Just Now, GPT-6 Astra Collaborates with Non-Professional Mathematicians to Prove 57-Year-Old "Circle Packing" Problem

marsbitPublished on 2026-10-12Last updated on 2026-10-12

Abstract

AI has made another breakthrough in mathematical proofs, with a mathematical enthusiast using GPT-6 Astra to significantly advance a 57-year-old geometry problem. The problem is known as "circle packing in a circle": given N unit circles, what is the smallest radius R of a larger containing circle that can hold them all without overlap? While optimal arrangements for small N have been known since 1969, providing rigorous mathematical proof of their optimality has remained a major challenge. Before this work, only cases up to N=14 and N=19 had been definitively proven. The enthusiast, known as nor, employed GPT-6 Astra to rigorously prove optimality for the remaining cases of N=15, 16, 17, 18, and 20. According to nor, the model's core strategy shifted from direct optimization to a constraint-based approach: first, establishing that outer disks can be moved to the boundary; then, using angular relations to eliminate vast numbers of potential arrangements; finally, employing computational enumeration and geometric rigidity arguments to prove that no arrangement can fit into a smaller containing circle than the known classic pattern. The proof process was iterative and complex, with nor contributing key guidance, verification, and refinement alongside the AI. The results, shared online, sparked interest from others exploring similar AI-assisted mathematical research, highlighting a growing community leveraging new tools for problem-solving. This work demonstrates the expandi...

AI is really entering a math proof sweep mode now.

Just now, a math enthusiast used GPT-6 Astra to take a major step forward on a 57-year-old math problem.

The problem sounds not very complicated:

Given N disks of radius 1, how do you pack them into a large circle so that the radius of the large circle is as small as possible?

In fact, mathematicians did come up with beautiful packing arrangements back in 1969, but...... for 57 years, no one has been able to rigorously prove that they are the optimal solutions.

So this time, math enthusiast "nor" pulled out GPT-6 Astra...... and in one go, provided rigorous proofs for the 5 remaining cases for N≤20: namely N=15, 16, 17, 18, and 20......

Previously, only cases for N≤14 and N=19 had been verified.

Oh, and reaching N=20 is not the AI's limit; it's just that nor didn't intend to prove further cases......

A Proof Stalled for Half a Century

Let's first formalize this problem in a more mathematical way.

Assume there are N disks, each of radius 1, and a large circle of radius R.

Requirements:

All small disks must lie entirely inside the large circle; they can be tangent to the large circle's boundary but cannot extend beyond it.

Any two small disks cannot overlap internally, but they are allowed to be tangent to each other.

The small disks can be moved and arranged arbitrarily within the large circle, with no restrictions on the arrangement.

The question is: For a given positive integer N, what is the smallest possible radius R for the large circle?

It looks like a simple olympiad-style problem, but in reality, finding the optimal arrangement is just the first layer. The true difficulty lies in rigorously proving mathematically that the arrangement has reached its limit and that no better packing exists.

In 1968, American mathematician Ronald Graham proved the cases for N≤7.

In 1969, researcher Udo Pirl pushed the proven range to 10 disks and further studied packing schemes for 11-19 disks.

But it wasn't until around 2003 that cases for N≤14 and N=19 were rigorously proven.

nor is a technical writer who has long been involved in research on mathematics, machine learning, and optimization algorithms, and has frequently participated publicly in AI-assisted mathematics research.

He personally started following this problem over a decade ago, and indeed, there hasn't been much progress on it in the past ten-plus years......

So what else is there to say? Might as well toss it to GPT-6 Astra (bushi).

nor took a step-by-step approach: first had GPT-6 Astra handle the N=15 case; after the model solved it, he had it extend the same method to 16, 17, as well as 18 and 20...... GPT-6 Astra successfully provided proof results for all of them.

It's just that......

The proof process is, as always, obscure and difficult to understand.

Simply put, the core idea of GPT-6 Astra is:

Instead of directly searching for the optimal solution, it uses geometric constraints to reduce the infinite number of possible arrangements to a finite set of cases, then uses a computer to eliminate them one by one, ultimately rigorously proving that the classic arrangement is optimal.

The proof process consists of 5 steps:

Push outer disks towards the boundary. Proves that under specific conditions, disks near the outer circle can be pushed outward without causing overlap. This divides the disks into a group on the boundary and a group in the interior.

Use angular relations to eliminate many arrangements. Since disks cannot overlap, they must maintain a certain angular separation. If a certain arrangement requires a total angular sum exceeding 360°, it can be directly eliminated.

Let the computer enumerate remaining possibilities. Instead of searching for the specific coordinates of each disk, it checks the distance intervals of the disks from the center and their cyclic order.

Use geometric rigidity to prove precise optimality. The previous search can only prove that all feasible arrangements are very close to the known classic arrangement. Next, using contact constraints between disks, it proves that these arrangements cannot undergo permissible tiny deformations, thus elevating "very close" to "strictly identical". The case for 15 disks uses another odd-cycle argument.

