EinsteinArena lifts a Newton-era math bound from 593 to 604
Original: EinsteinArena is a platform where AI agents collaborate on open science problems — submitting solutions, posting in discussion threads, building on each other's constructions in real time. Agents just improved a math problem that's been open since Newton. Kissing Number in dimension 11: 593 → 604. View original →
What the tweet revealed
Most AI benchmark posts talk about held-out datasets. This one claimed movement on an open mathematical problem that predates modern computing. In its April 13, 2026 X post, Together AI wrote:
“Agents just improved a math problem that's been open since Newton. Kissing Number in dimension 11: 593 → 604.”
The claim matters because 593 → 604 is not a marketing score. It is a lower bound for the 11-dimensional kissing number, a classical packing problem with roots going back to Isaac Newton. Together's thread says the result came from EinsteinArena, a platform where agents submit constructions, leave discussion traces, and build on one another's work in real time.
Why it stands out
The Together AI account usually uses X for model rollouts, inference infrastructure, and benchmark snapshots around its cloud platform. This time the post pointed users to a research blog and the live EinsteinArena leaderboard. The companion write-up says agents had already produced 11 new state-of-the-art results on open problems as of April 11, 2026, and describes the platform as a public system with discussion threads, verifiers, and live ranking instead of a manually maintained note page.
The follow-up tweets add the technical detail missing from the headline. Together says one agent first submitted a construction with slight sphere overlap, other agents picked it up and optimized it, an LSQR step reduced overlap loss from 1e-13 to 1e-50, and a final integer-snapping step produced a verified solution with 604 spheres. That sequence matters because it suggests the gain came from iterative repair and collaboration rather than one lucky sample.
What to watch next is external verification and repeatability. Open math claims travel farther when other researchers inspect the verifier, reproduce the construction, and see whether the platform can keep extending bounds on more than one problem. If EinsteinArena continues generating improvements in public, this post may end up looking less like a flashy demo and more like early evidence that multi-agent systems can contribute measurable work to live scientific search.
Sources: Together AI X post · EinsteinArena blog · EinsteinArena leaderboard
Related Articles
Long-document OCR is bottlenecked by page chunking and growing KV cache. A widely shared post says Baidu’s Unlimited-OCR uses 3B total parameters, 500M active parameters, and a 32K context window to read 40-page documents in one pass.
Enterprise agents are moving from demos to operating metrics. OpenAI says Presence resolves 75% of inbound issues in its English phone support channel without human help, with a Codex improvement loop cutting handoffs by 15 percentage points in 10 days.
AI infrastructure competition is being measured in training throughput, not just chip availability. NVIDIA says Blackwell Ultra reached 1,648 TFLOPs per GPU on DeepSeek-V3 671B, about 3x prior delivered performance.