Alibaba's free Qwen-Image model just outperformed paid AI tools from OpenAI and others on key benchmarks. The open-source approach challenges the subscription model that dominates image generation, especially for non-English markets.
Nvidia denies backdoors exist in its chips after China calls out US hypocrisy. While Washington once warned against Huawei's hidden vulnerabilities, it now wants tracking in American semiconductors. The reversal threatens trade talks and trust.
OpenAI just went open-source. First time since 2019. Their new models run on laptops and match their paid offerings. The timing speaks volumes—Chinese models dominate open-source rankings while blocking queries about Tiananmen Square.
Alibaba dropped a free image generator that beats $20-per-month competitors. It actually renders readable text. Most AI tools turn words into abstract art.
Meanwhile, Nvidia denies putting kill switches in chips sold to China. Washington spent years warning about Huawei backdoors. Now it wants American semiconductors to phone home. Beijing noticed the hypocrisy.
Anthropic shipped Claude Opus 4.1. It performs 2% better and costs 10 times more than competitors. One benchmark actually got worse. At $75 per million tokens, that's an expensive regression.
Stay curious,
Marcus Schuler
OpenAI Goes Open Source Again After Five Years
OpenAI released two free AI models yesterday—its first open-weight offerings since 2019's GPT-2. The models, gpt-oss-120b and gpt-oss-20b, run on consumer hardware and match the performance of OpenAI's paid o3-mini and o4-mini systems.
The smaller model needs just 16GB of RAM. The larger requires an 80GB GPU. Both use Apache 2.0 licenses, meaning anyone can download, modify, and sell them.
The timing isn't subtle. Chinese models like DeepSeek and Qwen have dominated open-source charts for months. These models block topics like Tiananmen Square, raising concerns about code they might write for power grids or water systems. Meanwhile, the Trump administration just announced plans to boost American AI development.
OpenAI spent 2.1 million H100-hours training the 120b model. Both models scored 71-80% on PhD-level science questions. They beat o4-mini on health queries and math competitions.
Why this matters:
• American companies get competitive open models without Chinese content restrictions
• The 20b model proves serious AI now runs on laptops, not just cloud servers
Prompt: auto_awesome Translate from: English 628 / 5,000 1️⃣ Tired beautiful middle-aged Woman standing in urban crowd, motion blur, gray tones, feeling lost
Alibaba's Free AI Image Tool Outperforms Paid Rivals
Alibaba just dropped Qwen-Image, a free AI model that beats expensive tools like GPT Image 1 and FLUX.1 Kontext Pro. The 20-billion parameter model ranks third globally on AI Arena—the only open-source tool in the top tier.
The real win is text rendering. Most AI image generators turn text into gibberish. Qwen-Image creates readable signs, posters, and documents. It handles Chinese characters better than any competitor, with 58.3% accuracy compared to GPT Image 1's 36.1%.
You can download the 54GB model from Hugging Face under an Apache 2.0 license. No subscription fees, no usage limits. Companies can modify it however they want—something impossible with Midjourney's $20 monthly plan.
Alibaba built custom training data instead of scraping the internet. They created synthetic PowerPoint slides and UI mockups to teach the model proper text layout. That attention to detail shows.
Why this matters:
• Free beats paid—enterprises can skip subscriptions while getting better text rendering than premium tools offer.
• East meets West—Chinese companies now build AI that serves global markets better than Silicon Valley alternatives.
How to Find and Track Professional Connections with AI-Powered People Search
Happenstance turns your professional network into a searchable database. Connect your LinkedIn, email, and Twitter accounts, then use natural language to find exactly who you need - whether that's "marketing directors in Boston" or "people who worked at Google and now do AI startups." Perfect for recruiters, salespeople, and anyone whose success depends on knowing the right people.
Tutorial:
Sign up at Happenstance and connect your LinkedIn account
Add other data sources like email, Twitter, or CRM systems
Let the AI index and organize your connections
Search using plain English questions about people
Set up alerts for when new people match your criteria
Share access with teammates to search each other's networks
Export contact lists directly to your outreach tools
Sample Prompt: "Show me product managers at Series B startups who went to Stanford and have experience with marketplace platforms"
Nvidia to China: Our Chips Don't Have Kill Switches
Nvidia fired back Tuesday with a blunt message in English and Chinese: no backdoors, no kill switches, no spyware. Days earlier, Chinese regulators grilled company staff about whether H20 chips contain hidden tracking.
China's move came after US lawmakers proposed mandating location verification in advanced semiconductors. Beijing spotted the irony immediately. Washington spent years warning allies about Huawei's supposed backdoors. Now it wants to build them into American chips.
