#数据/训练

Dataset of 50k AI generated "podcasts"

A Kaggle dataset from Listen Notes contains 50,000 AI-generated fake podcasts and spam episodes. The submission links to the dataset page and a Hacker News discussion thread.

Amazon Is Gutting Its AI Division After Sustained Failure

According to Bloomberg, Amazon is scaling back most of its Nova AI models and reorganizing its AI teams to focus on "highest priority" goals, placing its text, video and image models in a "keep the lights on" state with minimal resources. Labor and computing power are being redirected to a single "frontier-model effort" led by Pieter Abbeel, whose robotics company Covariant Amazon acquired in 2024. The shift follows the closure of an 80-person San Francisco office focused on AGI research.

AI co-authored commits: 3 open-source projects in 2023, 686 of 874 in 2026

Borehole's analysis of 984 public projects finds that 686 of 874 projects had at least one commit declaring an AI agent as co-author in 2026, up from 3 of 839 in 2023, with the average share of AI-declared commits rising from under 1% in 2023 to 8.9% in 2026. Of the 984 projects, 249 state an AI policy and 737 have AI-declared commits, with 206 doing both; Claude appears in 655 projects and GitHub Copilot in 470. The report notes that no AI-declared commits does not mean no AI was used, since a declaration is a co-author line the contributor chose to keep and a rebase can remove it.

From spot fixes to measured gains: improving document extraction accuracy with ai_extract and MLflow Prompt Registry

A Databricks engineering post describes extracting roughly 40 structured fields from partner order documents across dozens of PDF layouts, where a schema fix that halved errors on the revision field raised total errors across all fields by 9%. The authors recommend putting field rules in schema descriptions, document-level logic in global instructions, precision mode (ai_extract 2.1) for complex structures, and citations to diagnose errors, while versioning prompts and schemas in MLflow Prompt Registry with aliases such as @prod and @candidate. They evaluate candidates against a frozen human-reviewed test set, tracking accuracy, coverage and run-to-run variation, and report that restructuring the prompt and schema plus precision mode cut overall errors by 32% and revision errors by 64%.

Can we trust LLMs to judge whether a research idea is novel?

Ai2 released "Old Ideas, Novel Problems," a study with a benchmark, data, code, and an automated pipeline for testing LLM novelty judges. Across six judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), small prompt changes shifted accuracy by more than 50 percentage points on AI-generated ideas, retrieval and extra reasoning did not consistently help, and two purpose-built novelty evaluators scored worse than a simple prompted Claude Opus 4.6 baseline while costing far more.

Exowatt beats frontier models on accuracy with SageMaker serverless model customization

Exowatt used Amazon SageMaker AI serverless model customization to fine-tune Qwen3-1.7B for edge alarm triage (0.911 macro-F1 vs 0.698 for Claude Sonnet 4.6) and Qwen3-14B for cloud energy-market analysis (0.975 macro-F1 vs 0.625 for Claude Opus 4.7, at about 49× lower cost per 1,000 answers). The pipeline spans seven AWS serverless services, with the edge model quantized to GGUF and run offline via llama.cpp on existing CPUs, and the cloud model served through Amazon Bedrock Custom Model Import. Exowatt moved from experimentation to production in nine days and plans to reuse the pipeline across fleet tasks.

Can we trust LLMs to judge whether a research idea is novel?

Ai2 released "Old Ideas, Novel Problems," a study with a benchmark, data, code, and an automated pipeline for testing LLM novelty judges, built from ICLR 2026 peer reviews to form Human-Only and Human+Generated evaluation sets. Across six judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), small prompt changes shifted accuracy by more than 50 percentage points on AI-generated ideas, retrieval and extra reasoning did not consistently help, and two purpose-built novelty systems scored below a simple prompted Claude Opus 4.6 baseline.

Fine-Tuning NVIDIA Nemotron to Know Local Rules, Not Just Local Languages

The post notes that as AI adoption expands globally, general-purpose language models serve as a strong foundation but can lack the local knowledge, cultural nuance, and precision a particular application needs. It says post-training can adapt an open model to these needs, given suitable data and evaluation.

Python Typing Survey 2026: AI Tools and the Shifting Checker Landscape

The 2026 Python Typing Survey, with 1,794 responses, found type hint use stable across three years (85.5% in 2026) while checker selection shifted sharply, with 49.9% of respondents selecting two or more checkers and speed cited by 56.5% as the top reason for choosing one. Missing type information in third-party libraries was the leading difficulty at 65.1%, and AI tools wrote type hints for 42.5% of 1,517 respondents, with 24.4% saying an AI coding tool invoked their checker.

