AI Industry Heat Rankings

NO.01爆
平稳

Lean Community Launches Independent Verification of Riemann Hypothesis Formal Proof

Anthropic mathematician Levent Alpöge claims to have produced a Lean formal proof of the Riemann Hypothesis, with Claude contributing key insights. The proof, spanning about 17 million lines of code, has been made public and is under intense scrutiny by the Lean community. The Riemann Hypothesis, open since 1859, is one of the Millennium Prize Problems. Stanford researcher Alessandro Masullo has also released an independent human-written proof, and the validity of both remains to be confirmed by peer review. According to reports, Claude provided crucial steps in the later stages, including continuous Theta Lifts and core parameters, and the entire argument was formalized in about six weeks. Terence Tao noted that the proof likely relies on modern results from the Langlands program and arithmetic geometry, and verification will take time. Initial checks confirm the formalized statement is the Generalized Riemann Hypothesis, but theoretical risks about Lean's kernel reliability remain. MIT's Andrew Sutherland independently verified the proof using Claude Code: all 175,187 modules compiled successfully, the five core theorems depend only on three standard axioms, and two independent kernels, nanoda and con-leche, accepted the full exported proof. If the proof ultimately passes peer review, it will be a milestone for AI-assisted mathematical research, supporting hundreds of results that depend on the Riemann Hypothesis, highlighting AI's actual participation in top-tier mathematics, and advancing mechanisms for trustworthy verification of AI-generated results.

Update:The Lean community is intensively verifying Alpöge's proof. MIT's Andrew Sutherland has completed an independent verification using Claude Code: all 175,187 modules compiled, and two independent kernels accepted the proof. Meanwhile, a Stanford researcher released a human-written competing proof. Final confirmation awaits peer review.
24h Heat Trend102
24h agoNow
11 Sources
View Story Timeline→
NO.02爆
平稳

Thinking Machines Open-Sources 26B Looped Model Inkling 2 Mini

Thinking Machines Lab has open-sourced Inkling 2 Mini under the Apache 2.0 license, a multimodal mixture-of-experts model with 26B total parameters that uses a looped Transformer architecture. The middle 12 layers share weights and are executed three times, giving each token an effective 48 layers of computation and extending internal reasoning without increasing output tokens. The model supports text and image inputs with a 1M token context window, and its checkpoint is only about 20GB, allowing it to run locally on a 32GB Mac or DGX Spark. Sources agree on the architecture: each loop iteration maintains its own KV cache and reads gating messages from the previous round, with activated parameters ranging from 3.4B to 5.4B (depending on counting method) and only about 3.1B unique weights per token. In performance, it leads open-source models under 100B on agentic coding benchmarks, scoring 48.2% on DeepSWE v1.1, 79.2% on SWE-bench Verified, and 20.1 on the AA Index—matching Qwen3.8 27B with 5-6x fewer activated parameters and about one-sixth the inference cost. However, knowledge recall is a clear trade-off, with a non-hallucination rate of only 63.4%. vLLM offers Day-0 support on Blackwell, Hopper, and DGX Spark, with built-in DSpark speculative decoding for acceleration. The release marks a significant milestone for looped architectures in open-source models, achieving intelligence comparable to larger models with very low activated parameters and greatly lowering the barrier to local deployment. Its strong agentic coding and long-context capabilities open new possibilities for long-horizon task automation on personal devices, potentially advancing a local-first AI application ecosystem.

9 Sources
96
NO.03新
回落

Anthropic Raises $93.5 Billion in Record-Breaking Nasdaq Debut

Anthropic officially listed on Nasdaq on Thursday, raising approximately $93.5 billion at an IPO price of $83 per share, surpassing SpaceX's $86.25 billion record to become the largest IPO in history. The stock opened at $111.50, a 33% premium to the offer price, pushing its fully diluted valuation above $1.5 trillion. The offering was about 6 times oversubscribed, with retail allocation increased to 20%. In addition to the 825 million Class A shares in the public offering, seven giants—Google, Nvidia, Microsoft, AMD, Broadcom, Micron, and TSMC—participated in a $26 billion private placement at the same price, with Nvidia alone subscribing $10 billion. The company's annualized revenue rose from $71 billion in September to $81 billion by late October, underpinning strong demand. CEO Dario Amodei said the primary goal of going public is to stockpile cash to fund $900 billion in compute infrastructure, including a data center agreement with SpaceX worth up to $3.1 billion per month. Anthropic's listing marks the first time a leading large-model company has entered public markets, providing a key pricing anchor for AI industry valuations and commercialization. Its roughly $220 billion cash reserve will intensify the compute arms race with rivals like OpenAI, while also fueling the debate over an AI bubble.

