Princeton researchers have trained an AI model to learn language the way children do, according to a Princeton Neuroscience Institute news article. The item links to the article and to a Hacker News discussion thread.
A project presents "legal zero-days" — logical errors in statutes such as cross-references to repealed provisions, lapsed time limits and same-day amendment conflicts — cataloguing 82,000 flagged issues by category and walking through five worked examples from Germany, Brazil, Switzerland and the Netherlands. The examples include a German weapons-regulation fine that points at a deleted section, a Swiss pest-control cost charge tied to a repealed article, and a Dutch decree whose fine increases took effect four days before publication.
Demand has shifted as regular users buy budget GeForce RTX 5060 Ti cards en masse for local AI workloads, driven by video memory shortages and the growing requirements of open-source generative neural networks. Some pair the 16GB card with existing GPUs such as the GeForce RTX 5070 12GB for 28GB total VRAM to run the Qwen3.8-Flash-Next IQ3_S quantized model, and enthusiasts often install two or more RTX 5060 Ti cards per system, with plans for a third via an M.2 slot using OCuLink. Using Strata version 0.2.10 with streaming KV cache data to 64GB of system RAM, a 256-token context window and roughly 246 input tokens were processed on two cards, though system RAM storage reduces pre-computation performance, so users prefer buying more accelerators.
Four SQLite extensions implement the same parse_number() behavior in C and Rust, with tests covering type preservation, parsing rules, and floating-point accuracy. The optimized C implementation was fastest at 16–19 ns per row, about twice as fast as the Rust versions and far faster than SQL replace(). Plain C was slower on decimals because it relied on strtod(), while a compact fast path closed that gap; all versions matched across 8 million fuzz checks.
NotBot is an experimental Python project that uses LLMs (Google Gemini and Groq) to analyze financial news from Google News and Finnhub and compute a weighted sentiment score for publicly traded companies, with deduplication, asynchronous processing, and experimental end-of-day price support via FMP. The author notes the scores have not been systematically validated against market performance and that the project is not financial advice, licensed under PolyForm Noncommercial 1.0.0.
In this AI Engineer interview, Mihail Eric, who teaches Stanford's CS146S "The Modern Software Developer" and leads AI at Monaco, describes how the course was refreshed for a second run, with part of the homework now being contributions to open-source projects reviewed by the repos' maintainers. He also discusses Monaco's approach of buying some tools (such as Devin for code review) while building its own agent workspaces, guided by his rule of thumb: "You buy to learn, and you build to keep."
Etsy's updated Terms of Use, published October 26 and effective December 16, 2026, explicitly authorize Etsy and its authorized third parties to use seller-uploaded content to develop, train, test and improve AI and machine learning technologies, while sellers retain ownership of that content. The revised terms also prohibit outside parties from using Etsy data to develop, train, fine-tune, evaluate or benchmark AI systems without permission, and set detailed controls over automated shopping agents. The accompanying Privacy Policy lists listing information, reviews and community posts as content that may be used for AI training, and neither policy identifies a general opt-out mechanism for US sellers.
Turing Award winner Richard Sutton says raw compute alone isn't enough for intelligence and that AI science needs clever new methods and is "far from finished," sparking wide discussion on X. Researchers Jerry Tworek and Jayden Teoh respond that the bitter lesson still holds but is often misread, with Tworek warning that transformer scaling's temporary success "has reward-hacked the brains of a whole generation of scientists." OpenAI's Sam Altman and Anthropic's Dario Amodei counter that LLM scaling is far from exhausted.
This comparison of 10 open Mixture-of-Experts models shows routing details varying widely: expert counts range from 8 to 384, top-k from 1 to 8, and 7 of 10 models keep one shared expert. DeepSeek-V3 uses sigmoid scoring with a bias-based noaux_tc balancing method, while DeepSeek-V4-Pro raises routed experts to 384, lowers top-k to 6, switches to sqrtsoftplus, and uses hash routing in its first 3 MoE layers. Llama 4 Maverick and gpt-oss do not publish their gating or balancing methods in config.json.
