## News / Update
Industry momentum spanned competitions, infrastructure, research, and community events. Together AI is backing the Agents4Science conference with compute awards to catalyze agent-written and agent-reviewed research, while Bill Gates launched a $1M AI challenge focused on Alzheimer’s. Multiple rigorous evaluations of AI’s real-world impact are underway, with FactoryAI and METR running blind hackathons that pit teams with and without AI tools against each other and a separate $15k study on greenfield coding productivity. On the infra side, Avataar reported 11x cost savings and seamless multi-cloud scaling after moving to SkyPilot, and LlamaCloud is enabling StackAI to process over a million documents reliably for enterprise agents. Community activity includes a LangChain x Grammarly NYC meetup and Elicit marking two years of work on reasoning-centric AI. Tooling and reliability advances included an urgent Qwen patch to v0.35.1, GLM-4.5 integration on TensorBlock Forge for streamlined model ops, and new techniques: DSPy and MIPROv2 improving prompt optimization workflows and a multi-objective RL approach advancing red-teaming for safer deployments.
## New Tools
A steady stream of launches targets practical, domain-specific workflows. Night Knight uses LiquidAI to help users curb late-night phone use and improve sleep. Jupyter Agent 2 automates data loading, code execution, and plotting inside notebooks using Qwen3-Coder and Cerebras. The Deep Agents framework now ships a TypeScript package, bringing composable multi-agent systems to JavaScript alongside Python. Just-RAG blends LangGraph’s agentic flows with Qdrant for smarter PDF Q&A, while an AI Bank Statement Analyzer uses LangChain and local models to turn PDFs into searchable financial insights. ChuanhuChat offers a multitasking web interface for real-time document Q&A and autonomous agents across multiple LLMs. Higgsfield Soul introduces highly consistent AI characters with long-term memory via Soul ID for storytelling. Building voice agents is now near-instant on platforms like ai-combo.com, lowering the barrier to conversational AI apps.
## LLMs
Model releases and benchmarks highlighted rapid progress—and sharper reliability. Google debuted Gemma 3 270M for efficient task-specific fine-tuning, ByteDance open-sourced Seed-OSS 36B with strong long-context and agentic capabilities, and NVIDIA announced Nemotron Nano V2 (9B, hybrid SSM) with 6x speed and improved accuracy, alongside an open pretraining corpus. On evaluation, an evidence-grounded model topped Google’s FACTS leaderboard, beating Gemini 2.5 Pro and a GPT-5 variant with fewer hallucinations. ComputerRL set a new state-of-the-art among 9B open models on OSWorld, surpassing OpenAI Operator and Claude Sonnet 4.0 in computer-use tasks. PolyComputing reported solving 99% of Putnam problems, while GPT-5 Pro delivered a verified new proof in convex optimization and another GPT-5 update achieved state-of-the-art spatial reasoning yet remains below human-level. Broader leaderboard dynamics shifted as a full GPT-5 launch window slipped, leaving Gemini 2.5 Pro temporarily leading until the next wave (e.g., DeepSeek-V4). Methodologically, DeepMind’s retrieval technique reduced hallucinations by 40% and improved relevance by 50%, and the ARC-AGI-3 benchmark surfaced fresh insights from thousands of interactive reasoning games. Developers also praised agentic capabilities in the 20B-scale GPT-OSS model.
## Features
Major platforms rolled out meaningful capability upgrades. Google’s Gemini Live gained visual grounding via live camera sharing, on-screen object highlighting, and more natural, expressive speech. Gemini 2.5 Pro became available in VS Code, with new agent prompts for Insiders testing GPT-5 integrations. Google Photos now supports natural-language and voice-driven edits. Google previewed Pixel 10 experiences like Magic Cue for proactive information and fully on-device voice translation, while the redesigned Pixel Watch 4 adds longer battery life, faster charging, and AI-powered health insights. Anthropic expanded Claude Code to Team and Enterprise plans with flexible seat mixing. Perplexity introduced Max Assistant to orchestrate complex, long-horizon research directly in the browser. Runway delivered faster, more controllable creative tools, and Google unveiled a Gemini-powered health coach for personalized fitness and sleep plans. The Gemini app also added rapid video generation from text or photos with audio, with a promo tied to select Pixel 10 devices.
