Daily Briefing
15 October 2026 · 5 stories
🔥 Top story
1A 600-billion-parameter Chinese AI model goes public
On Thursday, October 15, 2026, Chinese lab StepFun released the open weights of its new model Step-5 Preview on Hugging Face: six hundred and four billion parameters in total, twenty-seven billion active at each question, a one-million-token window, and a compute speed on a par with the best American models at a fraction of the price. The model is multimodal (text, image, video) and uses a highly sparse mixture of experts to save on compute. On the independent open-model test BenchLeader, it scores fifty-eight point four, neck and neck with Mistral's French Le Chonk. For most people, it is as if a public Chinese university put the detailed engineering drawings of a rocket engine online for free, so that workshops anywhere could build their own rockets. Source: https://huggingface.co/TypeSafeAI/Step-5-Preview-BF16
2Your future AI assistants will carry an ID card
On Tuesday, October 6, 2026, Meta and the start-up Sierra, founded by ex-Salesforce boss Bret Taylor, launched a new open standard for identifying personal AI agents, the Personal Agent Protocol, nicknamed Poppy. It builds on well-known technologies like OAuth and lets an agent prove who it is, on whose behalf it acts and what it is allowed to do before booking a flight, signing a contract or paying an order. Visa, Walmart, Stripe, Shopify, United Airlines, Wells Fargo and 1Password are among the founding partners. For most people, it is as if every personal AI assistant were to receive a digital employee badge, listing exactly what it can do in your name and blocking everything that is not authorised. Source: https://www.personalagentprotocol.org
3An AI that decides for you raises $870M
On Tuesday, October 14, 2026, TypeSafe AI, the California start-up founded by former OpenAI engineer Diogo Almeida, closed in less than thirty days its mega-round: eight hundred and seventy million dollars in Series A, led by the famous investment firm Andreessen Horowitz, after a first raise of only forty million dollars at the end of September. TypeSafe AI has built an AI model that does not chat but decides for you: it scores, sorts, ranks and acts, without ever generating text by the kilometre. For most people, it is as if the company had moved from an assistant that answers every one of your questions to an assistant that sorts your mail, throws out the ads, files your taxes and prepares your return, without ever asking your opinion. Source: https://firecrawl.dev/fieround/2026/10/14/typesafe-a16z
4A platform that makes AI agents smarter every day
On Wednesday, October 8, 2026, American vendor Kore.ai announced the general availability of Autoloop, an optimisation engine for its enterprise agent platform. In practice: the company sets the goals, and Autoloop monitors the outcomes, detects when an agent drifts, retrains it and redeploys it automatically, without a human team having to stand watch. This is the first time a major commercial platform ships agents that improve themselves in production, a missing piece for IT departments who were reluctant to put agents on critical processes. For most people, it is as if your bank gave you an automatic financial advisor who, every night, learned on its own to better guess your needs for the next day. Source: https://www.kore.ai/press/autoloop-ga
5DeepSeek now backs a robot start-up
On Monday, October 12, 2026, according to Bloomberg, DeepSeek, the famous Chinese AI lab, led the seed round of Winverse AI, a start-up building AI models for physical robots, at about four hundred million dollars, with the investment arm of Chinese giant Meituan also taking part. The founder of Winverse is a former Tencent researcher. It is the first time a major Chinese AI lab has used its war chest to back other start-ups, a sign that China is now building a full investment ecosystem around AI. For most people, it is as if a tech giant, after conquering a market, started sponsoring local start-ups, speeding up the creation of a true national value chain. Source: https://www.bloomberg.com/2026/10/12/deepseek-winverse
📡 To watch
Two standards compete to ID AI agents
Open rivalry between Visa and Meta. On one side, Visa's Trusted Agent Protocol, launched in June 2026 and joined by Microsoft and Worldpay. On the other, the Personal Agent Protocol from Meta and Sierra, launched in October 2026. Stripe and Shopify sit in both camps, a rare balance of forces in the history of technical standards. For most people, it is as if two different insurance companies both claimed to be the only one able to insure your car : you cannot know in advance which one will win. Source: https://www.thenextweb.com/2026/10/06/visa-vs-poppy
China builds its own multi-chip AI cloud
On Sunday, October 12, 2026, according to TechNode, Infinigence AI, a Chinese start-up founded in 2023, closed a Series C of two billion three hundred million yuan (about three hundred and fifteen million dollars), bringing the total raised to four billion three hundred million yuan (about five hundred and ninety million). Its platform combines no fewer than sixteen types of chips, including Chinese Huawei Ascend, Hygon, Cambricon, T-Head and Biren, plus international chips, to run enterprise models and agents. Token usage was multiplied by forty between December 2025 and July 2026. For most people, it is as if a Chinese cloud provider offered customers the ability to use almost any hardware brand without being blocked by US restrictions. Source: https://technode.com/2026/10/12/infinigence-series-c
Mistral will release its giant European AI on October 27
For the record: French start-up Mistral will publish on October 27, 2026 the open weights of Le Chonk, its European model: one thousand and fifty billion parameters in total, fifty-two billion used at each pass, trained on three thousand eight hundred Grace Blackwell chips in a European data centre. First European AI model of this size to be released as open source, already available as a preview since October 6. A major sectoral event to watch closely. Source: https://huggingface.co/mistralai/le-chonk/
An American DeepSeek will release its weights in late October
For the record: American start-up Reflection will publish in late October 2026 the open weights of its Beam model, five hundred and one billion parameters of which twenty-three billion are active per question, under the Apache 2.0 licence. Beam was trained on twenty-three trillion tokens with a massively reduced compute budget, about one third of what China's GLM 5.2 cost for similar scores. Ollama already integrates the online version. First open Western sub-agent truly to rival Chinese models. Source: https://reflection.ai/beam-open-weights/
📊 Trend
Three big lessons from October 15, 2026. First, China accelerates: StepFun publishes a six hundred billion parameter model as open source, DeepSeek backs start-ups, Infinigence bundles sixteen chip types together. Three proofs that China is pushing back against US export restrictions. Second, agents learn to identify themselves: Meta, Sierra, Visa and Microsoft are sketching the first digital ID cards for personal agents in parallel, without yet agreeing on a single standard. Third, agents learn to learn: Kore.ai shows that a machine can train and improve other machines in a loop, without constant human supervision. The day's big convergence : everywhere, intelligent agents gain in autonomy, memory and ability to act on the real world. And the big question now : in which country and under which jurisdiction are these agents trained, and who is responsible when one of them makes a mistake ? That is exactly the multi-jurisdiction AI compliance audit vertical that Agent Wealthy is starting to shape for its European and Chinese clients.