Research updates reveal the next wave of product capabilities and infrastructure needs. This hub follows papers, benchmarks, datasets, lab systems, releases, and open reproductions, focusing on which results may reach model training, agent systems, robotics, or developer tools.
The Soofi consortium unveils Soofi S, a 30B Mixture-of-Experts model trained on 27 trillion tokens, focused on German and English, for industrial applications requiring control and transparency. The model is currently in testing with partners and not yet publicly available.
Soofi S is a 30B MoE model trained on 27 trillion tokens, optimized for German and English.
Designed for industrial use cases including technical documents, code generation, and agentic AI.
Teya is an open‑source AI family agent that turns a cheap Android phone into a wall‑mounted smart home hub. It understands context, remembers personal facts, and performs tasks like shopping lists, calendar management, timers, reminders, expense tracking, and safe calling. Privacy is built in: all data stays on device, and conversation transcripts are never saved.
Runs on a cheap Android phone (Android 8.0+), no server needed.
Voice‑controlled, recognizes individual family members, and remembers personal details.
SlopSift uses a custom-trained dependency parser to detect canned arguments, unsupported claims, and filler in writing. It runs locally, respects privacy, and offers CLI and agent integration for automated linting.
Employs a small dependency parser to analyze word relationships and identify structural issues.
Fully local: model and rules run on-device without uploading data.
The AI boom is increasingly financed by debt, but investor demand is falling as hyperscalers accelerate bond issuance. Amazon's recent bond sale required higher yields due to lower demand, with order coverage dropping. AI bond supply is surging while investors demand wider spreads. Meanwhile, the breakthrough performance of Chinese AI model Kimi K3 raises concerns about the sustainability of US AI spending, potentially leading to an economic slowdown.
Since early 2025, Alphabet, Meta, Amazon, and Oracle have issued over $300 billion in bonds.
Investor demand for AI bonds is declining; Amazon's bond orders fell from 3.2x to 2.5x coverage.
PixelUp is a lightweight AI video upscaler for Windows that runs entirely offline. It uses FSRCNN and ESPCN deep learning models to upscale low-resolution videos to high-definition quality with fast processing and hardware acceleration. The software offers one-time purchase pricing ($19) with no subscriptions, batch processing, lossless audio syncing, and full privacy protection.
100% local processing, no internet or account required, ensuring data privacy.
Utilizes optimized FSRCNN and ESPCN models for rapid detail reconstruction.
This article explores how to develop reasoning models with multiple effort modes, covering the evolution from o1 and DeepSeek-R1 to GPT-5.6, and key techniques such as RLVR training, inference scaling, think tokens, and reasoning mode toggles.
Reasoning models output intermediate reasoning traces, distinguishing them from conventional LLMs.
RLVR training rewards only final answer correctness, not the reasoning trace.
AI company logos commonly feature circular gradients with central openings, humorously compared to anuses. The article analyzes the design psychology, unintended biomimicry, and copycat effect behind this trend, and reviews tech design history.
Many AI logos share circular, gradient, and central opening features, jokingly called 'butthole style'.
The trend stems from circular design psychology, unconscious biomimicry, and industry mimicry.
Terry Tao uses analogies from chess and mathematics to argue that the best way to improve is to continually challenge yourself with problems slightly above your current ability. He warns against the extremes of playing it safe (leading to stagnation) or aiming too high (leading to frustration), and advocates a balanced approach of tackling problems just outside one's range, collaborating with others, teaching, and using simplified models.
The optimal growth strategy is to consistently aim for problems slightly beyond your current capability.
Avoid the extremes of only easy tasks or only impossibly hard ones.
Google Cloud's generative-ai repository ships the Always-On Memory Agent, a reference implementation that treats memory as a running process. Built on Google ADK and Gemini 3.1 Flash-Lite, it uses no vector database and no embeddings. Instead, an orchestrator routes to Ingest, Consolidate, and Query sub-agents that read, connect, and write structured memory into SQLite 24/7.
Always-On Memory Agent is a lightweight background process that runs 24/7, using Google ADK and Gemini 3.1 Flash-Lite.
It eliminates vector databases and embeddings, relying on an LLM to write structured memory to SQLite.
