# Anthropic Lands John Jumper as Export Controls and GLM-5.2 Reframe the AI Race

*By AI High Signal Digest • June 20, 2026*

Anthropic landed one of DeepMind’s most prominent scientists, GLM-5.2 kept narrowing the gap with closed models, and U.S. export controls reshaped access to Anthropic’s frontier systems. The brief also covers new research on memory and transparency, notable product launches, and key industry infrastructure moves.

## Top Stories

*Why it matters: today’s biggest shifts were in talent, open-model capability, and who controls access to frontier systems.*

- **Anthropic hired John Jumper from Google DeepMind.** Jumper, who shared the 2024 Nobel Prize in Chemistry with Demis Hassabis for AlphaFold, said he is leaving DeepMind after nearly nine years to join Anthropic [^1][^2]. Hassabis said AlphaFold changed the world and showed what AI could do for science and medicine [^3].
- **GLM-5.2 kept gaining ground as an open-weight alternative.** Design Arena moved it to **#1** at 1360 Elo ahead of the now-unavailable Claude Fable 5, and one observer noted it is the first non-multimodal model to lead the design category [^4][^5]. Jeremy Howard said it was at least as good as Opus 4.8 and GPT-5.5, while being fast, inexpensive, nuanced, and strong on long context [^6].
- **Export controls are now directly shaping model availability.** Andrew Ng said U.S. Commerce restrictions on Anthropic’s Mythos and Fable require licenses for any foreign national, including Anthropic employees, which led Anthropic to disable Fable worldwide [^7]. He added that the move is already pushing more countries to think seriously about AI sovereignty and open-source alternatives [^7].

## Research & Innovation

*Why it matters: the strongest technical updates focused on memory, transparency, and training agents on richer feedback.*

- **AtomMem** stores small atomic facts instead of coarse summaries, organizes them into hierarchical event structures and temporal user profiles, and uses an associative memory graph at retrieval time [^8]. The paper reports state-of-the-art results on the LoCoMo multi-session benchmark while staying cheap enough to deploy [^8].
- **A transparency audit of DiffusionGemma** found that, although text diffusion models are harder to inspect than today’s LLMs, their intermediate states remain interpretable and recover many benefits of chain-of-thought monitoring for safety work [^9][^10].
- **Recent agentic RL work is pushing past action masking.** Posts summarizing ECHO and PaW argue that models should train on both action tokens and environment feedback tokens; the setup uses RL on actions and SFT on tool responses, with reported large performance gains [^11][^12][^13].

## Products & Launches

*Why it matters: shipping products are getting more multimodal, more open, and more useful in production workflows.*

- **Together AI added OpenAI’s GPT Image 2** to Serverless Inference, with 95%+ multilingual text rendering, support for up to 16 reference images, and native 1K, 2K, and 4K outputs for design, marketing, e-commerce, and editorial use cases [^14][^15][^14].
- **Magnitude** launched as a coding agent that runs entirely on open models; its launch post claims **60% lower cost** than Claude Code with no drop in performance [^16].
- **LiteParse v2.1** was released as an Apache 2.0, model-free PDF-to-markdown parser that its creators say is faster and more accurate than other open-source parsers on three benchmarks, while staying competitive with some frontier VLMs on text- and table-heavy documents [^17][^18][^17].

## Industry Moves

*Why it matters: competition is spreading beyond model quality into recruiting, large-scale deployment, and inference speed.*

- **The talent market stayed unusually fluid.** Posts this week tracked Noam Shazeer’s move from Google DeepMind to OpenAI, John Jumper’s move to Anthropic, and Barret Zoph leaving OpenAI again [^2]. François Fleuret argued that this kind of turnover has helped keep information flowing and competition high across the sector [^19].
- **Shopify said one internal AI/ML team is clustering billions of products for agentic commerce,** and that its ICML lineup will cover search, recommendations, Sidekick, SimGym, Flow, ads, financial services, and the global product catalog [^20]. Mikhail Parakhin added that the company is serving **2.2 trillion requests** while improving SimGym [^21].
- **Modal and Z Lab co-released six DFlash speculators for Alibaba Qwen 3.x,** claiming **over 1,000 output tokens per second** for Qwen 3.5 122B-A10B on a B200 [^22][^23].

