Orchestration models 🤖, DeepMind exodus 👋, loop engineering 🔄
TLDR AI’s June 22 issue spotlights Sakana Fugu’s system for coordinating expert models, Mercury 2’s diffusion-based fast reasoning, and Nobel laureate John Jumper’s switch from DeepMind to Anthropic. It also examines audits of diffusion models, the US export controls halting Claude deployments, advances in loop engineering and robotics, and includes a European AI doomsday scenario alongside industry job notes.
- John Jumper (AlphaFold co-creator, Nobel laureate) is leaving DeepMind after nine years to join Anthropic, underscoring AI talent wars.
- US export controls have halted Anthropic's Claude Fable 5 and Mythos 5 deployments over a misidentified "jailbreak" that was actually a routine code-fix request, with a week of no resolution.
- New orchestration tools (Sakana Fugu/Fugu Ultra) and diffusion-based fast inference (Mercury 2 at ~1,000 tokens/sec) show diverging directions in model architecture—multi-agent coordination vs. raw speed.
- Loop engineering is emerging as a shift from one-shot AI coding prompts to iterative cycle-test-reprompt workflows, paralleled by NVIDIA's ENPIRE robotics framework automating similar refinement loops.
Sakana Fugu and its bigger sibling, Fugu Ultra, let you hit one OpenAI-compatible endpoint and tap a team of specialist models under the hood. They decide on the fly whether a request can be handled directly or needs multiple experts, then manage who does what and stitch the outputs back together. You get model selection, delegation, verification and synthesis without writing orchestration code yourself.
Inception Labs’ Mercury 2 pumps out around 1,000 tokens per second by borrowing diffusion tricks from image generators. It isn’t built for the hardest reasoning tasks but shines when you need huge volumes of straightforward text fast. It’s cloud-only, served via API.
John Jumper, who helped build AlphaFold and shared a Nobel Prize for protein-structure predictions, is leaving DeepMind after nine years to join Anthropic. DeepMind has struggled to turn its AI coding tools into business hits and this departure highlights the talent war in AI research.
On the audit front, researchers found Google’s DiffusionGemma remains as transparent as its predecessor, Gemma, despite a diffusion-based design. They flagged gaps between visible variables and hidden algorithm steps, and dug into oddities like non-chronological reasoning and token smearing.
Anthropic’s Claude Fable 5 and Mythos 5 got paused because the White House slapped export controls on them. A supposed jailbreak turned out to be a simple code-fix request, but regulators insist it needs “fixing.” It’s been a week with no progress.
In robotics, NVIDIA’s ENPIRE framework automates robot policy improvement by looping through resets, evaluations and refinements. On the software side, AI coding is shifting from one-shot prompts to loop engineering—cycle, test, re-prompt until you hit your goal.
Morph LLM speeds up codegen by training on code rather than general web text, achieving a 3Ă— boost in speculative decoding. Autoresearch tweaks GPU kernels to hit 162 tokens/sec on budget Nvidia and AMD cards, swapping NVLink for TCP-based cache sharing and slashing time-to-first-token by 84%.
TLDR is hiring a Senior Product Marketing Manager (remote, $180–225k base plus $40–50k bonus). And in a provocative think-piece called Europe 2031, Brussels strategists warn that underinvestment in datacenters could leave Europe trailing the US and China, fueling populism, financial chaos and cyber-attacks.