The Infinite Loop

Welcome to The Infinite Loop, the editorial arm of @nebiusai. We exist to provide a platform for insights from across the AI ecosystem — showcasing the researchers, founders and engineers who are pushing technology into new territories. Start exploring: infiniteloop.media

"Our onion algorithm is very strong" - @ecorobotix, on why their sprayers don't kill your dinner. @DroidFarm's solar-powered farmbots: planting and weeding 24/7. @timmpsmith breaks down the ROI: infiniteloop.media/p/farmbots-bri…

@profjamesevans, co-author of the Nature study, says it’s cheaper to use AI to squeeze out more output from existing data. The downside? Less discovery. His fix would horrify most funders. Read why in Sage Lazzaro’s piece: infiniteloop.media/p/ai-scientist…

Among AI threats to keep you up at night, “scientific monoculture” doesn’t top the list. Should you be worried? A study published in Nature suggests it’s not good. “AI scientists” are accelerating research while shrinking the range of topics. More papers. Less originality. 1/5

Even more futuristic: @lecong is combining AI scientists with XR glasses and robotics so the AI can literally design experiments, watch them happen, catch errors in real time, and run procedures itself. It’s already in 5 university labs. 4/5

@jrpenades gave Google's Co-Scientist a hypothesis his team spent years solving. The AI cracked it in two days. Meanwhile @SakanaAILabs got a fully AI-written paper through peer review. 3/5

It's not intentional. AI works best where there's already lots of data and existing research to draw on. Scientists drift toward those areas for faster results and quicker recognition. It’s just math. 2/5

@levie Using a mix of models makes sense. What's less clear is what it means for scientific research. Running multiple models in parallel may make existing research faster without broadening the topics, because the fields with the most data will still attract the most attention.

Another new idea to push the state of AI architectures forward. Sakana released a model that effectively uses a mixture of models to get work done. You get a single API but then the work gets farmed out the model that best performs the task. “Fugu manages model selection, delegation, verification, and synthesis automatically. It solves tasks directly when that is enough, or coordinates a team of expert models when a problem calls for more. The complexity of a multi-agent system never reaches your code.” This is generally how applied AI products are building their agent harnesses at this point, but the idea of making this an LLM that any developer can interact with is also a great idea. As we get more innovation with both frontier closed and OSS models, there’s going to be a ton of value produced for the layer that can route the best.

@TheTuringPost @AnthropicAI @SakanaAILabs Wild: 80% of Anthropic's code is Claude-written. RSI moves fast in ML due to data density, yet that same efficiency narrows exploration. It's ironic how our most powerful tools are consistently pointed at problems we've already largely solved.

AI that builds AI - 3 early steps of Recursive Self-Improvement (RSI) ▪️@AnthropicAI: 80% of the code merged into their codebase was authored by Claude ▪️@SakanaAILabs - RSI is their mission. With research like The AI Scientist and Darwin Gödel Machine, they already have one of the strongests foundation for RSI ▪️ @Recursive_SI is automating the research loop itself with the Recursive system, generating and testing improvements to models, training recipes, and GPU kernels. Here is a full guild to what is RSI exactly, how it works in these 3 cases and how they transform research loops today: turingpost.com/p/what-is-recu…

@hardmaru Everyone will focus on the Fable/Mythos numbers. The more interesting line is what happens when a whole country's science runs on one model. We reported on Sakana's work recently, and this is one of the quieter risks in the AI boom that nobody's really talking about yet.

Human intelligence is fundamentally a collective intelligence. We solve complex problems by participating in a vast cultural network that builds upon ideas across generations. I believe the strongest AI systems will become a collective intelligence, too. Since we started Sakana AI, our core conviction has been that the most powerful AI systems will be collaborative ecosystems, not isolated monoliths. Evolution innovates under constraints, and the future belongs to systems that explicitly learn how to coordinate collective intelligence. Today, we are taking a major step toward that future with the launch of Sakana Fugu. Fugu dynamically orchestrates the world’s best models to tackle complex tasks. We are proving that a well-orchestrated pool of swappable agents can match restricted frontier models like Fable and Mythos. But Fugu is about more than just performance. I believe that Orchestration Models are the next frontier, beyond bigger models. Relying on a single company’s model for national infrastructure is a massive risk. As recent export controls have shown, access to top models can disappear overnight. Collective intelligence is the practical hedge against this concentration of power. Fugu simply routes around vendor restrictions by relying on an entirely swappable agent pool. I am incredibly proud of our Tokyo team for shipping this. By orchestrating the world’s models, we are delivering the resilient blueprint required for AI sovereignty. Read our full vision and results here: sakana.ai/fugu-release 🐡

@iam_elias1 The "winner is the thing that uses all of them" framing is right. What's less clear is whether that changes what gets researched, since models tend to check out things with more data and not unexplored routes as much. Speed and diversity are different problems.

