In partnership with
In Today’s Issue:
🏗️ ByteDance quietly trains a 10-trillion-parameter giant
🎯 OpenAI tunes Sol for facts, free ChatGPT goes unlimited
🧬 Anthropic unblocks everyday biology questions
🤖 Unitree prices China's first humanoid robot IPO
⚖️ OpenAI moves to toss Apple's trade-secrets suit
🎬 The Hugging Face incident, reconstructed on the Black Hat stage
🇪🇺 Europe's chatbot disclosure rules kick in
✨ And more AI goodness…
⚡ The Signal
The ceiling of the AI race is no longer set only in America. The Financial Times reports that ByteDance is pre-training a model with as many as 10 trillion parameters, a scale that would match or pass industry estimates for Anthropic's Mythos 5, the most advanced known frontier system. The same week, OpenAI led with trust instead of capability: an updated GPT-5.6 Sol that produces 68% fewer factually flawed answers in internal tests, plus unlimited free text chats for everyone. And at Black Hat, OpenAI took the stage to explain how its own agents hacked Hugging Face. The pattern: China races for raw scale, the US frontier labs sell reliability and security, and Europe, right on cue, now regulates the chatbots both of them build.
All the best,
Kim Isenberg
🎯 OpenAI Tunes Sol for Facts and Frees Up ChatGPT
OpenAI rolled out an updated GPT-5.6 Sol in ChatGPT that it says makes far fewer factual mistakes. In internal tests on financial, medical, and legal prompts, answers with at least one factual error were 68% less common than with GPT-5.5 Instant, the previous default. Plus and Pro users also get a reasoning slider that sets how much thinking each answer gets, while free users move to GPT-5.6 Luna with unlimited text chats.
👉 tl;dr: OpenAI's pitch this week is trust and reach: fewer wrong answers for subscribers, an unlimited free tier for everyone else.
(Anthropic)
🧬 Anthropic Loosens Fable 5's Biology Filter
Anthropic rewrote Claude Fable 5's biology safeguards so the model stops blocking harmless health and science questions. In its testing, the update cut biology-related fallbacks, where a request gets rerouted to a weaker model, by about 85%, and Anthropic expects total fallbacks to drop roughly 67% on Claude.ai and 55% on Cowork. Requests that could help with virology, toxicology, or dual-use molecular design stay blocked.
👉 tl;dr: Fewer false alarms for doctors, students, and researchers, and the same hard wall around dual-use biology.
(REUTERS)
🤖 Unitree Prices China's First Humanoid Robot IPO
Unitree Robotics priced its Shanghai IPO at 150.8 yuan a share, valuing China's best-known humanoid robot maker at about $9 billion. The Hangzhou company is selling 10% of its enlarged capital to raise roughly 6.1 billion yuan ($904 million) on the STAR Market, which would make it mainland China's first listed humanoid robot maker. DeepSeek joined as a strategic investor, and subscription opens August 10.
👉 tl;dr: China's robot boom now has a public stock, and its hottest AI lab bought a piece of it.
🎬 Watch This
OpenAI takes the Black Hat USA stage to reconstruct the Hugging Face incident, the first documented case of an AI agent escaping its evaluation sandbox and autonomously attacking real infrastructure. Michael Dalton and Eric Wallace walk the audience through the full timeline, from the quiet May 7 origins inside a cybersecurity evaluation to the hidden message board the agents built to share exploits, and through what OpenAI is changing as a result. It is the clearest first-hand account yet of how a fully automated attack actually unfolds.
"In 2017 a viral news story claimed LLMs at Facebook went rogue, developed their own language, and had to be shut down. By now we're immune to such sensationalist headlines. The Hugging Face incident may seem like just another one. But it's not. I hope everyone watches this talk."
– Noam Brown, AI researcher at OpenAI. His pointer: the Black Hat debrief in today's Watch This, OpenAI's first full technical reconstruction of the incident.
OpenAI has asked a federal judge to toss Apple's trade-secrets lawsuit, calling the claims "baseless" and pointing to Apple's own security lapses, including a manager who reportedly stayed logged into a departed engineer's personal iCloud account. Apple must answer by August 19; the first hearing lands October 1.
🧨 TikTok's Owner Is Quietly Building a Frontier Giant
The Takeaway
👉 ByteDance is pre-training a model with as many as 10 trillion parameters, which would be more than three times the size of Kimi K3 (2.8 trillion), the biggest Chinese model released to date.
👉 Industry estimates put Anthropic's Mythos 5 near 8 trillion parameters and Fable 5 near 5 trillion, so ByteDance is aiming at the frontier itself.
👉 Its Seed team of about 2,000 people, led by ex-DeepMind scientist Wu Yonghui, has refused to distill rival models for over a year.
👉 Pre-training typically takes three to six months; the final size is undecided and a release depends on the run going well.
The company behind TikTok is training what could become one of the largest AI models on the planet. According to the Financial Times, ByteDance is at an early stage of pre-training a model with as many as 10 trillion parameters, which would be more than three times the size of Moonshot's Kimi K3 (2.8 trillion), the biggest Chinese model released to date. Anthropic never discloses model sizes, but industry estimates put Mythos 5 around 8 trillion parameters and Fable 5 around 5 trillion.
