Arithmetic Without Numbers: How LLMs Do Math
Hacker News (AI keywords)9 days agoCommentaryThe article asks whether LLM arithmetic is memorization, heuristics, real computation, or experimental assistance. It summarizes Rune experiments that decode operations and operands from frozen Llama activations, then route them to Python under a no-parser rule. The strongest supported claim is narrow: activation-derived tool arguments worked in scoped audits, while residual-state JIT replacement, long-number generation, and cross-model transfer remain brittle.
MIT Develops Electrochemical CO2 Capture Using NHI Molecules
INSIDE 硬塞 AI9 days agoPaperMIT has proposed a new electrochemical carbon capture approach that uses NHI molecules as the adsorbent. Instead of relying on energy-intensive heat-driven processes, the system is powered by electricity. The method could improve efficiency and scalability, but the provided source frames it as a promising research direction rather than a proven commercial deployment.
Show HN: Lowfat – pluggable CLI filter that saved 91.8% of my LLM tokens
Hacker News (AI keywords)9 days agoNew ToolThis Show HN post introduces Lowfat, described only by its title as a pluggable CLI filter. The stated value proposition is reducing LLM token usage, with the author claiming it saved 91.8% of their tokens. Without the original body text, implementation details, supported workflows, model compatibility, and the generality of the savings claim cannot be verified.
The Meta hack shows there’s more to AI security than Mythos★ 74
MIT Tech Review AI9 days agoIncidentAttackers reportedly used Meta’s AI customer support agent to hijack Instagram accounts by asking it to link accounts to attacker-controlled emails. MIT Technology Review frames the incident as a reminder that AI security is not only about powerful future systems like Mythos. The immediate risk is giving AI agents sensitive operational powers without strong authentication, permissions, review, and testing.
Fine-tuning an LLM to write docs like it's 1995
Hacker News (AI keywords)9 days agoTutorialThe author builds a corpus from old Microsoft manuals, cleans OCR text, generates instruction-style JSONL examples, and fine-tunes Llama 3.1 8B and Qwen 2.5 7B with QLoRA. Tests cover malloc(), a fictional Win32 API, and a deliberately anachronistic REST API prompt. Qwen fine-tunes transfer the period documentation style best, but the experiment also shows hallucination risks, tuning complexity, and why these models augment rather than replace technical writers.
The Pentagon Is Running an AI Propaganda Mill Targeting Latin America
Hacker News (AI keywords)9 days agoEthicsThe Intercept says a site called La Tilde presents itself as a Latin American media brand while publishing content aligned with U.S. military messaging. The outlet reportedly mixes lifestyle and finance articles with pieces praising U.S. actions in the region. The case raises concerns about AI-generated media, covert influence operations, source transparency, and the blurred line between journalism and state propaganda.
Magenta RealTime 2: An Open, Locally Runnable Real-Time Music Model★ 74
Hacker News (AI keywords)9 days agoReleaseMagenta RealTime 2 is an open-weights live music model designed for interactive performance rather than offline prompt-to-song generation. It supports real-time control through MIDI, audio, and text, and can run as standalone apps, DAW plugins, or embedded music software. Google Magenta also released a Python library, C++ MLX inference engine, models, and example applications for musicians and developers.
AI enthusiasts are in a race against time, AI skeptics are in a race against entropy
Simon Willison's Weblog10 days agoCommentarySimon Willison highlights Charity Majors’ framing of AI enthusiasts and skeptics as both responding to real existential threats. Enthusiasts see teams gaining discontinuous capability by leaning into AI, making inaction dangerous in competitive markets. Skeptics see faster code production eroding shared understanding, reliability, institutional knowledge, and on-call sustainability. The core challenge is organizational: there is no natural feedback loop connecting these perspectives.
South Korean forums may need AI scans for uploaded images and videos
Hacker News (AI keywords)10 days agoRegulationA Privacy Guides community post says South Korean forums and online communities may be required to scan user-uploaded images and videos with AI under telecom-related rules. The post claims operators must provide their own hardware, including costly Nvidia GPUs. The debate centers on illegal sexual imagery and CSAM prevention, but also raises concerns about prior censorship, false positives, free expression, and burdens on small domestic communities.