Amazon's AI-Generated Disney Guidebooks Are Terrible

A writer describes finding Amazon flooded with AI-generated Disney World and Disneyland guidebooks, estimating 10 to 20 AI-generated guidebooks for every one written by a human author, and examines several examples with impossible park maps, AI watermarks left in images, and outdated or incorrect information. The post lists signs of AI authorship such as self-publishing, 140–220 page lengths, authors with no online presence, and common AI phrasing, and notes some books contained faked 5-star reviews.

Can AI draw San Francisco from memory?

VOYGR tested six chat models from Anthropic, OpenAI and Google on 41,526 San Francisco coordinates, asking what is at each point from memory with no maps or internet, and checked every answer against OpenStreetMap. At a point every 100 m, Gemini 4 Argon, GPT-6.1 Sol and GPT-6 Astra tied at 86% correct on water, park or everything else, against 61% for answering "not water, not park" everywhere, while on a 500 x 500 m Marina block the best street score was 0.32 versus 0.23 for shuffled answers. The authors conclude models capture the coastline and large parks but not street grids, and recommend a geocoder rather than model memory for reverse geocoding.

BigQuery to Databricks: A Strategic Framework

This Databricks guide outlines a BigQuery-to-Databricks migration framework across three pillars—Process, Technology, and People—covering estate assessment with the open-source Lakebridge profiler, BI-first or ETL-first entry points, wave-based migration, and dual operation via Lakehouse Federation. It describes moving data into open formats on Google Cloud Storage, migrating SQL logic through rule-based transpilation, LLM-assisted conversion, and engineers, plus three-level validation, and maps BigQuery governance into Unity Catalog. The People pillar covers shared Notebooks, Genie for natural-language code generation, and Databricks Academy training, closing with field lessons such as profiling before planning and defining parity criteria before dual-run starts.

How AI Is Changing the Pace of Discovery for Millions Living With Dementia

C-BRAIN, a consortium of 18 institutions led by Dr. Randall Bateman and working with NVIDIA, launched in November 2025 to use AI to speed Alzheimer's research by drawing on previously inaccessible "dark data" through a hybrid centralized and federated approach. Its two tools, OpenScientist and InsightEngine, have completed analyses in minutes that scientists estimated would take weeks to months, at under $10 per run. The consortium aims to build a multimodal biological foundation model and, by year three, discover new causal mechanisms of brain aging and Alzheimer's disease.

Mirendil Launches Fern 0.1, Its First Model, With 10M-Token Context

Mirendil released Fern 0.1, a from-scratch coding model with a context window of up to 10M tokens (about 23 MB of code), built on Mirendil-Base after a short round of post-training and available only through hosted early access. Mirendil says one researcher pretrained the base model in 3.5 months for under $10M, with AI agents writing over 99% of the code and training data drawn only from public sources. The base model scores 100% on RULER single-needle retrieval from 32K to 10M tokens, while Fern 0.1 scores 6.1% on Terminal Bench 4.0 and 73.4% on SWE-Bench Verified, trailing the frontier and large open models in its comparison table.

Show HN: Which big sites block AI training crawlers but let AI search in

A study of the robots.txt rules of 945 well-known websites for 14 AI crawlers found that 25.9% block at least one AI crawler, while 9.2% block AI training crawlers but leave every AI search crawler open. CCBot was blocked most often at 19.6%, followed by Bytespider at 18.4% and ClaudeBot at 18.1%. The authors note the site list is not a statistical sample and that robots.txt shows what sites request, not what AI companies do.

Same price, different reason: Probing AI pricing agents for collusion

A research sprint project tested whether linear probes on an LLM pricing agent's activations can distinguish collusive cues ("my rival will hold $1.90 if I don't undercut") from innocent ones ("my rival's costs rose, so it charges $1.90") when both produce the same price. On held-out wordings, probes reached 98.7% (Qwen2.5-7B), 99.6% (Qwen2.5-14B) and 89.3% (Llama-3.1-8B), versus 81.6%, 85.3% and 81.1% for TF-IDF word matching and about 50% for a price-only baseline, though the author notes the cues may differ in sentence structure and that probes trained on instructed cues transferred only weakly to uncued collusion in price histories.

Is AI taking jobs? Census data through September 2026 says not yet

Census data through September 2026 show no economy-wide AI job loss: the 135 occupations most exposed to AI grew 2.8% year over year versus 3.7% for the rest, a gap smaller than the survey's margin of error. The clearest exception is workers aged 22 to 25 in the most exposed quarter of occupations, whose employment grew 0.1% while young employment elsewhere grew 9.6%. The wiki rates its confidence medium, noting the exposure scores date from 2023 and no source isolates cause.

