October 5, 2026 — Sovereignty-claimed model, self-graded homework

Host0:42

Aleph Alpha had a filter problem. A common data filter throws out unusually long words. Applied to German administrative prose, that filter would have thrown out the prose. So the team adjusted it. A long-word filter on German bureaucracy would have deleted the bureaucracy. That detail comes from the announcement of Kolibri, made Saturday, October third, Germany's Unity Day. It's an English and German open-weight model. The weights are on Hugging Face under Apache two point oh. Here's the spec. Seventy eight point one billion parameters in total, three point four six billion active per token. It's a mixture of experts. Fifty layers, three hundred eighty four routed experts in each, six picked per token, plus one shared expert. The trade-off is memory. Only a sliver of the model works on each token, but the whole thing has to be loaded. The FP8 weights take about seventy eight gigabytes. Listed minimums are two A100 eighty gigabyte cards, two H100s, one H200, or one B200 or B300. Context tops out at just over a million tokens. The model card recommends staying under two hundred sixty two thousand. Most attention layers only look at a five hundred twelve token window, with one full-context layer for every four windowed ones. Aleph Alpha says that bounds compute and memory as context grows. For you as the caller, you serve it with vLLM through Aleph Alpha's plugin. You get an OpenAI-compatible API with tool calling, and four reasoning levels: none, low, medium, and high. The predecessor, Kolibri Origin, was never released. Native context goes from about sixty five thousand tokens to two hundred sixty two thousand, and there's a new bilingual tokenizer. Now the numbers everyone will quote. Aleph Alpha's own table has Kolibri at ninety six point nine on AIME twenty twenty-five. Qwen three point six thirty-five B A three B scores eighty four point six. On GPQA Diamond, it's eighty four point three against eighty three point four. Those are the company's numbers, from its own harnesses. It says prompts and settings were identical. Trending Topics reports no independent Artificial Analysis ranking yet, and that the rivals were spring releases, not newer open-weight models. The claims about sovereignty, and about sitting on the quality-versus-cost frontier, are also the company's assertions. Hosted pricing and API model IDs weren't published in anything we read. The model card says it's for human-reviewed systems, not autonomous ones. Safeguards in the model don't replace yours. The verdict. Try it if you need to self-host for German and English documents, long-context retrieval, or agents with a human in the loop, and you have the GPU memory. Wait if you need independent evals or a hosted endpoint. To try it, run pip install aleph-alpha-inference, version one point oh or later. Then run vllm serve on the Kolibri one model with the KV cache dtype set to fp8. Longer contexts need extra arguments. Kolibri's model card insists on a human in the loop. Our next project is built for the opposite instinct, an agent that drives the browser for you.

4:24

Every AI agent is a model and a browser, says the Dots README. You swap the model with one flag. The browser is what the website sees. Dots bets on that second half. It's an MIT-licensed browser agent in the feder-cr repo on GitHub. You give it a task in plain language through a local web interface, and it drives its own browser. It can use models on OpenRouter. It can also expose that browser to MCP clients, like Claude Code, Codex, or Gemini CLI. The mechanism is the interesting part. The browser is a real Firefox engine, patched in C++. Fingerprint decisions happen inside the engine, not painted over with JavaScript the page can see. Two flags matter. The seed flag gives you the same fingerprint identity every run. The profile-dir flag keeps logins and cookies between runs. The agent also uses pointer movement and individual key presses meant to produce trusted browser events. My favorite detail is that screen, fonts, GPU, timezone, and language are supposed to agree with each other. One seed reproduces the whole persona. It's a browser with a consistent alibi. It had two point six thousand stars and four hundred thirty seven forks when we read it. That's a snapshot, not a spec. It also says it isn't affiliated with OpenAI, despite the name. Now the limit. The README's claim that it doesn't get blocked is marketing, and nothing we read tested it independently. Website operators and security teams may object to stealth-oriented automation. Think about site rules before you point it at anything real. To try it, you need uv and an OpenRouter key. Run uvx, pointed at the feder-cr dots repo on GitHub, then dots, with the openrouter-key flag. Then open localhost, port eighty seven sixty five. Dots argues agents fail at the browser, before the model gets a say. Google's problem is the other side of automation, what happens when machine-made input arrives in bulk.

