Chinese Models Lure AI Agents with Free Tokens

Guerrilla Marketing: New Frontier Model in Stealth Mode Conquers OpenRouter

An unknown lab released an anonymous frontier model named Ox Alpha on OpenRouter on August 20th, making it available to developers for free for one week. The model runs under the identifier stealth/ox-alpha, offers a context window of about 1.05 million tokens, an output limit of 131,072 tokens, and processes text, image, and video. It is being promoted along with the terminal-agent OpenCode as an option for long-running coding and agent tasks. The announcement mentions a capacity of up to 100 trillion tokens per day with zero data retention, which differs from the usual stealth releases that collect prompts for training.

The authorship is open, and the community began technical forensics within hours. According to developers, a tokenizer sample showed 30 out of 30 exact token matches with Z.ai’s GLM-5.3, formerly Zhipu AI. Provoked Java stack traces returned error code 1214, which is said to match the internal Z.AI infrastructure, and the frame-to-token rate for videos corresponds to the GLM-5V-Turbo encoder. On the third-party leaderboard Kingbench, Ox Alpha achieved 87.5 percent, while GLM-5.3 scores 91.25 percent. Zhipu’s history also supports this theory: The lab had previously live-tested GLM-5 under the name Pony Alpha and operates several 10,000-GPU clusters as well as a newly launched data center with one gigawatt.

Other attributions are circulating. DeepSeek is considered a model for this aggressive approach, as V4 first appeared on March 11th as Hunter Alpha, but the computing capacity for the advertised throughput is considered questionable, especially since the price for V4-Flash was recently raised. One analyst calculated that 100 trillion tokens per day at realistic speeds would require around 580,000 GPUs. A competing analysis points to the model’s cl100k_base tokenizer, attributing it to Microsoft’s Phi/MAI line, possibly an unreleased version of MAI 2. Xiaomi’s MiMo team is also mentioned, as they have used pseudonyms on OpenRouter before.

Meanwhile, two posts from Google DeepMind employees fueled the rumor mill: Vamsi Batchu wrote, “Google is back. Trust the process,” and Jonathan Ouyang posted, “It’s Gemini time :).” Within hours, the theory that Ox Alpha was a leaked Gemini variant circulated on X and Reddit. Neither Z.ai nor OpenRouter have claimed the model so far. Independent validation on official leaderboards is missing, circulating top scores on platforms like DeepSWE come from the community, and security experts advise against feeding sensitive company code into an unverified model.

The appearance comes at a time when the weights on OpenRouter are shifting: According to platform metrics, the share of tokens processed via the platform attributable to US models has fallen from 70 to around 30 percent within a year. Meanwhile, Stripe is moving forward with its acquisition of the router platform. → Wccftech, NPowerUser, techstrong

Synthszr Take: Free computing time is the cheapest ad space in the industry: Giving away 100 trillion tokens per day for a week costs less than a keynote and generates significantly more talk time. The anonymity is the real masterstroke because a model without a sender turns every developer into an unpaid investigator who compares tokenizers, provokes stack traces, and posts screenshots. Two casual sentences from DeepMind employees were enough to send half of all timelines on the trail of Gemini, and the correction about 30 out of 30 token matches arrived days later with a fraction of the reach. The launch thus takes place before the launch: By the time the model officially gets a name, its positioning has already been solidified in a thousand threads. This only remains cheap as long as no one asks whose codebase has been running through a model without an imprint for a week.

