TechCrunch Disrupt 2026 will be held October 13-15 at Moscone West in San Francisco, bringing together more than 10,000 founders, investors, and tech leaders, with the full breakout session agenda now announced.
Mistral released a preview of Mistral Large 4, with 1 trillion total parameters and 49B activated parameters, trained on its own 3,800 NVIDIA Grace Blackwell GPUs. It is available via the Mistral API, with open weights promised by the end of the month.
Over the past year, AI labs including OpenAI and Anthropic have announced breakthroughs on multiple long-unsolved math problems, even a famous Millennium Prize Problem. These achievements should have been celebrated, but instead sparked a strong backlash in the math community. OpenAI says it will consult top mathematicians to avoid repeating the mistakes, but whether it delivers remains to be seen.
OpenAI has released 722 mathematical manuscripts covering 372 families of results, which AGMAI says include solutions to "hundreds" of open problems, produced by an unreleased frontier model. The manuscripts were published in a GitHub repository with paper revision and citation protocols, along with model reasoning summaries and compute estimates; OpenAI says each result corresponds on average to three hours of ChatGPT Pro thinking time.
Why it matters: The report gives the scale of 722 manuscripts, the number of problems, and compute estimates, so readers can gauge the actual content of these mathematical results and the academic controversy around them.
A benchmark tested whether a local Qwen3.8-27B-Q4_K_M.gguf model can produce exact results of positive integer addition using only English words, with 5,070 no-reasoning cases and 169 reasoning-enabled control cases. With reasoning off, numerical accuracy was only 23.57% — 97.04% for 1-3 digit problems but 6.44% for 10-13 digits, with 96.17% format compliance; with reasoning on, it answered 167 of 169 attempts correctly.
Thore Graepel, a former core member of DeepMind's AlphaGo team, argues in an essay that large language models are essentially fast-associative next-token prediction, that their chains of thought are not an independent reasoning mechanism, and that they do not reason in the scientific sense.
The Latent Space podcast interviewed MIT PhD student Alex Zhang about his work on GPU Mode, KernelBench, and Recursive Language Models (RLM).
A research team from Carnegie Mellon, MIT, NYU, and Stanford has developed an AI called Ataraxos that defeated Pim Niemeijer, considered the strongest Stratego player in history, with 15 wins, 1 loss, and 4 draws, using only 16 GPUs and a few thousand dollars for training. Stratego is an imperfect-information game with large amounts of hidden information and very long game timescales; even DeepMind had previously failed to build a machine that could reliably beat top human players.
Ataraxos, developed by a research team, defeated four-time world champion and widely acknowledged greatest Stratego player Niemeijer with an 85% effective win rate over three weeks of play, and won 38 of 40 games in an exhibition match at the 2025 world championship, using less than $8,000 of compute to crack imperfect-information games.
Why it matters: The article compares Ataraxos and DeepNash training costs and methods, giving readers insight into the latest progress in imperfect-information game research.
Google DeepMind released Gemini 4 Argon, focused on coding, enterprise knowledge work and cyber defense, ranking first in 13 of 19 benchmarks. Its output cap rises from 64K to 1M tokens, priced at $4/$20 (limited-time half price $2/$10), currently available only to government users of the Fairwind program and trusted cyber defenders.
Google released Gemini 4 Argon, its first frontier model in more than seven months, with introductory pricing of $2 per million input tokens and $10 for output, and about a 95% discount on cached input.
Google released its new frontier model Gemini 4 Argon, saying it delivers frontier performance in software engineering, enterprise knowledge such as law and finance, and complex workflows like cybersecurity defense.
CoreWeave announced that its cloud platform now offers NVIDIA Vera Rubin NVL72 systems with Spectrum-X 102.4T Ethernet networking. Cognition became the first customer to run production workloads on it, with measured total token throughput for SWE-2 inference up to 4.8x higher than GB200 NVL72.
GPT-6.1 Sol is now generally available on Amazon Bedrock, focusing on stronger reasoning for agent coding, computer use, and professional workloads. According to OpenAI, its DeepSWE v1.1 score matches GPT-6 Astra at about one-fifth the per-task cost and beats GPT-6 Sol's best score by 6.4 percentage points; it also outperforms GPT-6 Sol on transparency, user intent, and explicit constraint evaluations.