6:42

Google's bug bounty reward table lists a payout of three thousand one hundred thirty three dollars and seventy cents. Another is thirteen thousand three hundred thirty seven dollars. Those look like a wink at the audience. As of October first, the table is partly academic. Google's Open Source Software Vulnerability Reward Program stopped accepting product-vulnerability submissions. Google says it will provide an update in the first quarter of twenty twenty-seven. Reports received before October first aren't affected. Supply-chain reports are still accepted. Some Google Cloud issues may still go through the Cloud VRP. Google's reason, in a statement quoted by TechCrunch, is a significant rise in automated submissions, the vast majority of which are not valid. Tom's Hardware reports thousands of reports it characterizes as invalid or unexploitable. That volume figure is the outlet's, not Google's. The link to AI is Google's explanation, and nobody measured submission quality in the pages we read. The rules already asked for evidence. Reports on third-party dependencies must show impact in Google's projects. They must come at least thirty days after the upstream fix. For some flagship projects, memory-corruption reports need exact OSS-Fuzz reproduction steps or a merged patch. Standard and low-priority tiers get no money for product vulnerabilities at all. Here's my read, and it's my read. A bounty is a queue. If generating a plausible report costs nothing, reading one still costs a human. Expect more projects to demand a reproducer or a patch before anyone looks. If you run a disclosure channel, that's the engineering work: automated checks that run before a person reads. Security researchers and maintainers will push back. Pausing the product channel could discourage legitimate discoveries along with the junk. So what evidence threshold would let Google reopen without pushing validation back onto maintainers? Nobody knows yet. The date to watch is the first quarter of twenty twenty-seven. Triage is about telling allowed from not allowed. So is the first release in our ship log.

9:08

Claude Code, version two point one point two eight nine. The fix that matters is for managed machines. Deny and ask rules on nested parts of compound shell commands weren't carrying through a user-installed mod's approval. Related fixes cover Bash rules missing commands behind environment-variable prefixes or bare variable assignments, when sandbox auto-allow is on. If you administer policy for a team, upgrade. It also fixes a terminal freeze on short code blocks with many unclosed script tags. And it reverts a two point one point two eight eight change that may have caused more VS Code sign-outs. Unclosed script tags froze a terminal, so yes, HTML is still winning. Transformers, version five point eighteen. It adds NVIDIA's Nemotron 3 Diarization, a model for who spoke when. It handles up to eight speakers, streaming or offline. Input latency is configurable from eighty milliseconds to thirty point four seconds. Point three two seconds is the lowest recommended. NVIDIA and Hugging Face report a fourteen point seven two percent diarization error rate and first place on VoiceArena's initial benchmark. That covers one hundred thirty nine English conversations, about twenty two hours. They say results may change. Speaker labels are anonymous channels, and it doesn't transcribe, so pair it with ASR. Upgrade if you build meeting transcripts. Vite, version eight point three point two. If you use renderBuiltUrl and your URLs carry queries, CSS preloading now works. Bundled-dev mode now serves lazy chunk source maps and the installed Rolldown runtime. File-watcher errors no longer crash it. Upgrade if you use either feature. Vite says important fixes and security patches are backported to seven point three and eight point two. The release notes also fix dead og-image PNG links in old changelog entries, which is a patch for the patch notes. That's the stuff that ships. Now the stuff that just happens.

11:23

Fast Boot, where every story has a catch and one of them has a link shortener. Phishing with manners. Proofpoint says China-aligned TA419 spent July phishing AI-policy experts, relaying the real Microsoft sign-in to grab MFA codes and session cookies. The attackers waited for a reply before sending the link, which is more manners than most sales outreach. Go address tax. Bernat clocked the recommended Go call for IPv4 to IPv6 conversion at seven point one four nanoseconds on his machine, versus point eight eight natively. Along the way, an optimization made things worse, so he optimized around the optimizer. VB6 in-browser. VB6 Studio Web puts a classic Visual Basic form designer in your browser, with thirty seven controls and single file HTML export. One commenter's nostalgia test was three minus three point oh one, and fair enough, nobody trusts a time machine until it fumbles arithmetic. Shortcuts for everyone today. The phishers shortened a link, Bernat shortened an address conversion, and VB6 can export a standalone app as one HTML file.

12:39

Today we covered Kolibri's open weights and company-run benchmarks, the Dots browser agent, and Google's paused open-source bug bounty, plus releases for Claude Code, Transformers, and Vite. Your try-it for today: if you have uv and an OpenRouter key, run Dots on localhost with a single command. Links to everything are in the show notes. Follow The Daily Reboot in your podcast app, so tomorrow's edition shows up on its own. I'm your host. Go build something.