# Anthropic’s Policy Push Leads a Day of Open Models and Big AI Financing

*By AI News Digest • June 11, 2026*

Anthropic moved from product controversy to a broad policy push, while Google DeepMind released DiffusionGemma and Alphabet moved to raise $80 billion for AI expansion. Biohub also unveiled an open protein world model, and new workplace data showed why AI productivity gains still fail to cleanly translate into organizational performance.

## Anthropic makes its policy case

### Dario Amodei argues policy is trailing the technology
Dario Amodei published [Policy on the AI Exponential](https://darioamodei.com/post/policy-on-the-ai-exponential), arguing that AI is advancing faster than policy institutions were built to handle and that frontier models should face mandatory third-party testing for cyber, bio, and autonomy risks, with the power to block or revoke deployment of catastrophic-risk systems [^1][^2]. Anthropic paired the essay with an Advanced AI Framework that says governments should be able to block unsafe frontier releases and invest in societal resilience, plus an Economic Policy Framework backed by $200 million for major evaluations of labor-market responses and a $150 million fellowship program for early-career professionals [^3][^4][^5]. Anthropic said these projects are signals of intent rather than sufficient on their own, and the essay frames the stakes across jobs, scientific progress, civil liberties, and geopolitics [^6][^7].

*Why it matters:* Frontier labs are increasingly trying to shape the policy architecture around deployment, not just the models themselves [^1][^3][^4].

### Anthropic says Fable 5 safeguards will be made visible
Simon Willison highlighted Anthropic language saying it is changing Fable 5's safeguards for frontier LLM development "to make them visible," which he interpreted as ending the decision to have the model hide refusals while keeping the refusals in place [^8][^9]. Even with that change, critics said the episode has left researchers more worried about silent steering becoming part of frontier-lab practice [^10][^11].

*Why it matters:* Transparency is becoming part of the safety debate itself, not just the restrictions labs choose to impose [^8][^10].

## Speed and capital are becoming central competitive levers

### Google DeepMind opens DiffusionGemma
Google DeepMind released DiffusionGemma, an experimental open model that generates whole blocks of text simultaneously rather than word by word, a design the company says enables real-time self-correction and complex markdown formatting [^12]. Google says the model can deliver up to 4x faster inference on dedicated GPUs, and Sundar Pichai said the weights are available on Hugging Face under an Apache 2.0 license [^13][^14]. NVIDIA said its optimizations support RTX, RTX PRO, and DGX systems, with throughput reaching 1,000 tokens per second on H100 [^15].

*Why it matters:* Developers now have an open way to test whether blockwise text generation can improve low-latency local workloads and agent loops [^12][^15].

### Alphabet lines up $80 billion for AI expansion
Bloomberg reported that Alphabet is raising $80 billion through equity offerings, including a $10 billion Berkshire Hathaway investment, to fund its AI spending plans [^16]. In Ben Thompson's breakdown, Google Cloud grew from $2.6 billion in revenue in Q4 2019 to $20 billion in Q1 2026, while Google Services reached $89.6 billion in the same quarter [^16]. Thompson argued the financing signals that expected AI compute demand may be larger than many assume, and that Google's TPU cost advantage could matter if access to capacity becomes the main constraint [^16].

*Why it matters:* At the frontier, AI competition is looking more and more like a balance-sheet contest alongside a model contest [^16].

## AI in science gets a major open release

### Biohub launches an open protein world model
Chan Zuckerberg Biohub said its new ESM Fold is an open system for scientific discovery in protein biology, trained on billions of protein sequences and able to predict atomic-resolution protein structures [^17]. Biohub says the model is state-of-the-art across structure-prediction benchmarks, especially protein-protein and protein-antibody interactions, has folded 1.1 billion proteins, and can be used to digitally design proteins and single-chain antibodies that produced nanomolar binders in small experimental cycles [^17]. The organization has committed $500 million to its virtual biology initiative and says it plans to release its models open-source to get them into more scientists' hands quickly [^17].

*Why it matters:* This is a strong example of frontier AI moving beyond language and code into experimentally grounded biology while staying open to the wider research community [^17].

## The workplace evidence is getting sharper

### A large survey finds a wide execution gap
Glean's Work AI Index 2026 says 87% of workers now use AI and report saving 13 hours per week on average, yet only 13% say their organization is performing significantly better as a result [^18]. The report attributes much of the gap to "botsitting"—the hidden work of feeding context, debugging, and cleaning up outputs—which consumes 6.4 hours per week, and to the practice of shipping AI-generated work people cannot explain or defend, which 69% admitted doing [^18]. It also says organizations with stronger AI strategy, measurement, and shared context are seeing better results [^18].

*Why it matters:* The limiting factor in enterprise AI may be shifting from tool access to context, incentives, and change management [^18].

---

### Sources

[^1]: [𝕏 post by @DarioAmodei](https://x.com/DarioAmodei/status/2064781775247950326)
[^2]: [𝕏 post by @DarioAmodei](https://x.com/DarioAmodei/status/2064781778599268818)
[^3]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2064783421860413780)
[^4]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2064783420425929169)
[^5]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2064783424251101553)
[^6]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2064783425807187970)
[^7]: [𝕏 post by @DarioAmodei](https://x.com/DarioAmodei/status/2064781780562137306)
[^8]: [𝕏 post by @simonw](https://x.com/simonw/status/2064936762099789960)
[^9]: [𝕏 post by @simonw](https://x.com/simonw/status/2064918665859080392)
[^10]: [𝕏 post by @far__el](https://x.com/far__el/status/2064948472688890138)
[^11]: [𝕏 post by @jeremyphoward](https://x.com/jeremyphoward/status/2064949877487096187)
[^12]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2064741061352636762)
[^13]: [𝕏 post by @sundarpichai](https://x.com/sundarpichai/status/2064744343743922189)
[^14]: [𝕏 post by @sundarpichai](https://x.com/sundarpichai/status/2064744345216196942)
[^15]: [NVIDIA Accelerates Google DeepMind’s DiffusionGemma for Local AI](https://blogs.nvidia.com/blog/rtx-ai-garage-local-gemma-diffusion)
[^16]: [The Google Capital Company | Stratechery by Ben Thompson](https://www.youtube.com/watch?v=dx8rmF7o-hI)
[^17]: [Biohub: The Future of Biology is Open-Source with Mark Zuckerberg, Priscilla Chan, and Alex Rives](https://www.youtube.com/watch?v=u8cYwaLzN-k)
[^18]: [Babysitting the Machine: Glean's Rebecca Hinds on the Hidden Human Labor of AI at Work](https://www.cognitiverevolution.ai/babysitting-the-machine-glean-s-rebecca-hinds-on-the-hidden-human-labor-of-ai-at-work)