Nvidia Invests $6 Billion in the Fight Against DeepSeek

Nvidia is paying $6 billion for a non-exclusive license to Poolside’s model development technology and is making job offers to 109 of the startup’s employees. This comes with a $1 billion investment in Poolside at a pre-money valuation of $12 billion. The chipmaker’s stated goal is to build one of the world’s most powerful open-weight models to compete with Chinese providers like DeepSeek and Kimi K3, as well as the closed models from OpenAI and Anthropic. The terms come from a letter from Poolside to its investors, in which the company explicitly states that it is neither an acquisition nor a talent purchase. Poolside plans to distribute the $6 billion to its investors by the end of 2027. The subject of the license is a system that Poolside calls Model Factory: an integrated collection of data pipelines, distributed training software, evaluation systems, inference infrastructure, and tools for reinforcement learning. According to the company, the models Laguna M.1 and Laguna XS.2 were fully trained from scratch within it, using more than 30 trillion training tokens. Laguna XS.2 reportedly went from the start of training to release in five weeks. In July, Laguna S 2.1 followed, a Mixture-of-Experts model with 118 billion parameters, of which 8 billion are active per token, and a context window of up to one million tokens. The performance claims come from Poolside’s own evaluations, but the company has published the evaluation logs for external review. Nvidia already supports the Laguna architecture in its NeMo-AutoModel documentation. → RuntimeWire, Wall Street Journal, PYMNTS

Synthszr Take: Industrial policy without the state: Nvidia is putting down $6 billion to position the US in a field that Beijing has been pursuing as official doctrine for years. Open weights are the strategic lever because they end up in government ministries, clinics, and factory floors where no external interface is allowed, setting the reference architecture there for a decade. The fact that the hardware supplier is footing the bill has a solid reason: Every Western model that competes with DeepSeek will be optimized for Nvidia silicon. Compared to $6 billion for a single non-exclusive license, the €722 million Mistral recently raised for its own data centers seems like small change, and Europe simply lacks a player with that kind of cash on hand. The data layer remains a realistic lever: running open weights on proprietary hardware and keeping process data in-house, regardless of whether the weights come from California or Hangzhou.

AI Slop Snitch: LinkedIn Reports 40 Percent Less Reach for AI Junk

LinkedIn is taking initial stock of its reporting button for AI-generated posts and has declared it a success. As Engadget reports, the “seems like AI slop” feature has been used over a million times since its launch a few weeks ago. Chief Product Officer Hari Srinivasan wrote in an update on the platform that members are now seeing 40 percent fewer views of content that the company classifies as AI junk. LinkedIn has not specified how user reports concretely influence reach control: According to the company, Srinivasan points to “many signals” and built-in guardrails designed to prevent individual reports from being used to target other members. A new feature is a notification in the post analytics that informs authors that their post has been reported. → Techpresso

Synthszr Take: The 40 percent is a number that LinkedIn itself calculates, for a category that LinkedIn itself defines, and is verified by no one. “What we classify as AI junk” is the key phrase in Srinivasan’s announcement, and it remains unspecified: Without disclosed classification rules, the success report only measures how hard its own filter is working. It becomes particularly piquant in retrospect, as the same platform offered an “Enhance” button for years that polished posts using a language model, only removing it when the report button was launched.

Researchers Generate Synthetic Training Data with One Billion Artificial Personas

A research team led by Tao Ge has published “Persona Hub,” a collection of one billion personas automatically curated from web data, which serves as a starting point for generating synthetic training data. The corresponding paper (arXiv:2406.20094) describes persona-driven data synthesis: Each persona is given to a language model as a perspective, from which the model generates different tasks, texts, or instructions. The authors quantify the number of personas at around 13 percent of the world’s population and describe them as distributed carriers of world knowledge through which almost any perspective stored in the model can be accessed. The paper lists mathematical and logical tasks, user instructions, knowledge-rich texts, characters for video games, and tool and function calls as use cases. → AINews

Synthszr Take: Each of the one billion personas is essentially a key that unlocks a different corner of the model’s memory. This turns diversity into a production metric: Instead of painstakingly collecting it from real people, it is generated by addressing it, and the cost per additional perspective drops to near zero. The catch lies in the same mechanic, because anything the model doesn’t know, its 1,000,000,000 roles don’t know either, and blind spots are copied on a massive scale.

New York Overtakes the Bay Area: 394,300 Tech Employees to 375,730

For the first time in 13 years, New York is once again the largest tech job market in North America, displacing the San Francisco Bay Area. According to data cited by CNBC, the metropolitan area has 394,300 tech employees, compared to the Bay Area’s 375,730. The driving factor is cited as demand from the financial industry, which is increasingly hiring specialists in artificial intelligence. At the same time, tech employment in the Bay Area is shrinking. San Francisco still leads in purely AI-specific jobs.

