A Mysterious Free AI Model Is Impressing Developers, but Nobody Knows Who Built It
A new AI model called Ox Alpha has appeared anonymously on OpenRouter and is attracting attention for its coding and reasoning abilities, despite being completely free and having no publicly identified creator.
- Ox Alpha is listed as a “stealth model” from an anonymous provider. It is designed for coding, complex reasoning, long-running agentic tasks, and workflows that combine text with visual information.
- The model is reportedly free for a week with near-unlimited usage, while its provider claims to have enough capacity for up to 100 trillion tokens per day.
- One leading theory is that Chinese AI lab Z. ai may be behind Ox Alpha. Developers have pointed to similarities with Z. ai’s GLM-5 model and the company’s previous use of the anonymous name “Pony Alpha” when testing models.
- Another analysis suggested that Ox Alpha’s tokenizer could instead be connected to Microsoft’s MAI model family.
The identity remains far from confirmed. Still, it has already attracted interest from developers and prominent tech figures like Stripe CEO Patrick Collison, who tested the model and described it as ‘very impressive.’
LinkedIn’s ‘Seems Like AI Slop’ button is a hit
More than one million people have used LinkedIn’s “Seems like AI slop” button since it launched in July. The company also says users are now seeing 40% fewer views on posts it classifies as AI-generated content. LinkedIn is adding a new message to notify creators when members flag their posts as potentially AI-generated.
Hugging Face is exploring a possible sale that could value the AI startup at $13 billion or more
The company, which hosts open-source AI models and datasets, is working with a bank to gauge interest from potential buyers. That would represent a major jump from its $4.5 billion valuation in 2023.
Dr Dre calls AI in music ‘a new tool for creativity’
Dr. Dre says he has started using AI in his music production and does not view the technology as a threat to artists. Instead, he compared it to earlier innovations such as drum machines and synthesizers, calling AI a new tool for creativity.
Sam Altman says the economy is adapting to AI slower than he expected
Altman says he misjudged how quickly AI would change the economy after the launch of GPT-4 in 2023. He expected companies to move away from established software and adopt new AI-powered tools faster, but said many businesses are still sticking with familiar products and workflows.
Chinese hackers are using AI in cyberattacks, researchers say
State-backed hacking groups from China have more than doubled their attacks since adopting AI for tasks such as malware development and reconnaissance, according to Taiwanese security firm TeamT5. The researchers said DeepSeek is particularly popular among these groups, while other cases have involved ChatGPT and Claude Code.
Michael Polansky Is Training AI on Living Human Skin
Michael Polansky, the co-founder of AI and biology startup Outer Biosciences and Lady Gaga’s partner, has spent years developing a system that keeps donated human skin alive for weeks so AI can study how it responds to different compounds.
- Outer Biosciences uses human skin that would otherwise be discarded after surgery. The company receives the tissue through vetted tissue suppliers, with donor consent and samples de-identified before they arrive.
- The company has developed a system that can keep the tissue viable for up to a month, far longer than the usual few days available for testing. This gives researchers time to study slower processes such as inflammation, pigmentation, collagen changes, and skin repair.
- The living tissue is used as part of an AI feedback loop. The model predicts which untested chemicals could improve a specific skin function, researchers test those compounds on the tissue, and the results are fed back into the model to improve its next predictions.
- AI has dramatically increased the company’s discovery pace. Polansky says the company went from finding only a few promising compounds through an earlier research process to generating a new candidate roughly every six weeks, with six leads currently active.
Outer Biosciences currently focuses on cosmetic ingredients rather than drugs, planning to license or sell promising discoveries to beauty and pharmaceutical companies that can turn them into consumer products. Four of its six current leads are considered likely to reach commercialization, according to Polansky.
OpenAI’s custom chip beats Nvidia’s best hardware
OpenAI is trying to break its addiction to Nvidia, and its first internal benchmarks suggest it might actually pull it off.
- Tests show Jalapeño handles up to 1.9x more AI work per kilowatt than Nvidia’s flagship GB300 while pulling only 500 to 700 watts.
- Response speeds dropped by up to 3.6x, meaning zero lag when millions of people hit the model at once.
- OpenAI designed this exclusively for running models rather than training them, co-developing the silicon with Broadcom for deployment by 2026.
After spending billions on third-party GPUs, leading AI labs are moving hardware design in-house to slash data center energy bills and control their own supply chains.
