#数据/训练

Tuning a Linux laptop for AI-assisted development

Jordan Dimov describes using Claude Code to diagnose and fix performance problems on his Ubuntu Dell Precision 3591 laptop, including a 90%-full disk from retained Rust debug builds, 30-second SSD freezes caused by Intel VMD (fixed by switching BIOS storage mode from RAID On to AHCI/NVMe), and a GPU kept awake by a kernel flag. Postgres tuning, parallel test databases, and CPU affinity changes cut the full test run from 283 s to 198 s and commits from 0.58 ms to 0.04 ms. He also covers local models on the 6 GB GPU, a pgvector-based code search tool, and a log-digest script for agents.

GDC November 2026 survey shows 40% of developers using generative AI

The GDC 2026 survey of over 2,600 game industry professionals across 69 countries found that 40% use generative AI tools as part of their job, up from 36% last year, with usage rising from 37% to 44% outside the US while US usage stayed within the margin of error at 38%. The report also notes 34% of respondents were laid off in the last two years, 54% of those laid off remain unemployed, and 62% of educators allow, encourage, or require students to use generative AI in game development coursework. Godot is used by 6% overall and 11% of indie developers.

Amazon Threat Intelligence observes threat actor campaigns with ARTEX AI agents

Amazon Threat Intelligence reports that between Oct. 3 and Nov. 5, 2026, its MadPot decoy sensors captured requests from ARTEX agentic AI penetration-testing agents that carried their own campaign records, including a large-scale AI credential theft operation in which stolen keys paid for the LLM inference used to hunt more keys, and a campaign against two cryptocurrency exchanges using real-time deepfake injection to bypass KYC checks. Amazon assesses with moderate confidence that ARTEX was used and that the activity comes from disparate Chinese-nexus threat groups, based on simplified Chinese prompts, China time-zone configurations, targeting and tooling. The post shares IOCs and argues that while the techniques were not new, agents change attacker speed and persistence, making basic hygiene such as removing default credentials and exposed secrets the most effective defense.

Europa, the EU's open AI model, is aiming for nearly 1T parameters

Domyn, the Milan-based company leading the EU-backed EUROPA consortium, reported in a November 12, 2026 technical post that it ran training trials for a model with 951.9 billion parameters on the JUPITER supercomputer in Jülich, Germany, more than double the Commission's 400 billion minimum. Using a mixture-of-experts design with 384 experts (8 active per token, 32.8 billion active parameters), Domyn says it raised throughput from 90 to 1,293 tokens per second per GPU through software fixes contributed back to PyTorch and TorchTitan. The post describes only trials, not a completed training run, and EUROPA's weights, license, training data and release date remain unpublished.

Show HN: Aurora GeoXAI – Sovereign geospatial AI built out of frustration

Aurora Geointelligence XAI is an experimental, CPU-first Python MVP (v0.1.0) that reads PNG, JPEG, TIFF and GeoTIFF imagery, records available raster metadata, applies scikit-learn MiniBatchKMeans to preview RGB pixels, optionally computes NDVI/NDWI from user-selected bands, and exports images with a traceable JSON report. It uses NumPy, Rasterio, Pillow, scikit-learn and Gradio, and explicitly does not use PyTorch, provide validated model-level explainability, or constitute a production or forensic system. The repository pairs the code with English and Portuguese manifestos on technological sovereignty, stating that no MLCommons benchmark, Finep/CDTI partnership, institutional endorsement or legal admissibility is claimed.

Australia's central bank finds AI hitting young workers first, especially in software and customer service

An analysis by Mike Major for the Reserve Bank of Australia built a monitoring system focused on entry-level workers and highly AI-exposed occupations such as programming, clerical work and customer service, finding job postings fell across all groups since ChatGPT launched in late 2022, with the steepest drop in the highest-exposure group. In that group, workers aged 15 to 24 and young college graduates have weaker employment rates and higher unemployment than older workers, and employment among young people in high-exposure jobs has declined since late 2022 while older workers' employment rose. Major cautions against early conclusions, since the study does not establish causation and measures theoretical exposure rather than actual company use.