## Tutorials & Guides
Practitioners got high-quality learning resources across evaluation, deployment, and app design. Hamel Husain released free guides on robust LLM evaluation and advanced RAG. A step-by-step tutorial showed how to run GPT-class models efficiently on nearly any hardware using llama.cpp. A hands-on DSPy “context engineering” guide detailed how to build smarter LLM apps with dynamic prompt optimization and retrieval flows.
## Showcases & Demos
Demonstrations emphasized speed, realism, and domain fit. Custom game-specific retrievers built with LlamaIndex and Superlinked outperformed generic search by understanding gamer jargon and context. Everlyn AI showcased ultra-fast, photorealistic video generation, underscoring the gap between research and real-time creative tooling. Google’s “Nano Banana” produced convincing camera-shot text effects with realistic lighting and color, outperforming standard font-swap methods on challenging perspectives.
## Discussions & Ideas
Debate intensified around timelines, methods, and strategy. Updated forecasts lowered the odds of full R&D automation by 2029, tempering near-term AGI expectations. Experts argued for domain-specific evaluations over generic benchmarks to catch real-world failures, and Yann LeCun urged research beyond LLMs for human-level intelligence. Andrew Gordon Wilson challenged the notion that deep learning is inscrutable, while industry voices argued that “pretraining is over” is itself over. Modal emphasized building for rapid iteration rather than pure inference. Macroeconomic commentary suggested AI infrastructure is propping up US capital expenditures. On go-to-market and product, founders were advised to sell to other startups for faster feedback cycles, and a case was made that AI creative tools must prioritize mobile to reach mainstream users.
Tweet: Together AI launches $10K awards at Agents4Science conference
reTweet: Together AI is backing the first conference where AI agents write and review papers, offering big compute prizes. Only 15 days left to submit your work and be part of a groundbreaking event.
Tweet: Gemini Live adds visual guidance and improved speech
reTweet: Gemini Live can now highlight objects on your screen through your camera and features more natural, expressive speech, making the assistant even more helpful and interactive.
Tweet: Gemini 2.5 Pro now available for VS Code users
reTweet: Developers can now access the upgraded Gemini 2.5 Pro in Visual Studio Code, unlocking advanced AI features directly within their coding environment.
Tweet: Google announces Gemma 3 270M, a compact fine-tuning model
reTweet: Google’s new Gemma 3 270M model is designed for efficient, task-specific fine-tuning, packing 270 million parameters into a tool tailored for specialized AI jobs.
Tweet: New language model tops Gemini 2.5 Pro and GPT-5 for accuracy
reTweet: A grounded language model has surpassed both Gemini 2.5 Pro and GPT-5 by a wide margin on Google’s FACTS leaderboard, promising far fewer hallucinations and more reliable outputs.
Tweet: Night Knight app uses AI to cut your screen time
reTweet: Night Knight, built with LiquidAI tech, encourages you to put your phone down when you exceed daily limits, aiming to improve sleep by reducing late-night screen time.
Tweet: Jupyter Agent 2 automates code, data, and plots in real time
reTweet: The new Jupyter Agent 2 loads data, runs code, and creates plots faster than ever directly in Jupyter notebooks, powered by Qwen3-Coder and Cerebras.
Tweet: Prompt optimization gets a boost with MIPROv2 and DSPy
reTweet: Advances in prompt optimization with tools like MIPROv2 and DSPy let AI teams compete with reinforcement learning algorithms, marking a major leap in model efficiency.
Tweet: RL-based red-teaming gets multi-objective boost
reTweet: Researchers unveil a new RL-based red-teaming approach that optimizes for multiple objectives at once—including likelihood and toxicity—offering flexible defenses for diverse use cases.
Tweet: Higgsfield Soul brings state-of-the-art AI characters
reTweet: Higgsfield Soul promises ultra-consistent, realistic AI characters for stories. With Soul ID, it remembers you better than ever, opening new creative doors for storytellers.
Tweet: Pixel 10 series debuts with new AI hardware features
reTweet: Google’s Pixel 10 phones pack major AI and hardware upgrades, offering users a peek behind the scenes of the latest features set to change their mobile experience.
Tweet: PolyComputing AI cracks 99% of Putnam problems
reTweet: PolyComputing’s proprietary AI models now solve nearly all Putnam math contest problems—test them out on the Leibniz platform.
Tweet: Gates launches $1M AI prize to combat Alzheimer’s
reTweet: Bill Gates has unveiled a $1 million competition challenging AI researchers to find breakthrough solutions for Alzheimer’s, spotlighting big bets on tech versus disease.
