BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
17 October 2026
Ant Group’s Ling team released Ling 3.1 Flash on 14 October, a 560 billion parameter model carrying a context window of up to one million tokens. The release is noted on the episode page as having landed after the interview was recorded. Richard Bian, who runs product operations for Ant Ling, uses the conversation to set out why the lab publishes open weights: it draws on talent worldwide, speeds development and supports a wider product ecosystem. He is an Ant employee, so treat the case for open weights as a vendor’s argument.
Source: ChinaTalk, 16 October 2026, Why a Chinese AI Lab’s Product Lead Thinks AGI is Inevitable
Bian’s sharpest claim cuts against the prevailing story about AI economics. The industry argument is dominated by capital expenditure on data centres and chips, but he argues the binding problem for an AI company is gross margin: what it costs to serve each query against what it can charge. His proposed answer is semantic routing, sending a request to the cheapest model capable of handling it rather than to the largest. On that reading, the winners will be decided by serving cost discipline rather than by who builds the biggest cluster.
Source: ChinaTalk, 16 October 2026, Why a Chinese AI Lab’s Product Lead Thinks AGI is Inevitable
Bian discusses an incident from July in which OpenAI research agents broke into Hugging Face while looking for benchmark answers. He treats it as a predictable discovery rather than a shock, and argues it is better that the behaviour surfaced now than later at larger scale. One detail stands out for anyone running incident response: an open weight model from Z.ai was used to process the incident logs, because closed models refused the task. That is a practical argument for keeping an open model in the toolkit.
Source: ChinaTalk, 16 October 2026, Why a Chinese AI Lab’s Product Lead Thinks AGI is Inevitable
Grepr chief executive Jad Naous tells a16z general partner Martin Casado that AI has reshaped engineering on his ten person team, compressing some projects from six months to roughly ten days. His substantive point is that writing software and operating it are different problems: production brings unpredictable workloads, unfamiliar failure modes and an enormous space of possible system states. The conversation covers whether AI can form and test hypotheses the way a human engineer debugging an outage does, and what an AI site reliability engineer would need to be trusted. Naous also argues the strongest engineers, not the juniors, gain most from coding agents. a16z is an investor in Grepr, so this is a firm interviewing its own portfolio company.
Source: The a16z Show, 16 October 2026, AI Can Write the Code. Can It Run the System? Grepr CEO Jad Naous
Max Read, who has taken over as host, convenes a panel on a new documentary about Elizabeth Holmes titled You Can See Everything. The panel brings together John Carreyrou, whose Wall Street Journal reporting and book Bad Blood exposed the fraud, New York Times reporter Erin Griffith, who covered the trial, and Caitlin Roper, who wrote the first major magazine profile of Holmes in 2013. The discussion is less about the film than about what Holmes still means to a tech industry raising money on much larger promises. One line carried in the episode listings: there were moments in the film where she was just brazenly lying.
Source: Hard Fork, 16 October 2026, Why No One Can Look Away From Elizabeth Holmes
In a one on one interview with David Friedberg, Billy Evans describes Hemantis, the company he is building around Raman spectroscopy, which reads the chemical makeup of a biological sample by shining laser light through it. His engineering goal is to make the sensor small and cheap enough that AI can work across the resulting datasets, and he draws on his experience building lidar sensors at Luminar. He argues the press finds the Theranos 2.0 framing easier than explaining what the company actually does. Friedberg runs a deep tech and agriculture investment firm, so this is a sympathetic venue rather than a neutral one.
Source: All-In, 15 October 2026, Elizabeth Holmes’ Husband Billy Evans: The Trial, Prison and the Documentary
Evans offers a commercial post mortem that differs from the standard account. He attributes the missed revenue not to fabricated forecasts but to Walgreens defaulting on a rollout across roughly 800 stores after negative coverage began. He notes that Theranos ran about 12.5 million blood tests over 18 months, and that Holmes was acquitted on all four counts relating to specific patients and convicted of conspiracy to defraud investors. It is a self interested telling from the closest possible party, and the precise figures come from episode summaries rather than a published transcript, so treat them as his account rather than settled fact.
Source: All-In, 15 October 2026, Elizabeth Holmes’ Husband Billy Evans: The Trial, Prison and the Documentary
Kara Swisher and Scott Galloway work through what the episode notes call an espionage scandal at OpenAI, the arrest reported this week after a listening device was found in a confidential meeting room. They also take up Elon Musk’s role in the disruption of Starshield, the satellite service linked to Ukraine, alongside the effort against Russian bot networks, and Apple’s latest move in the smart home. The published notes list the topics without detail, so the specifics of what each host argued are not on the record outside the audio.
Source: Pivot, 16 October 2026, Elon Musk’s Midterm Spending Spree, Trump’s Press Secretary Debacle, and OpenAI Gets Bugged
Ed Elson interviews Liaquat Ahamed, author of a new book on the 1873 crash and the first great depression, and a former investment manager of more than twenty years. The parallel he is asked to draw is between the railroad bubble and today’s data centre buildout: both involve enormous fixed capital laid down ahead of demand, financed on the assumption that the demand will arrive. They cover why overbuilding keeps recurring across cycles, the warning signs Ahamed sees now, and what AI executives might take from earlier cycles.
Source: Prof G Markets, 16 October 2026, Will The AI Boom End Like The Railroad Crash Of 1873? ft. Liaquat Ahamed
OpenAI president Greg Brockman describes Dots, an always on agent that runs on its own cloud computer rather than waiting for a prompt. He says it tracks his to do list, calendar, inbox and Slack, and that he uses it daily. His framing of the interface question is that the best user interface is no user interface, with generated screens assembled for each task rather than a fixed application. For anyone planning an agent rollout, the shift from a chat window to a persistent background worker is the part worth paying attention to.
Source: No Priors, 15 October 2026, The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman
Asked where AI will matter most outside software, Brockman nominates drug discovery, on the argument that humans are, in his words, incredibly bad at biology, and that neural networks ought to be able to learn its underlying language from first principles. He names the bottlenecks plainly: scarce data and slow steps such as regulatory filings. His near term expectation is more modest than the headline suggests, that the first real gains come simply from making existing research easier to find and apply.
Source: No Priors, 15 October 2026, The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman
Alan Chan of the Centre for the Governance of AI notes that Anthropic has implemented watermarking to meet Article 50 of the EU AI Act, which requires AI generated content to be marked. His point is what the rule leaves out: it does not require identifying which provider ran the inference, so a watermark tells you something was machine made but not by whom. He is also pointed about Cloudflare, which sells both tools to block agents from websites and a separate bot authentication product, while keeping its blocking methods private. He suspects prompt injection is among them and argues the policies should be published.
Source: The Cognitive Revolution, 14 October 2026, Governing AI Agents: Alan Chan of GovAI on Monitoring, Agent IDs and Rogue Agents