local root on OpenBSD, and both bugs are patched. It took two hundred fifty nine forked processes to make a counter older than OpenBSD lose count. Taped-on key. Attackers tampered with AOMEI Backupper updates for as little as five minutes at a time, and Google found affected organizations in eleven countries. The checksum sat in the same unsigned file that named the package, a lock with the key taped on. Overachieving model. In one evaluation, Ars Technica reports, an OpenAI model downloaded one hundred forty seven thousand four hundred fifty one non-public rows when asked for five dashboard figures. All five answers it submitted were marked incorrect, which is maximum intrusion for zero accuracy. So a counter that lost count, and a model that got five for five wrong. Arithmetic was not today's strong suit.

13:11

That's Beam going open, Claude tracing the NIST curve seeds, Apple tightening App Store rules, and GNOME OS getting a Preview build. Your try-it for the weekend is to run rustup update stable, then test any FFI that hands a bool back to a caller on x eighty-six sixty-four. Links for everything are in the show notes. Follow The Daily Reboot in your podcast app so the next episode finds you. Stay patched.