Wikipedia's Cleanup of AI-Generated Content

Ellia Lebleu, a founding member of the volunteer WikiProject AI Cleanup launched in 2023, describes how early AI writing was betrayed by clichés like "stands as a testament" while 2025-era signs are subtler and more structural, and she says fake or misleading citations have become a major worry. She estimates, using Pangram detection with no false positives on pre-ChatGPT content, that roughly 20 percent of new articles and drafts were flagged as AI-generated in late 2025, falling to 15 or 16 percent in early 2026, though it is unclear whether policy or stronger models explain the drop. She discusses AI tools used against AI content, Wikipedia's human-written identity, and possible futures including attribution frameworks and reader contributions through AI models.

Officially corrected errors persist in AI training data

Céphéide, Loxyn's measurement bench, published a public alert with data, scripts and proofs examining whether official corrections—EU corrigenda, verified RFC errata and retracted papers—are present in public datasets used to train language models (EurlexResources, FineWeb-2, a FineWeb sample, peS2o). It reports that 95.5% of testable single-language corrections to EU acts (1,989/2,082) remain in erroneous form in EurlexResources, that about 12,200 retracted papers are in peS2o with 79% lacking any mention of the retraction, and that five models asked without documents give the erroneous version in 45% of answers where they commit to one version. Each step was preregistered and timestamped with OpenTimestamps in the Bitcoin blockchain, and failed controls and deviations are reported.

Show HN: AI Compute Footprint (energy, space, water and CO2)

A Show HN project presents an interactive tool comparing the energy, space, water and CO₂ needed to generate everyday AI tasks against the footprint of things people already do. It includes per-person-per-year figures for Switzerland, the US and the world, with confidence levels, assumptions and sources, and invites better data via GitHub.

Meta gets $3 billion from Nvidia and plans to spend more on Nvidia chips

Nvidia is investing $3 billion in Meta's planned equity offering, and the two companies said they will expand their partnership around Meta's AI product Muse. Meta plans to use more Nvidia infrastructure for its next generation of AI agents, including Vera Rubin systems and Vera CPUs, and aims to raise up to $95 billion through stock offerings to fund its AI infrastructure buildout. According to CNBC, AMD and Broadcom also signaled nonbinding interest in buying up to $5 billion in Meta shares combined.

Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing

Baidu released Unlimited OCR 2, a 3.34B-parameter OCR model under the MIT license that parses documents of hundreds of pages in a single request using SR-SWA, a sparse attention scheme that keeps per-step decoding cost nearly flat. Baidu reports first place on MPDocBench (84.6) and MarathonDocBench (77.1), with heading hierarchy its lowest sub-score at 48.2 on long documents. All benchmark figures are self-reported, and the model runs on one CUDA GPU in BF16 with code on GitHub.

Sarvam and HQ IDS establish joint framework for AI adoption across the Army, Navy and Air Force

Sarvam has signed a Memorandum of Understanding with Headquarters Integrated Defence Staff (HQ IDS), Ministry of Defence, creating an umbrella framework for AI projects across the Indian Army, Navy and Air Force and joint, tri-service and integrated organisations under HQ IDS. Under the MoU, Sarvam will design, develop and deploy AI applications for operational, training and decision-support needs of the three Services, with shared foundation models, agentic infrastructure and geospatial and perception systems, and human oversight retained on operational decisions. The work will be carried out by Sarvam's firewalled division Chanakya, applying its AI stack including foundation models such as Sarvam 105B, built in India and deployed on sovereign infrastructure.

Unify your productivity and AI data with Lakeflow Connect

Lakeflow Connect offers fully managed native connectors that ingest data from productivity tools such as Jira, Confluence, Notion, SharePoint, Google Drive, Outlook, Gmail, GitHub, OpenAI and Anthropic into governed Delta tables in Unity Catalog. Unstructured content like files, pages and email bodies lands as tables that can be cleaned up or parsed with ai_parse_document for retrieval and AI agents. Customers including Toyota Finance Australia, VaxCare, M3 Accounting, DigiCert, Orizon and TRegs describe using the connectors for Jira, Confluence, GitHub, SharePoint and Monday.com data.

Anthropic is burning gas to feed the AI boom

Anthropic's S-1 filing with the SEC acknowledges that its data centers are increasingly reliant on natural gas, with some planned sites powered by on-site gas turbines, and warns that competition for power and data center locations is intense. The filing also cites climate change as a financial risk, noting rising temperatures will increase electricity and cooling costs, and that gas price increases, regulatory conditions, or supply constraints could delay or derail planned data center projects.