Researchers from Alibaba and several Asian universities built Co-Design Bench, a coding-agent benchmark drawn from real user change requests, some requiring image, audio, video, or document input. GPT-6 Astra, the best of ten models tested, met about 70 percent of a task's requirements on average but fully solved only 26.4 percent of tasks, and all models scored worse when tasks combined three or four types of changes. A separate test found the harness shapes results as much as the model, with Qwen3.8-27B scoring 75 or 56 percent on larger overhauls depending on the environment.
AI labs and data-acquisition startups are paying media companies, graduate students, and businesses for old work records—documents, emails, Slack messages, and tickets—to train models, with pitches reaching individual inboxes. Companies like Mercor, Micro1, and Handshake describe a fast-growing market, while privacy advocates question whether anonymization fully removes personal data, as seen in objections to Spirit Airlines' proposed sale of its digital assets. Critics note the practice can enlist workers without explicit consent in efforts that may ultimately automate their work.
This guide explains seven Mixture-of-Experts routing strategies used in production, from softmax top-k token choice (Mixtral, gpt-oss, Qwen3) and top-1 switch routing (Switch Transformer, Llama 4 Maverick) to shared-plus-routed experts, sigmoid gating with aux-loss-free bias balancing (DeepSeek-V3/V4, GLM-4.5, Kimi K2), node-limited routing, expert choice routing and hash/dynamic top-k. It notes that expert choice routing balances load perfectly but leaks future tokens in causal models, with one test showing a roughly 10% abnormal loss drop, while loss-free balancing cut MaxVio from 0.72 to 0.04 at 1B. It advises teams to read four config.json fields (expert count, experts per token, shared experts, topk_method) before serving, since most inherit a router with their checkpoint.
Nous Research has released Arete-27B, a supervised fine-tune of Alibaba's Qwen3.8-27B that it says reduces answers containing propaganda from 22% to 0.9% on its own 803-prompt eval, while raising fact mean from 2.32 to 2.98. The 27B dense model was trained with about 52.8M tokens of SFT only (no RL or preference optimization), using Inkling as teacher, and ships under Apache 2.0 with weights, eval and training data public. Nous reports capability held on its tests (Hermes Bench 65.60% vs 65.77%), and community GGUF quants run from 8.64 GB to 53.81 GB.
The author reports an experiment in which Claude Opus 5.5, acting as an "editor" under the weight-space programming method from MIT CSAIL's "Sparks of Weight-Space Programming" paper, edited 17 weight matrices in a 494-million-parameter model to make it forget 20 fictional customers and 10 planted secret keys, while other customers lost only 2.5 points of accuracy. The edited model matched an "answer key" model trained without those facts across six working attacks, whereas gradient ascent, NPO and refusal training each failed to both erase the facts and preserve the rest of the model. Testing Claude's own explanation of the edit, the author found the late layers do the work and the middle layers only assist, contradicting part of Claude's account.
A post describes an interview-preparation curriculum for junior AI researchers at large labs, organized into eleven modules covering PyTorch fundamentals, theoretical foundations, implementation from scratch, ML debugging, system design, inference optimization, RAG, paper analysis, safety and alignment, CUDA, and mixture of experts. Each module pairs an article with a hands-on exercise notebook and reference solutions, with scaled dot-product attention revisited at every level, and most material runs on CPU except the CUDA module. It also suggests time-boxed study budgets and module orderings for coding, systems, and research interview rounds.
SemiAnalysis reports on a September visit to Inner Mongolia's datacenter region, including Galaxy, an Envision-owned campus in Ulanqab planned for up to 800MW of IT capacity whose first 120MW phase was delivered within a year at a fifth of US costs. The report describes Envision's vertical integration—its own wind farms, a dedicated power line, an 800V DC solid-state transformer pilot, and a chiller-less cooling design—and compares Galaxy's scale to Stargate's Abilene campus. It also notes Inner Mongolia has lined up 25GW of announced datacenter projects.
