A University of Texas at Austin team's automated system, Darmok, solved all seven British cryptic crossword puzzles error-free at the 2026 Times Crossword Championship, beating the human champion by 3 minutes 13 seconds in the Final, the first automated system to reach superhuman performance in this event. Notably, the primary model provider slowed sharply during the Final, and backup routing supplied 17 of the 30 answers, underscoring the value of multi-model redundancy for latency-critical tasks.
A University of Texas at Austin team's automated system, Darmok, solved all seven British cryptic crossword puzzles error-free at the 2026 Times Crossword Championship, beating the human champion by 3 minutes 13 seconds in the Final, the first automated system to reach superhuman performance in this event. Notably, the primary model provider slowed sharply during the Final, and backup routing supplied 17 of the 30 answers, underscoring the value of multi-model redundancy for latency-critical tasks.
A new University of Maryland paper finds that frontier models, given only a single text and a zero-shot prompt with no tools or candidate lists, detect AI-written prose as well as trained detectors like Pangram while explaining their reasoning. More strikingly, they infer authorship: Claude Fable 5.2 identified the reviewer in 56% of 398 anonymous peer reviews and the account behind 69% of comments from Redditors with over 500 posts a year. The anonymity assumptions behind blind review and pseudonymous forums are now weak.
A new University of Maryland paper finds that frontier models, given only a single text and a zero-shot prompt with no tools or candidate lists, detect AI-written prose as well as trained detectors like Pangram while explaining their reasoning. More strikingly, they infer authorship: Claude Fable 5.2 identified the reviewer in 56% of 398 anonymous peer reviews and the account behind 69% of comments from Redditors with over 500 posts a year. The anonymity assumptions behind blind review and pseudonymous forums are now weak.
A new Columbia and MIT paper reports that Claude autonomously discovered a faster method for computing selected entries of a product of two long, thin matrices, which classic reductions turn into speedups for 3SUM and all-pairs shortest paths. 3SUM now runs in O(n^1.9992) and APSP in O(n^2.9995, tiny gains that nonetheless refute the long-standing 3SUM and APSP hypotheses. The method was found in a single AI session using 16M output tokens with no human input and was formally checked in Lean 4.
A new Columbia and MIT paper reports that Claude autonomously discovered a faster method for computing selected entries of a product of two long, thin matrices, which classic reductions turn into speedups for 3SUM and all-pairs shortest paths. 3SUM now runs in O(n^1.9992) and APSP in O(n^2.9995, tiny gains that nonetheless refute the long-standing 3SUM and APSP hypotheses. The method was found in a single AI session using 16M output tokens with no human input and was formally checked in Lean 4.
The open source LookAtMyQuant project introduces a post-training quantization compiler that allocates bit budgets per stored unit based on measured output damage, using Fisher information and trellis codes. Its second version compresses Qwen3.5-4B to 1.23 GB (~2.3 bits per weight), solving 330 of 540 exact-answer math problems versus 328 for the same-size Bonsai 2 Ternary 4B, with lower KL divergence (0.45 vs 0.59 nats). The full recipe and roughly $34 of GPU cost are published under the MIT license.
The open source LookAtMyQuant project introduces a post-training quantization compiler that allocates bit budgets per stored unit based on measured output damage, using Fisher information and trellis codes. Its second version compresses Qwen3.5-4B to 1.23 GB (~2.3 bits per weight), solving 330 of 540 exact-answer math problems versus 328 for the same-size Bonsai 2 Ternary 4B, with lower KL divergence (0.45 vs 0.59 nats). The full recipe and roughly $34 of GPU cost are published under the MIT license.
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.
Following the release of prompt templates for GPT-6.2 Astra and Sol, OpenAI removed temperature and top_p parameters from the API, and editing conversation history now discards reasoning steps. Arena.ai evaluations show the model's coding rank jumped to 3rd, while chat rank fell to 11th, with the overall ranking rising only one spot.
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.
GroupAssistBench has drawn attention since its release, with the finding that stronger memory recall does not equal stronger service capability becoming a focal point. The team has open-sourced the paper, code, and dataset, which may spur more optimization solutions for group chat agent memory.
An NVIDIA paper proposes compiling an LLM's reasoning trace into small parameterized Python functions that run on CPU, invoking a fine-tuned 1B model only for lightweight text tasks such as value extraction. The cache matched or beat the source model on 7 of 8 math and science benchmarks, scoring 93.3% versus 83.3% on AIME '25, and covered at least 90% of rephrased or re-numbered variants at 16x to 291x lower cost.
An NVIDIA paper proposes compiling an LLM's reasoning trace into small parameterized Python functions that run on CPU, invoking a fine-tuned 1B model only for lightweight text tasks such as value extraction. The cache matched or beat the source model on 7 of 8 math and science benchmarks, scoring 93.3% versus 83.3% on AIME '25, and covered at least 90% of rephrased or re-numbered variants at 16x to 291x lower cost.
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.
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.
Samsung Electronics has locked up most of its memory chip capacity for 2027 through 2029 under long-term agreements, with some customers pushing supply guarantees out to 2031. AI data center expansion has driven shortages across HBM, commodity DRAM and enterprise NAND, with major makers meeting only 60% to 70% of demand and 2027 commodity DRAM wafer capacity effectively sold out, strengthening memory vendors' pricing power.
Samsung Electronics has locked up most of its memory chip capacity for 2027 through 2029 under long-term agreements, with some customers pushing supply guarantees out to 2031. AI data center expansion has driven shortages across HBM, commodity DRAM and enterprise NAND, with major makers meeting only 60% to 70% of demand and 2027 commodity DRAM wafer capacity effectively sold out, strengthening memory vendors' pricing power.
In her final speech before stepping down, ECB President Christine Lagarde called on the EU to use public funds, including the EU budget, to absorb first-loss risk on large chip fab and data center projects so as to crowd in private capital through cross-regional public-private partnerships. She cited an estimate that lifting Europe's share of global AI compute from about 5% to 15% by 2030 would require roughly EUR 1.3 trillion in investment.
In her final speech before stepping down, ECB President Christine Lagarde called on the EU to use public funds, including the EU budget, to absorb first-loss risk on large chip fab and data center projects so as to crowd in private capital through cross-regional public-private partnerships. She cited an estimate that lifting Europe's share of global AI compute from about 5% to 15% by 2030 would require roughly EUR 1.3 trillion in investment.
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.
Following its release, Inkling 2 Mini has been integrated into Arena.ai's Agent Arena, Code Arena, and Text Arena, with leaderboard results forthcoming. vLLM offers Day-0 support, and Hugging Face co-released the model with mixed-precision weights and DSpark draft model acceleration, while the community actively tests its agentic workflow performance on local devices.
US tech stocks suffered their worst weekly loss since the pandemic as AI valuation doubts, an SEC accounting review of Nvidia and war-driven oil costs weighed on sentiment. Nvidia plunged 14% on Friday, the US semiconductor index fell 10.7% with losses across Sandisk, Micron and Broadcom, and Anthropic dropped 9.1% a day after its market debut. Investors are now questioning whether elevated chip and memory valuations remain justified.
US tech stocks suffered their worst weekly loss since the pandemic as AI valuation doubts, an SEC accounting review of Nvidia and war-driven oil costs weighed on sentiment. Nvidia plunged 14% on Friday, the US semiconductor index fell 10.7% with losses across Sandisk, Micron and Broadcom, and Anthropic dropped 9.1% a day after its market debut. Investors are now questioning whether elevated chip and memory valuations remain justified.