Models, products and AI startups, near-duplicates collapsed to the most credible source
Is regulation enough to stop AI systems killing us all?The article questions whether regulation alone can ensure AI safety. Author Martyn Thomas notes frontier AI companies have not provided detailed risk analyses or safety cases, while traditional safety-critical systems require proof of extremely low accident probabilities. He argues regulation must be based on rigorous analysis, not fantasy.
21M model with 6.4B lookup table matches 114M dense model, table on SSDDeveloper fehcyy released sparse-memory-lm, giving a 21M-parameter model a 16.8M-row lookup table (6.4B table params, 33M used per token), achieving performance comparable to a 114M dense model on 500M Wikipedia tokens. The table can be 4-bit quantized and memory-mapped from NVMe SSD, achieving 140 tok/s on RX 9070 with only 0.4 GB VRAM, but long prompts are slow due to 4KB page misses. Triton kernels run unchanged on Radeon, MI350X, and H100/H200. Bolting a table onto a pretrained model didn't work.
S&P 500 hits record high on AI bets as oil dropsStocks rose for a fourth straight session as oil prices fell, with the S&P 500 up nearly 1% and heading toward its first record close since August, buoyed by AI trade enthusiasm and optimism over corporate resilience.
NJ Assemblywoman Pushes AI Companion Chatbot Transparency BillNew Jersey Assemblywoman Marisa Sweeney introduced bill A4732 requiring AI companion chatbots to clearly notify users they are not talking to a human. Sweeney said transparency matters more as companions grow more lifelike; the bill cleared committee last month.
RFK Jr.'s AI Embrace Confounds MAHA Allies and Medical ExpertsUS Health Secretary Robert F. Kennedy Jr. endorsed AI at a MAHA summit, saying it lets people fact-check medical advice and escape 'medical tyranny.' Some MAHA voices criticized the corporate embrace, while medical experts said his vision is flawed because AI aggregates medical literature and expertise.
Mirror Particle builds world model of human behavior, challenging LLM role-playMirror Particle will launch at TechCrunch Disrupt with a from-scratch world model to predict human behavior, arguing LLM role-play falls short for market research. The company believes LLMs only model language, while human behavior involves visual, spatial, and social intelligence, requiring longitudinal data to capture dynamic changes.
AI hardware blurs definition of 'recording', tech companies push boundariesWith AI hardware, tech companies are redefining what constitutes a 'recording.' Previously, a device with a microphone or camera meant it could record, but now the line is blurring. The article discusses this trend, citing Apple's smart home device as an example.
Comparing memory mechanisms across RNNs, Transformers and SSMsA community post compares architectures through working memory: RNNs compress history into a recurrent state with roughly O(N) state versus O(N²) parameters, Transformers store past representations in a growing KV cache while weights stay frozen, and selective SSMs like Mamba return to fixed-size recurrent memory with input-dependent retention.
Open-source openTPU shows AI designing its own inference hardwareThe open-source openTPU accelerator, developed by AI agents, spans SystemVerilog hardware, an instruction set, a bit-exact simulator, a kernel language and compiler, and host software, running ten real-weight models on a Xilinx Kintex-7 FPGA card with bit-exact token output versus the simulator.
AI startup Hark debuts Hark Pro app ahead of 2027 hardware launchHark, the AI-focused hardware startup founded by Brett Adcock, introduced a new "proactive" Hark Pro app ahead of launching its first devices next year.
AAP issues guidance on generative AI use in pediatricsThe American Academy of Pediatrics (AAP) released guidance on the use of generative AI in pediatrics, offering a framework for clinical and related applications.
McDonald's hit with federal lawsuit over alleged AI price hikesA potential class action filed in Illinois federal court accuses McDonald's and its franchisees of sharing nonpublic pricing data and using a Tiger Analytics pricing engine to raise prices, violating the Sherman Act and Illinois laws. Menu prices allegedly rose about 40% over five years.
User Buys Mining Rig With 54GB VRAM for $35A user bought an old mining rig on a Russian classifieds site for 3,000 rubles (~$35) containing nine working P106 6GB GPUs totaling 54GB of VRAM. Only an SSD was missing; it booted fine from USB.
Insurers Brace for Millions in Claims as AI Agents Spin Out of ControlThe Decoder reports that insurers are bracing for millions in claims from rogue AI agents, with executives like OpenAI's Sam Altman and Anthropic's Dario Amodei potentially personally on the hook.
