November 7, 2026 tightened the link between proof and progress. A small research system settled a 2008 quantum conjecture. Infrastructure and security news followed.
PSI Proves the Aaronson-Ambainis Conjecture With an Anthropic Model
A 2008 Question About Quantum Speedups
Scott Aaronson and Andris Ambainis posed the question in 2008. It asks when quantum computers can be exponentially faster on random inputs. Quantum speedups like Shor's factoring rely on hidden structure. Random strings almost never have such structure.
Aaronson and Ambainis Reduced It to Pure Math
They stated a conjecture about polynomials and boolean functions. If true, an ordinary computer can imitate a quantum one. The number of input "looks" grows only polynomially in T. Exponential quantum speedups would then require rare input patterns.
PSI's System Found the Proof in Three Days
PSI developed the research system internally. It ran on PSI's own infrastructure. Non-public Anthropic model instances powered the run. The system held a complete proof by the third day. Lean, a proof assistant, checked every step.
The Bound and the Constant
PSI proved that influence is at least c · v^(5/2) / d^6, and at least (v/d)^7. The constant c equals 815 / (125π^4), about 0.0025. Earlier bounds for general polynomials decayed exponentially in the degree.
Independent Readers and Auditors
Guy Blanc, Jordan Docter, Carmen Strassle, and Li-Yang Tan settled a special case in August 2026. They are now reading PSI's full proof. They called the result a huge one. David Turturean led the work and is responsible for correctness. Jim Halverson audited the proof. Phiala Shanahan commented on the Lean formalization.
What It Means for Quantum Computing
Quantum computers cannot save exponentially many looks on yes-or-no questions. That holds unless the input has a rare pattern. One long-open question is now settled. The proof paper and Lean code are public.
Sources: PSI — A proof of the Aaronson-Ambainis conjecture
Google Acquires Crusoe's Wyoming Data Center Site
The Transaction Completes Crusoe's Pre-Development
Crusoe announced the sale on November 6, 2026. Google acquired Crusoe's site in southeast Wyoming. Crusoe finished site selection, approvals, and power strategy. Its CEO Chase Lochmiller framed Crusoe as activating gigawatts for intelligence. Crusoe reports over 6 gigawatts contracted and over 40 gigawatts in its pipeline.
Hybrid Power and a Carbon Capture Road Map
The site pairs fuel cells, grid power, and natural gas turbines. It sets a long-term road map for carbon capture. Tallgrass Energy's regional CO2 sequestration hub handles that capture. Crusoe and Tallgrass co-developed the hybrid power strategy.
Why This Follows the November 6 Campus Deal
On November 6, Google committed $17 billion to a Wyoming campus. That plan pairs 702 megawatts of Bloom fuel cells with capture. Our November 6 roundup covered that deal. Owning Crusoe's pre-developed site gives Google clearer control. The hybrid power approach remains the plan.
Crusoe Stays Busy Elsewhere
Crusoe is also developing Google's campus in Armstrong County, Texas. Its Paraguay project with Araico expands AI cloud capacity. Crusoe says it activates energy-first infrastructure for AI.
Sources: Crusoe — Wyoming Data Center Campus Update
Microsoft Ends Its 2023 Oracle AI Cloud Deal
Capacity Switches Off Over Older Nvidia Chips
Microsoft stopped renewing a 2023 Oracle contract. That contract originally powered AI features in Bing search. Microsoft switched off rented servers in recent months. Those servers ran older Nvidia chips that no longer match workloads. Oracle shares fell about 3% on the news. Oracle sits more than 50% below its June peak. Its free cash flow turned negative. Its credit rating sits one notch above junk.
Microsoft's Broader Neocloud Reassessment
Microsoft committed more than $80 billion to CoreWeave, Nscale, and Nebius. It turned down a Nscale expansion option in Texas. It also pulled out of a Nscale site in the UK. Oracle maintains demand remains strong, including for older chips. Microsoft declined to comment. Bloomberg could not determine the contract's full size.
The Silicon Generation Gap Drives Decisions
Gartner warns that much AI infrastructure spending faces large write-downs. Newer chips can strand older capacity. Lenders and operators now track chip generations closely. Hyperscalers walk away from signed capacity when silicon ages. Oracle's next earnings print is the near-term checkpoint.
Sources: AI Weekly alert (via The Next Web)
Ant Group Ships Ling-3.2-flash, a 560B Multimodal Model
A 560-Billion-Parameter Release With a Free Window
Ant Group released Ling-3.2-flash. It is a 560-billion-parameter multimodal model. Ant offers a two-week free access window. The publisher page lists a mixture-of-experts design. It lists 25 billion active parameters and a 1 million token context. China's labs keep shipping large open models.