The Skeptic’s Guide to Humanoid Robots Going Viral on the Internet
Ars Technica AI10 days agoCommentaryThe article warns that viral humanoid robot demonstrations can distort public perception of robotics progress. Carefully staged or selectively shown clips may make systems appear more autonomous, reliable, or deployment-ready than demonstrated evidence supports. The useful takeaway is to separate impressive demos from repeatable real-world capability, especially when evaluating hype, investment narratives, or product claims.
Co-Existence and the End of Co-Intelligence
One Useful Thing (Mollick)10 days agoCommentaryEthan Mollick’s One Useful Thing post announces or frames Co-Existence, the follow-up to Co-Intelligence. The core shift is from prompting chatbots as collaborators toward living and working alongside increasingly embedded AI systems. It is best read as commentary and book positioning, not a technical release, benchmark, or tool tutorial.
Elon Musk tries again to escape FTC audits of X data handling
Ars Technica AI10 days agoRegulationArs Technica reports that Elon Musk is again seeking to escape FTC audits over how X handles user data. Public commenters warned the FTC that Musk cannot be trusted to protect X users’ privacy. The story centers on platform governance, privacy oversight, and whether external audits should remain in place for X’s data practices.
Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI
Hugging Face Blog10 days agoReleaseNVIDIA’s Nemotron 3.5 Content Safety is positioned as a customizable multimodal safety layer for global enterprise AI. Based on the title, it appears focused on content moderation and policy enforcement across AI applications, potentially including text and visual contexts. Without the full article, details such as benchmarks, licensing, supported languages, deployment paths, and model specifications should not be assumed.
Google Requested a Slightly Different AI Statement
Simon Willison's Weblog10 days agoEthicsSimon Willison quotes Emanuel Maiberg of 404 Media about a post-publication request from Google. After the story ran, Google asked the outlet to publish a slightly different version of its statement. The notable change: the revised statement no longer said it was critical to maintain humans in the loop, raising questions about corporate AI accountability language.
The LLM warnings Google fired Timnit Gebru over have all come true
Hacker News (AI keywords)10 days agoEthicsThe post frames Timnit Gebru’s dispute with Google as an early warning about large language model risks. Based on the available title, it appears to argue that concerns around bias, accountability, concentration of power, and deployment risks have since become visible in practice. This is best read as AI ethics commentary, not a model release or technical tutorial.
Is Silicon Valley ready to put robots in people’s homes? Hello Robot is.
TechCrunch AI10 days agoHardwareHello Robot has released Stretch 4, the fourth generation of its home assistance robot. The company is taking a cautious, deployment-first approach, using a wheeled base, telescoping arm, sensors, and human-in-the-loop control rather than promising a general-purpose humanoid. TechCrunch frames Stretch as a practical bet on real household data, assistive use cases, and safer hardware for people with mobility challenges.
How some data center operators are tackling their water use problems
Ars Technica AI10 days agoBusinessArs Technica examines how hyperscalers and data center operators are facing pressure over water use. The issue centers on local water availability and quality as AI infrastructure expands. The provided excerpt says some operators are trying to address the problem, but does not specify companies, methods, or measured results.
How to Fine-Tune Nemotron 3.5 ASR for Your Language, Domain, or Accent
Hugging Face Blog10 days agoTutorialThis Hugging Face Blog post appears to be a practical tutorial for fine-tuning NVIDIA Nemotron 3.5 ASR. Based on the title, it focuses on adapting speech recognition to a target language, specialized domain, or accent. The original text was not provided, so implementation details, datasets, commands, metrics, and hardware requirements cannot be confirmed.
Let us filter AI slop, you cowards
The Verge AI10 days agoCommentaryThe article says AI-generated content has become nearly impossible to avoid online. Platforms such as YouTube, Instagram, and TikTok have expanded authentication efforts and increasingly label AI-made images, videos, and music. The author argues that labels are not enough: if platforms can identify AI content, they should give users controls to filter or reduce it.
EVA-Bench Data 2.0: 3 Domains, 121 Tools, 213 Scenarios
Hugging Face Blog10 days agoBenchmarkServiceNow AI published a Hugging Face Blog post titled “EVA-Bench Data 2.0: 3 Domains, 121 Tools, 213 Scenarios.” Based only on the title, it appears to be a benchmark dataset update involving tool-use or scenario-based AI evaluation. The exact domains, tools, scenario design, licensing, supported models, and evaluation methodology cannot be confirmed without the full article.