Why it matters: The official announcement gives specific comparisons against GPT-6 Sol and GPT-6 Astra, helping readers evaluate its cost-effectiveness for agentic work scenarios.
NVIDIA has released the open-source tabular foundation model Kumo Tabular, now on Hugging Face. Given labeled tables, the model predicts labels for new rows in a single forward pass with no training, tuning, or feature engineering; it supports classification and regression, is Transformer-based with column, row, and in-context attention, and was pretrained entirely on synthetically generated tables.
Why it matters: The official post covers the architecture, pretraining recipe, and four benchmark results of a zero-finetuning tabular foundation model, letting readers judge whether it can replace traditional gradient-boosted-tree workflows.
xAI's Grok 4.7 is now available on Amazon Bedrock, offering a 500K token context window and four reasoning effort levels: low, medium, high, and xhigh.
Why it matters: The article covers how to access it on Bedrock, reasoning effort configuration, and cost trade-offs, so readers can directly plan their usage and spend.
AWS showed how to use the open-source RL framework SkyRL on SageMaker HyperPod to train the Qwen3-VL-8B vision-language model on a visual maze navigation task. Starting from a VisGym SFT checkpoint, GRPO post-training raised the solve rate on a fixed set of 64 mazes from 43.75% to over 95%.
The AWS team detailed five techniques NarrateAI uses to achieve production-grade LLM quality assurance on Amazon Bedrock: adaptive pipeline routing, cross-account multi-model failover, real-time streaming evaluation, a composite evaluation framework, and data accuracy validation.
Latent Space's swyx announced a revamp of AINews, which has run for 3 years and has over 200,000 subscribers, planning to merge AINews v3 with the tens-of-thousands-member Latent Space Discord, migrate the site to Beehiiv and launch a new homepage.
Anthropic released Claude Opus 5.5, and about an hour later OpenAI launched GPT-6 Sol and GPT-6 Luna.
Why it matters: The author hands-on tested models from all three providers and compiled a full price table, letting readers compare the new round of price cuts and real-world performance across tiers.
Latent Space released a long-form interview with John Platt, inventor of two textbook algorithms including SVMs, who now leads Empirical Research Assistance (ERA) research at Google.
Last week TypeSafe AI released Jev, a new category of model it calls a System One model, which takes text input but outputs classifications, yes/no decisions, scores, and confidence as floating-point numbers. Jev charges only for input at $0.042 per million tokens, below GPT-5 Nano's $0.05/million, supports evaluating multiple questions in parallel, and suits tasks like classification, ranking, and search reranking.
NVIDIA Vera Rubin NVL72 made its debut in MLPerf Inference v6.1 preview submissions, delivering up to 3.7x the throughput of GB300 NVL72 on Qwen3-VL and 2.5x on DeepSeek-R1.
OpenAI has released GPT-6 Astra, positioned as its most powerful business-focused model, with advanced reasoning and computer use capabilities, plus stronger writing and design judgment.
Why it matters: The official launch of a business-oriented model highlights reasoning, computer operation, and writing/design judgment, letting readers assess where it fits in their workflows.
OpenAI shared an AI-generated solution approach to the Navier–Stokes Millennium Prize Problem, including a solution write-up and a Lean formalized proof. The material comes from an official OpenAI announcement; the feed does not include the full text, so specifics of the proof and methods should be checked via the original link.
Why it matters: OpenAI has publicly released an AI-generated approach to the Navier–Stokes Millennium Prize Problem with a Lean formalized proof, letting readers track AI's actual output in mathematical proof work.
Google Research released TimesFM-3, a time series foundation model with 330M parameters, pretrained on over 1 trillion timepoints of real and synthetic data, and the first to natively support zero-shot multivariate forecasting.
IBM released the Granite 4.2 reasoning model family, comprising 3B, 8B and 30B dense decoder models built on the Granite 4.1 base (pretrained from scratch on about 15T tokens, with a five-stage strategy extending context to 512K), post-trained with SFT and multi-stage GRPO reinforcement learning, all open-sourced under the Apache 2.0 license.
