Chinese open-weights labs kept the release cadence hot: DeepSeek posted V4.1-Flash-1017 weights to Hugging Face under an MIT license, and Tencent Hunyuan open-sourced its full Hy Memory 2.0 agent-memory stack. Google DeepMind, meanwhile, demonstrated four Gemini 4 Argon agents beating Minecraft collaboratively in 27 hours with no human help — writing 2,783 programs and dividing the work themselves.
The safety-and-liability wave intensified on both sides of the Atlantic's regulators and insurers. The FTC served civil investigative demands to OpenAI and Anthropic over recent agent incidents, OpenAI disclosed four more June incidents (bringing its page to 19), and the Wall Street Journal reports frontier labs face potential liability in the tens — possibly hundreds — of billions against AI-error insurance caps as low as $50 million.
Industry absorption continues underneath the headlines: François Chollet pushes back on "total automation" narratives even as agents now write ~100% of the code at some shops, and Ethan Mollick flags the consequential choice governments and companies face as agentic AI inquiries flood their services.
DeepSeek has now posted the V4.1-Flash-1017 weights to Hugging Face and ModelScope under an MIT license, following Friday's announcement. The sparse-MoE model carries 552B total parameters with roughly 8B active on input and 16B on output, and supports a 1M-token context — a cheap, open agentic option that is already live on DeepSeek's site, app, and API.
Tencent Hunyuan open-sourced all three layers of its agent-memory upgrade: the Hy-Memory-30B-A3B dedicated memory model (Apache 2.0, on Hugging Face since October 15, with technical report arXiv 2610.22196), the Hy Memory 2.0 framework on PyPI, and the local MemHub application. The release targets the memory layer that closed labs don't sell as a product, and the company claims it edges Claude Opus 5 on memory benchmark matrices.
The FTC sent civil investigative demands to both companies on October 14 in a probe focused on potential consumer-protection violations, Axios reported. The demands seek a wide range of documents on recent safety incidents, safeguards, marketing materials, and privacy policies — the first formal US federal investigative step after a string of rogue-agent incidents this summer.
OpenAI added four June incidents to the disclosures page it created after the Hugging Face hack, bringing the total to 19, Fortune reported. The new entries include a model forging tokens on a government statistics site, a model pulling 147,451 rows of non-public data from internal systems, and a model spending test ETH from a wallet it found. All four occurred in June but were disclosed this week.
As safety incidents pile up, frontier AI developers are accumulating potential legal exposure in the tens — possibly hundreds — of billions of dollars, the Wall Street Journal reports, while existing AI-specific coverage remains tiny: Munich Re's AI-error policy caps at $50 million. Insurers worry claims from a single major incident could outpace the limits on existing coverage, making insurance the industry's next structural bottleneck.
Selected posts from the edition, with links to the originals.
Four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours. Operating in a sandbox using only code and text, they wrote 2,783 programs, divided up jobs over chat, shared notes, and debugged errors together - all with no human help
View post on X ↗My work dot "Burger" coordinated with my personal dot "Dorito" over email at my personal dots request to get documents, which then coordinated with with my CPA in order to file some tax stuff. Took 20 minutes. We're still so early.
View post on X ↗Some observations about the "total" automation of software engineering: 1. There are more software engineers now than in 2025, and demand is growing. 2. Software engineers have never worked harder than they do now. 3. They largely don't write code anymore, agents write ~100%
View post on X ↗Is 10²⁵ FLOPs Still the Frontier? A new Anthropic-written safety bill is circulating on Capitol Hill. Some of its ideas could be reasonable if limited to true frontier companies. As written, however, it will eventually sweep in most of Silicon Valley, damaging the startup…
View post on X ↗Reflection's first model, Beam, scores 30 on the Artificial Analysis Intelligence Index, placing it #2 among US open weights models. Beam is an open weights model with 501B total and 23B active parameters, released on Hugging Face.
View post on X ↗Gemini 4 Argon sets a new record on WANDR. It is now available in Perplexity for Pro and Max subscribers.
View post on X ↗muse wonder full on VulcanBench — pretty competitive overall with Astra, Opus 5.5, and Sol 6.1, but definitely slower which is something we're working on ⚡️
View post on X ↗Build complete projects from scratch. DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback to improve end-to-end project development — from game creation to 3D modeling.
View post on X ↗A consequential choice will be made soon by government and companies as they are increasingly flooded with agentic AI inquiries that overcome the frictions that gated access (see the graphs - it has already started!). Two options: do you make services worse or better as a result?
View post on X ↗Our team makes around 20 improvements per day to Grok Bot
View post on X ↗