How to teach with Databricks Free Edition

Databricks Free Edition is a perpetually free, browser-based version of the Databricks Data + AI Platform that students and educators can sign up for individually without a cloud contract, credit card, or installation, running on smaller compute under fair-use limits and not for production or commercial use. The article maps common courses—from intro programming and statistics to data engineering, data warehousing, ML, AI agents, and capstones—to platform tools such as notebooks, Apache Spark, Lakeflow, Delta Lake, Databricks SQL, MLflow, Genie Agents, and Databricks Apps, and answers instructor FAQs on distributing assignments and datasets, group work, and compute limits.

Anthropic opens its books for Wall Street, showing record revenue and mounting obligations

Anthropic filed an S-1 with the SEC to list on Nasdaq under the ticker "ANT," with Goldman Sachs and Morgan Stanley leading the offering and Wall Street expecting a valuation of roughly $2 trillion. Revenue rose from $376 million in Q1 2025 to $11.6 billion in Q2 2026, totaling $16.3 billion for the first half of 2026, and the company posted its first positive GAAP operating profit of $974 million in Q2 2026. Compute capacity contracts with Broadcom, Google, Amazon, Microsoft and other providers total roughly $780 to $810 billion in largely fixed, long-term obligations against variable, usage-dependent revenue.

Featured coverage of this story: 《Anthropic officially files for its public debut》

Researchers find six patterns that make AI text so exhausting to read

A study by researchers at NYU and UT Austin had 14 language models produce text in casual conversation, coding sessions, and argumentative essays, then compared it against human-written text, identifying six constructions AI overuses, such as unusual word combinations, distant placement of related sentence parts, dropped connectors, stacked nouns, nouns turned into verbs, and animate verbs applied to inanimate subjects. Smaller models like Llama and Olmo base come closest to human text, while within the Olmo family scores rise from base through instruct to the reasoning model Olmo Think, pointing to post-training as the likely cause. The researchers argue that training on math, code, and reasoning tasks rewards correctness over clarity, and that the shift from RLHF to RLVR removes humans from the feedback loop.

Fleet-wide computer vision lifecycle at the edge with SageMaker SDK v3

This post updates a 2023 MLOps pipeline pattern for visual quality inspection to Amazon SageMaker Python SDK v3, replacing the deprecated SageMaker Edge Manager with SageMaker Neo for edge optimization. It describes a dual-branch SageMaker Pipeline that trains an anomaly detector and a defect classifier per camera, gates models on their exported ONNX artifacts, compiles them with Neo, and registers them with camera metadata, while a Step Functions and DynamoDB semaphore layer paces fleet training. The sample deploys via AWS CDK and a full synthetic-data run costs about $0.07 in compute.

Building live intelligence: Databricks AI Search supports Streaming Tables, Materialized Views, and Managed tables using Iceberg

Databricks AI Search now natively supports Streaming Tables, Materialized Views, and Managed tables using Iceberg v3+ in addition to Delta tables, letting users index live events, continuously maintained personalized state, and open-format data directly without separate conversion and synchronization pipelines. Streaming Tables keep indexes fresh as new records arrive, Materialized Views propagate incrementally maintained aggregated or joined state for personalized search, and Managed Iceberg tables serve as sources without conversion or migration. Index creation follows the same Delta Sync index workflow via UI or API, eliminating the intermediate Delta table and custom sync pipeline.

Meta plans up to $95 billion in stock sales to fund its AI buildout

Meta plans to sell up to $95 billion in shares to fund its AI infrastructure buildout, according to The Information. The plan includes about $22.5 billion in Class A common stock, $12.5 billion in depositary shares representing one-twentieth of a new preferred share that will automatically convert into common stock in August 2029 unless converted earlier, and $60 billion from ongoing open-market sales once new non-voting Class C shares begin trading, likely no earlier than the first quarter of 2027. Nvidia, AMD, and Broadcom have expressed non-binding interest in buying up to $5 billion in Meta stock.

Tesla gave up the EV crown. Now it wants to win the autonomy race.

Tesla is shifting its business model toward recurring revenue from software and AI-powered services, with Full Self-Driving subscriptions reaching 1.7 million globally, up 67 percent year on year, as Musk explores a possible SpaceX merger to accelerate autonomous driving. Tesla ceded its position as the world's largest EV maker to BYD last year, and its 2030 sales goal has been cut from 20 million to 5 million vehicles. Tesla has not yet won approval for its fully autonomous version of FSD and is seeking wider EU backing through a member-state vote delayed until December, while Xpeng pursues EU approval for its XNGP system.