If you run local models or build AI applications, keep an eye on inference-specific architecture: hardware tailored purely for running models will soon make API costs significantly cheaper across the industry.
Why the M5 Ultra is Apple’s privacy workaround for Europe
Apple is handing enterprise buyers a hardware loophole to run massive AI models offline without sending a single byte to foreign cloud servers.
- The new M5 Ultra Mac Studio packs up to 512GB of unified memory at 1.2TB/s bandwidth, letting desktop machines run models with hundreds of billions of parameters strictly locally.
- Apple launched its first 2nm silicon with the baseline M6 inside a revamped Mac mini, doubling up on AI-focused Neural Engines.
- The timing hits right as global DRAM shortages push desktop costs up, making half-terabyte setups a costly luxury despite high demand.
Strict EU privacy regulations make shipping sensitive medical, legal, or financial prompts to third-party data centers a legal liability for big organizations.
If your team handles sensitive enterprise data, test running open-weight local models on high-memory Apple silicon — it completely sidesteps API compliance reviews and keeps your data strictly under your roof.
Anthropic lets you edit what Claude remembers about you
Anthropic is turning Claude into a persistent assistant, but giving AI a long memory comes with major privacy trade-offs users will have to manage manually.
- Claude and Claude Cowork now save your style guides, code preferences, and project context across web, mobile, and agent sessions without needing constant re-prompting.
- The system automatically blocks financial records and government IDs, but users can explicitly force it to store sensitive personal data like health specs or dietary needs.
- Anthropic lets you edit, delete, or wipe individual memory snippets from a centralized dashboard, addressing growing privacy concerns around data persistence.
Most conversational interfaces treat every chat session as a blank slate, forcing developers and power users to paste repetitive system prompts into every new window.
Check your memory dashboard in settings immediately after updating to curate what context Claude keeps active — cleaning out stale project instructions early prevents your system prompt from getting cluttered over time.
Apple’s new Macs are built for running AI locally
Apple announced new Mac Mini and Mac Studio models built for local AI work. The Mac Mini now offers the M6, Apple’s first 2-nanometer chip from TSMC, alongside an M5 Pro option.
Apple says the M5 Pro Mac Mini processes LLM prompts 8.5 times faster than older Pro models. Linked M5 Ultra Mac Studios can pool memory and run trillion-parameter frontier models together.
The Mac Mini now starts at $899, up from $599 before this summer’s memory-driven hike. Running agents at home keeps getting easier and the machine that does it keeps getting pricier.
Musk wants an Nvidia supercomputer in orbit
Nvidia announced SpaceXAI will base its first Starmind satellite on a space optimized Vera Rubin NVL72 rack, the same 72 GPU system Nvidia sells for data centers on Earth.
Musk set the timeline himself on X, targeting Q4 2027 for launch and significant scale in 2028. Nvidia’s press release skipped dates entirely and filed the plan under forward looking statements.
SemiAnalysis pegs orbital compute at over 4x ground cost today and sees parity around 2040. Sam Altman called the idea ridiculous in February, and the cost math still backs him, for now.
Meta to launch OpenClaw rival Hatch
Meta plans to launch Hatch, a consumer AI agent platform derived from OpenClaw, as early as late August, The Information reported. A new model codenamed Watermelon follows in October.
Internal documents show Hatch navigating DoorDash, Etsy, Reddit, Yelp, and Outlook, with a customizable dashboard for agent-built skills. Meta has weighed a premium tier costing up to $199.99 monthly.
Meta shipped three Muse Spark releases since April while pushing open weights on the side. Meta stock is down 15 percent this year, which explains the sudden interest in subscription revenue.
Trump wants $103k for every foreign tech hire
DHS published a proposed rule charging $103,265 per H-1B petition under the annual 85,000 cap, on top of existing fees. Universities, hospitals and nonprofit research employers escape it.
DHS says the fee would fund immigration work across four federal agencies. The proposal follows a judge striking down Trump’s $100,000 H-1B proclamation in June, and public comments stay open until September 24.
Computer jobs make up nearly two thirds of H-1B approvals, so this lands squarely on tech. Cato notes the last six-figure fee cut filings almost 90%, a strange way to recover costs.
Parents can now watch their teen’s Uber ride live
Uber launched live video streaming for teen rides, using the driver’s selfie camera. Only linked parent or guardian accounts can watch, and both the driver and teen get notified when someone tunes in.
