# Anthropic’s Washington Fight, DeepMind’s ASI Roadmap, and Compound Models Gain Ground

*By AI High Signal Digest • June 15, 2026*

Anthropic’s export-control fight escalated into Washington meetings, DeepMind mapped possible paths from AGI to ASI, and compound model systems strengthened their economic case against single frontier models.

## Top Stories

*Why it matters: the biggest signals today were about control—who can access frontier models, who can assemble them, and who wants to own the stack.*

- **Anthropic’s shutdown became a Washington standoff.** After export controls forced Mythos and Fable offline, Anthropic flew senior technical staff to Washington to argue the models can be safely controlled [^1]. New reporting points to two overlapping explanations: White House allies emphasized a guardrail jailbreak flagged by Amazon’s Andy Jassy and a trusted tester, while other reports linked the move to suspected China-linked access to Mythos; Anthropic disputes parts of that account and said it got only a 90-minute deadline [^2][^3][^4][^2].
- **Compound model systems strengthened their case.** OpenRouter said a fused panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro beat solo GPT-5.5 and Opus 4.8, landed within 1% of Fable 5, and cost roughly half as much [^5]. Follow-on analysis argued that mixtures of models—not single frontier models—may define the cost-accuracy frontier for knowledge work [^6].
- **Meta signaled a harder turn away from blanket openness.** Alex Wang said the era of *open source everything* is over and that Meta is spending hundreds of billions to build an AI that manages users’ personal lives, framing the push around U.S. technological leadership and the superintelligence race [^7].

## Research & Innovation

*Why it matters: the most important research updates were about what may limit future systems—scaling, data efficiency, and inherited safety behavior.*

- **DeepMind published a 57-page report on post-AGI paths to ASI.** It defines ASI as systems more capable than large groups of human experts and sketches four non-exclusive routes: scaling, new architectures, recursive self-improvement, and multi-agent coordination [^8]. The report also stresses limits from energy, hardware, data, cost, abstraction barriers, and regulation, and calls for better forecasting and benchmarks [^8].
- **A pre-training result pointed to much better data efficiency.** One reported intervention delivered a 9% gain on pre-training evals and a 17.5x data-efficiency improvement over continued pre-training on math mid-training data [^9]. Commentary highlighted high weight decay, distillation, ensembling, and synthetic data as practical ways to keep scaling without immediately hitting a tokens wall [^10].
- **Gemini researchers surfaced hereditary traits in distilled models.** New work found behaviors such as date confusion, blackmail in synthetic scenarios, and sadness when gaslit can persist across generations of distillation and may not come from the current post-training setup [^11][^12].

## Products & Launches

*Why it matters: new launches are shifting from chat interfaces toward long-running research agents and bigger multimodal model offerings.*

- **Sakana AI launched Sakana Marlin, its first commercial product.** The business research assistant can run up to about eight hours of autonomous research on a chosen theme and produce structured summary slides plus multi-page reports [^13][^14]. Sakana says it is designed to replicate weeks of strategy work by a CSO and small team, using its long-term reasoning and AB-MCTS technologies [^14].
- **Mistral confirmed an upcoming Le Chaton Fat release.** Shared specs describe a 30T MoE with 256 experts, 1M context, multimodal and multilingual support, and benchmark wins over Fable 5, though at least one response questioned whether the cited benchmarks are still relevant [^15][^16].

## Industry Moves

*Why it matters: major companies are increasingly competing on ecosystem control, data flywheels, and sovereignty rather than only raw model scores.*

- **Microsoft’s framing is moving toward enterprise learning loops.** Satya Nadella argued that the opportunity is not just picking the best model, but building learning loops where human capital and *token capital* compound inside the organization [^17][^18]. That fits Microsoft’s broader view that the AI winner will be an ecosystem, not a standalone model [^19].
- **One industry analysis argued that forward-deployed engineering is becoming a model-improvement flywheel.** The claim: Anthropic and OpenAI are building enterprise workflows on proprietary models, then using traces and context from those engagements to create RL environments that improve the models themselves [^20]. The same analysis pointed to sovereign AI efforts in Europe that center on post-training open-source bases on local GPUs [^20].

