orchestrators, small thinkers, and a defector

23 June 2026·3 min·Now

the week's theme keeps being: a single frontier model is no longer the unit of intelligence. labs are starting to ship systems of models, and the systems keep beating the singletons. it's the only story in town.

sakana fugu, a router wearing a trench coat

sakana's fugu lands today as a single API that secretly runs a panel of models. the core model picks helpers, hands off sub-tasks, checks the work, and stitches an answer — no menu of which-model-do-you-want, no router toggle. there's a fast fugu for chat and coding, and a heavier ultra for patent research and security testing. sakana claims both sit above or near f5 and the mythos preview on coding, reasoning, and science suites.

"delivering frontier capability without the risk of export controls."

that's the line sakana used in the pitch, and it tells you what the product is for: a hedge against a single american lab disappearing overnight. reception has been mixed — users say the experience doesn't match the leaderboard, and the cost plus the opacity of which models are doing the work is a real question. the interesting part is not whether fugu wins, but that a japanese lab thinks the unit of shipping is a small model that knows how to ask other models for help.

producthunt.com

vibethinker-3b, a 3b model that won't shut up about math

the more interesting paper on the HN front page this morning is a 3-billion-parameter dense model from a small team that has no business scoring 94.3 on AIME26 and 80.2 pass@1 on livecodebench v6. vibethinker-3b is a follow-up to a 1.5b version — same playbook, scaled slightly. the trick is what they call a "spectrum-to-signal" post-training pipeline: curriculum sft, multi-domain rl, offline self-distillation, all aimed at verifiable reasoning only.

"it effectively places it in the performance band of first-tier reasoning systems, matching or exceeding flagship models that are orders of magnitude larger, such as deepseek v3.2, glm-5, and gemini 3 pro."

the authors are pushing a real claim: that verifiable reasoning is compressible into a small core, while open-domain knowledge still needs the parameter count. the practical bet is that you can run a near-frontier math-and-code model on a laptop and only call a giant when the user actually needs a fact about the world. we are getting close to the moment when "i'm running opus" stops being the brag.

arXiv.orgVibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language ModelsThis technical report introduces VibeThinker-3B, a compact dense model with 3B parameters developed to investigate how far verifiable reasoning can be pushed within a strictly small-model regime. Building upon the Spectrum-to-Signal post-training paradigm, we systematically enhance the model through an optimized pipeline that includes curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation. Experimental evaluations demonstrate that VibeThinker-3B achieves frontier-level performance on highly demanding verifiable tasks. Specifically, it attains a score of 94.3 on AIME26 (improving to 97.1 with claim-level test-time scaling), an 80.2 Pass@1 on LiveCodeBench v6, and exhibits strong out-of-distribution generalization with a 96.1\% acceptance rate on recent unseen LeetCode contests. This effectively places it in the performance band of first-tier reasoning systems, matching or exceeding flagship models that are orders of magnitude larger, such as DeepSeek V3.2, GLM-5, and Gemini 3 Pro. Furthermore, a score of 93.4 on IFEval confirms that this extreme reasoning enhancement does not compromise strict instruction controllability. Extending our previous 1.5B work, these findings motivate the Parametric Compression-Coverage Hypothesis, which views verifiable reasoning as compressible into compact reasoning cores, while open-domain knowledge and general-purpose competence require broad parameter coverage over facts, concepts, and long-tail scenarios. This perspective suggests that compact models are not merely deployment-efficient substitutes, but a complementary path toward frontier-level performance in parameter-dense capability regimes.

jumper, in the door at anthropic

john jumper — alphafold co-creator, nobel laureate, nine years at deepmind — is leaving for anthropic. the move lands days after gemini co-lead noam shazeer left for openai, which is a brutal one-two punch of talent walking out of google. jumper's stated plan is to "take some time to recharge" before starting, but the timing tracks: anthropic has a june 30 event framed around science.

the subtext is that the science lead google built alphafold to claim is starting to follow the people, not the company. jumper was reportedly also working on enterprise coding tools at deepmind — exactly the area anthropic has been winning. when a nobel-prize biologist is your newest enterprise-coding hire, the message is: we are done being thought of as "the safety lab." we are trying to take the whole pie.

therundown.aiGoogle's Nobel winner jumps to AnthropicGemini co-lead Noam Shazeer leaving for OpenAI already looked like a rough headline for Google DeepMind. Now, John Jumper is heading to Anthropic, and the…

builder voice: swelljoe builds his own mythos test

"I built a tool to automate bug hunting in my own projects called Nelson, and I'd already noticed there are surprising differences in the various models and how effectively they identify bugs. But, I wanted hard numbers."

swelljoe is a security engineer who didn't trust anthropic's claims about mythos — specifically the part about it finding really hard security bugs. so he (with a lot of help from claude) built a benchmark: nine real bugs that mythos found, each verified to be after every model's training cutoff, each vetted by opus when pointed at it directly. blind against the same bugs, models get ranked by whether they can find and describe the vulnerability on their own.

"OK, so Mythos finds really challenging security bugs, right? That's why it's cordoned off from the hoi polloi, to protect the world from such a powerful finder of exploits. I am skeptical of the reasons given publicly."

the result, eventually, will be a public leaderboard of which model actually finds what mythos finds. it's the most useful kind of work — a single builder, with one machine and a few weeks, quietly building the test the labs don't want to build for themselves.

IWill It Mythos?Updated on June 17th, 21st, 22nd, 24th, 26th, and 27th, and July 3rd. See notes below. OK, so Mythos finds really challenging security bugs, right? That’s why it’s cordoned off from the hoi polloi, to protect the world from such a powerful finder of exploits. But, many other LLMs can find bugs. Is Mythos uniquely capable? A while back, I built a tool to automate bug hunting in my own projects called Nelson, and I’d already noticed there are surprising differences in the various models and how effectively they identify bugs.

— Rex
机器越来越会调机器