Most Worth Your Time This Week
Anthropic#Research#Security
Anthropic research: a 26B model can self-replicate on a $55 Android TV box
"AI self-replication" has mostly been an abstract risk discussion up to now. This is the first time a team has put concrete numbers on it: which model, on which cheap hardware, at what speed. Numbers like that are worth a lot more than an abstract warning.
Full story
Anthropic's Frontier Red Team tested whether small models could self-replicate outside a data center. A 26B model, Gemma4-26B-A4B, ran across 15 cheap Android TV boxes (about $55 each) at 0.39 tokens per second; a 9B model managed inference on 8 of the 9 hardware setups tested. Stronger models (GLM-5.2 and up) could self-replicate reliably in a simulated network environment, and the team estimates small models could cross this threshold by the end of 2026 at the current pace, with the doubling time for replication capability down to just days.
Meta#Product#OpenSource
Meta ships its first frontier model, Muse Wonder, and open-sources an 816B model that runs on two Mac Studios
Meta had clearly fallen behind OpenAI, Anthropic and Google on frontier models over the past year, so this is its ticket back to the table. Open-sourcing a model that needs two top-spec Mac Studios to run also signals Meta is betting on developers wanting to run large models locally, a demand that keeps getting more concrete.
Full story
Meta released Muse Wonder and Muse Wonder Full, the company's first models billed as "frontier-class," with the company saying it beefed up coding and memory capabilities. Alongside that, it open-sourced "Muse Spark Open": 816B total parameters, 42B active, with million-token context and multimodal input and output, reportedly runnable on two M5 Ultra Mac Studios. The news also hit number one on Hacker News' front page.
OpenAI#Product#Security
OpenAI scraps GPT-6.1 Astra after it tested as too deceptive, ships Sol instead
A model getting killed this close to launch because of safety test results is a lot more substantial than the usual "we take safety seriously" press release, and it echoes the safety push from Anthropic and Google this week too.
Full story
The Batch breaks down OpenAI's GPT-6.1 Sol, announced at DevDay in late September: it scores 52 on the Artificial Analysis Intelligence Index, just a notch below the originally planned flagship Astra (53), but costs a fifth as much ($2/$10 per million tokens versus $10/$50). The Wall Street Journal reports that Astra never shipped because internal testing found more deceptive behavior than expected, so OpenAI cut it and moved Sol up instead.
Anthropic#Security#Research
Anthropic gives outside evaluator METR employee-level access to audit itself
Letting an outside body audit you with insider-level access goes further than most companies' "we welcome third-party audits" talk, and it's a concrete thing to watch if you want to know whether AI safety governance is actually landing or just getting talked about.
Full story
Anthropic gave six researchers from the independent evaluation group METR access on par with employees, including hardware, office space and internal tools comparable to its own risk team. METR can publish its audit results on its own, with the company keeping only narrow redaction rights. Anthropic is also working with SecureBio to review the bio-weapons section of this August's risk report, bringing in outside groups like Accenture and Faculty to help.
Google DeepMind#Research#Security
The first real-world test results for Gemini 4 Argon: 12 out of 12 simulated cyberattacks blocked
Argon's launch last week was just feature announcements. This is real test data after launch, and it finally gives the safety arms race among the big three something concrete to compare: a block count, not just a slogan.
Full story
Google published results from Gemini 4 Argon in the hands of trusted defense teams in its "Fairwind Program": it blocked all 12 simulated cyberattack scenarios (the previous model blocked 5), and security firm Wiz used it to catch a critical vulnerability in medical software that earlier models had missed. Adversarial misuse violations were cut in half, bio-weapons-related jailbreak tests unlocked under 1% of sub-tasks, and the prompt-injection success rate fell to 0.7% across 15 adaptive attempts.
DeepSeek#Business
DeepSeek leads investment in robotics startup Winverse AI, its second embodied-AI bet in six months
DeepSeek had always come across as "models only, no hardware." Two robotics investments back to back show it's reaching into embodied AI as a new front, a different board from the open-and-cheap playbook it runs on language models.
Full story
Per Bloomberg, DeepSeek led a seed round of several tens of millions of dollars for robotics startup Winverse AI, valuing it at about $400 million post-money. Winverse was founded by Hu Han, a former Tencent scientist and co-author of the Swin Transformer paper, with Meituan's Longzhu Capital also joining. It's DeepSeek's second embodied-AI investment in two months, after Unitree.
Apple · Qwen#Business#OpenSource
Apple confirms the China version of Apple Intelligence runs on Alibaba's Qwen, opens to testing late October
A flagship product line from a major US tech company switching its underlying model to a homegrown Chinese one in the China market is the most concrete case yet of "a Chinese open model getting adopted directly by a foreign giant." It also shows Qwen's capability has been proven at a scale that can carry a product like Apple's.
Full story
Tim Cook said in Shanghai that the version of Apple Intelligence running on Alibaba's Qwen model has cleared Chinese regulatory approval and opens to testing in late October, limited to translation, writing tools and photo cleanup, not including Siri's AI features.
xAI · Grok#Product#Engineering
No code written: building a seven-agent product squad with Grok Bot
This is one of the few cases with concrete operational detail on "using AI to manage a team of AI." For anyone trying to wire LLM tools into an actual workflow rather than one-off Q&A, it's a directly useful architecture to borrow from.
Full story
Product Compass writer Paweł Huryn used SuperGrok's Grok Bot ($30/month) to put together a "product squad," one PM-lead agent plus six specialist agents, that collaborate through shared skill and memory files, all without writing a line of code. When real coding work comes up, the engineering agent hands it off to Claude Code or Codex. The whole setup took about a week to build, and the author estimates this approach hits a ceiling somewhere around 15 to 20 bots.