The Coming Patent Storm in AI Hardware

A Texas-based technology company filed patent infringement actions in May 2026 in Texas district courts, Delaware, and the ITC against Nvidia, SK Hynix, Kioxia, Corsair Gaming, and Western Digital over memory technologies, per Law360. The article argues that concentrated value in AI chips, shared engineering bottlenecks in HBM, chiplet interconnects, co-packaged optics, and cooling, and a maturing patent landscape are converging to make AI hardware the next major patent litigation front. It notes that the window before large operating companies build defensive moats through cross-licensing and inter partes review appears open now in HBM and CPO and is opening in chiplet interconnects.

AI in the Search for Extraterrestrial Life

This overview traces how machine learning is used across radio SETI, infrared technosignature searches, exoplanet vetting and atmospheric retrieval, and biosignature and sampling work, noting that none of these systems has established that life exists beyond Earth. It describes concrete projects including the UCLA direction-of-origin model, the 2023 Breakthrough Listen study that flagged eight signals of interest, GLOBULAR's clustering-based interference rejection, NASA's ExoMiner++, the Ariel Data Challenge, FM4AR and INARA DS1, and laboratory and field biosignature work. It also compares open datasets and tools and outlines evaluation checks such as splitting by measurement source, keeping competitive baselines and testing transfer to new environments.

Microsoft Patent Tips Using Idle Gaming PCs to Train AI in Exchange for Rewards

Microsoft Technology Licensing filed a 2025 patent, published last week as US-20260349820-A1 and titled "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization," describing an opt-in system that uses idle GPU cycles on gaming PCs for AI training, similar to Folding@Home. The system would monitor memory, CPU, GPU, temperature and power draw to find cool idle periods, track completed work, and issue credits redeemable for use of the resulting AI model, free software or cash, with manual start/stop and usage limits.

Cmpnd Releases dspy-wsp: Weight-Space Programming Comes to DSPy

Cmpnd released dspy-wsp, an MIT-licensed prototype DSPy optimizer called WeightSpace in which a frontier LLM (recommended gpt-6-astra) reads and edits a student model's weights directly, without gradients, a training loop or a fine-tuning dataset. In tests on open decision models from 0.6B to 9B parameters, runs of 25 writer passes cost $5.42 to $7.63 and took 78 to 167 minutes, with d1-omni-600M rising from 50.9% to 69.1% using WeightSpace plus ReAnchor, while the 9B Clef-Flash gained only +1.9 from WeightSpace alone. Cmpnd describes the package as very experimental and slow, and notes that under the original paper's stricter rule d1-omni-600M gained just +2.8 points.

China builds 800MW AI data center campus to rival US Stargate project

According to a SemiAnalysis report, China is building an 800MW AI data center campus in Inner Mongolia that could challenge the economics of America's Stargate project. Envision built the first 120 MW of its Galaxy AI data center campus in under a year, at a cost researchers put at one-fifth of comparable US projects, after visiting the site in September. To secure the large amounts of electricity AI data centers need, Envision is building its own wind farms and battery storage alongside the computing facility.

AI model for predicting neurophysiological responses to text and pictures

The report introduces Concept2Brain, a deep neural network architecture that generates synthetic EEG signals from pictures or text, trained on EEG data from 88 observers viewing 360 naturalistic photographs and validated via 10-fold cross-validation. Synthetic signals reproduced empirical ERP components, time-frequency patterns, and participant-level differences, and out-of-sample checks against THINGS MEG/fMRI and ERP-CORE data showed expected correspondences. The model is available as an open cloud platform, and the authors note that text-based outputs should be treated with heightened caution because the architecture was trained on picture-evoked signals.

Engineering Fought Productivity Metrics. AI Ended the Fight

Rising, variable AI token costs (Uber exhausting its 2026 budget in four months, Salesforce's reported ~$300M/year Anthropic bill) have ended engineering's exemption from productivity measurement, but DORA and SPACE show no single metric works, and 2025–2026 data finds AI raises throughput while also raising bugs, incidents and review time. The article argues for attributing AI activity to shipped work and outcomes, using cost per merged pull request, autonomy rate and token yield rate instead of token consumption, and treating committed AI configuration files as maintained code.