San Francisco City Attorney David Chiu has sent letters to Apple and Google demanding the removal of several AI-powered 'nudify' apps that can create nonconsensual intimate images. The letters cite California's deepfake laws and call for better screening. Meanwhile, concerns about Grok generating CSAM add pressure on app stores.
San Francisco City Attorney demands Apple and Google remove nudify apps that violate deepfake laws.
Researchers found 70% of tested face-swapping apps can be used for nudification.
A shift from copilot-style AI to agentic AI is transforming marketing. Autonomous agents now execute multi-step campaigns, optimizing budgets and channels without human intervention. While powerful, success requires human oversight of goals and boundaries.
Agentic AI systems autonomously execute marketing goals, unlike copilot tools that require prompts.
Examples include Salesforce Agentforce, HubSpot Breeze, Adobe Agent Orchestrator, and Braze Operator.
Sakana AI's Error Diffusion is a local learning rule that trains neural networks without weight transport or backpropagation while obeying Dale's principle. It uses a dual-stream architecture with excitatory and inhibitory pathways, and modulo error routing to scale to multi-class classification, achieving 96.7% on MNIST and 61.7% on CIFAR-10. The innovations show task-dependent importance, and the method extends to reinforcement learning via ED-PPO, outperforming BP-PPO on some tasks.
Error Diffusion trains Dale-compliant networks without backpropagation or weight transport.
Modulo error routing scales ED beyond binary classification to MNIST and CIFAR-10.
Anthropic will include Claude Fable 5 in all Max and Team Premium plans at 50% limits starting July 20, and offer a one-time $100 credit to Pro and Team Standard users, reversing its earlier plan to remove the model from subscriptions due to competitive pressure from GPT-5.6 Sol and others.
Claude Fable 5 becomes permanent in Max and Team Premium plans (50% limits).
Pro and Team Standard users get ongoing usage credits plus a $100 one-time credit.
Retriever launches agentic dataset enrichment running in your browser, allowing you to enrich contact lists from any webpage you're logged into, such as Luma event pages, with LinkedIn profiles, work emails, and more, then score and contact top prospects—all for about $1.25 per 500 records.
Works on any webpage you have open, using your existing logins to access attendee lists or employee directories.
From a single prompt, it extracts data from the page, matches against pre-indexed datasets, and performs live scrapes to fill in missing fields.
As AI tools proliferate, the definition of critical thinking needs expansion. This article breaks it down into reflection and judgment, highlights intellectual humility, and argues that education must cultivate the ability to make sound judgments under uncertainty.
Critical thinking involves two steps: reflection and judgment; digital environments erode the space for reflection.
Intellectual humility is crucial — recognizing the limits of one's understanding.
Tabstack is a Mozilla-backed platform that offers a unified API for extracting structured data, conducting research with citations, and automating browser tasks, without managing LLMs, browsers, or pipelines. It emphasizes privacy (no training on data, data purged) and uses the open-source browser engine Pilo to reduce token consumption.
Tabstack provides endpoints like /extract/json, /research, and /automate for data extraction, research Q&A, and browser automation.
All calls run on Mozilla-backed infrastructure; data is never used for model training and is promptly purged.
PenEcho is an open-source shared canvas that integrates AI for handwriting, equations, diagrams, and spatial context. It operates through a browser canvas, server validation, and multiple executors (OpenAI API, Codex CLI, Claude CLI) to generate editable drafts. Users can move, resize, accept, or discard each AI suggestion. The canvas supports a 20,000 x 20,000 logical size with sparse rendering, local snapshots, and various configuration options. Requires Node.js 18.17+ and an API key or authenticated CLI tools. The article covers installation, executor selection, security, and cost estimation.
PenEcho is an open-source AI-powered shared canvas for handwriting, equations, and diagrams.
It captures content on the browser canvas, validates via server, and generates drafts using AI executors.
Contrary to popular belief, AI hasn't shifted the bottleneck from coding to code review. The real constraint is downstream deployment batches, where changes accumulate after review. Over 90% of teams ship in batches, and speeding up code review only worsens the actual bottleneck.
The perceived shift to code review is a myth; the real bottleneck is deployment batches.
More than 90% of teams deliver changes in batches, with most having 2-10 pending changes.