Why it matters: The official disclosure covers the full training pipeline from pretraining and SFT to multi-stage GRPO and agentic RL, transferable to similar model-building workflows.
Hugging Face proposes Quantization-Aware Healing (QAH), compressing GPT-OSS 120B to a 60B-parameter structure and quantizing it to MXFP4, then distilling directly from the original pre-compression model rather than from a recovered bfloat16 checkpoint.
Import AI issue 470: METR research finds AI brings significant speedups to cybersecurity discoveries, modest contributions to math, but no measurable acceleration yet for optimizing AI research algorithms. Teams from multiple universities propose the SPADE framework, which co-evolves environment synthesis with agent capabilities via self-play, reaching an average score of 58.3 on a suite of game environments with Qwen3-30B-A3B, 5.3 above the baseline. This issue also introduces Hawkeye, for building better GPU kernels.
DiG-bench (Discovery in Games) has been released, a benchmark of 70 games for testing AI's ability to autonomously discover hidden rules through exploration; 21 of the games are publicly available.
Google Research released the WikiProfile benchmark (2,150 Wikipedia facts, each with 10 tasks) and a knowledge profiling framework. After evaluating 13 LLMs, it found that frontier models have 95–98% of facts encoded but still fail to directly recall 26–34%, and 11–12% still fail even with thinking enabled.
Microsoft Research introduces CARE-X, a research model and unified chest X-ray VLM built on SigLIP2-so400M and Phi-4-mini-instruct (3.8B), jointly trained with auxiliary heads to provide report generation and structured predictions with confidence, and aligned to clinical rewards using DAPO reinforcement learning.
At FMS 2026, Marvell unveiled a new-generation AI memory infrastructure portfolio spanning server-level AI storage, rack-level CXL memory expansion and pooling, and cabinet-level optical shared memory.
TrendForce analysts explain three paths tech giants are taking to compress the KV Cache through architectural and algorithmic innovation, with Google and DeepSeek respectively reworking attention mechanisms via GQA and MLA.
Google Research published a reinforcement learning-based quantum error correction control framework in Nature, letting an agent continuously tune thousands of control parameters using error detection events to counter drift during computation. In experiments injecting artificial drift on the Willow superconducting processor, RL control improved logical stability by 3.5x and further reduced the logical error rate by 20% on top of expert calibration.
Thinking Machines has released the open-source multimodal model Inkling on Hugging Face, with about 1T total parameters (975B, a MoE with 41B active), natively accepting image, text, and audio inputs, supporting 1M context, trained on 45 trillion tokens, and shipping BF16 and NVFP4 weights.
Why it matters: The official blog provides architecture details, VRAM requirements for each tier, and deployment configurations, letting readers evaluate multimodal deployment costs and model selection.
Hugging Face published part three of its Profiling in PyTorch series, using the profiler on an NVIDIA A100 to compare traces of six attention implementations: naive attention, in-place masking, and SDPA math/efficient/flash/cuDNN.
Why it matters: The original benchmarks kernel-level differences across six attention implementations with the profiler, teaching readers to read from traces how in-place saves a copy and what each SDPA backend really costs.
Mistral released the Apache-2.0 open-source model Leanstral 1.5, with 119B total parameters and 6B activated parameters, trained via mid-training, supervised fine-tuning and CISPO reinforcement learning.
Why it matters: The official post discloses the full training pipeline, benchmark figures and real code verification cases, letting readers gauge how practical formal verification models are for mathematical proofs and open-source bug finding.
The Berkeley AI Research (BAIR) lab releases its 2026 PhD graduate showcase, highlighting graduates such as Baifeng Shi, Charlie Snell, and Hanlin Zhu and their work in robotics and embodied intelligence, LLM reasoning and test-time scaling, generative models, AI safety and human-AI interaction, and speech and medical AI.
A Google Research paper at COLM 2026 found that enabling chain-of-thought reasoning allows Gemini-2.5 (Flash and Pro) and Qwen3-32B to recall simple single-hop facts they nearly always fail to answer when reasoning is turned off.