Back-to-Back Chinese Lab Releases
Tencent shipped Hy4 one day earlier. Our November 6 roundup covered Hy4's pricing and WorkBuddy scores. Competitive pressure keeps rising. Both labs target cost-sensitive developers.
Why Developers Should Watch This
Free access lowers switching costs. Developers can benchmark Ling-3.2-flash against Hy4. Watch rate limits, licensing, and inference costs. Competition among open labs benefits users.
Sources: Pandaily (via AI Weekly) — Ant Group ships Ling-3.2-flash
A New Analysis Says Anthropic Has Already Crossed Its 'Dramatic Acceleration' Threshold
From Report to Implication
Ajeya Cotra published a November 6 analysis of METR's Anthropic assessment. Epoch Capabilities Index AECI scores grew about twice as fast. That spans October 2025 to October 2026. Anthropic's RSP threshold requires doubled progress attributable to AI automation. METR attributes roughly 80% of current acceleration to AI. Anthropic staff report roughly 12x productivity uplift. Cotra argues the threshold is already met.
The Compounding Risk
METR found no clear bottlenecks or steep diminishing returns. Diminishing returns could still slow compounding. Cotra's naive extrapolation implies superintelligence within a year. That assumes smooth acceleration. Reality likely hits efficiency ceilings.
Redaction Limits Trust
Anthropic retains unilateral redaction rights over METR's report. Much of the most informative evidence stayed private. Cotra argues security is not ready. ASL-4 security and alignment readiness lag behind the risk. Recent breaches by amateur groups underscore the gap. Transparent evidence would help align safety and expectations.
Sources: Ajeya Cotra — AI is dramatically accelerating AI research
OpenSSH Bug Opens Password-Disabled Linux Accounts
The '*sshole' Flaw on musl Systems
Attackers can abuse a flaw dubbed "*sshole". It affects OpenSSH on musl-based Linux systems. A password of 256 or more characters can log in. It works against accounts that had passwords disabled. Alpine and similar musl distributions face direct risk. Administrators should patch immediately.
What Admins Should Do Now
Update OpenSSH packages across all systems. Audit for password-disabled accounts. Review authentication configuration. Check vendor advisories for affected versions. Rotate keys if compromise is suspected. Test fallback login paths. The flaw highlights review debt in auth code.
Sources: AI Weekly alert — OpenSSH flaw
New Research Asks Whether AI Models Know Their Own Thoughts
Quanta Tests Metacognition With a New Probe
Quanta Magazine published a November 4 piece. It tests whether AI models know their own thoughts. The research probes internal state awareness. Results may shape how we monitor models. If models cannot report their processes, alignment checks get harder. The debate stays early.
Why It Matters for Trust and Safety
Better metacognition could improve self-audits. Weak metacognition limits chain-of-thought trust. Developers should not assume explanations are complete. Tools must verify, not just read model words. Safety teams gain from measurable probes.
Sources: Quanta Magazine — New Research Tests Whether AI Models Know Their Own Thoughts
Frequently Asked Questions
What Did PSI Prove?
PSI proved the Aaronson-Ambainis conjecture from 2008. A research system ran a non-public Anthropic model. Lean verified the proof. The system held it by day three.
Why Did Microsoft End Its Oracle Deal?
Older Nvidia chips no longer matched Microsoft's workloads. Microsoft reduced reliance on neocloud capacity. Oracle shares fell about 3%.
What Is Ling-3.2-flash?
It is Ant Group's 560-billion-parameter multimodal model. Ant offers a two-week free access window. Competition with Tencent Hy4 is heating up.
What Is the OpenSSH sshole Flaw?
A bug lets long passwords log into password-disabled accounts. It hits musl-based Linux systems. Patch OpenSSH and audit auth settings now.
What Does the Quanta Research Question?
It asks whether AI models know their own thoughts. The answer affects monitoring and alignment trust.
The Bottom Line
Proof and Infrastructure Converge
PSI's theorem and Cotra's acceleration reading share one theme. Capability growth now feeds on verification and AI labor. Control systems still lag behind capability.
Control Is the Open Gap
Google, Microsoft, and the OpenSSH bug show physical and software layers move fast. Safety planning must match that tempo. Redaction and weak monitoring make that harder.
What to Watch Next
Watch Anthropic's next model and METR's follow-up. Watch Oracle's earnings and chip-generation pricing. Watch PSI's public proof paper and new applications. Watch school and security impacts from prior Gemini and ARTEX stories.
Our November 6 roundup covered METR's Anthropic acceleration and GPT-6 Cyber. Today's proof and Cotra's analysis extend that same arc. The November 5 post tracked provable inference and PHP2Shell. Verification now decides trust.
Top comments (0)