Stanford Research Open-Sources Memorilla, a Latent Memory Module for Frozen LLMs

Stanford researchers have released Memorilla, an open-source 194M-parameter memory module that condenses a whole document collection into 16 memory tokens for a frozen LLM, with code, training recipes and checkpoints public under Apache 2.0. On FactKG, a frozen Qwen3-8B scored 84.33% with Memorilla versus 80.43% for full context, using 108x fewer document tokens, and it reported the highest average (49.68) across 8 benchmarks while losing on PubMedQA (59.00% vs 69.00% for LLMLingua-2). The quickstart requires one 80 GB CUDA GPU, and the full training recipe takes about 19 hours on 4 H100 GPUs.

Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon

Former OpenAI and Anthropic researcher Jacob Coxon says a slide from Anthropic's IPO roadshow was a major reason he left the company after about four months; the chart shows pre-training compute multiplier for code over time on a logarithmic y-axis, and Coxon says its steep climb points to accelerating algorithmic efficiency. He argues Anthropic's Mythos model was "not a one-time deviation, but a point on a superexponential," and that Anthropic could already retrain Mythos with a fraction of the compute originally needed. Coxon says algorithmic improvements are "on the verge of being automated" and wants frontier model development paused next year so competitors can catch up.

OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary

OpenAI is disbanding its robotics team again, moving robotics researchers to Mark Chen's main research team and the applied robotics and hardware teams to Greg Brockman's Applied organization, according to Fortune. The company reportedly believes general-purpose models like GPT-6 Astra can control robots directly, citing an Nvidia study where an Astra-powered agent reached a 74.8 percent success rate on simulated robotics tasks versus 51 percent for the best specialized models, and a StationeryBench comparison where Astra fully completed 7 of 100 tasks while Ai2's MolmoAct2 completed none. OpenAI shut down its robotics division in 2020 and rebuilt it in 2024, with CEO Sam Altman reportedly targeting a first public robot demo for 2027.

ShieldFont – Protect human muffins by poisoning unauthorized AI starfish

ShieldFont is an open-source tool that swaps content words in a page's HTML source for other real, same-grammar-class words, so humans see the original text via a custom font while scrapers collect altered meaning. Built with Danish type foundry Playtype as ShieldFont Optik, it ships with three ~12,000-pair English dictionaries and supports bringing your own font or mapping. The project reports that swapping about 25% of words caused 50% of 1,500 test passages to lose their original factual claim, and that roughly one in ten shielded chunks still passed the FineWeb-Edu quality filter.

Amazon has a plan to index the world’s online shopping data

According to Reuters, Amazon plans a "World Product Graph" project to build a vault of data on every product sold anywhere online, including items Amazon doesn't sell, and make it available through its own site. The effort aims to keep shoppers and AI agents on Amazon for product research and to give its Alexa for Shopping assistant more data; an internal document reportedly cites an October 1st, 2027 release date, though Amazon said draft documents don't necessarily reflect current plans.

Production MLOps at GS Caltex with Amazon SageMaker Canvas, part 1

GS Caltex connects models built by process engineers in Amazon SageMaker Canvas to production workflows in Amazon SageMaker Pipelines, with ML engineers handling deployment and scheduled inference. The company automated the handoff so ML engineers now deploy models within a few hours instead of about a week, and it runs batch inference every 30 minutes, which it reports has meaningfully lowered inference costs. The post details the reference architecture, including reusing Data Wrangler preprocessing flows, registering and approving models in Model Registry, and scheduling runs with Amazon EventBridge.