Malwarebytes' 2026 report reveals that 85% of people can no longer distinguish real from AI-generated content, 50% have encountered AI-driven scams, with Gen Z most at risk. People are retreating from online sharing due to AI threats, but few take protective actions. The report also uncovers moral contradictions: many fear deepfakes yet find using AI for personal purposes acceptable.
85% of respondents say it's now hard to tell what's real, up from 66% in 2025.
50% of adults have encountered an AI-driven scam, with Gen Z exposure at 67%.
LangChain has released a hosted version of an open-source extraction service that supports extracting structured data from PDF, HTML, and text files. The service is free to use but not intended for production workloads or sensitive data. It allows users to define extraction schemas, add few-shot examples, and switch between different LLM models. With a simple frontend, developers can quickly experiment and integrate the service into their own LangChain workflows.
LangChain launched a hosted version of an open-source structured data extraction service with a simple frontend.
Supports PDF, HTML, and text files; users can define custom schemas and provide few-shot examples.
In 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon proposed the Dartmouth Summer Research Project on Artificial Intelligence, marking the birth of AI as a field. The proposal defined AI's goal: to make machines use language, form abstractions, solve human problems, and improve themselves.
Proposed in 1955 by McCarthy, Minsky, Rochester, and Shannon as the founding document of AI.
First use of the term 'artificial intelligence', setting the goal of simulating human intelligence in machines.
Oversikt.se is a public data and AI evidence engine for Swedish politics, offering interactive visualizations of taxes, budgets, party positions, and public opinion to enhance political transparency. Users can input their salary to see personal tax breakdowns and track government expenditures in real time.
Oversikt.se visualizes Swedish tax, budget, and political data, allowing personalized queries.
The platform integrates an AI evidence engine to help the public understand party budget proposals and their impacts.
This article explores the concept of bio-metals and focuses on a study that discovered metallic structures in the mouth of an ancient creature, revealing potential insights into biomineralization and natural materials.
Bio-metals are metallic elements found within living organisms, often with unique properties.
A new study examines mysterious metallic remains in the maw of an ancient organism.
Nurses at Kaiser Permanente report that workplace surveillance, including AI monitoring of call times and empathy, is undermining patient care and causing staff stress.
Nurses face criticism for calls over 15 minutes.
AI systems track call length, predict unproductivity, and rate empathy.
This tutorial demonstrates how to build an agentic event venue operator with persistent memory and operational context using MongoDB Atlas, Voyage AI embeddings, LangGraph, and optional Langfuse tracing. The demo scenario is the MongoDB Open, a fictional tennis tournament, where the agent handles weather disruptions, distinguishes visitor segments, and makes real-time decisions under capacity constraints. The article covers architecture, setup, UI walkthrough, memory store, vector search, hybrid search, and visual RAG.
Tutorial builds an agentic event venue operator with persistent memory and operational context, going beyond simple chatbot demos.
Uses MongoDB Atlas as the operational and memory layer, combined with Voyage AI embeddings and LangGraph workflows.
Zyphra released ZUNA1.1 on July 16, 2026, under the Apache 2.0 license. The 380M masked diffusion autoencoder reconstructs, denoises, and upsamples scalp-EEG across arbitrary channel layouts. It accepts variable-length inputs from 0.5 to 30 seconds, against ZUNA1's fixed five seconds. Reported NMSE holds or improves while the input range widens.
ZUNA1.1 accepts variable-length inputs from 0.5 to 30 seconds, tokenized into 0.125-second segments.
It uses a transformer encoder-decoder with 4D RoPE and rectified flow objective.
The Port Index is a free, searchable reference that aggregates 3,804 seaports and 9,640 airports worldwide, offering key details like depths, runways, codes, and coordinates from public-domain datasets—no signup required.
Free index of 3,804 seaports and 9,640 airports across 195 countries
Provides channel depths, vessel limits, runway lengths, and more
In a recent benchmark, GPT-5.6 Sol Ultra autonomously constructed a complete Chrome V8 exploit chain from scratch by analyzing security-fix patches, ultimately popping a calculator. Other frontier models like Sol Medium and Grok 4.5 stalled early. The author argues that exploit development as a human skill is now obsolete.