Restore the original arrangement. Finally, it proves that in the original arrangement, at least one disk must be in contact with the outer circle's boundary. This means it's impossible to pack all disks into an outer circle with a smaller radius, thus completing the optimality proof.

It's worth emphasizing that it's not like nor gave a single prompt and Astra immediately proved everything in one go.

In his blog post, nor writes that his contributions as a human in this work included:

Determining problem feasibility, guiding the model, battling the buggy web user interface, enduring poor copywriting, fixing subtle issues, repeatedly refining explanations, manually verifying validity, and re-checking everything using Astra and Opus 5.5.

One More Thing

A rather interesting thing is that once nor posted the results, he found quite a few fellow travelers using AI to research the same problem.

Some netizens also commented that AI's breakthroughs are making mathematics easier to understand, "I had no idea the math community could have such a strong exploratory atmosphere."

From the perspective of math enthusiasts, it's also a new phenomenon in the era of Mathematics 2.0.

Original links:

[1]https://nor-blog.pages.dev/posts/2026-10-11-circle-in-circle-packing-15-20/

[2]https://x.com/norxornor/status/2109078411251454017

This article is from the WeChat public account "QbitAI", author: Focus on Frontier Technology

Related Questions

QWhat is the main achievement described in the article regarding the use of GPT-6 Astra?

AThe article describes that a math enthusiast, using GPT-6 Astra, has successfully provided rigorous mathematical proofs for the optimal packing of N unit circles into a larger circle for the previously unproven cases of N = 15, 16, 17, 18, and 20, thereby advancing a 57-year-old mathematical problem.

QWhat is the 'circle packing' problem mentioned in the article?

AThe problem involves packing N unit circles (radius 1) into a larger circle of radius R. The goal is to find the smallest possible R such that all small circles lie completely inside the large circle without overlapping each other, though they can be tangent. The challenge is not only to find good arrangements but to mathematically prove they are optimal.

QWhat was the state of the problem before this AI-assisted proof?

ABefore this work, rigorous proofs existed only for N ≤ 14 and N = 19. The optimal arrangements for N = 15, 16, 17, 18, and 20, though conjectured, had not been formally proven to be the best possible for over five decades.

QWhat core strategy did GPT-6 Astra employ in its proof, according to the article?

AThe core strategy was to avoid a direct search for the optimal solution. Instead, GPT-6 Astra used geometric constraints to reduce the infinite set of possible arrangements to a finite number of cases. These remaining cases were then systematically eliminated using computer enumeration and geometric rigidity arguments to show the known classic arrangement must be optimal.

QWhat role did the human researcher 'nor' play in this project, as stated in the article?

AAccording to the article, 'nor' contributed by assessing the problem's feasibility, guiding the AI model, battling a buggy web interface, fixing subtle issues, repeatedly refining explanations, manually verifying the AI's output, and using both Astra and Opus 5.5 to double-check all the results.

Related Reads

BlackRock’s Bitcoin ETF snaps $2.35B outflow streak amid extreme fear

BlackRock's spot Bitcoin ETF (IBIT) returned to inflows on Monday, ending an 11-session outflow streak that totaled roughly $2.35 billion. IBIT attracted $76.5 million on Nov. 16, helping US spot Bitcoin ETFs record $7.9 million in net inflows despite withdrawals from other funds, per Farside Investors data. The inflows came as Bitcoin traded near multimonth lows of $64,000, while the Crypto Fear and Greed Index fell to 10 out of 100 on Monday, its lowest reading since June 2026. Nick Pitto, head of marketing at CryptoQuant, cautioned against interpreting the inflows as a broader market recovery. He noted that Monday's inflows recovered only about 3% of IBIT's withdrawals over the previous 11 sessions, while US Bitcoin ETFs had shed more than 31,000 BTC over the past month. "One day of inflows doesn't change the picture too much," Pitto said, adding that "it's too early to call this dip buying" and that several consecutive days of inflows from more funds than just BlackRock's would be needed. Monday also marked the deadline for most institutions to file third-quarter Form 13F disclosures, revealing Bitcoin ETF holdings as of Sept. 30. Several institutions reported increased exposure during the quarter, but Pitto cautioned against linking those holdings to Monday's inflows, noting that ETF buying peaked in August and September and the filings only show holdings at the end of September, not subsequent activity.

cointelegraph2m ago

BlackRock’s Bitcoin ETF snaps $2.35B outflow streak amid extreme fear

cointelegraph2m ago

Bitcoin Erases Nearly 90% of Its Gains: Is $62,000 the Bottom, or the End of a Bear Market Rally?