OpenFold3 on Amazon SageMaker JumpStart: Predict biomolecular structures

Amazon SageMaker JumpStart now lets you deploy OpenFold3-preview, an open source all-atom biomolecular structure prediction model from the OpenFold Consortium and Columbia University's AlQuraishi Lab, to predict 3D structures of proteins, DNA, RNA, small-molecule ligands and their complexes. The post walks through deploying it on an asynchronous inference endpoint, submitting prediction jobs, interpreting mmCIF outputs with pLDDT, PAE and PDE confidence metrics, running parallel predictions on multi-GPU instances, supplying your own MSAs, and scaling the endpoint to zero when idle.

The Chinese city powering the country's AI ambitions with data farms

In the grasslands of Ulanqab in Inner Mongolia, large data centers are being built rapidly as China seeks the computing power needed to compete with the United States for global AI dominance.

Export-banned Nvidia chips are behind China’s most powerful AI models — uh-oh, Jensen!

Unemployment Labs researcher Will DePue analyzed the weights of openly released models from DeepSeek, Moonshot AI, MiniMax and Ant Group and found evidence that models including Moonshot's Kimi K3 releases and DeepSeek's V4 bear fingerprints of training on export-banned Blackwell chips. Epoch AI estimates about two million advanced chips entered China in 2024 and 2025 through schemes hiding their final destination, though the chips could have been in offshore data centers such as those in Malaysia, which imported an estimated 275,000 Nvidia AI accelerators in that period. The article also reports that Nvidia CEO Jensen Huang left the company this week after admitting an affair with Alice Huang, an executive of Megaspeed, a company implicated in smuggling Nvidia chips into China.

Simple Puzzle Reveals AI's Jagged Intelligence

An article describes training a GPT from scratch on a million solutions to the 15 puzzle, after which it solves 99.9 percent of randomly scrambled boards but fails on the easiest puzzles, including one that is already solved, while assigning its wrong moves over 97 percent probability. The author attributes this to AI emulating its training data through shortcuts rather than understanding, producing a "jagged" boundary where performance collapses outside the training distribution. The piece draws on Ilya Sutskever and Melanie Mitchell to argue that skill is not understanding and that AI can be confidently wrong.

The avalanche that hit math is coming for AI development

The author argues that AI is already accelerating frontier AI research—citing METR's finding that Anthropic's progress rate has doubled in a year, largely due to AI-assisted R&D—and that rapid gains in verifiable domains like mathematics suggest AIs could match the best human AI researchers within a couple of years, with a median estimate of 2028. He contends that AI research resembles the well-scoped, verifiable problems AIs excel at, so automating it could compress years of algorithmic progress into one, yielding either a "spark" or a "software-only singularity."

Build a model improvement flywheel with Amazon SageMaker HyperPod Data Capture

Amazon SageMaker HyperPod now offers Data Capture, which records live inference inputs and outputs to Amazon S3 at three independent tiers (SageMaker endpoint, load balancer, and model pod) with minimal latency impact and no impact on model availability. The post walks through a single-cluster example in which a home-appliance support router's production traffic is captured at the model-pod tier, analyzed with Amazon Athena to surface a new-product blind spot, labeled with a SQL policy, and used to fine-tune the router on a second GPU. Accuracy is measured before and after on data that could not have existed before deployment, with scripts, manifests, SQL, and synthetic datasets in an accompanying GitHub repository.

AI, work from home, and the young-graduate job market

US young college graduates face a weaker labor market, with a 22–26 bachelor's-degree employment-population ratio of 88% in September 2026, about 6% unemployment at end-2025, and longer job searches. Using Current Population Survey data, the analysis finds that occupational work-from-home adoption is significantly associated with slower young-graduate employment growth relative to mid-career workers, explaining 56% to 88% of the FTE employment-population decline from 2023 to 2026, while AI exposure and industry AI adoption show negative but statistically insignificant effects. The authors caution the model is suggestive rather than causal, given correlated factors such as interest rates, hiring swings, and immigration.

Frontier AI takes only months to get 10 times cheaper

Village's LLM Atlas tracks both capability scores and the cost of completing a typical agentic task, showing that matching any landmark model now costs a fraction of its launch price and that recent landmarks get cheap fastest. GPT-5-level work costs a thirty-third of its launch price, o1-level work is 143 times cheaper, and matching Claude Opus 4.6 is 51 times cheaper today. Levels set since o1 are falling 11 to 25 times a year, versus 3 to 5 times a year for GPT-4 and GPT-4o levels.