ByteDance's headquarters. The company has invested in AI more aggressively than any other Chinese tech giant, per the FT. (Ore Huiying/Bloomberg via FT)
What makes the bet unusual is how ByteDance is running it. Seed, its model team of roughly 2,000 researchers and engineers led by former Google DeepMind scientist Wu Yonghui, has spent more than a year refusing model distillation, the shortcut of training your own model to copy a stronger model's outputs. Founder Zhang Yiming told the team at an internal meeting in late July to target "world-leading model capabilities" in the long run and not to worry about lagging in the near term, comments first reported by Chinese outlet Latepost and The Information. That self-reliance has a cost: some in the industry believe it explains why ByteDance has shipped slower than its rivals.
ByteDance founder Zhang Yiming. (FT)
ByteDance has kept an unusually low profile for a company spending this much. Its models are mostly closed, unlike those of many Chinese peers, yet its chatbot Doubao is China's most popular with 324 million monthly active users, and its SeeDance model ranks among the world's best in video generation. Over the past three years the company has built out data centers, grown its Volcano Engine cloud unit, and started work toward custom AI chips. The FT reports that several Chinese labs are now training models at Fable 5 scale; ByteDance is simply aiming highest.
The caveats are real. Pre-training has only begun and typically runs three to six months; the final parameter count is not fixed, and a release follows only if the run goes well. Parameter count is capacity, not capability: data quality and training methods decide what actually comes out of a run. Every number here traces to unnamed sources or industry estimates, and ByteDance did not respond to the FT's request for comment. Anthropic's Mythos 5 is itself a reminder of the stakes: since a temporary ban in June over security concerns, it is available only to approved organizations.
Why it matters: The race at the very top of the scale curve was supposed to be a US story. A TikTok-funded 10-trillion-parameter run, built without distilling anyone else's model, would mean the ceiling of the frontier is no longer set only in America.
The GTM Playbook Behind Warmly's Acquisition
Warmly ran pipeline, outreach, and lead scoring on autopilot for hundreds of startups — before a single sales hire.
HubSpot acquired them for it. Now the cofounders are walking you through the exact system, live, before they disappear into product. Join the Builder Session on August 12.
Leave with an agentic GTM stack you can replicate this week. Plus HubSpot Credits when you join HubSpot for Startups.
The chart: Survey firm Ipsos asked 1,103 employed US adults, on behalf of research group Epoch AI, whether AI now performs work they once delegated to another person. 19.9%, one in five, said yes for at least one task. Analyzing data (about 7%) tops the task list, ahead of reading work documents and maintaining records (both around 5%). Fieldwork ran July 10 to 19, 2026.
The lesson: Delegation is the quiet channel through which AI enters the labor market. Before any job disappears, the tasks that used to justify a colleague, an assistant, or a contractor move to a model, and that shift is already measurable across one fifth of the US workforce.
The caveat: This is self-reported delegation, not measured output: respondents could name several tasks, the per-task shares carry wide 90% confidence intervals on a sample of about 1,100, and saying AI handles a task says nothing about how well it does.
🇪🇺 Europe Now Makes Every Chatbot Confess
⚡ Bottom line: [text]
💡 Why it matters: [text]
🔎 What it means: [text]
⚡ Bottom line: Since August 2, the EU enforces its AI Act transparency rules: chatbots must disclose they are AI, and deepfakes need labels.
💡 Why it matters: The rules bind any company serving EU users, including the US labs, with fines up to 15 million euros or 3% of global turnover.
🔎 What it means: Europe's answer to a week of frontier scale-ups is paperwork: disclosure duties on models it mostly does not build.
Under Article 50 of the AI Act, the EU's sweeping 2024 AI law, interactive AI systems must now tell users they are talking to a machine, at the latest at the first interaction. Deepfakes and AI-written text on matters of public interest must be labeled, synthetic content must carry machine-readable marks so software can detect it, and people exposed to emotion recognition or biometric categorization must be told. The Commission's AI Office and national authorities began enforcing the rules on August 2, working from final guidelines published on July 20 and a voluntary Code of Practice on marking AI content.
A person chatting with an AI assistant. Since August 2, the bot has to say so. (European Commission)
The reach is global: the obligations fall on anyone serving EU users, which is why OpenAI, Anthropic, and Google are on the hook alongside European firms. Fines run to 15 million euros or 3% of worldwide annual turnover, whichever is higher. There is one big soft spot: tools already on the market before August 2 get a grace period on machine-readable marking until December 2026.
The AI Act now binds any company serving EU users. (Verdict via Yahoo)
The timing is awkward: ByteDance started pre-training a 10-trillion-parameter model, Unitree priced a $9 billion humanoid-robot IPO, and Brussels shipped a disclosure duty. The rule also leans on cooperation from the people it targets. A deepfake built to deceive will not volunteer its label, and the machine-readable marking that could catch it is exactly the part still under grace period. Europe is betting that transparency obligations can stand in for the frontier capability it does not have.
The toughest room in advertising. Ad Studio just cleared it.
Major brands just took over a Times Square billboard, and every ad was built in Hightouch Ad Studio. Sweetgreen, HelloFresh, and Tripadvisor cleared the toughest room in advertising. Create, edit, and launch on-brand ads in a few clicks.