GPT-5.6 Sol Ultra completed a 9-step exploit chain in three days, including Maglev type confusion, sandbox read/write, sandbox escape, UAF, and code execution.
Sol Medium and Grok 4.5 failed to advance beyond sandbox primitives; Sol Ultra used 74 sub-agents and 2.1B tokens at a cost of ~$1,597.
Linus Torvalds defends the use of AI coding tools in Linux development, calling AI a pragmatic tool based on technical merit. He acknowledges AI isn't perfect but urges critics to first look at human shortcomings. Despite studies showing decreased productivity with AI tools, Torvalds emphasizes their utility and reveals he uses 'vibe coding' tools in his hobby projects.
Torvalds says AI is a useful tool and criticism should be based on technical merit, not fear.
He acknowledges AI's imperfections but notes human maintainers also have flaws.
Amazon Quick is an AI assistant that helps sales reps spend more time selling by automating CRM updates, prospect research, email drafting, and more. It covers the entire sales cycle from lead scoring to CRM automation.
Amazon Quick automates lead scoring and prioritization using CRM and other data.
It enables personalized outreach with context-aware email generation.
Chai Discovery Inc. announced a $400 million Series C funding round, tripling its valuation to $3.8 billion. The company develops AI models to predict biochemical interactions, and its latest model Chai-3 achieves 35-40% hit rates for molecular targets. It has secured partnerships with Pfizer, Eli Lilly, and Novartis, though no AI-discovered drug has yet been approved despite significant investment.
Chai Discovery raises $400M Series C, valuation jumps to $3.8B
New AI model Chai-3 doubles success rates for molecular interaction targets to 35-40%
This issue of The Download covers the hype and misinformation around perimenopause, China's new open-source AI model that narrows the gap with the US, and other tech stories including Trump Media's monetization, an atmosphere on an Earth-like planet, brain implants restoring feeling, and more.
Perimenopause discussions are more open but increasingly filled with misinformation and unsupported treatments.
A Chinese startup released the world's largest open AI model, competing with US models and impacting stocks.
Dutch climate activist group Extinction Rebellion claimed responsibility for an attack on a datacenter construction site in Amsterdam, throwing water balloons filled with an acidic mixture aimed at degrading concrete and steel. The facility, built by Pure Data Centres Group, is reportedly fully leased to Microsoft. The group says the action protests datacenters and AI worsening the climate crisis and Israeli actions against Palestinians. The builder says the attack had no impact and plans legal action.
Extinction Rebellion threw water balloons containing hydrogen peroxide, acetic acid, salt, and acrylic paint at a datacenter site.
The activists claim datacenters and AI exacerbate the climate crisis and are linked to Israeli actions.
Meta's new Muse Spark 1.1 model is now available on Databricks via Model Provider Services (MPS) in Unity AI Gateway. This service allows organizations to register providers once in Unity Catalog, eliminating API key sprawl and centralizing governance through familiar permissions, rate limits, and guardrails. Additionally, every request is automatically tracked with token usage, latency, cost attribution, and audit logs for end-to-end observability.
Access Meta's new Muse Spark 1.1 model on Databricks through Model Provider Services in Unity AI Gateway.
Register providers once in Unity Catalog to centralize access, rate limits, and guardrails.
Gen Z's vocal backlash against AI, from booing commencement speakers to online criticism, reflects a growing generational divide. Studies show younger generations are skeptical about AI's benefits, while baby boomers embrace it. The article argues that young people face an existential crisis and seek to reclaim agency over a future that feels predetermined by algorithms.
Gen Z boos speakers who praise AI at graduations, signaling strong pushback
Gallup study finds Gen Z unconvinced AI enhances creativity or critical thinking
Simon Willison developed a tool to detect and highlight common clichéd phrases often found in LLM-generated text, such as 'no fluff, no filler, no jargon'. The tool runs in the browser, provides pattern counts and navigation, and aims to reduce frustration with formulaic AI writing.
Simon Willison created the LLM cliché highlighter to identify overused phrases in AI-generated content.
The tool highlights patterns like 'no X, no Y' chains and 'you already know'.
The author recounts their journey from AI skeptic to enthusiast, building an LLM-driven MMO game (SAO: Slop Art Online) and encountering latency issues. They devised a hybrid NPC AI approach combining behavior trees with LLM decision-making, which inspired them to create SLOP, a protocol for agent-application interaction that features contextualized actions and state projections.