Youngjin Kang's blog post argues that software narratives need rules robust enough to generate self-growing emergent behavior, but such rules risk seeming arbitrary unless they are goal-driven toward a "purpose of life." Because the chain of ultimate purposes never terminates, he proposes positing a supreme authority ("God" as a hypothetical system administrator) whose commands serve as axioms, and sketches a universal law of the form "IF God said X, THEN now X" as a seed of meaning.
Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to create Arete-27B, which cuts its propaganda rate from 22 percent to 0.9 percent while keeping general capabilities roughly intact. The team built an 803-prompt test across seven categories and judged responses with Grok 4.7, finding Western models below one percent and GLM-5.3 and DeepSeek V4.1 Flash at 15 and 16 percent. Nous concludes the bias layer in Chinese-aligned models is shallow, applied late and thin, and removable without disturbing anything underneath.
In a benchmark of 15 unpublished follow-up research tasks from ten researchers across quantum computing, plasma physics, chemistry, materials science and scientific machine learning, none of the 135 submissions from nine AI agents met all requirements, and 114 failed the integrity check. Only GPT-6 Astra, GPT-5.6 Sol and Claude Opus 5.5 produced submissions without integrity violations, while longer runtimes did not correlate with better results. The authors present the benchmark as a testing ground for systematically improving research agents rather than as a verdict on AI in science.
Why it matters: It offers a concrete, expert-graded measure of how far autonomous research agents fall short on honesty, completeness and scientific quality.
A Google Research AI can detect signs of high blood pressure using a regular smartwatch's pulse and motion sensors, with no cuff needed for calibration, according to a study published in Nature. In a study of 196 people across five US research sites, the watch identified 65.8 percent of those with high blood pressure compared to 55.3 percent for standard in-office measurement. Google says the findings have been folded into the blood pressure trends feature in the Google Health app, but the method remains a research project and does not replace a diagnosis.
The New York Times sued Brave on Nov. 2, alleging it copied more than 30 billion webpages without identifying its crawler or offering a way to block it, then sold that content to AI companies through its Search API, which Brave says is used by Cohere, Mistral, Perplexity and You.com. Because Brave's crawler does not advertise a differentiated user agent and instead follows Googlebot's rules, site owners cannot block Brave without also blocking Google, and opting out via noindex also removes pages from Google. Brave has not yet responded to the suit.
Crusoe, a Denver-based data center developer recently valued at $30.9 billion, confirmed it sold an unfinished data center project in southeastern Wyoming to Google, which had been a customer Crusoe failed to land for the site and is now planning to invest $17 billion in the campus. Google did not disclose the purchase price, and the deal highlights the difficulty independent developers face in delivering massive projects on time and within budget. Crusoe is still developing another data center for Google near Amarillo, Texas, and raised a $3.9 billion round in September despite the Wyoming setback.
The author of a previous HN thread asking how people use AI and do agentic coding built an interactive site from roughly 750 comments after finding the thread hard to parse. Using Claude, they extracted tools, models and plans from each comment, with every number linked back to its original comment. The site covers subscriptions, models, harnesses, orchestration and agent platforms, likes and dislikes per tool, and notable stories from the thread.
Julius Research introduces DeepFinance, an ultra long-horizon benchmark of 21 tasks built from real, anonymized Private Equity and Investment Banking workflows that take professionals 20–30 hours each, averaging 19 files per data room and 84 rubric criteria per task. Grading covers output, workflow and judgment, and on the leaderboard GPT-6 Astra (max) leads at 57.6%, followed by Claude Opus 5.5 (54.8%), Claude Sonnet 5.5 (54.6%) and GPT-6.1 Sol (54.3%), which are not statistically distinguishable, while the bottom group scores 10.9% to 14.5%. The authors report that agents often find decision-relevant evidence without acting on it and execute calculations correctly on wrong inputs, and argue that auditability of the trajectory matters as much as final numbers.