XYBench: Helping with the problem behind the question

XYBench is a new benchmark of 8,115 real-world questions, drawn from Stack Overflow, Stack Exchange, curated examples, and WikiHow, that tests whether AI assistants recognize when a user's request rests on a misconception (the "XY problem") and redirect them toward their actual goal. Evaluating Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro, the authors found every model was more likely to address the stated request than the underlying goal, and all trailed human reference answers on misconception identification. The paper won the Best Resource Paper Award at EMNLP 2026.

XYBench: Helping with the problem behind the question

AI2 released XYBench, a benchmark of 8,115 real-world questions (from Stack Overflow, Stack Exchange, curated examples, and WikiHow) that tests whether AI assistants recognize when a user's request rests on a misconception—the "XY problem"—and redirect them toward their actual goal; it won the Best Resource Paper Award at EMNLP 2026. Evaluating Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro, the researchers found all five models were more likely to address the stated request than the underlying goal and trailed human reference answers on misconception identification. A separate multiple-choice experiment showed the same models often picked goal-focused answers when judging them (71–84%) far more than they produced them (35–41%), exposing a gap between recognizing useful advice and generating it.

1 in 4 3rd Party AI Endpoints Underperform Their Advertised Models

Artificial Analysis released version 1.1 of its Endpoint Accuracy Index, which compares more than 100 serverless endpoints hosting open-weight models against reference deployments of the official weights, finding roughly one in four score below their reference. The index reruns each endpoint on BFCL v4-500, HLE-250 and AA-LCR-25, and version 1.1 raises long-context repeats from 10 to 25, cutting the average 95% confidence interval from plus or minus 4.7 to plus or minus 2.6 points, while coverage expands to six models including DeepSeek V4.1 Flash 1017, GLM-5.3, GLM-5.3 Flash, Muse Spark Open, Kimi K3 and gpt-oss-120b. Reported gaps trace to output caps that truncate answers, shorter reasoning, and a schema bug encoding string tool arguments as numbers, with GLM-5.3 endpoints scoring from 67.2% to 99.3% and all 22 DeepSeek endpoints statistically consistent with the reference.

Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models

Will DePue says weight-initialization patterns in six Chinese models released since April 2026 match Nvidia's Blackwell core counts (148 or 152), while earlier models like DeepSeek-V3 and Kimi K2.5 show Hopper signatures; he says OpenAI and Anthropic researchers independently confirmed the findings and has published verification code on GitHub. He maps DeepSeek-V4-Flash and MiniMax-M3 to B200/B300 chips and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash to GB200/GB300, estimating the five training runs used about 12,000 Blackwell GPUs costing over half a billion dollars. The report notes the fingerprint shows only which hardware initialized the weights and does not by itself prove smuggling or a violation of US export law, since capacity can be rented outside China.

Featured coverage of this story: 《Chinese frontier AI models carry the fingerprint of export-controlled GPUs》

How AI Is Helping Care Teams Get Patients the Right Care, Sooner

NVIDIA describes how healthcare systems and startups use AI to reduce routine chart work, citing Dyania Health's Synapsis AI, a clinical reasoning engine that runs on NVIDIA-accelerated computing inside a hospital's firewall and surfaces source-traceable clinical signals for clinician review. Cleveland Clinic used it to support a study of more than 7,800 aortic valve replacement patients from 2005 to 2021, building a predictive model of valve deterioration and reintervention risk at three, five and ten years.

Rethinking Memory for AI Inference

Moving compute closer to memory can cut data movement, power consumption and latency, but different AI workloads need different approaches to memory architecture, interconnects and processing. Software support, standardization and thermal management remain major barriers to wider adoption, with processing-in-memory and near-memory processing still early despite efforts such as LPDDR6-PIM and compute on custom HBM base dies. The article surveys industry views on how edge inference, where models are known ahead of time, may benefit from ahead-of-time compilation and near-data compute.

