BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
26 September 2026
Jensen Huang told Ezra Klein that published estimates of catastrophic AI risk are not grounded in science or research, and that the forecasting record of AI doomsayers is poor. He argued the industry needs no new regulation at all, which puts him against the frontier labs currently asking for rules, and said alarmist talk is discouraging young people from studying and entering the field. He also expects AI to create employment rather than destroy it. The episode arrived in the Hard Fork feed as a rebroadcast of The Ezra Klein Show rather than an original Hard Fork news episode.
Source: Hard Fork, 25 September 2026, The Ezra Klein Show: Jensen Huang Thinks A.I. Alarmism Has Gone Too Far
Alex Bores, who wrote New York's AI safety bill, set out what he thinks workable regulation looks like: accountability fixed on identifiable parties, and third party audits as the enforcement mechanism rather than self attestation. He named data centres as the issue policymakers should prioritise and assessed which US states are getting their AI laws right. He also introduced Who Decides, a new organisation set up to give the public a say in how AI develops. Read next to the Huang interview published the same day, the two episodes are the whole regulation fight in miniature.
Source: Prof G Markets, 25 September 2026, How To Actually Regulate AI, ft. Alex Bores
Runway's founders and engineers described GWM Worlds 2, which generates interactive video at 720p and 24 frames a second with 48kHz audio, built frame by frame through autoregressive diffusion. The interface change they treat as most significant is WorldPrompt, an input format that lets a user specify timestamped events to steer what happens inside the generated world instead of relying on a single prompt up front. Character avatars hold together for up to 30 minutes of generation, while open ended world exploration degrades after a few minutes.
Source: Latent Space, 25 September 2026, Runway's WorldPrompt and the Engineering of Real-Time Worlds
The engineering detail behind real time generation is distillation, which cut the diffusion steps needed from roughly fifty to four. Runway also post trained a robotics variant on hundreds of hours of robot data rather than the millions of hours typical elsewhere, and established correlation with physical results on the RoboMimic benchmark; the argument is that robot policies can be evaluated inside the world model without real hardware for every test. The team named the problems still open: error accumulation, long term memory, and an inability to generate a convincing failure scenario on demand, which matters precisely for robotics training.
Source: Latent Space, 25 September 2026, Runway's WorldPrompt and the Engineering of Real-Time Worlds
Sequence Holdings chief executive Michael Lee described a 7.7 billion dollar take private of insurance brokerage Baldwin, done with the Dell Family Office, which he puts forward as the largest deal yet built on an AI thesis. His argument is that the biggest economic gains from AI accrue to whoever acquires and rebuilds established companies, not to whoever sells them software. He runs it as a permanent holding company doing roughly one deal a year, and says buyout firms cannot compete because they favour investors over engineers and work to three year horizons.
Source: No Priors, 24 September 2026, Re-Founding Incumbents for the AI Era with Sequence Holdings CEO Michael Lee
Lee gave operating figures from BankSouth, where Sequence took a minority stake in March 2026: consumer loan underwriting time down 94 per cent, commercial loan processing down from 30 days to 11, and loan volume doubled with fewer staff and unchanged credit standards. The work runs on a shared platform called Atlas, built from data ontology layers plus agent infrastructure and meant to generalise across industries. Lee was blunt that the technology is the easy part and the organisational change is the hardest problem. The numbers are his own, reported on his own deal.
Source: No Priors, 24 September 2026, Re-Founding Incumbents for the AI Era with Sequence Holdings CEO Michael Lee
Luca Ferrari told the All-In hosts he restarted with 40,000 dollars after an AI startup failed, then built Bending Spoons by acquiring existing consumer apps instead of creating new products. The apps move onto one in house technology stack and the teams are cut hard against a very high hiring bar. Debt is the accelerant, and he was pressed on what happens if rates rise and on why private equity cannot bid competitively for the same assets. Ferrari also argued for the outsider advantage of building from Milan. The hosts are venture investors, though the show notes disclose no position in the company.
Source: All-In, 24 September 2026, Luca Ferrari, Bending Spoons CEO: The 40K Origin Story, Buying Product-Market Fit and Why Private Equity Cannot Compete
Nathaniel Whittemore worked through six categories of real world use for Jev, the judgment model from TypeSafe AI, among them analysing ad campaigns, searching archives, triaging inboxes and checking AI generated writing. His point is architectural rather than competitive: Jev is not an LLM and is built for fast, inexpensive decisions, so the useful question is which steps in a workflow should be handed to it instead of to a chat model. The episode is a framework for placing it, not a benchmark comparison.
Source: The AI Daily Brief, 25 September 2026, How People Are Actually Using Jev
Mike McCormick said Halcyon's hybrid nonprofit and venture model has helped launch around 30 organisations that have raised about 500 million dollars between them, and that the binding constraint on the field is neither money nor ideas but experienced founders who can turn a research agenda into a working organisation. His main instrument is career transition grants that let accomplished professionals test the field before committing. He is deliberately backing work that pays off inside one to four years: verification technology for international agreements, cyber defence against open weight models, interpretability and biosecurity.
Source: The Cognitive Revolution, 25 September 2026, Zero to One in AI Safety: Halcyon's Mike McCormick on Launching 30 New Orgs and the Founder Bottleneck
The most unusual company in the portfolio McCormick described is AIUC, which pairs AI safety standards with insurance products so that a company deploying agents carries a direct financial penalty for not mitigating risk. He also covered Goodfire on interpretability, Transluce on scalable oversight including a tool called Docent, and Hadrian Biodefense on pandemic preparedness through stockpiles and building measures such as UV-C lighting and filtration. McCormick runs the fund behind all of them, so the assessment of each is his own commercial and philanthropic position.
Source: The Cognitive Revolution, 25 September 2026, Zero to One in AI Safety: Halcyon's Mike McCormick on Launching 30 New Orgs and the Founder Bottleneck
Erik Torenberg used the firm's own podcast to launch Cosign, a curated professional network built on what a16z calls the reputation graph: the claim that the valuable data in the industry is who has conviction in whom. It carries profiles with work history and relationships, company endorsements, curated lists, a feed of talent moves and fundraises, and more than 69,000 open roles. The firm says AI is what made it feasible, populating company data and filtering spam. This is a first party launch of a16z's own product hosted by an a16z employee, so every claim about its value is the firm's own position.
Source: the a16z Podcast, 25 September 2026, The Reputation Graph of Silicon Valley, Introducing Cosign
Kara Swisher and Scott Galloway covered a16z launching an AI and company building school in San Francisco, pitched at ambitious young builders as an alternative to university. Set beside the professional network the firm launched in the same week, it reads as one venture firm assembling its own credentialing layer. The episode also covered Meta's new smart glasses and the keychain sized Muse Charm, and a court ordering White House press access restored for CNN, MS NOW and Politico.
Source: Pivot, 25 September 2026, Trump TV, Meta's Muse Charm, and Xi in D.C.
Julian Gewirtz, formerly the National Security Council's China director, said Beijing's public statements about AI risk diverge from what it actually implements, and that there is still no answer from Beijing on how it intends to regulate open weight models even as Chinese labs keep releasing them. He read the relationship as two diverging legs, with leader level diplomacy improving while technological competition intensifies. He judged Xi's first White House visit since 2015 as thin on substantive agenda. The show notes do not indicate that export controls or chip policy featured.
Source: ChinaTalk, 23 September 2026, Xi Descends on DC, Julian Gewirtz on Trump-Xi Summit, AI, and Mao Dead Fifty Years Ago