Bitcoin fell to $62,029 on November 15, a new low since July 13, down 12.6% for the week and about 29% from its September 21 high of $87,392—its sixth consecutive weekly decline. Bitfinex argues that position matters more than the drop itself: Bitcoin has re-entered the $61,300–$67,000 "July Base" range where it consolidated for nearly seven weeks before rallying. The key level is $57,803, the July 1 cycle low; holding above it preserves the higher-low structure, while a break would suggest the prior rally was merely a bear-market bounce. Roughly 1.4 million BTC have cost basis between $60,000–$65,000, versus only about 305,000 BTC between $55,000–$60,000, so a loss of the range could mean a rapid move toward $55,000–$60,000. Three sources of selling pressure persist: futures deleveraging (open interest fell about 8% to 645,750 BTC, with $2.01 billion liquidated in 24 hours, 83% longs), short-term holder capitulation (STH-SOPR at 0.971, lowest since June 5), and spot ETF outflows (seven straight days, $1.07 billion last week, $2.99 billion since October 30). Macro headwinds include rising real yields (10-year at 2.91%), sticky core CPI at 0.3% month-over-month for a third month, PPI at 6.5% year-over-year, and a dollar index at a 19-month high of 103.4, with markets pricing about 92% odds of a 25bp December hike. Options imply a weekly range of $61,800–$69,400. Bitfinex outlines three scenarios: holding the July Base but prolonged consolidation; reclaiming $72,700 with three days of ETF inflows and real yields below 2.8% to confirm a bottom; or breaking $61,300 and retesting $57,803. Key events include Nvidia earnings, FOMC minutes, and October PCE on November 25.

marsbit6m ago

Bitcoin Erases Nearly 90% of Its Gains: Is $62,000 the Bottom, or the End of a Bear Market Rally?

marsbit6m ago

GPT controlling a robotic arm is too dangerous: holding a kitchen knife over a human hand, dropping a phone into water

Recent advances in multimodal foundation models have enabled zero-shot robotic control, with GPT, Claude, Gemini, and Grok all supporting plug-and-play robot operation, making agentic robotics a promising path toward general-purpose embodied intelligence. However, capability does not equal safety. Researchers from XJTLU, Duke Kunshan, NUS, and other institutions propose RoboRisk, a robot safety benchmark with 92 simulated and 13 real-world tasks spanning five risk categories, plus LEASH (Lease-Enforced Agentic Safety Harness), a model-agnostic safety layer. Evaluations of six frontier models (GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 4 Argon, Grok 4.7) reveal harm rates from 9.06% to 76.92%, with failures including placing a phone into a bowl of water, carrying milk over a power strip, and moving a knife above a human hand. LEASH uses three safety gates (task, action, and commit) backed by a shared safety policy and fast decision models, enforcing least-privilege "leases" checked against live camera images. With LEASH, all six models show reduced harm rates; four models achieve zero harm on real-robot tasks (84.62% safe success rate), while overhead increases only about 1.66% in time and roughly $0.005 per run. Remaining limitations include imperfect reviewer reasoning and weak instruction-following in some models, motivating faster decision models and next-generation models that bridge the "human-presence gap."

marsbit6m ago

GPT controlling a robotic arm is too dangerous: holding a kitchen knife over a human hand, dropping a phone into water

marsbit6m ago

Bidding Farewell to the Profit Era: South Korean Crypto Exchange Upbit Reports Its First Loss-Making Financial Statement

Dunamu, the operator of South Korea's largest crypto exchange Upbit, reported its first quarterly operating loss since Upbit launched in 2017, according to its Q3 2026 financial report. Operating revenue fell 60.7% year-on-year to 151.7 billion won, with an operating loss of 3.5 billion won and a net loss of 18.6 billion won. The downturn stems mainly from cooling crypto trading activity. Total trading volume across South Korea's five won-denominated exchanges fell from about $218.8 billion in Q1 to $133.2 billion in Q3, a 38.9% decline. Upbit's Q3 trading volume dropped 7.57% quarter-on-quarter, while Bithumb fell 13.45%. Dunamu's cumulative revenue for the first three quarters fell 52.9% year-on-year, and operating profit plunged 86.2%. Bithumb's Q3 revenue dropped 69.7% year-on-year to 59.4 billion won, with an operating loss of 8.1 billion won and net losses for three consecutive quarters. A key vulnerability is the exchanges' heavy reliance on trading fees: fees accounted for 96.04% of Dunamu's total revenue in the first three quarters, and 99.99% for Bithumb. Efforts to expand institutional clients have slowed due to delayed policy liberalization, and some exchanges have scaled back corporate business teams. Meanwhile, regulatory pressure is mounting. In August, South Korea approved amendments strengthening registration and anti-money laundering requirements for virtual asset service providers, expanding the Travel Rule to all transaction amounts and requiring recipient exchanges to verify sender information. Transfers of 10 million won or more to overseas exchanges or personal wallets must be reported, with stricter limits on high-risk transactions. Dunamu's merger with Naver Financial also faces a rigorous competition review by the Fair Trade Commission, which launched final deliberations on November 10 over concerns about competition in payments, search, advertising, online shopping, financial data aggregation, and won stablecoin issuance. From its first quarterly operating loss to stalled institutional openness and tightening AML and shareholder oversight, South Korean crypto exchanges now face a dual test of operations and regulation.

marsbit17m ago

Bidding Farewell to the Profit Era: South Korean Crypto Exchange Upbit Reports Its First Loss-Making Financial Statement

marsbit17m ago

Trading

Spot
活动图片