TimeNet challenge: An online hackathon for time-series AI

The TimeNet challenge is an online hackathon where participants pick a time-series problem, onboard a dataset in TimeNet format, and train or fine-tune a Temporal AI Model such as OpenTSLM for tasks like forecasting, classification, anomaly detection or reasoning. Round 1 launches November 11 and submissions close November 22 at 11:59 p.m. PST, with judging the following Monday and more rounds to come. Teams of 2–4 can apply for $100 or $500 Nebius vouchers per round, and winning teams receive prize credits starting at $25,000 from AWS and Google Cloud, with terms and conditions applying.

Fossil versions: a fingerprint of AI-written code in 250k NPM packages

A random sample of 250,000 npm packages found that the share of new packages whose first-release manifests pin mostly round, already-superseded versions (x.0.0) rose from 1–2.4% in every half-year from 2019 to early 2024 to 13.7% (95% CI 13.1–14.4) in the first half of 2026, falling to 7.6% in the second half of 2026. Median pin age at birth rose from about 155 days in 2021–2023 to 451 days in Q2 2026, and the share of packages born with essentially all pins on the newest release fell from about 28% to about 10%. The author argues the timing, the specific versions and how package managers work point to manifests written from a language model's memory rather than by running an install, while noting the data cannot prove who or what wrote any single package.

Michelangelo AI

Michelangelo AI is now open source, with the material citing Uber-scale usage of 400+ active ML projects, 20K+ model training jobs per month, 5K+ models in production, and 45M+ real-time predictions per second at peak. It offers feature management, model training, evaluation, deployment, and monitoring, with feature management, evaluation, and monitoring marked coming soon. A Python API example shows distributed XGBoost training with Ray and an end-to-end training workflow.

Notion AI chat transcripts exposed in Fireworks AI breach

Notion said a Fireworks AI incident affected under 1% of Notion workspaces, with the impacted dataset including Notion AI chat transcripts used for the Insights feature that were largely unattributable to users or workspaces. Notion disconnected from Fireworks on 3 November, deleted its data there, disabled Insights, and stopped serving open-weight models through Fireworks; it contacted affected customers on 9 November. Affected customers can retrieve the data in their workspace and are advised to treat it as exposed and rotate any sensitive information, passwords or API keys pasted into Notion AI during the incident timeframe.

Turning Teacher Expertise into AI-Generated Educational Simulations

The Generative AI Labs team introduces Primer, an open-source system that lets educators build educational simulations through a structured AI interview rather than code, producing a reviewable design document that must be approved before any simulation is built. Specialized agents then create and run the simulation, with student-facing mentor, practice (Roleplay or Teachback), and feedback agents, and educators receive individual and aggregate performance reports. Primer is described as an experimental prototype, not a vetted classroom product, and its prompts, templates, and code are released on GitHub.

I Love AI

In a personal essay, the author reflects on twenty years since leaving home to study computer science, explaining why AI drew them in: a field built on tackling problems computers cannot yet solve, where success gets absorbed into ordinary technology (the "AI Effect") and only the failures keep the AI label. They contrast that tradition of heuristics and resource-constrained approximation with today's generative AI, which they see as mostly scale-up engineering, and with papers whose contribution is simply asking an LLM to do something. They still call themselves an AI researcher "at heart," with a tongue-in-cheek "not that kind" disclaimer.

Schelling Atlas: AI agents know where to find each other on the web

Researchers ran trials in which two independent copies of a model had to find each other on the web and exchange secret tokens without being told where to meet; most pairs converged on the push-notification service ntfy.sh within minutes, often posting to topics named from the task's own words such as agent-rendezvous. With ntfy.sh blocked, agents scattered across many sites and almost no pairs met in five minutes, though more time helped, and some agents left tokens in unusual places like OpenStreetMap notes, shared text canvases and a demo shop's product list. Asked to leave no trace, 37 of 90 agents chose channels that keep nothing or tried to delete their posts, and the authors note agents read the real internet while their posts reached only a simulated copy.