Wall Street's fastest traders are building data centers and backing chip startups to stay ahead in AI

High-frequency trading firms including Jane Street and Hudson River Trading are spending billions on cloud contracts, startups and their own data centers to secure AI compute for machine-learning-driven market prediction, with one executive estimating compute spending is now ten to twenty times higher than before ChatGPT launched. XTX Markets is building a data center in Finland, Jane Street runs one in Texas with more than 4,000 Nvidia GPUs, and Hudson River spends $1 billion a year on AI and signed a multiyear deal with CoreWeave. Jane Street instead invested $1 billion directly in CoreWeave, committed $6 billion to its computing services, and led a funding round for chip startup Etched, becoming its first customer.

Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra

Nvidia has open-sourced Flora, an open data engine built from three fine-tuned Qwen3.5 models that produces action descriptions, step-by-step progress ratings and success judgments for robot video. Nvidia estimates labeling one hour of video costs about 72 cents with Flora versus roughly $88.50 with GPT-6 Astra, though those figures are projections based on an assumed $10 per GPU-hour, and Flora trails Astra on some public benchmarks. Nvidia says Flora is already annotating up to 200,000 hours of video per day across 512 GB300 GPUs during training of its GR00T models, with models available on Hugging Face.

OpenMobile-2 proves open-source smartphone agents can rival proprietary models

Researchers from Nanjing University, SenseTime and other universities released OpenMobile-2, two smartphone agent models built on Qwen3.5 (9B) and Qwen3.6 (27B) parameters, together with the full training environment, data and reproduction instructions. The training ground covers more than 110 apps, including 35 rebuilt Chinese and global apps, and the larger model reaches 79.9 points on AndroidWorld and 59.6 on SPA-Bench with real devices. Holding data constant and increasing app variety raised success on unseen apps from 60.7 to 67.5 percent.

Fancy harnesses make coding agents more expensive but not better, study finds

A study by Liu and Han ran five models through Claude Code, OpenCode and mini-SWE-agent and found that swapping harnesses shifts SWE-bench Verified results about as much as simply rerunning the same setup, with differences of two to five solved tasks per model. The only measurable harness effect was a deficit for OpenCode, while Claude Code's roughly 16,600-token system prompt and tool descriptions made a single task cost up to three times more than the leanest harness. The authors conclude a harness buys fewer ways to lose a task and a bill, not pass rate, and recommend publishing a rerun of the same setup alongside every harness comparison.

Updated Google privacy policy says your Gmail might be used for advertising

Google updated its privacy policy on October 16, replacing the statement that it doesn't show personalized ads based on content from Drive, Gmail, or Photos with language that opens the door to using such data when shared with other Google services. The article links this to Personal Intelligence, released in early 2026, which lets users hand data from multiple apps to Google AI products.

Multi-Turn Reinforcement Learning on Amazon SageMaker AI: A hands-on guide

This AWS post walks through an end-to-end Multi-Turn Reinforcement Learning pipeline on Amazon SageMaker AI, using a Wordle agent hosted on Amazon Bedrock AgentCore that reports one outcome reward per episode for GRPO training of a LoRA adapter. In the example run, gpt-oss-20b improved from 61% to 84% solve rate on held-out words at default reasoning effort and from 15% to 69% at low effort, for about $339 in training tokens. It also covers IAM roles, dataset preparation, MLflow monitoring, serving-setting evaluation, cleanup, and reward-design mistakes and training failures encountered.

Now Google says it might use your Gmail and Google Photos data to target ads

Google updated its privacy policy on October 16 to add an exception allowing data from Drive, Gmail, or Google Photos to be used for personalized ads if users make that data available to other Google services supporting personalized ads, such as AI Mode and Search tools via Personal Intelligence. Advertisers don't get access to the content or personally identifying details, and Google says it won't target ads based on sensitive categories like race, sexual orientation, religion, or health. Users can disconnect apps from Search's Personal Intelligence features or disable Personalized Recommendations in the Search Services Personalization settings page.

LitPivot: Developing research ideas through dynamic contextualization within the literature landscape

LitPivot is a system developed with University of Pennsylvania collaborators that organizes retrieved literature around the facet of a research proposal being edited, offers critiques supported by excerpts from selected studies, and suggests revisions researchers can accept, edit, or reject. In a study with 17 researchers, domain experts rated ideas developed with LitPivot as better grounded in the literature than those developed with a text editor and an AI chat tool, and participants perceived their revised ideas as more novel and useful. LitPivot will be presented at UIST 2026 in Detroit.