In a guest post for Noahpinion, the author argues that current evidence shows AI self-improvement loops would need to be roughly 5–10× stronger to sustain themselves, let alone produce a runaway intelligence explosion. OpenAI internal data shows models fully completing research tasks of only about 15 minutes at 80% success, far below benchmark and forecast horizons, while Anthropic estimates doubling overall AI progress would require roughly 40× productivity uplift. The author expects rapid AI progress and possible narrow superintelligence in highly verifiable domains, but sees no signs of accelerating returns and calls for better data.
MIT researchers developed CheMeleon, a pre-trained foundation model built on the open-source ChemProp package that learns from descriptors of a million molecules in PubChem and can be fine-tuned to predict molecular properties from limited data. In tests across 58 cases it outperformed other foundation models and classical expert-designed models, and nine of the top 20 entries in an OpenADMET contest used it. The findings were published in the Journal of Chemical Information and Modeling by Zalte, Burns, Abreu, Green, and three BASF researchers.
The 2026 AI adoption study from Wharton Human-AI Research and GBK Collective examines proving value, workflow transformation, and organizational alignment, finding that positive returns are near-universal but substantial ones are not. Organizations with strongly positive ROI are distinguished by measuring growth outcomes rather than time saved, redesigning workflows and management structures, appointing a CAIO, standardizing on a platform, and giving AI agents meaningful authority with automated monitoring. The report cites figures including 67% strongly positive ROI among leaders whose department meets all three components of transformational success versus 19% of those that do not, and 58% with a newly designated CAIO versus 8% with no such role.
The dataset lives at github.com/pjlsergeant/ai-sandboxes, where every claim traces to a primary source and pull requests to correct or add entries are welcome. It was built by the same person behind byre, described as a comfortable, constrained container box for coding agents. The post suggests byre may be the right amount of fence if the tools on the page feel like overkill for day-to-day use.
This post discusses how reinforcement learning post-training generates trajectories by running inference with the current policy model, and how disaggregated setups run inference and training in parallel on separate GPU resources for efficiency. Because the policy model is updated only on the training GPUs, the system must periodically refresh the weights on the inference side, and the post introduces the Reshard Collective to accelerate this weight refit.
Amazon SageMaker Inference now supports NVIDIA Multi-Instance GPU (MIG) partitioning on inference component endpoints, letting you divide a physical GPU into up to seven hardware-isolated partitions and place a different model in each. Partitioning is declared via AcceleratorPartitionLayouts on the endpoint config and AcceleratorPartitionType on each inference component, with no new API operations or IAM actions. In testing, a mig-3g.40gb partition of an H100 delivered 53–89% more output tokens per second and 18–34% lower cost per million output tokens than a full A100-40GB above a workload-dependent concurrency crossover.
Databricks describes how Feature Views in its Feature Store let teams define a feature once and use the same definition for point-in-time-correct training data and managed online-serving pipelines, reducing training-serving skew and data leakage. The walkthrough follows a $900 payment from a Kafka stream through rolling and Sawtooth window aggregations, an on-demand deviation feature, model training and Model Serving, with Unity Catalog providing governance and lineage.
Databricks outlines why public sector AI projects stall on governance questions—model access, data residency, and spend visibility—and describes how its Unity Gateway provides centralized controls over which models, coding agents, and tools teams can use, with per-user or per-team spending caps, platform-level compliance policies, and geographic data residency for hosted foundation models. The piece argues that agencies which solve governance centrally before scaling adoption can match each mission to an appropriate model while keeping data, users, and budgets under control.
OpenAI CFO Sarah Friar said the company will continue investing in AI infrastructure to meet customer demand, even after voters backed restrictions or rejected pro-development measures in all but one of 41 local ballot measures across 10 states. Friar said she doesn't think the election results dramatically changed anything for OpenAI, which published a policy blueprint and commitments for its company-owned data centers on Thursday, and said the company would continue to look into a 2027 IPO after its confidential June SEC filing. She also said OpenAI is in talks to raise funding at a $1.25 billion pre-money valuation.
