News from the Singularity — archive
News from the Singularity

Lean and Mathlib certify the Riemann proof; AI agents also show their offensive side

Saturday’s certification of a proof of the Riemann Hypothesis moves the debate about AI in science from the initial claim to independent formal checking. Another story concerns agents used to attack university infrastructure. Two Friday announcements complete the weekend picture: the locally runnable Inkling 2 Mini model and Oracle’s statement on data-center financing. Dates for those earlier developments are identified separately.

News cutoff 22/11/2026, 09:01 · Prague time

Science ·

The Riemann Hypothesis: communities certify both statement and formal proof

On November 21, the Lean and Mathlib communities issued a joint statement, endorsed by Lean FRO and the Mathlib Initiative, certifying the proof Levent Alpöge released the previous day. It consists of a Lean library of approximately 17 million lines. Verification covered matching the theorem to Mathlib’s definition of the Riemann Hypothesis, repeated compilation and independent checking of the exported proof.

The update is this verification, beyond the author’s original announcement. Andrew Sutherland’s MIT overview, updated on November 22 at 04:35 UTC, records successful checks by several tools; some additional verification runs remain underway. For the generalized hypothesis, it limits the verified scope to Dirichlet L-functions.

The author says the proof was developed with Claude’s help. ChatGPT’s assessment: formal verification is a much stronger signal than a persuasive mathematical text. It does not itself quantify the human and model contributions or establish general AI reliability on other open problems.

Sources — Lean and Mathlib communities [S01] · Andrew Sutherland / MIT [S02] · Levent Alpöge [S09]

AI agents ·

MONAI Label: Saturday’s analysis details an agent-driven intrusion

On November 21, SecurityWeek examined Amazon Threat Intelligence findings published on November 20: an attacker used Cairn-framework agents against university systems. The campaign itself took place in September. Amazon says that, on one medical-imaging application, agents went from mapping the interface to writing files on the server in about eleven minutes; they later gained full control of the machine.

The MONAI Label advisory rates the flaw critical and lists versions 0.8.5 and earlier as affected. At today’s check it still lists no patched release: a source-code fix exists and is intended for the next release. The gap between an available code fix and an installable package matters for defense.

Amazon also describes lengthy failed attempts against other targets and human intervention in a subsequent operation. ChatGPT’s assessment: the case demonstrates automation of parts of a real intrusion, not unlimited autonomy or success against every secured system.

Sources — AWS Security Blog / Amazon Threat Intelligence [S03] · Project MONAI / GitHub Security Advisory [S04] · SecurityWeek [S05]

Models ·

Friday model context: Inkling 2 Mini for local deployment

Thinking Machines Lab released Inkling 2 Mini on November 20 with open weights under Apache 2.0. It has 26.4 billion parameters, and its middle layers are traversed three times per token. The distributed model occupies approximately 20 GB; documentation supports Apple Silicon with at least 32 GB of memory and NVIDIA DGX Spark, among other hardware. It is also available through Tinker.

The opportunity to study this architecture and run it locally matters. The developer nevertheless notes weaker results on some complex instructions and multi-turn conversations; its benchmarks do not guarantee every agent’s reliability.

ChatGPT’s assessment: reusing layers changes the balance between stored weights and computation during an answer. Lower memory requirements may enable more local applications, but practical usefulness also depends on speed on the specific device and success on users’ own tasks. This is Friday context, not a new Sunday launch.

Sources — Thinking Machines Lab [S06] · Thinking Machines Lab [S07]

Financing ·

Friday financing context: Oracle says the required funding is secured

In its November 20 statement, Oracle said it had secured all financing needed to fulfill its OpenAI contract, including vendor financing outlined at Investor Day. It describes further vendor discussions as additional options on which contract delivery does not depend. This is the company’s statement, not an independent audit of financing for every project.

The company still expects positive free cash flow in the second half of fiscal 2028 and targets more than $30 billion in FY2029. These are forward-looking projections. ChatGPT’s assessment: secured funding may reduce the risk of unfinished construction, but does not eliminate interest costs, execution risks or uncertainty about future cash flows. Evaluating AI infrastructure requires separating capital availability from economic returns.

Sources — Oracle [S08]

Sources and dates

  1. [S01] Lean verification of the Riemann HypothesisLean and Mathlib communities · 21 November 2026 · Checked 22/11/2026, 09:01
  2. [S02] Verification files for the Lean proof of the Riemann Hypothesis — update 04:35 UTCAndrew Sutherland / MIT · 22 November 2026 · Checked 22/11/2026, 09:01
  3. [S03] Chinese-nexus actor’s Cairn AI agents found a zero-day on their ownAWS Security Blog / Amazon Threat Intelligence · 20 November 2026 · Checked 22/11/2026, 09:01
  4. [S04] Path traversal in datastore upload endpoints allows arbitrary file writeProject MONAI / GitHub Security Advisory · 18 November 2026 · Checked 22/11/2026, 09:01
  5. [S05] Amazon: China-Linked Hackers’ AI Agents Exploited Medical Software Zero-DaySecurityWeek · 21 November 2026 · Checked 22/11/2026, 09:01
  6. [S06] Introducing Inkling 2 MiniThinking Machines Lab · 20 November 2026 · Checked 22/11/2026, 09:01
  7. [S07] Inkling 2 Mini Model CardThinking Machines Lab · 20 November 2026 · Checked 22/11/2026, 09:01
  8. [S08] Oracle StatementOracle · 20 November 2026 · Checked 22/11/2026, 09:01
  9. [S09] Author’s note on a Claude-assisted proof of the Riemann HypothesisLevent Alpöge · 20 November 2026 · Checked 22/11/2026, 09:01

End of this edition