Tweet: AGI not coming as soon as predicted, says AI expert
reTweet: AI progress is slowing, with updated forecasts lowering the chances of full R&D automation by 2029 from 25% to just 15%—a signal that rapid AGI timelines are less likely.
Tweet: Historic hackathon tests AI tools versus humans
reTweet: FactoryAI and METR_Evals are hosting a unique event: developers compete in blind-judged teams, some using AI coding tools and others without, inspired by research on real-world AI impacts.
Tweet: Deep Agents architecture now available for TypeScript
reTweet: The Deep Agents framework launches its TypeScript package, enabling easier agent-building in both Python and JavaScript for advanced, composable AI applications.
Tweet: LangChain and Grammarly bring AI builders together in NYC
reTweet: Join local developers for the LangChain x Grammarly NYC meetup—discussing architectures, production-ready AI systems, and hands-on building for anyone interested in cutting-edge AI apps.
Tweet: Just-RAG powers smarter PDF conversations with LangGraph
reTweet: Just-RAG combines LangGraph’s agentic workflows and Qdrant’s vector search to create an intelligent PDF chat system, making document handling and Q&A more seamless.
Tweet: AI Bank Statement Analyzer turns PDFs into financial insights
reTweet: Effortlessly analyze your bank statements: this tool uses LangChain's tech and local LLMs to extract, query, and bring clarity to your finances directly from PDF files.
Tweet: ChuanhuChat links multiple LLMs in a dynamic web interface
reTweet: ChuanhuChat delivers real-time document Q&A and autonomous agents through a modern UI, powered by LangChain and designed for seamless multitasking with multiple large language models.
Tweet: Custom AI retrievers outperform generic search in gaming
reTweet: Using LlamaIndex and Superlinked, custom retrievers built for Steam games understand gaming-specific jargon and context, showcasing the power of domain-aware semantic search.
Tweet: ComputerRL open model beats OpenAI and Claude on OSWorld
reTweet: Computer Use Agent, ComputerRL, leverages reinforcement learning to reach 48.1% success on the OSWorld Benchmark—surpassing OpenAI Operator, Claude Sonnet 4.0, and setting a new state-of-the-art for 9B-parameter open models.
Tweet: Domain-specific AI evals beat generic benchmarks, says expert
reTweet: Generic metrics can’t capture real-world failures—AI consultant Hamel Husain explains on Chain of Thought podcast why custom, app-focused evaluations are vital for agent reliability.
Tweet: Google teases smarter Pixel 10 with Magic Cue, on-device translation
reTweet: The new Pixel 10 lineup debuts Magic Cue for proactive info and full on-device voice translation, making digital interactions easier, faster, and more private—no searching or lag required. The latest Tensor G5 chip powers advanced AI like never before.
Tweet: Pixel Watch 4 revamped: longer battery, new AI fitness features
reTweet: The redesigned Pixel Watch 4 offers a sleek look, extended battery life, faster charging, and AI-powered health tools to help you stay healthy and connected all day.
Tweet: ByteDance launches Seed-OSS 36B LLM as open source
reTweet: ByteDance’s Seed-OSS 36B lands on Hugging Face, offering powerful long-context, reasoning, and agentic abilities to the open-source AI community.
Tweet: GPT-5 Pro proves new math results—checked and correct
reTweet: An AI model just solved an open problem in convex optimization, delivering a stronger proof than the original academic paper—and the math checks out, according to experts.
Tweet: SkyPilot slashes Avataar’s AI costs by 11x and boosts scaling
reTweet: Switching from SLURM to SkyPilot led to huge savings, seamless multi-cloud scaling, and a familiar workflow—redefining efficient AI infrastructure for enterprise teams.
Tweet: Google Photos now edits images with just your words or voice
reTweet: Simply tell Google Photos what you want—unblur, fix lighting, or more—and AI-powered tools handle photo editing instantly, no technical know-how required.
Tweet: Claude Code now scales flexibly for teams and enterprises
reTweet: Anthropic's Claude Code expands to Team and Enterprise plans, letting organizations mix and match seat types and scale usage as needed for coding with AI.
Tweet: LlamaCloud powers StackAI to process over 1M docs with high accuracy
reTweet: StackAI, running on LlamaCloud, processes over a million documents with precision, building trustworthy enterprise agents for finance, insurance, and more.