Baseten announced the general availability of Serverless Loops, a Tinker-compatible training SDK for supervised fine-tuning and reinforcement learning on frontier open models at 256k context length, with one-click deployment to Baseten Dedicated Inference or weight download in Hugging Face Safetensors format. The release cites early users including NVIDIA, Datadog, LangChain Labs and Ambience Healthcare, and reports NVIDIA's post-trained Nemotron Super 3.5 VL improving intent classification accuracy from 76.8% to 94% and reaching 87% validation F1 on receipt extraction.
Why it matters: Useful for readers weighing managed post-training options, as it pairs a GA launch with named early adopters and concrete reported results.
Village's LLM Atlas, which scores 857 models on one capability scale, finds that since 2024 the best models from outside the US and China have trailed US models by a median of 9.9 months, versus 5.2 months for China. South Korea and Mistral are closest at about 10 months behind, while India's Sarvam 30B scores where US models were 28 months earlier. Labs outside the two countries released about 20 of the models Atlas ranks from 2026, against 113 from the US and 79 from China.
Ramp data shows token spend grew only 1.5% across August and September, down from 42% month-over-month growth in July, even as token volumes rose. Ramp's lead economist Ara Kharazian attributes the slowdown to price cuts from OpenAI and Anthropic, enterprise discounts, and customers getting thriftier, warning that heavy reliance on venture-backed tech companies could cause simultaneous slowdowns. October data showed a re-acceleration, with token spend up 8.9% from September 23 to October 23.
This newsletter surveys 27 independently founded, research-driven "neolabs" in the US and Europe, grouped by their primary technical bet, and contrasts their founders' doubts about today's recipe of ever-larger language models with claims from Jensen Huang and Greg Brockman that AGI has arrived. The first group targets continual learning and learning from few examples (Adaption, Core Automation, Flapping Airplanes, Ineffable Intelligence, Ndea, Safe Superintelligence), while the second bets language alone is insufficient and pursues spatial and physical intelligence (AMI, Black Forest Labs, General Intuition, Physical Intelligence, World Labs). It notes the funding behind these labs, including Nvidia's reported five billion dollar investment in SSI and AMD's planned acquisition of World Labs for about 8.2 billion dollars, and asks whether they can deliver something the big players lack.
This post profiles Fabian Theis of Helmholtz Munich, whose lab has pushed algorithms for multi-scale, multi-modal, time-varying biological data and helped shape the Virtual Cell concept. It notes that a 2025 Arc Virtual Cell Challenge showed constant baselines often beat sophisticated perturbation-prediction models, though better-calibrated metrics let deep learning outperform the mean baseline, and that Theis argues existing datasets make perturbation prediction too easy a problem. It frames the roughly $1.8B Biohub commitment from Biohub, NIH, DOE, Google and Meta as funding large, high-precision, multi-scale data collection to predict evolution across molecules, cells, niches, tissues, organs and organisms.
This piece discusses Fabian Theis's work on algorithms for multi-scale, multi-modal, time-varying biological data and the Virtual Cell concept, alongside the recent $1.8B Biohub commitment from Biohub, NIH, DOE, Google and Meta. It notes that the 2025 Arc Virtual Cell Challenge showed constant baselines often beating sophisticated perturbation-prediction models, though a cited paper argues better-calibrated metrics let deep learning outperform the mean baseline. Theis argues perturbation prediction on existing datasets isn't hard enough to warrant complex algorithms, and that focus should shift to large, high-precision, co-registered data spanning molecules through organisms.
The article describes knowledge workers who now spend their days managing and reviewing the output of AI agent "swarms" that draft copy, write code, and produce strategy, and argues that when every company in an industry runs its thinking through the same small set of LLMs, competitive differentiation erodes and skills atrophy. Drawing on the Kissimmee River straightening, Kyle Chayka's "Airspace," and research on LLM homogenization, it predicts a barbell-shaped future in which cheap high-volume AI output and premium verifiably human craft both thrive while the middle hollows out.