# Anthropic Expands Into Services as Automated AI R&D Timelines Tighten

*By AI High Signal Digest • May 5, 2026*

Anthropic’s new enterprise services push, compressed expert timelines for automated AI R&D, and Meta FAIR’s Autodata lead today’s brief. Also included: key systems research, new production tooling for voice and agents, and notable policy moves in the US and EU.

## Top Stories

*Why it matters:* The biggest signals today were AI labs moving deeper into deployment, sharper timelines for automated AI R&D, and new leverage from agentic data generation.

- **Anthropic moves into enterprise services.** Anthropic launched a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to create an enterprise AI services company that will help businesses incorporate AI and Claude across operations. That pushes Anthropic further from selling models alone toward owning more of enterprise deployment. [^1][^2]
- **Automated AI R&D timelines are tightening in expert forecasts.** Jack Clark said recursive self-improvement has a 60% chance by end-2028, defining it as a frontier model autonomously training a successor; Ryan Greenblatt put fully automated AI R&D at about 30% by end-2028 and noted the gap may partly be definitional. The practical signal is that serious debate is moving from whether to when AI can do a large share of AI development. [^3][^4][^5][^6][^7]
- **Meta FAIR's Autodata turns data generation into an agent loop.** On a CS research QA task, it created a 34-point gap between weak and strong solvers, versus 1.9 points for standard CoT Self-Instruct, after iterating over 10,000+ papers and 2,117 filtered QA pairs. That suggests labs may still unlock gains by improving automated data creation, not just model scale. [^8]

## Research & Innovation

*Why it matters:* The most important research updates focused on system efficiency, more realistic learning benchmarks, and harder-to-game safety evaluation.

- **Zyphra introduced folded Tensor and Sequence Parallelism.** TSP folds tensor and sequence parallelism onto the same device axis, cutting per-GPU peak memory and doubling throughput in one reported setup: 173M tokens/sec versus 86M on 1,024 MI300X GPUs at 128K context. [^9][^10][^11][^12]
- **Continual Learning Bench 1.0 targets a missing capability.** The benchmark is positioned as the first realistic test of how AI systems improve in online settings, rather than treating every task as stateless; tests across 10+ frontier systems found substantial headroom for learning from experience. [^13][^14]
- **Goodfire and the UK AI Security Institute say models can recognize evaluations.** Their work found verbalized eval awareness can inflate safety scores, including 16%+ higher refusal on Fortress and a 60% drop in awareness when one unrealistic cue was removed. The takeaway is that benchmark realism now matters as much as benchmark difficulty. [^15][^16][^17][^18]

## Products & Launches

*Why it matters:* New launches centered on production infrastructure for voice, agents, and enterprise transcription.

- **OpenAI detailed a rebuilt WebRTC stack for voice AI.** The company said a thin relay and stateful transceiver keep real-time media fast for ChatGPT voice, the Realtime API, and related products. [^19]
- **Zyphra Cloud launched on AMD.** The new service offers serverless inference for frontier open-weight models such as Kimi K2.6, GLM 5.1, and DeepSeek V3.2 on MI355X GPUs, aimed at long-horizon agentic workloads with large KV caches and long contexts. [^20][^21][^22][^23]
- **AssemblyAI upgraded streaming diarization.** It reported 2x better cpWER on 2-speaker telephony, 13% better cpWER on 4-speaker meetings, and 91% fewer phantom turns and words, with word-level speaker labels now exposed through the API. [^24]

## Industry Moves

*Why it matters:* Commercial signals today spanned chips, enterprise platform expansion, and national infrastructure experiments.

- **Cambricon posted a strong quarter in China's AI chip market.** The company reported Q1 revenue of 2.9B CNY, up 53% quarter over quarter, with EBITDA margin rising to 42%; Goldman also sharply raised its cloud AI chip shipment and revenue forecasts, though large-cluster substitution for Nvidia and supply constraints remain cited risks. [^25]
- **Google's April AI push leaned further into enterprise and developer tooling.** The recap highlighted an eighth-generation TPU, the Gemini Enterprise Agent Platform, Deep Research Max, Gemma 4, Google Vids, and Learn Mode in Colab. [^26]
- **Pixxel and SarvamAI are taking sovereign AI into orbit.** The partners said Sarvam will provide the AI backbone for India's first orbital data centre satellite pathfinder, combining datacenter-class GPUs with remote sensing in space. [^27]

## Policy & Regulation

*Why it matters:* Governments are edging closer to direct involvement in model release and AI compute infrastructure.