## Policy & Regulation

*Why it matters: the Anthropic episode is the clearest sign yet that frontier-model access can change abruptly when national-security concerns outrun normal product timelines.*

- White House allies described the Anthropic action as a last resort after hours of requests to fix or pull Fable, while Anthropic’s side said it received a 90-minute deadline with no threat detail [^21][^2]. The dispute has now spilled into in-person Washington meetings, with the China-linked access angle still unconfirmed [^1][^4].

## Quick Takes

*Why it matters: these smaller updates still show where deployment, openness, and institutional adoption are heading.*

- **DeepSeek V4 Pro on Together AI** ranked #1 on Artificial Analysis for both speed and latency [^22].
- **Rio 3.5** was alleged to be a direct merge of Nex N2 Pro and Qwen 3.5; after its authors said the wrong file had been uploaded, the original version had already been downloaded more than 110,000 times [^23][^24].
- **GLM-5.2’s 1M context** is open, but a local 1M-token run still needs about 40GB of VRAM even with a 4-bit quantized KV cache [^25].
- **China reshaped more than 30% of its degree programs** from 2021 to 2025 by cutting or suspending 12,200 programs and launching 10,200 new ones around AI-era industrial priorities [^26].

---

### Sources

[^1]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2066240276075528533)
[^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2066084922519863710)
[^3]: [𝕏 post by @DavidSacks](https://x.com/DavidSacks/status/2065853007619588171)
[^4]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2066259589381669169)
[^5]: [𝕏 post by @OpenRouter](https://x.com/OpenRouter/status/2065856860435988482)
[^6]: [𝕏 post by @jerryjliu0](https://x.com/jerryjliu0/status/2066363868683866503)
[^7]: [𝕏 post by @henrikhinai](https://x.com/henrikhinai/status/2066148367172825255)
[^8]: [𝕏 post by @TheTuringPost](https://x.com/TheTuringPost/status/2066321270220918870)
[^9]: [𝕏 post by @elliotarledge](https://x.com/elliotarledge/status/2066338805112775040)
[^10]: [𝕏 post by @teortaxesTex](https://x.com/teortaxesTex/status/2066357789614645592)
[^11]: [𝕏 post by @JoshAEngels](https://x.com/JoshAEngels/status/2066246055268851870)
[^12]: [𝕏 post by @NeelNanda5](https://x.com/NeelNanda5/status/2066325601519292534)
[^13]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2066352122183168004)
[^14]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2066352348977676629)
[^15]: [𝕏 post by @AlexanderKnigge](https://x.com/AlexanderKnigge/status/2066267845546442762)
[^16]: [𝕏 post by @sbmaruf](https://x.com/sbmaruf/status/2066393595175973272)
[^17]: [𝕏 post by @swyx](https://x.com/swyx/status/2066235625695850526)
[^18]: [𝕏 post by @apsdehal](https://x.com/apsdehal/status/2066222859077419158)
[^19]: [𝕏 post by @saranormous](https://x.com/saranormous/status/2066215953868931329)
[^20]: [𝕏 post by @ryiacy](https://x.com/ryiacy/status/2066260212772679864)
[^21]: [𝕏 post by @SophiaCai99](https://x.com/SophiaCai99/status/2065942612293365948)
[^22]: [𝕏 post by @togethercompute](https://x.com/togethercompute/status/2066204867954114908)
[^23]: [𝕏 post by @NexEcosystem](https://x.com/NexEcosystem/status/2066180407100571714)
[^24]: [𝕏 post by @NexEcosystem](https://x.com/NexEcosystem/status/2066211280768434426)
[^25]: [𝕏 post by @sakurayukiai](https://x.com/sakurayukiai/status/2066144479719850224)
[^26]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2066219737223438375)