The feature runs in nine markets including Atlanta and Cleveland, expanding nationwide in coming weeks. Uber says nobody retains access to the video once the ride ends.
Anyone else riding along appears on the feed with no way to opt out. Uber calls the video encrypted but would not tell TechCrunch whether that means end-to-end.
Bilibili Makes a Global Push
Chinese video giant Bilibili is expanding beyond its home market, relaunching its international app and courting Western creators as it looks to build a global alternative to YouTube.
- Bilibili has relaunched its international Android app, with an iOS version expected to follow and an English-language website also in development.
- The company has removed its identity verification requirement for overseas creators, eliminating the need to provide documents such as a passport before uploading content.
- Bilibili is actively recruiting international creators and has already encouraged major names, including MrBeast, to publish content on its platform.
- The company is hiring community teams across Los Angeles, London, Mexico City, São Paulo, Istanbul, and Tokyo, while also developing AI-powered moderation for its international platform.
- Bilibili already reaches more than 376 million monthly active users on its core platform and is pitching global creators on monetization, brand partnerships, and access to its young audience.
Bilibili already has the scale and creator ecosystem of a major video platform. Removing signup barriers and building dedicated international infrastructure could turn it from a largely China-focused service into a serious new destination for creators worldwide.
OpenAI Calls for Stronger AI Safeguards
OpenAI is urging lawmakers to strengthen rules governing frontier AI, including tougher monitoring and cybersecurity requirements as increasingly capable models create new safety risks.
- OpenAI wants California to strengthen SB 53, the state’s landmark frontier AI safety law, despite previously opposing the legislation.
- The company is calling for mandatory monitoring of frontier models during training and evaluation to detect potentially serious incidents before deployment.
- It also wants stronger cybersecurity protections across the entire model-development lifecycle, from research environments to deployed systems.
- At the federal level, OpenAI supports requirements including independent safety audits, incident reporting, whistleblower protections, and government testing of the most advanced models.
- The shift follows recent safety concerns around increasingly capable models, including OpenAI’s decision to temporarily slow frontier training while strengthening alignment, containment, and chain-of-thought monitoring.
OpenAI is increasingly arguing that frontier AI labs shouldn’t be left to police themselves. If adopted, these safeguards could shape how the most powerful AI models are developed and regulated across the US.
SpaceX Is Taking Nvidia AI to Orbit
SpaceX is expanding its NVIDIA partnership, with plans to use NVIDIA Vera Rubin systems on Earth and eventually take the same AI computing technology into orbit.
- SpaceXAI will deploy NVIDIA Vera CPUs for agentic AI workloads and expand the infrastructure powering Grok using the Vera Rubin platform.
- The first-generation Starmind AI satellite will use a space-optimized version of NVIDIA’s Vera Rubin NVL72 rack-scale system.
- Elon Musk says the first Starmind satellites are targeted for Q4 2027, with the orbital computing network expected to reach significant scale in 2028.
- SpaceX plans to build its future AI infrastructure exclusively around NVIDIA systems, covering both terrestrial data centers and its planned orbital compute network.
- SpaceX expects its AI computing capacity to grow from more than 2 gigawatts by the end of 2026 toward roughly 10 gigawatts by the end of 2027.
SpaceX is betting that the next frontier for AI infrastructure could be space itself. If orbital computing works at scale, it could open a massive new market for NVIDIA while offering an alternative to the power and land constraints facing AI data centers on Earth.
AI Uncovers Hidden Patterns in Breast Cancer
Researchers at the University of Southampton have developed an AI tool that can map previously difficult-to-see cellular patterns inside breast tumors, potentially revealing clues about how cancers grow and spread.
- The open-source AI platform, called CenSegNet, analyzes tiny cellular structures called centrosomes, which help organize cell division.
- Researchers analyzed more than 330,000 centrosomes across 911 tumor samples from 127 breast cancer patients.
- The AI identified two distinct types of centrosome abnormalities: cells with too many centrosomes and cells with unusually enlarged ones.
- Higher levels of enlarged centrosomes were associated with more aggressive tumor characteristics, including higher tumor grade and lymph-node involvement.
- Researchers say these spatial patterns could eventually help identify high-risk patients, predict disease progression, and guide more personalized treatments.
Instead of treating a tumor as one uniform mass, CenSegNet can reveal how cellular abnormalities differ from one region to another. That could give researchers a more detailed picture of cancer behavior and potentially lead to better biomarkers and targeted treatments, though the technology is not yet ready for routine clinical use.