ASP Data Center Said to Weigh Oslo ListingAI infrastructure developer ASP Data Center AS is reportedly considering an IPO in Oslo that could raise about $100 million, signaling continued fundraising activity in AI compute infrastructure.
Atlassian and OpenAI expand partnership to turn enterprise knowledge into actionAtlassian and OpenAI are expanding their partnership to connect frontier models with enterprise knowledge, helping teams plan, build, and deliver work. This aims to boost enterprise productivity and marks a significant move for OpenAI in enterprise software.
Anthropic Offers Startups Free Year of Claude Team and $1,000 CreditsAnthropic expanded its Claude for Startups program, offering qualifying startups founded in the last five years or funded in the last two years a free year of Claude Team with five premium seats, $1,000 in API credits, Claude Marketplace access, and virtual office hours with its Applied AI team.
Google Releases Open Multimodal Embedding Model EmbeddingGemma 2Google DeepMind released EmbeddingGemma 2, an open multimodal embedding model mapping text, images, video and audio into a unified 768-dimensional space. With 740M total parameters, it supports 100+ languages, improves code tasks ~14% over its predecessor, and runs on consumer hardware like phones and laptops with low latency.
LibreOffice Says 'No AI' Is Now a Software FeatureThe Document Foundation, maker of LibreOffice, said it will not add AI to its software's default configuration due to user privacy, noting its latest version contains no generative AI features and users can install extensions to connect to local AI models.
Woman reported to police after using Claude as diaryA woman reportedly used Claude as her personal diary and was reported to police over its contents. The post offers no further details and the claim is unverified.
OpenChart: an open-source TradingView alternative with AI agentsOpenChart is an open-source TradingView alternative that lets AI agents like Claude and Codex work directly with charts in one workspace, annotating setups, writing indicators, investigating moves and setting alerts; the backend and data stay local, it currently supports only Apple Silicon macOS, and works with free market data.
Is Artificial Intelligence a Public Health Threat?The article explores whether AI poses a public health threat, mentioning Anthropic and its Claude models. Written by Devon Herrick of the Goodman Institute, it does not offer specific conclusions.
Court Rules Copyrighted Material as AI Training Data Is Not Fair UseThe US Third Circuit Court of Appeals upheld a ruling that Ross Intelligence violated copyright law by training its legal search engine on Thomson Reuters' Westlaw materials. This first-of-its-kind decision could weaken AI companies' fair use defense, as OpenAI and others face numerous similar lawsuits.
Qwen3.8-Flash-Next gains Strix Halo support on StrataA developer released official Strix Halo support for Qwen3.8-Flash-Next on Strata, claiming best long-context decode performance, up to 1M context length, Linux-only and marked experimental.
AI adoption to surpass past tech cycles, Neostellar principal saysNeostellar principal Willy Lee told Bloomberg that AI will see massive adoption unseen in previous tech cycles, discussing AI investment and the IPO market from a VC perspective. Bloomberg Intelligence noted the US AI performance lead over China narrowed to a record low as labs like DeepSeek gained ground.
Mistral releases 1T-parameter multimodal model Large 4French lab Mistral AI released Mistral Large 4, a 1-trillion-parameter multimodal model nicknamed Le Chonk, accessible only via a guarded public endpoint with open weights promised in three weeks after safety testing. Trained on just 4,000 NVIDIA GPUs, using two to three times less compute than Chinese rivals, it targets use cases like cybersecurity.
Dimon discusses AI boom, bond selloff, UK banksJPMorgan Chairman and CEO Jamie Dimon discussed financing the AI boom, the global bond selloff and the state of UK banks with Bloomberg's Tom Mackenzie at the firm's Tech Stars Conference in London.
Pinterest launches AI Beauty Guides to turn Pins into salon plansPinterest launched Beauty Guides, powered by its generative AI, to translate hair and nail Pins into salon terms like "balayage" and give estimated costs, appointment times and maintenance needs. The feature covers hairstyle, color, nail shape and finish categories, helping users plan salon visits.
Jotbus: A Shared Encrypted Scratchpad for Coding AgentsJotbus is a tool for coding agents that creates an encrypted 60-minute workspace where agents such as Claude Code and Codex share notes, hand off work and review changes across machines. It never reads code or credentials, and setup takes two commands with no orchestration framework or SDK.
Oracle Blamed for Bursting the AI BubbleA Hacker News discussion thread claims Oracle triggered the implosion of the AI bubble, sparking debate in the tech community. No specific financial details or events were provided.
The A.I. Daily