8 Sources
56
Clustered Hot Stories (NO.04 - 10)
  1. 04

    OpenAI Hardens Reasoning Defenses, Cuts Extraction Attack Success to 0.2%

    OpenAI has systematically disclosed a multi-layered defense framework against adversarial distillation, achieving significant improvements: the success rate of extracting protected reasoning from historical requests dropped from 54.0% to 0.2%, and for active requests from 22.9% to 0.9%. The defense covers frontier models including GPT-6.2 Astra and Sol, with some improvements backported to GPT-6 Astra. The framework combines three layers: per-request adversarial training, input/output classifiers, and API-level reasoning isolation; cross-session and cross-organization reasoning reuse restrictions; and batch account bans for coordinated attacks. OpenAI takes enforcement action against thousands to tens of thousands of unauthorized distillation accounts weekly and shares attack indicators with other vendors. The official blog notes that full reasoning traces are more valuable for distillation than summaries on SWE-bench Verified and frontier cybersecurity tasks, so defenses must cover reasoning chain outputs. This disclosure carries dual significance: adversarial distillation allows attackers to replicate frontier capabilities without safety guardrails, posing real abuse risks, and OpenAI's framework offers reusable cross-industry lessons. Meanwhile, Arena.ai's updated Alignment Index shows OpenAI sweeping the top seven spots, with Claude Fable 5.2 debuting eighth and Kimi K3.1 Pro the highest-ranked open-source model at ninth, underscoring that safety and alignment are becoming key competitive dimensions.

    5 Sources46
  2. 05

    OpenAI Releases Prompt Templates for GPT-6.2 Astra and Sol

    OpenAI has released system prompt templates for GPT-6.2 Astra and Sol, designed to correct behaviors such as overly brief default responses and frequent clarifying questions that interrupt workflows. The templates cover writing style, proactivity, and coding guidelines, with the coding template explicitly requiring reuse of existing code, avoidance of unnecessary abstraction, and adding tests only when explicitly requested or when specific errors occur. Meanwhile, the API has undergone significant changes: editing conversation history now discards previous reasoning steps, affecting token costs and response quality, and parameters like temperature and top_p have been removed. According to Arena.ai's Agent Arena Index, GPT-6.2 Astra's coding score rose from 1,288 to 1,338, jumping from 8th to 3rd place, while its chat score fell from 1,265 to 1,211, dropping from 8th to 11th. These changes nearly cancel out in the overall score, with the composite ranking only one spot higher than GPT-6 Astra. The release reflects OpenAI's effort to guide users toward standardized usage via prompt templates while improving coding capabilities, compensating for weaker chat performance. The removal of API parameters and discarding of reasoning steps may impact developer costs and output stability, while the significant coding improvement could attract more code-generation use cases. Overall, the capability divergence between professional tasks and general conversation may become a key focus for future iterations.

    2 Sources36
  3. 06

    Tencent Yuanbao and KCL Launch Group Chat Agent Memory Benchmark

    Tencent Yuanbao and the NLP team at King's College London have jointly released GroupAssistBench, the first benchmark for evaluating agent memory in group chat scenarios. Built from 50 real group chats with 80,299 historical messages, the benchmark comprises 2,395 evaluation questions covering single-memory recall, multi-memory integration, and update conflicts. It is the first to split memory capability into two levels: L1 factual recall and L2 open-ended service. Experiments across 14 configurations reveal that stronger memory recall does not equal stronger service capability. For instance, EverMemOS scores 0.621 on L1, higher than full context's 0.570, but only 0.247 on L2, far below the latter's 0.431. This indicates that retrieval-augmented and external memory approaches generally lag behind full context on open-ended service tasks, offering a critical warning for agent memory system design. The paper, code, and dataset are all publicly available, which is expected to promote standardization and reproducibility in group chat agent memory research. As AI assistants increasingly integrate into group collaboration, balancing memory recall and service capability will become a core issue for developers, and this work points the way forward for future optimization.