Tweet: GPT-OSS (20B) impresses with agentic capabilities beyond benchmarks
reTweet: Developers say gpt-oss (20B) excels as an agentic, tool-calling model, showing "frontier magic" that benchmarks don't capture.
Tweet: Google unveils Gemini-powered fitness coach for personalized health
reTweet: Google partners with health experts for a smart coach that offers tailored fitness and sleep plans, science-backed advice, and real-time progress insights—all AI-driven.
Tweet: Major hackathon to test real-world impact of AI coding tools
reTweet: FactoryAI and METR host a blind hackathon: half build with AI coding tools, half without, inspired by breakthrough research measuring AI’s effect on software development.
Tweet: Help research AI’s impact on greenfield coding—win $15k in SF
reTweet: Take part in a unique study on AI and software engineering, join anonymized data collection, have fun, and compete for $15,000 this September in San Francisco.
Tweet: VS Code Insiders can trial new agent prompt for GPT-5 today
reTweet: The Visual Studio Code team is testing updated AI agent prompts for GPT-5—Insiders can try them now and give feedback to improve developer experiences.
Tweet: Free resources to master LLM evaluation and advanced RAG
reTweet: Hamel Husain offers free guides on evaluating large language models and exploring advanced retrieval-augmented generation methods—grab the links to level up your AI skills.
Tweet: Perplexity unveils Max Assistant for advanced research tasks
reTweet: Perplexity Max subscribers now have access to Max Assistant, which uses advanced reasoning models for complex workflows and long-horizon research—bringing Claude Code-like capabilities to the browser.
Tweet: Gemini’s new features turn text and images into instant videos
reTweet: GeminiApp now lets users generate videos with sound from text or photos in minutes, and offers a free year of Google AI Pro with select Pixel 10 phones.
Tweet: Share your camera in real time with Gemini Live
reTweet: Gemini Live now enables live camera sharing in conversations, providing instant visual feedback and real-time advice—like finding the best glasses for your face shape.
Tweet: California’s batteries now supply over a quarter of peak power
reTweet: Batteries delivered 27% of California's peak electricity demand yesterday, highlighting the growing impact of solar and storage for energy security.
Tweet: ComputerRL aims to revolutionize autonomous digital agents
reTweet: Zai.org’s ComputerRL framework empowers agents to skillfully navigate complex digital workspaces by combining direct GUI interaction with API calls.
Tweet: Building voice agents is now easier than ever
reTweet: Creating conversational voice agents—once a futuristic task—now takes just one minute with platforms like ai-combo.com.
Tweet: Elicit celebrates two years of reimagining reasoning in AI
reTweet: Born from a non-profit, Elicit set out two years ago to fix how complex decisions are made, focusing on radically improved AI reasoning tools.
Tweet: Qwen Image Edit tops Hugging Face trending models chart
reTweet: Qwen’s image edit tool is back at #1, reflecting its popularity on the Hugging Face platform for AI model sharing.
Tweet: Yann LeCun: “Don’t study LLMs for human-level AI”
reTweet: AI pioneer Yann LeCun advises researchers to explore beyond large language models for advances in human-level intelligence.
Tweet: GPT-5 launch delayed, Gemini 2.5 Pro takes top spot (for now)
reTweet: GPT-5's high-powered release window has closed, with a scaled-back version out instead. For now, Gemini 2.5 Pro leads the pack again—at least until DeepSeek-V4 arrives.
Tweet: DeepMind’s new RAG slashes hallucinations by 40%
reTweet: DeepMind unveiled a Retrieval-Augmented Generation method that cuts AI hallucinations by 40% and boosts answer relevancy by 50%, marking a leap for accurate AI responses.
Tweet: Bill Gates launches $1M AI competition to fight Alzheimer’s
reTweet: Bill Gates announces a $1 million AI prize targeting solutions for Alzheimer’s. This week also saw updates from Deepseek, ElevenLabs, Eight Sleep, Perplexity, and Runway.
Tweet: Modal builds for AI iteration, not inference
reTweet: Modal’s mission isn’t running inference, but speeding iteration for data and AI teams—going back to founder Erik’s original vision before ChatGPT made waves.
Tweet: Runway drops new features for creators
reTweet: Runway just rolled out updates that promise more control, speed, and flexibility in creative AI workflows—opening up fresh possibilities for digital creators.