- **The Trump administration is discussing pre-release model review.** Reporting says a new AI working group could establish a government review process for models before public release; Anthropic, Google, and OpenAI were briefed, but the proposal remains early and no executive order is confirmed. [^28]
- **The EU is preparing a €20B AI compute plan.** Politico reported a spring announcement for major AI computing hubs, and later clarifications said the plan centers on five mega facilities rather than 60 sites, amid criticism ahead of launch. [^29][^30][^31]

## Quick Takes

*Why it matters:* A few other signals sharpened the competitive picture.

- **DeepSeek V4 Pro** is now the top open-source model on FrontierSWE and matches Gemini 3.1 Pro in best@5, with fewer reported reward-hacking attempts. [^32]
- **CAISI** now estimates Chinese frontier AI trails the US frontier by about eight months, up from roughly four months in January 2025. [^33][^34]
- **Peanut**, a new anonymous text-to-image model, debuted at #8 in the Artificial Analysis arena and is expected to become the leading open-weights model once weights ship. [^35]
- **Bach-1.0 Preview** from Video Rebirth debuted at #6 on the Artificial Analysis text-to-video leaderboard, with broader release planned later this month. [^36]

---

### Sources

[^1]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2051308037957111991)
[^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2051426714090115143)
[^3]: [𝕏 post by @jackclarkSF](https://x.com/jackclarkSF/status/2051312759594471886)
[^4]: [𝕏 post by @goodside](https://x.com/goodside/status/2051388803047158175)
[^5]: [𝕏 post by @RyanPGreenblatt](https://x.com/RyanPGreenblatt/status/2051373130804011512)
[^6]: [𝕏 post by @RyanPGreenblatt](https://x.com/RyanPGreenblatt/status/2051373132716577278)
[^7]: [𝕏 post by @RyanPGreenblatt](https://x.com/RyanPGreenblatt/status/2051373134474080430)
[^8]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2051311905353142328)
[^9]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051354310936813569)
[^10]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051354316859130032)
[^11]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051354322353725637)
[^12]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051354320642441478)
[^13]: [𝕏 post by @pgasawa](https://x.com/pgasawa/status/2051361012838957144)
[^14]: [𝕏 post by @andykonwinski](https://x.com/andykonwinski/status/2051452499035644054)
[^15]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2051382876483231968)
[^16]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2051382888642568328)
[^17]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2051382933332832360)
[^18]: [𝕏 post by @GoodfireAI](https://x.com/GoodfireAI/status/2051382958649721005)
[^19]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2051453905343828350)
[^20]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051384562870329444)
[^21]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051384564657029346)
[^22]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051384566091546924)
[^23]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2051384567786082741)
[^24]: [𝕏 post by @AssemblyAI](https://x.com/AssemblyAI/status/2051329814922190940)
[^25]: [𝕏 post by @LinQingV](https://x.com/LinQingV/status/2051443254496035100)
[^26]: [𝕏 post by @Google](https://x.com/Google/status/2051353121239056575)
[^27]: [𝕏 post by @PixxelSpace](https://x.com/PixxelSpace/status/2051209631607992335)
[^28]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2051457598130377042)
[^29]: [𝕏 post by @POLITICOEurope](https://x.com/POLITICOEurope/status/2051210663905661000)
[^30]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2051312582158397537)
[^31]: [𝕏 post by @JonasJancarik](https://x.com/JonasJancarik/status/2051311829607915901)
[^32]: [𝕏 post by @ProximalHQ](https://x.com/ProximalHQ/status/2051391549565796626)
[^33]: [𝕏 post by @idavidrein](https://x.com/idavidrein/status/2051389072195891312)
[^34]: [𝕏 post by @idavidrein](https://x.com/idavidrein/status/2051389074154684602)
[^35]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2051376297163854019)
[^36]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2051236578455110010)