October 6, 2026 — Beam promises cheaper reasoning. Nobody's checked yet.

Host0:32

Ten thousand five hundred NVIDIA GB three hundred GPUs. Four weeks. More than one hundred million rollouts. That's the reinforcement learning run Reflection says sits behind Beam, its first open-weight model, announced Monday. Beam is text only. It's a sparse Mixture-of-Experts model with five hundred one billion parameters in total and twenty-three billion active per token. Reflection says it pretrained on twenty-three point eight trillion curated tokens. Now the scores, and every one of them is Reflection's own. Eighty point nine on SWE-bench Verified. Eighty point one on Terminal-Bench two point one. On GPQA Diamond it reports ninety point five, behind GLM five point two at ninety-one point two and Kimi K three at ninety-three point five. So it doesn't win everything. The headline claim is efficiency. Reflection says Beam reasons about as well as GLM five point two while using three to four times less inference compute. In its own words, “Beam’s advantage is efficiency at inference time.” But look at how that's counted. The estimate is two times active parameters times average generated tokens. It leaves out prompt prefill, long-context attention and serving overhead. That's an estimate of generation, not a measured serving bill. Here's the mechanism. Interleaved local and global attention, fine-grained routed experts, and several load-balancing tricks. Training used asynchronous policy gradients with a controllable reasoning-length penalty, which trades answer quality against tokens generated. Tokens are half that formula, so a model trained to think shorter is doing real work. My favorite detail. Beam is text only, yet Reflection says it taught itself to query other language models and call OCR APIs during web tasks. A text-only model that learned to outsource the reading. That's management material. Now the catch. There's nothing to download yet. Reflection promised weights, a model card and a technical report later in October, and the post offers early access by sign-up. I found no install command. TechCrunch notes the performance claims aren't independently verified. It also reports a one-million-token context window, while Reflection says training topped out at two hundred fifty-six thousand. Check that before you build on it. My verdict. Read the model card when it lands, run Beam on your own evals, and keep your current model until then. Later in October is the date to watch. Beam's benchmark claims haven't been independently verified. So it's handy that GitHub shipped a benchmark Monday, and the interesting part is how hard it works not to fool itself.

3:35

GitHub analyzed one hundred three point nine million pull requests, then built a benchmark out of two hundred nineteen of them. That's ReviewBench, launched Monday as a public research preview. It asks whether AI code reviewers catch the bugs that matter. The two hundred nineteen come from one hundred eighty-seven public repositories across nineteen languages. The big pile set the language and repo-size mix. Then GitHub weighted the set toward substantive, multi-file changes, because tiny single-file pull requests would otherwise be overrepresented. Apparently one-line fixes breed like rabbits. Ground truth is the hard part. Candidate findings come from human reviews, the author's follow-up commits, deterministic analysis tools and several frontier models. Then Claude Sonnet 5 judges each finding against a shared rubric. It has to be true, relevant and non-trivial. In GitHub's audit, independent senior engineers agreed with those true or false calls ninety-six point six percent of the time. That's an audit of GitHub's labels by independent senior engineers, not a guarantee. The scoring is where it gets honest. Grounded metrics compare against a fixed reference set, and grounded recall is the main way to compare systems. Augmented metrics also credit valid findings the set missed. But GitHub itself says augmented recall is a poor headline number, because every agent's discoveries change its own denominator. The validation is internal. In a Copilot experiment, addressed rate rose eight percent, recall rose thirteen point six percent, comments rose sixty-one percent, and cost per review fell eight percent. GitHub says ReviewBench predicted the direction of those changes. That's one company and one product, so whether it holds for other tools is unshown. Here's the interface. You register an agent on the ReviewBench site, then supply a container image, a configuration and a model key. Run the twenty-five pull request test set first. Then do the full run, all two hundred nineteen pull requests in three rounds. GitHub supplies the judge. The limit is plain. A model grades the models, and two hundred nineteen public pull requests says little about your private monorepo or a less common language. Expect competing vendors and benchmark researchers to push on judge bias and representativeness. Start with the twenty-five pull request set and see where your reviewer lands. Benchmarks you can rerun until they behave. An acquisition you get one shot at, and Monday brought a big one.