The author's perspective shift from AI hater to AI advocate after using Opus 4.5.
Developed an MMO where NPCs are controlled by LLMs, leading to a hybrid AI architecture.
OpenTSLM is a multimodal LLM that treats time series as a native modality, enabling reasoning over raw multivariate signals alongside text. It outperforms baselines, including GPT-4o, on time series QA, activity recognition, sleep staging, and ECG QA. The model scales to multiple long time series with near-constant memory consumption. ECG reasoning validated by 7 cardiologists with 97% correctness. All code, datasets, and models are open-source.
OpenTSLM is a multimodal LLM that natively processes time series alongside text for reasoning.
It surpasses GPT-4o and other baselines on several time series tasks, even at 1B parameters.
GPT-5.6 Sol ranks first in Design Arena's Web Design leaderboard, outperforming its predecessor by 18 places. It actively avoids common AI design anti-patterns, combines strong templates with high personalization, and is faster and cheaper than competitors.
GPT-5.6 Sol ranks #1 overall, 18 places higher than GPT-5.5.
It explicitly avoids AI design anti-patterns like purple gradients and bento-box layouts.
Sarah Friar, CFO of OpenAI, introduces a practical AI scorecard to measure ROI through useful work, cost per successful task, dependability, and return on compute.
Sarah Friar introduces an AI scorecard to measure ROI
Four metrics: useful work, cost per successful task, dependability, return on compute
The article explores how AI-assisted development leads to 'single-mode burnout' by collapsing the cognitive modes of planning, implementation, and integration, leaving developers exhausted despite increased productivity.
AI-assisted development disrupts the natural rhythm of cognitive modes (planning, implementation, integration).
Implementation, which provided flow and cognitive reset, is replaced by supervisory tasks, leading to exhaustion.
This dataset provides a single-file, pre-embedded SQLite corpus of the EU AI Act (Regulation (EU) 2024/1689), chunked by legal structure with BGE-M3 dense embeddings, metadata, risk tier labels, and more. It is designed for local query and RAG research, with verified completeness and transparent derivation rules.
The author used AI coding agents to port the Python e-book reader epy to Rust, creating repy. The project took months instead of hours and garnered little attention, prompting reflections on the devaluation of software in the age of AI and the meaning of creation.
AI coding tools were used to port epy to Rust over several months, resulting in repy.
repy supports multiple formats, search, annotations, TTS, and is fully AI-generated.
This paper proposes ConFlow, a framework that incorporates constraint information directly into the flow matching training objective via differentiable barrier or cost functions and a conditional Gaussian Process, improving constraint satisfaction and trajectory quality in robot motion generation. Experiments on a two-robot navigation task demonstrate lower collision rates and higher trajectory quality compared to standard flow matching baselines.
ConFlow bridges the training-inference gap by integrating differentiable constraint functions into the training objective
Replaces standard Gaussian source distribution with a conditional Gaussian Process to handle smoothness and boundary conditions
This paper explores the feasibility of using brain signals via functional near-infrared spectroscopy (fNIRS) to modulate robot reinforcement learning. It compares agents trained on passive (observational) versus active (demonstrative) interaction tasks, and tests multiple methods for enhancing the RL algorithm with the neural signal, focusing on parameter augmentation rather than replacement. The results show that this framework is effective: the neural signal improves learning when augmenting trajectory priorities and state-action q-values. Additionally, the framework learns successfully from offline data, offering a practical alternative for settings where real-time BCI setups are impractical or only limited data is available.
fNIRS brain signals can enhance robot reinforcement learning
Comparison of passive and active interaction tasks
DiMaS is a distribution-matching steering strategy for flow-matching vision-language-action (VLA) models, enabling fine-grained behavioral control in robotic manipulation. It transports between representation distributions rather than shifting along a fixed direction, proving effective on two state-of-the-art VLAs. The study also examines transferability and explains why linear steering fails in visuomotor settings: behavioral features are linearly decodable but not linearly steerable.
DiMaS achieves fine-grained behavioral control by transporting between representation distributions instead of linear shifts.
It works on two SOTA VLAs, with analysis of how task similarity affects control transfer.