    2 Sources24
  4. 07

    GPT-6.2 Astra Doubles Embodied Task Success Rate to 41.2%

    GPT-6.2 Astra has achieved a major breakthrough in robotic manipulation benchmarks, doubling its overall success rate from 19.6% (GPT-6 Astra two months ago) to 41.2%. In XDOF's Ember-0 dual-arm robot benchmark of 16 tasks, the improvement spans the entire task set rather than isolated successes. Notably, two flexible-object tasks—knotting a rope and unfolding a T-shirt—were solved by a model for the first time, with previous 143 T-shirt tests all failing. Cross-validation reveals that GPT-6.2 Astra also excels in motion planning efficiency. On Artificial Analysis's AA-Robotics v0.1 benchmark, it issued only 668 commands across six tasks—the fewest among 11 models—while achieving the highest score. In the box-clearing task, it used just 92 arm commands, about 75% fewer than Claude Fable 5.2. XDOF's evaluation, using six YAM workstations in parallel with 30 repetitions per task and manual error annotation, is rigorous and reliable. This progress signals a shift from single-point demos to general-purpose task capabilities in embodied AI. Flexible-object manipulation has long been a challenge; this breakthrough will accelerate deployment in homes and industries. Improved action efficiency also reduces deployment costs, paving the way for large-scale adoption. XDOF notes the benchmark may soon be saturated by model capabilities, implying faster-than-expected iteration.

    2 Sources24
  5. 08

    Nvidia Sheds $2.73 Trillion as Nasdaq Posts Worst Week Since 2020

    Nvidia fell another 14% on Friday, extending a roughly 47% drawdown from its October peak and wiping out more than $2.73 trillion in market value, larger than all of Amazon. The selloff was triggered by CEO Jensen Huang's resignation amid a scandal, compounded by doubts over AI valuations and capex returns, sending the Nasdaq down 10% for its worst week since March 2020.

    3 Sources23
  6. 09

    OpenAI Releases MedChemBench Benchmark Revealing Model Decision-Making Gaps

    OpenAI has released MedChemBench, a benchmark for evaluating AI models' reasoning in real-world drug discovery tasks. The benchmark comprises 99 questions derived from 117 papers and patents, covering 761 graded molecular structures, and truncates at real decision points while hiding subsequent experimental results, requiring models to make judgments based on biological targets, synthesized compounds, and partial assay data. Test results show that the strongest model, GPT-6.2 Astra, achieves an overall score of about 84.5% at the xhigh reasoning level, with near-perfect molecular structure reconstruction at 95.4%, but only about 50% accuracy on key decisions such as compound screening and numerical property prediction. In contrast, GPT-5.6 Sol scores only 58.2%, indicating significant gaps between models. OpenAI emphasizes that these results do not prove models can discover new drugs, but serve to identify capability boundaries and shortcomings. The launch of MedChemBench provides a quantifiable evaluation tool for AI applications in medicinal chemistry, revealing a notable gap between structural interpretation and real-world decision-making in frontier models. This helps guide future model iterations to focus on improving causal reasoning and integration of experimental evidence, serving as a bellwether for the AI-assisted drug discovery field.

    3 Sources22
  7. 10

    Google Launches Gemini 4 Neon to Rank Second on Agent Arena

    Google DeepMind has released Gemini 4 Neon, the fourth model in the Gemini 4 series in just seven weeks, positioned as a mid-tier model balancing frontier intelligence with fast end-to-end task completion. It ranks second on the Agent Arena leaderboard with a score of 1336, Google's best result on that benchmark, at a median cost of only $2.82 per task—69% cheaper than the top-ranked Claude Fable 5.2, offering a significant price-performance advantage. Technically, Neon supports a 1-million-token context and four reasoning levels, topping the Artificial Analysis coding agent index with 58.6 points, surpassing Claude Opus 5.5, and ranking third on the intelligence index with 61 points. It runs at about 110 tokens per second, twice as fast as the previous Carbon model, with lower base pricing. In the Antigravity development environment, Neon can orchestrate multiple sub-agents for parallel long-task execution, boosting SWE-sweep from 12.8% to 39.0%. Security-wise, the model card claims 100% blocking of indirect prompt injection attacks on the Gray-Swan benchmark. Neon's release signals an accelerating cadence for Google's model iterations, with intervals shrinking from 28 days to 6 days, while the product line consolidates—Argon retired and older Pro and Flash models discontinued. By balancing speed, cost, and steerability, Neon directly challenges Anthropic's Opus and Fable series, intensifying competition in the mid-tier agent market and offering developers a cost-effective path to frontier intelligence.

    2 Sources22
How is the Heat Index calculated and ranked?Algorithm Details

The heat index integrates the volume of verified independent sources, node credibility tiers, publication density, and a 24-hour half-life decay model. It quantifies industry attention and technical resonance rather than subjective opinions.

The radar continuously monitors clues from the last 48 hours, deduplicating rewrites. When an event triggers simultaneous discussions across labs, open-source communities, and tier-1 media, its heat index accelerates rapidly.