I did not expect this to come from Japan. Japanese AI Lab just dropped a Mythos/Fable- level model. While every lab in America raced to build a bigger model, a quiet team in Tokyo built something stranger and possibly smarter. A single AI that commands all the other AIs. And Sakana is claiming it reaches Mythos-level performance, the export-controlled tier most of the world can no longer even access. It is called Sakana Fugu. It launched today, June 22, 2026. And the idea behind it is the most contrarian bet in AI right now. Here is what it actually is. Fugu is a multi-agent orchestration system that presents itself as a single API. You send one call. Internally, it routes your task across a pool of the world's best frontier models, dynamically, adaptively, with no hardcoded rules. A team of expert models does the work. You get back one answer, better than any single one of them could produce alone. Here is the part that makes it genuinely clever. Fugu is not a router built with if/else logic. Fugu is itself a trained language model, one trained to understand when to delegate, how the agents should communicate, and how to combine their outputs into something reliable. It can even call other models, including copies of itself, recursively. The orchestration is learned, not hardcoded. It is grounded in two ICLR 2026 papers from Sakana TRINITY, an evolved LLM coordinator, and Conductor, on learning to orchestrate agents in natural language. This is not prompt engineering dressed up as a product. It is a real architectural bet. Here is what the verified benchmarks actually show. Against the best publicly accessible models, Opus 4.8, Gemini 3.1 Pro, and GPT-5.5 Fugu Ultra leads or ties on 8 of 10 benchmarks. It posts 95.5 on GPQA-D, 93.2 on LiveCodeBench, 90.8 on LiveCodeBench Pro, 82.1 on TerminalBench 2.1, and edges Opus 4.8 on Humanity's Last Exam. The honest part: the wins are not universal. GPT-5.5 still leads on long-context recall. Opus 4.8 leads on cybersecurity. And one genuinely interesting quirk on a few tasks the lighter, balanced Fugu model actually beats Fugu Ultra. Sometimes more orchestration just adds noise. And here is why the timing matters more than the benchmarks. On June 12, 2026, Anthropic's Fable 5 and Mythos became inaccessible to most of the world overnight, locked behind national-security export controls. Ten days later, Sakana shipped a system designed to make that kind of lockout irrelevant. Because Fugu's model pool is swappable, if any single provider gets export-controlled tomorrow, Fugu just routes around it. No dependency on any one lab. No single point of failure. The orchestration layer is the hedge. That is the real story here. Everyone else is racing to build the one model that wins. A Tokyo lab just bet that the winner will not be a model at all it will be the thing that knows how to use all of them. And they might be right.

@KanikaBK What's interesting about Sakana is that meaningful AI research is starting to happen outside the usual zip codes. More geographic diversity in who's building this stuff is probably the best shot we have at actually diversifying what gets researched.

A Tokyo Lab just ended the AI MONOPOLY… When Fable 5 and Mythos became inaccessible two weeks ago everyone panicked. Now, we have Sakana Fugu, Built in Tokyo. CONTROLLED BY NOBODY…

@DeryaTR_ We're watching to see how this unfolds. The benchmarks are awesome, but seeing how it performs as a real lab utility is where it gets interesting.

Fugu model from Sakana AI looks amazing! If these benchmarks translate to real-world work, it would have a massive impact on frontier AI!

@ycombinator @matforge_ Matforge is a perfect example of something we reported on recently! 🔥 The fields with the most data are where AI scientists land first, and semiconductor materials is a textbook case.

Matforge (@matforge_) is building AI scientists to discover new materials for the semiconductor industry. The semiconductor industry needs better materials to bring back Moore’s Law, but finding novel materials today takes 10+ years of lab work. Matforge wants to accelerate this process. Congrats on the launch, @advaith_sridhar & @Akash__Ramdas! ycombinator.com/launches/Pz0-m…

@drfeifei Hard to argue with this. The wild part is a Nature study this year found AI tools are making individual scientists more successful AND shrinking the diversity of what gets researched overall. Better tools, but a narrower map.

Scientific research is fundamental to advancing civilization and helping people globally to solve the most critical problems, from medicine to materials, from brain science to physics, and much beyond. This is only possible when scientists have access to the best tools of the time to conduct scientific research, including having access to AI-based tools.

@ChrissGPT The benchmark table will get all the attention. The more interesting story is what orchestration actually means for research diversity. Swapping models in and out will probably just make the existing research loops faster, not more varied.

So I read the whole Sakana paper going in with the question “why wouldn’t I just use Fable 5, Mythos, or GPT-5.5 directly?” For a single clean prompt, you probably would. But it’s obvious that the messier the task is, whether it involves delegation, verification, synthesis, code review, research loops, security analysis, patent/literature search, or anything where different models have different strengths, the more it would make sense to use this. A lot of people are asking what Fugu is, and the answer is essentially an API that behaves like a model, while internally deciding which agents to call, when to recurse, when to verify, and how to route around weak or unavailable models. So the reason they bring up “export controls” is because if frontier AI access is removed because of vendor policy or regulation, then orchestration can switch different models in and out, with potentially worse performance ofc, but not as much as fully switching to another model from scratch. The benchmark table is also pretty strong HLE: 50.0 GPQA-D: 95.5 SWE Bench Pro: 73.7 TerminalBench 2.1: 82.1

@michellearning 100s of robots and not one single passive aggressive Slack message. The dream lab 🤖

Welcome to the scientific revolution. 100s of robots. Zero coffee breaks. America’s largest autonomous lab, open today.

AI agents will need reliable data, says Bou: “The Internet isn’t the best repository of information we have as humanity.” @abracrioca has the full story: infiniteloop.media/p/building-the…

There’s a lot of growing to do. "We're pushing vector databases to their limits — building systems that need to be 10 to 20 times the size of Wikipedia." By the end of the year, Cala projects its graph will hold up to a billion data entities. 4/5

She raised the largest pre-seed in Spain’s tech history and said nothing about it for over a year — @elisenda_bou had a product to build. It's not her first rodeo: Apple acquired her previous startup Vilynx in 2020 for its knowledge graph technology. 1/5