## Policy & Regulation

*Why it matters: government rules are now directly shaping access to frontier AI systems.*

- **Access restrictions remain uneven after the U.S. order.** Separate posts said roughly **200 organizations** still retain access to Claude Mythos, and that early users kept access mainly through Project Glasswing [^24][^25]. Trump later said he no longer viewed Anthropic or Dario Amodei as a national security threat and that the company had responded responsibly to the administration’s request [^26].

## Quick Takes

*Why it matters: these smaller updates still point to where capability is improving fastest.*

- **Ai2 released MolmoMotion,** a language-guided model for forecasting object 3D point trajectories from video that beats prior methods on motion forecasting, robot planning, and video generation [^27].
- **Datalab open-sourced lift,** a 9B document-extraction model that scored **90.2%** on its benchmark versus **91.3%** for Gemini 3.5 Flash [^28].
- **OpenAI Codex** can now hand off threads between local and remote hosts so work can move off a laptop and resume later [^29].
- **Figure** said robots now outnumber humans at the company [^30].

---

### Sources

[^1]: [𝕏 post by @JohnJumperSci](https://x.com/JohnJumperSci/status/2068001285173834106)
[^2]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2068037727572910234)
[^3]: [𝕏 post by @demishassabis](https://x.com/demishassabis/status/2068002732250640603)
[^4]: [𝕏 post by @Designarena](https://x.com/Designarena/status/2066940737011560652)
[^5]: [𝕏 post by @gneubig](https://x.com/gneubig/status/2067980056148717817)
[^6]: [𝕏 post by @jeremyphoward](https://x.com/jeremyphoward/status/2067757468189679764)
[^7]: [𝕏 post by @AndrewYNg](https://x.com/AndrewYNg/status/2068039709126017356)
[^8]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2067984002376749525)
[^9]: [𝕏 post by @JoshAEngels](https://x.com/JoshAEngels/status/2068034390245154877)
[^10]: [𝕏 post by @NeelNanda5](https://x.com/NeelNanda5/status/2068042997363769663)
[^11]: [𝕏 post by @cwolferesearch](https://x.com/cwolferesearch/status/2067996505231774168)
[^12]: [𝕏 post by @cwolferesearch](https://x.com/cwolferesearch/status/2067996508323037599)
[^13]: [𝕏 post by @cwolferesearch](https://x.com/cwolferesearch/status/2067996511296737424)
[^14]: [𝕏 post by @togethercompute](https://x.com/togethercompute/status/2068008498550444520)
[^15]: [𝕏 post by @togethercompute](https://x.com/togethercompute/status/2068008501440385388)
[^16]: [𝕏 post by @tomgreenwald](https://x.com/tomgreenwald/status/2068008885018062868)
[^17]: [𝕏 post by @jerryjliu0](https://x.com/jerryjliu0/status/2067679507126124858)
[^18]: [𝕏 post by @jerryjliu0](https://x.com/jerryjliu0/status/2068005414369906856)
[^19]: [𝕏 post by @francoisfleuret](https://x.com/francoisfleuret/status/2068210578565267741)
[^20]: [𝕏 post by @ShopifyEng](https://x.com/ShopifyEng/status/2067964144582557871)
[^21]: [𝕏 post by @MParakhin](https://x.com/MParakhin/status/2068007050345607314)
[^22]: [𝕏 post by @charles_irl](https://x.com/charles_irl/status/2068124629433262210)
[^23]: [𝕏 post by @akshat_b](https://x.com/akshat_b/status/2068127465004814495)
[^24]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2068038020394021000)
[^25]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2067876984206537188)
[^26]: [𝕏 post by @AndrewCurran_](https://x.com/AndrewCurran_/status/2068041439481901157)
[^27]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2067955723204477206)
[^28]: [𝕏 post by @VikParuchuri](https://x.com/VikParuchuri/status/2067941596306231421)
[^29]: [𝕏 post by @guinnesschen](https://x.com/guinnesschen/status/2068062280345162047)
[^30]: [𝕏 post by @adcock_brett](https://x.com/adcock_brett/status/2068040783295627609)