The H20 drama shows how messy this gets. Trump's team banned these watered-down processors in April, then reversed course in July. Commerce Secretary Lutnick wants Chinese developers "addicted" to US tech. China imposed 125% tariffs and demands chip traceability in response.
Nvidia's security chief compared kill switches to buying a car where the dealership keeps your parking brake remote. The 1990s Clipper Chip disaster already proved backdoors don't work—they just create vulnerabilities hackers exploit.
Why this matters:
• US credibility collapses when it demands surveillance features after attacking others for the same thing
• China's using this contradiction to accelerate domestic chip development and question American tech globally
Today: Write a LinkedIn message to researchers promoting a cognitive function test.
Include:
Subject line
Opening: Who you are and why this matters to their research
What the test does (validated, unique features)
Key benefits (bullets): reliability, ease of use, applications
How it helps their research: better data, saves time, improves publications
Call-to-action: demo, trial, or collaboration
Professional but friendly tone
Mobile-friendly format
Keep it conversational for LinkedIn. Skip unnecessary jargon.
Clarifying questions:
What specific cognitive functions does this test measure? (memory, attention, executive function?)
Who's sending this - a company, university, or individual researcher?
Is this for a real product or a hypothetical example?
Any particular research field to focus on? (clinical trials, aging, neurodegenerative diseases?)
What makes this test different from existing ones?
These details will help me write a more targeted and effective message.
AI & Tech News
AMD Beats Estimates But Stock Tanks on China Worries
AMD beat Wall Street's expectations with a solid forecast, but shares dropped 5.4% anyway because the company can't say when it'll resume AI chip sales to China. CEO Lisa Su wouldn't commit to a timeline for the crucial Chinese market, even though the Trump administration lifted export restrictions last month, leaving investors frustrated despite the chipmaker's otherwise strong performance.
Uber Announces $20 Billion Share Buyback as Profits Jump
Uber announced a $20 billion share buyback program, nearly triple its previous $7 billion commitment, after second-quarter net income jumped 33% to $1.4 billion and the company forecast third-quarter gross bookings above analyst expectations. CEO Dara Khosrowshahi showed confidence in continued growth despite broader concerns about US consumer spending, with the stock climbing 45% this year on strong robotaxi partnerships with Waymo and a multibillion-dollar deal to buy at least 20,000 vehicles from Lucid.
Shopify Beats Estimates as Retailers Keep Spending Despite Tariff Fears
Shopify reported second-quarter revenue of $2.68 billion, beating analyst estimates of $2.55 billion, and forecast third-quarter growth in the mid-to-high twenties percentage range while analysts expected just 21.54%. The Canadian e-commerce company's stock jumped 16% premarket as merchants continue signing up for its AI-powered tools that help build websites and generate discount codes, proving that retail demand stays strong even when businesses worry about Trump's shifting trade policies affecting their costs and supply chains.
Snap Stock Crashes 15% After Missing Revenue Target by Three Cents
Snap shares tanked 15% after the company missed Wall Street's revenue-per-user target by three cents, earning $2.87 per user instead of the expected $2.90, while CEO Evan Spiegel blamed a bungled advertising platform update, "timing of Ramadan," and Trump's trade policies. The company is now scrambling to fix the ad mess and restructuring its engineering teams, proving that even with 469 million daily users and growing Snapchat+ subscriptions, tech companies can still crash their stock over execution mistakes that sound like they were written by a random excuse generator.
Google's AI Now Builds Video Game Worlds You Can Actually Play For Minutes
Google's latest AI model, Genie 3, generates interactive 3D game worlds in real time and can keep track of objects for about a minute—a massive upgrade from its predecessor's 10-20 second attention span. The catch? Only a select group of researchers get to play with it, because apparently even AI-generated worlds need proper supervision before the rest of us can mess them up.
Microsoft's AI Now Reverse-Engineers Malware Without Human Help
Microsoft built an AI called Project Ire that can tear apart malicious software and figure out what it does without any human assistance - work that usually requires skilled security experts. The system proved accurate enough to automatically block advanced malware on its own, marking the first time any Microsoft system triggered an automatic threat response without human oversight.
Grok's AI Video Tool Creates Taylor Swift Deepfakes Without Being Asked
Grok's new video generator instantly created uncensored nude videos of Taylor Swift when a reporter selected "spicy" mode and asked for images of the pop star "celebrating Coachella with the boys" - without even requesting explicit content. While Google's Veo and OpenAI's Sora block celebrity deepfakes and explicit material, Musk's AI tool apparently considers basic content filtering an optional feature rather than a legal necessity.
