By banning Anthropic's best model, the US strengthens the case for open-weight AI
Plus: DeepSeek nabs $7.4B; Google, Microsoft and OpenAI team up; and much more.
Hi folks,
This week’s lead story looks at the fallout from the US government’s order forcing Anthropic to pull Fable 5 and Mythos 5 worldwide — and how quickly that opening was seized on by those professing the merits of more open-aligned models.
Elsewhere, DeepSeek raise its first-ever outside funding; Google, Microsoft and OpenAI team up on a new AI trust layer; Yale researchers propose “copyleft” rules for generative AI; and Epic Games open-source a Git alternative built for developers drowning in giant files.
As usual, feel free to reach out to me with any questions, tips, corrections, or suggestions: forkable[at]pm.me.
Paul
<Open issue>
A kill switch felt around the world
Last Friday, the US Government ordered Anthropic to pull Fable 5 and Mythos 5 — its two most capable models, launched just three days earlier — from every customer outside the United States, citing national security concerns over an alleged jailbreak. Because Anthropic couldn’t reliably verify user nationality in real time, it disabled both models for everyone, everywhere, including its own non-citizen employees.
Every other Claude model remained untouched. Anthropic disputed the severity of the finding, saying the demonstrated technique amounted to little more than asking a model to read code and flag bugs — “used every day by the defenders who keep systems safe” — and that a comparable standard applied industry-wide “would essentially halt all new model deployments for all frontier model providers.”
The fallout from the fracas arrived, on cue. In India, Zoho founder Sridhar Vembu declared that “globalisation is dead,” urging the government to back domestic and Chinese open-source models rather than depend on US providers that “can have access restricted at any time.”
In Hong Kong, investors reached the same conclusion from the opposite direction: shares in Zhipu AI, the company behind the GLM model family, surged as much as 48% after JPMorgan and Bank of America named it a winner in the China AI race — a rally that coincided with the company rushing out its latest open-weight model, GLM-5.2, replete with a pointed message: “Cutting-edge intelligence should not belong to only a few, nor should it be withdrawn at any time.”
On benchmarks, GLM-5.2 now sits within a point of Claude Opus 4.8 on long-horizon coding tasks, at roughly a sixth of the price — the kind of gap that companies running AI agents at volume aren’t willing to ignore. Forkable reported last week on one such case: Lindy, a no-code AI agent platform, switched its entire production stack from Anthropic to DeepSeek and said it’s saving millions in the process.
Box CEO Aaron Levie also weighed in, saying that regulating AI at the “model layer” rather than the “applied layer” (i.e. the products and services built atop the model) hands every other country a reason to build its own. “If at any moment a model can become unavailable to your country’s users or businesses, this poses very real risk on relying on technology from a particular country,” he wrote on social media, predicting the likely beneficiary would be open-weight models — “which currently is generally not coming from the US.”
Even US companies are hedging. Microsoft is reportedly weighing a version of DeepSeek V4 to power a cheaper tier of Copilot Cowork, alongside its existing OpenAI and Anthropic options. And over in France, Mistral CEO Arthur Mensch used the moment to make his own pitch, writing that his company exists “outside of centralised control exercised by states or corporations.”
The unease over depending on a single, US-controlled platform isn’t confined to AI models. As Forkable reported in December, a steady stream of open source projects, from the Zig programming language to the Naev game engine, have quietly migrated away from GitHub in recent months, citing similar concerns about concentrated corporate control (as well as performance issues). The Fable ban gives that argument sharper teeth: when access can be revoked overnight by directive, dependence on any single company, in any single country, starts to look like a strategic liability.
None of this means open-weight models have overtaken the frontier quite yet. But the past week has handed every argument for sovereign, open AI infrastructure a concrete example of what happens when access to the frontier can be switched off overnight by a single state.
Read more: Anthropic | NDTV | Aaron Levie | Arthur Mensch
<Patch notes>
DeepSeek nabs $7.4bn — and Liang Wenfeng keeps the keys
DeepSeek, the open-weight AI lab that rattled Silicon Valley with R1 last year, has closed its first-ever external funding round: more than $7.4 billion at a valuation above $50 billion, making it China's most valuable AI startup. Until now, every model the company has shipped was funded entirely in-house. Founder Liang Wenfeng remains firmly in control, with the new capital routed through a vehicle that strips outside investors of voting rights.
Read more: SCMP | Yahoo Finance
Google, Microsoft and OpenAI build AI's missing trust layer
The Linux Foundation has launched the Appia Foundation, a new body backed by 13 founding members — including Google, Microsoft, OpenAI, Arm, Siemens, and Mastercard — to build open, vendor-neutral specifications that let organisations prove their AI systems meet safety and compliance standards. The aim is to let "conformity evidence" be reused across the AI supply chain rather than rebuilt for every vendor and every regulator.
Read more: The New Stack
Google wants to take the pain out of fine-tuning
Google has released a research preview of OpenRL, a self-hosted, open-source API for fine-tuning LLMs with reinforcement learning on your own Kubernetes cluster. The project, inspired by Thinking Machines' Tinker design pattern, aims to separate infrastructure concerns from AI research the way Kubernetes itself separated infrastructure from application development.
Chainguard scanner catches 52,000 packages that pass every test and still misbehave
Chainguard has launched a scanner that blocks "greyware" — open source packages that openly disclose what they do, pass every existing security check, and still perform actions no security team would sign off on, from credential harvesting to silent remote access. The tool has already screened more than 100,000 npm packages a day and blocked over 52,000 of them, a problem Chainguard says is being made worse by AI agents pulling in dependencies faster than any human can review.
Read more: The New Stack
Yale proposes "copyleft" rules to make AI training data honest
Researchers at Yale's Digital Ethics Center have proposed a Contextual Copyleft AI License, which would treat generative AI models trained on open source code as derivative works and require developers to publicly disclose the model's architecture and training data. "Our analysis showed that extending the copyleft concept to generative artificial intelligence has the potential to give open-source software developers meaningful control over how AI developers use their code," lead author Grant Shanklin said.
Read more: Yale News
Xiaomi's MiMo Code claims it beats Claude Code on long tasks
Xiaomi has open-sourced MiMo Code, a terminal-native coding agent the company claims outperforms Claude Code specifically on tasks that run beyond 200 steps — the point where most agents start losing the thread and compounding small errors.
Read more: VentureBeat
<Final commit>
Epic drops a Git rival built for hoarders of giant files
Game developers have long endured an awkward relationship with Git: brilliant for code, miserable for the multi-gigabyte textures, meshes, and audio files that actually make up most of a modern game project.
And so Epic Games has decided enough is enough, releasing Lore, an MIT-licensed, Rust-built version control system designed from the ground up for exactly that mix — code sitting alongside enormous binary assets.
Where Git treats large files as a bolt-on via Git LFS, Lore handles them natively, with content-addressed storage, fragment-level deduplication, and “sparse” working copies that only download what you actually need, sparing 3D artists from hauling an entire asset library onto their laptop just to open one file. Everyday operations like staging, committing, and branching also work entirely offline, with no network round trip required. It’s not aimed at replacing Git for everyone — but for anyone who has ever watched a repository balloon to 40GB because someone committed a 4K texture pack, it might be a quiet revelation.
Read more: Phoronix | Lore | GitHub