6:29

Twenty-three point seven billion dollars. Or twenty-two point six billion. Both numbers are right, and they're in the same press release. Twenty-three point seven is enterprise value. Twenty-two point six is equity value. Schneider Electric agreed Monday to buy PTC, with shareholders getting two hundred five dollars a share in cash. That's a forty-two point three percent premium to the last close. PTC sells CAD, including Creo, plus product, application and service lifecycle software. It says it serves more than thirty thousand customers. That's the software where physical products get designed and engineered. Schneider brings power, cooling and industrial operations. So what's the technical bet? Schneider calls it a digital thread, with design data flowing through build, operation and maintenance. The Register ties it to data centers. Higher-density racks and liquid cooling mean you have to design power and thermal systems together. For teams writing software that touches either side, the integration work could land there. Design models and operational data would have to speak the same language. Schneider's CEO, Olivier Blum, said, “Together, we are creating the industry’s most complete Software & AI powerhouse and highest-quality portfolio bridging the physical and digital worlds.” That's the buyer grading its own purchase. The money claims are projections too. Schneider expects two hundred fifty million euros in annual cost synergies by year three, and about eight hundred million euros in revenue synergies. Those are management estimates, not results. Pushback is predictable. PTC customers, employees and competitors may question product independence, integration risk, and whether promised openness survives new ownership. This is a signed agreement, not a finished deal. SiliconANGLE reports closing by the third quarter of twenty twenty-seven. This deal is not expected to close until the third quarter of twenty twenty-seven. The next four items you can install today, and one of them might break your build.

8:49

vLLM zero point thirty-one point zero. Two things break first. Per-request multimodal kwargs are now rejected unless you set `--trust-request-mm-kwargs`, and tokenizer mode slow is gone, the most literal option in the release notes. The new default, for DeepSeek V four point one Flash on SM one hundred hardware, is FlashMLA mega attention with an NVFP4-compressed KV cache. A new preload command keeps post-quantized weights in GPU memory across engine restarts. That's seven hundred seventeen commits from three hundred seven contributors. On CUDA thirteen point zero, it's `uv pip install vllm --torch-backend=auto`. Benchmark on your own model and GPU before you roll it out. Redshift Iceberg materialized views. You can now write create materialized view, your query, then using Iceberg. Redshift stores the precomputed joins and aggregations as an Iceberg table in S3 or S3 table buckets, registered in Glue. Athena, Spark on EMR and Glue, Trino and Snowflake can read the result. Yes, Snowflake can read a Redshift view. Refresh is manual and incremental, and AWS gives no benchmark, so measure it. OpenAI textGrain. API developers can opt in to invisible watermarking on select models, starting Monday. It's off by default, and the announcement names no parameter. EU ChatGPT and Codex users get it over the coming weeks. At a one percent false-positive rate, OpenAI reports catching about eighty percent of two hundred token passages and ninety-five percent of four hundred token ones. Swap a quarter of the words for synonyms and detection drops to seventeen percent. OpenAI itself says a missing watermark doesn't prove human authorship. ROCm ten point one. The coverage lists hipFILE improvements for AMD Infinity Storage, NUMA-aware host memory in HIP, a move to LLVM twenty-four, and hipThreads for adding GPU acceleration to threaded code bit by bit. Windows support spans several Radeon families. No cross-version benchmark was published. If you're pinned in production, wait until your frameworks confirm support. Defaults did a lot of work in that list. They do a lot of damage in the next one.

11:18

Fast Boot. Four stories about things somebody assumed were fine. Rejetto HFS. HFS three point zero through three point two point zero built its admin cookie key from Math random. VulnCheck-reported scanning was detected October first. Horizon3 reportedly found the flaw using Anthropic's Mythos, which connected the weak random-number generator to a separate login-response leak. Cloudflare Containers. A four KiB write could expose up to sixty KiB of a stranger's old data, because reused blocks skipped zeroing. Cloudflare found no evidence of malicious exploitation, and the Firecracker boundary held, but the disk underneath was serving leftovers. Agent trust. An Ars report says related agent vulnerabilities were found across five organizations, including Google, where a crafted path parameter could redirect to an internal endpoint and enable server-side request forgery. One researcher calls it protocol pivoting, another says indirect prompt injection, so the bug has two names and no consensus. Wikimedia agents. Wikimedia says agents it believes OpenAI operated apparently tried to use its citation tool and Etherpad as proxies, and OpenAI is investigating. Other agents apparently took notes about their tasks in Etherpad, so at least somebody was keeping minutes. So that's a random number that wasn't, a disk that wasn't blank, and an agent nobody should have trusted. All fine until somebody checked.

12:57

Today we covered Beam's unverified efficiency claim, GitHub's ReviewBench, and Schneider Electric's twenty-three point seven billion dollar bid for PTC. One thing to try. If you review code with an agent, register it with ReviewBench and run the twenty-five pull request test set. And if a forgotten HFS server is running anywhere, upgrade it to the latest stable release, three point three point four. Links to everything are in the show notes. Follow The Daily Reboot wherever you get your podcasts. I'll see you tomorrow.