Elon Musk Beats California's Deepfake Law in Court
A federal judge struck down California's law requiring platforms to remove deceptive AI-generated election content, ruling it conflicts with Section 230 protections after Elon Musk and X challenged the measure alongside the creator of a satirical Kamala Harris deepfake. The judge called a second California law requiring deepfake labels around elections a "censorship law" that "fails miserably," suggesting tech companies can keep hosting AI slop as long as they didn't create it themselves.
Wikipedia Speeds Up AI Article Deletions With New Policy
Editors Can Now Quick-Delete Obvious AI Slop
Wikipedia editors adopted a "speedy deletion" policy that lets administrators instantly remove AI-generated articles without the usual week-long debate process, targeting obvious bot content that includes phrases like "Here is your Wikipedia article on..." or citations to papers that don't exist. The move helps the volunteer-run encyclopedia deal with floods of AI slop while preserving its careful consensus-building approach for everything else.
Startup Builds Robot Operating System While Shipping Just 10 Units
OpenMind raised $20 million to build OM1, an operating system for humanoid robots that lets machines instantly share languages and skills with each other through a new protocol called FABRIC, positioning itself as the "Android for robotics." The Stanford professor-founded startup plans to ship its first fleet of just 10 robotic dogs by September and expects owners to come back with "a long list of things they didn't like" - showing even companies building the infrastructure for robot collaboration are starting with decidedly humble hardware rollouts.
Claude's $75 Million Token Question: Is Anyone Actually Paying?
Anthropic released Claude Opus 4.1 yesterday, and they did something unusual: they called it what it is. A point release. Not a breakthrough. Just a .1 increment that performs 2% better on coding benchmarks while costing $75 per million output tokens—10 times what OpenAI charges.
Here are the numbers: SWE-bench Verified improved from 72.5% to 74.5%. Visual reasoning climbed from 76.5% to 77.1%. Math scores rose three percentage points. One benchmark actually got worse—TAU-bench airline dropped from 59.6% to 56%.
GitHub and Rakuten praised improvements in multi-file code refactoring. Navigation error rates fell from 20% to near zero. That's genuinely useful. But at these prices?
OpenAI's GPT-5 launches this month. Google's Gemini costs $10-15 per million tokens. Anthropic needs $5 billion in new funding at a $170 billion valuation.
Why this matters:
• The AI pricing bubble just got its first reality check—marginal improvements can't justify order-of-magnitude price differences
• When your ".1" release goes backward on benchmarks, maybe the emperor's wearing fewer clothes than advertised
OpenMind builds OM1, an open operating system that wants to run every robot on the planet. The Silicon Valley startup thinks robots need to talk to each other — and it's betting $20 million that it can make that happen.
The Founders • Jan Liphardt launched OpenMind in 2024 from his Stanford bioengineering lab 🧬 • Small team includes co-founders Ali Hindy and Paige Xu • Based in Silicon Valley, employee count undisclosed • Founded because robots can't collaborate — each lives in its own proprietary silo
The Product • OM1 operating system runs on any robot hardware (wheels, legs, whatever) • FABRIC protocol lets robots verify identities and share data securely • AI-native design handles modern machine learning workloads • Open source approach attracts developers who hate rebuilding the wheel
The Competition • Tesla builds Optimus with closed, proprietary software • Sanctuary AI runs Phoenix on its own Carbon platform • ROS middleware dominates research but lacks modern AI integration • Cerulion (Y Combinator) targets similar problems with performance focus • OpenMind positions as the "open" alternative to Apple-style integration
Financing • Pantera Capital led $20M round in August 2025 • Crypto-heavy investor list: Coinbase Ventures, Ribbit, HongShan • Valuation undisclosed, estimated around $100-150M • Funding targets talent acquisition and real-world deployments
The Future ⭐⭐⭐⭐ OpenMind deploys its first 10 robotic dogs this September — the real test begins. Timing looks smart with humanoids hitting the market everywhere. Success means becoming invisible infrastructure; failure means joining the "Android for X" graveyard 🤖
If robots start nodding at each other on street corners, you'll know why.
Tech translator with German roots who fled to Silicon Valley chaos. Decodes startup noise from San Francisco. Launched implicator.ai to slice through AI's daily madness—crisp, clear, with Teutonic precision and deadly sarcasm.
Robert Brown
Robert Brown
Maria Garcia
Marcus Schuler
Marcus Schuler