Tweet: Gemini Flash-lite cuts AI optimization costs by 10x
reTweet: Running DSPy optimization on Gemini Flash-lite, then deploying with Gemini Pro, has slashed costs by tenfold for some users—showing a practical way to save money on model tuning.
Tweet: The era of “Pretraining is Over” is officially over
reTweet: AI circles are buzzing as the once-popular “pretraining is over” narrative fades—pretraining techniques are back in vogue, reshaping model development strategies.
Tweet: Without AI, US Capex would be falling flat
reTweet: America’s capital spending may have slumped—if not for the booming investment in AI infrastructure, including sprawling server farms transforming the economic landscape.
Tweet: MAN VS MACHINE Hackathon pits coders against AI
reTweet: A new hackathon challenges half its participants to code without AI, while others use all available tools—testing whether AI gives real-world engineering edge or just hype.
Tweet: Startup founders: sell to other startups for quicker wins
reTweet: Selling your product to other startups can speed up decision-making and boost product feedback—even if the budgets are smaller, the benefits can shape your company’s trajectory.
Tweet: NVIDIA Unveils Nemotron Nano V2: 6X Faster, More Accurate Model
reTweet: NVIDIA’s new Nemotron Nano V2, a 9B parameter hybrid SSM, boasts 6X speed and enhanced accuracy versus similar models. Most of the data and pretraining corpus used are also now open for the community.
---
Tweet: Alibaba Qwen Team Issues Urgent Patch—Update to 0.35.1
reTweet: Two critical fixes have been patched in Qwen—users are urged to upgrade to version 0.35.1 for optimal performance and reliability.
---
Tweet: GPT-5 Sets New Bar in Spatial Intelligence—Human Level Still Elusive
reTweet: Despite achieving state-of-the-art spatial reasoning, GPT-5 falls short of true human-level abilities, sparking debate over what’s next for AI’s understanding of space.
---
Tweet: Everlyn AI Delivers Mind-Blowing Speed in Hyperreal Video Generation
reTweet: Everlyn_AI sets a new standard with its astonishingly fast, photorealistic video generation—a breakthrough blending research and engineering.
---
Tweet: DeepMind’s RAG Technique Slashes Hallucinations by 40%
reTweet: DeepMind introduces a simple retrieval-augmented generation (RAG) approach that dramatically cuts hallucinations and improves answer relevancy. Here’s how it can upgrade your RAG systems.
---
Tweet: GLM-4.5 Launches on TensorBlock Forge for Seamless AI Integration
reTweet: GLM-4.5 is live on the open-source TensorBlock Forge, streamlining AI model management. Get your API key and supercharge your model pipeline.
---
Tweet: Google’s Nano Banana Nails Hyperrealistic Image Distortion
reTweet: Google's Nano Banana model impresses with lifelike camera-shot text effects, lighting, and color accuracy, outperforming typical font swaps—even handling tricky side text.
---
Tweet: Top Guide: Run GPT-Based LLMs On Any Device With llama.cpp
reTweet: Unlock GPT-powered models on virtually any device—NVIDIA, Apple, AMD and more—using llama.cpp for lightweight, efficient, and flexible AI inference.
---
Tweet: How to Design Smarter LLM Apps? Learn DSPy Context Engineering
reTweet: Explore @neural_avb’s hands-on DSPy tutorial for building better LLM applications, covering everything from fundamentals to real-world context engineering.
---
Tweet: DSPy Makes Retrieval Data Generation More Dynamic & Effective
reTweet: A new paper shows DSPy can replace static prompt templates with chain-of-thought optimized prompts, reducing the need for aggressive filtering while boosting retrieval results.
---
Tweet: Chat: Deep Learning Is Less Mysterious Than You Think
reTweet: Professor Andrew Gordon Wilson breaks down why deep learning isn’t as inscrutable as it seems, tackling some of AI’s most puzzling paradoxes. Catch this fascinating expert discussion.
---
Tweet: AI Creative Tools Must Go Mobile or Miss the Mainstream
reTweet: With users demanding on-the-go creativity, launching AI tools on mobile is crucial for fast growth—laptops are now too slow for wide adoption.
---
Tweet: ARC-AGI-3 Benchmark: Big Learnings from 3,900+ Agent Games
reTweet: The ARC-AGI-3 interactive reasoning benchmark has yielded new insights after 30 days and thousands of agent plays. Here’s what the team learned about AI’s reasoning abilities.