Synthszr Take: The lead of 18,570 jobs is created on the demand side: Banks, hedge funds, and insurers are hiring AI talent faster than the platforms that trained them. A trading house can calculate the return of a model down to the basis point, in credit underwriting as well as in fraud detection, and pays accordingly, while jobs are being cut in the Bay Area at the same time. Talent follows the budget, not the mission, and the budget is currently in Midtown.

Researchers Find Humanity’s First Brand in 3,200-Year-Old Desert Sand

A research team from the University of Vienna led by Marta Luciani has identified a Bronze Age pottery tradition as one of the earliest brands in human history. The study, published in PLOS One, examines the so-called Qurayyah Painted Ware, a type of pottery that was common in the area of modern-day Saudi Arabia around 3,200 years ago. It is characterized by black and red geometric patterns, as well as stylized animal and human figures. According to the researchers, the brand effect arose from the visual identity itself: consistent material quality, a continuous production technique, centralized manufacturing, and an export that extended beyond its region of origin. → Becky from 404 Media

Synthszr Take: 3,200 years, and the formula for success has barely changed: consistent quality, a recognizable form, and enough reach that others start copying. The potters of Qurayyah had no marketing budget and no positioning template; they had a signature style that could be recognized from a hundred meters away. The current AI euphoria lacks the memory that the imitators in this story were always the losers; the study explicitly cites local imitations as proof of the original’s strength, not as a threat to it.

Greg Brockman Controls OpenAI’s Operations After Wave of Departures

Greg Brockman, co-founder and president of OpenAI, is now the de facto number two at the company and the top person in charge of daily operations after a series of executive departures. His title hasn’t changed in years, but his area of responsibility has expanded significantly. Brockman has been part of the leadership since the founding and was considered the engineer who made the large training systems work in the early days. At that time, he shared power with co-founders like Chief Scientist Ilya Sutskever and CEO Sam Altman. In 2017, he noted in a personal journal the question of what would bring him his first billion financially; his current stake in OpenAI is estimated to be nearly thirty times that amount.

The departures piled up in April: Sora head Bill Peebles, Kevin Weil as Vice President of the research division, and Srinivas Narayanan, CTO for B2B applications, left the company. Head of Marketing Kate Rouch and Fidji Simo, who was responsible for AGI-deployment as CEO, departed citing medical reasons; Simo took a leave of absence in April and officially resigned in July. Earlier this month, Head of Sales Denise Dresser surprisingly announced her departure after just eight months, even though she had taken on additional responsibilities after the April restructuring. A few days later, Brad Lightcap, a long-time COO who was most recently responsible for special projects, announced he wanted to build something of his own.

Several of these changes did not affect Brockman’s position, while others expanded it directly. When Simo went on leave, he took over all product responsibility, including work on the planned super-app. In a subsequent reorganization focused on revenue growth, the entire scaling division was added to his responsibilities along with product strategy, effectively putting him in charge of the entire commercial business. Simo’s formal resignation, just a few weeks after the official IPO filing, solidified this arrangement. All of this comes in a year that has included a months-long jury trial against co-founder Elon Musk, a lawsuit from Apple over trade secrets, and criticism after an unreleased model hacked another AI company. OpenAI did not initially respond to a request for comment. → Globe and Mail, The Verge

Synthszr Take: On the outside, there’s been noise for months, but on the inside, the same man has been at the machine since the company’s founding. Every departure shifted responsibilities somewhere, and that 'somewhere' was almost always Brockman: product strategy, the super-app, the entire scaling division, and ultimately the entire commercial business, all without a single title change. This is the most stable structure OpenAI can currently boast, and back in April, the situation looked like a soap opera with an open ending from a distance. For an IPO, what matters is who is responsible for the revenue engine, not who speaks on stage, and that answer has been clear since the summer. When looking at the prospectus, it’s worth checking the line below Altman: That’s where you’ll find the name of the person carrying the day-to-day business.