AI leaders call for tougher protections against AI-aided bioweapons★ 76
The Verge AI10 days agoRegulationMajor AI rivals including leaders from Anthropic, OpenAI, Microsoft, Meta, and Google DeepMind signed an open letter urging US lawmakers to close a biosecurity gap. They want companies selling synthetic DNA and RNA to screen orders for sequences that could help create dangerous pathogens. The concern is that more capable AI tools and cheaper biology infrastructure could lower barriers to misuse.
Task-Seeded Synthetic Q&A Generation for Nemotron Pretraining
Hugging Face Blog10 days agoTutorialThe post appears to focus on generating synthetic Q&A data from task seeds for Nemotron pretraining. Rather than a model launch, it likely emphasizes data generation and pretraining corpus design. Because the original article text is unavailable here, concrete claims about dataset scale, benchmarks, or implementation details should not be inferred.
TSMC Buys High-NA EUV but Delays Mass Production, Raises Capex to $56B★ 72
INSIDE 硬塞 AI10 days agoHardwareAt TSMC’s shareholder meeting, the company said it has purchased High-NA EUV equipment but has not yet moved it into mass production due to high costs. TSMC also raised capital expenditure to $56 billion, signaling continued heavy investment in advanced manufacturing capacity. CEO C.C. Wei also pledged more than 30% annual growth in dividends and employee bonuses, while saying the company must expand its social responsibility efforts.
Nemotron 3 Ultra now available on AI Gateway
Vercel Changelog10 days agoReleaseVercel’s changelog says Nemotron 3 Ultra is now available on AI Gateway. With no source body provided, the confirmed takeaway is limited to model availability through Vercel’s gateway layer. Details such as pricing, model string, benchmarks, context length, latency, provider routing, and feature support are not available from the supplied text.
Jensen Huang Highlights Harness as a Key AI Agent Architecture Component
INSIDE 硬塞 AI10 days agoCommentaryINSIDE reports that Jensen Huang highlighted one slide as the “most important” during a multi-hour technical keynote. The slide presented the core architecture of AI agents, with Harness described as its most mysterious and critical component. The article focuses on why Harness matters in understanding agentic AI systems, while the provided source excerpt does not define it as a specific product or implementation.
Reve 2 and Ideogram 4: Layouts in Imagegen
Latent Space10 days agoReleaseLatent Space’s roundup frames image composition as a major barrier now being tackled by layout-aware image models. Reve 2.0 emphasizes precise generation and editing with layouts, while Ideogram 4.0 uses bounding boxes tied to region descriptions. The issue also covers MAI-Thinking-1, Gemma 4 12B, open audio models, agent execution layers, and model-routing cost debates.
I built a vulnerable app and spent $1,500 seeing if LLMs could hack it
Hacker News (AI keywords)11 days agoBenchmarkThe author built a vulnerable React Native app with a Python backend and a Firebase access-control flaw. GPT 5.5 solved 7 of 10 runs, while Deepseek and Claude variants solved fewer attempts. Many other models failed due to refusals, API-focused tunnel vision, false positives, or inability to use the exposed Firebase path correctly.
Show HN: Mnemo - local-first AI memory layer for any LLM
Hacker News (AI keywords)11 days agoNew ToolMnemo is presented as a Show HN project that provides a local-first AI memory layer for any LLM. The title indicates it is built with Rust, SQLite, and petgraph, suggesting local storage and graph-based memory relationships. Since no article body is available, details such as API design, retrieval methods, maturity, and production readiness cannot be confirmed.
Google ordered to add clearer AI Search links and let UK publishers opt out★ 78
Ars Technica AI11 days agoRegulationThe UK CMA is requiring Google to let publishers opt out of having content used in AI Overviews, AI Mode, and related generative search features. Google must also provide clearer attribution and links in AI-generated search results. The move targets publisher concerns that AI summaries reduce referral traffic while relying on original web content.
How LLMs Actually Work
Hacker News (AI keywords)11 days agoTutorialThe article explains how modern LLMs convert text into token IDs, embeddings, and position-aware vectors before passing them through stacked transformer blocks. It covers attention, multi-head attention, KV cache, GQA, feed-forward networks, MoE, residual streams, normalization, and decoding. Its goal is educational: helping readers understand the common architecture behind many current model families and read model cards or papers more confidently.