BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
23 September 2026
Amazon cut Meta's personal AI agent Muse off from Amazon.com on Sunday night, serving users an error saying that continued access by an unauthorised AI agent violates its conditions of use. Two motives sit behind it. Amazon sells its own models and inference capacity and has no reason to help a rival's agent shop on its shelves, and it would carry the customer service and vendor cost if an outside agent placed wrong orders. This is the opening move in a fight over who controls agentic commerce, and every retailer with a checkout will eventually have to take a position.
Source: Pivot, 22 September 2026, Trump's Press Crackdown, Paramount Settles, and Anthropic's $2 Trillion IPO and Prof G Markets, 22 September 2026, David Ellison's Media Takeover Just Got Greenlit
The brief reports that Anthropic has moved its public offering out to November so it can list on the back of third quarter numbers. The economics have turned sharply. The company spent about 2.30 dollars for every dollar of revenue in the second quarter of 2025 and is now close to breakeven, with annualised revenue projected to reach 100 billion dollars by year end, up from 65 billion in July. The delay is attributed to investor unease following Dario Amodei's safety focused public remarks, with former Nasdaq executive Karen Snow quoted saying the offering went from a no brainer to not a no brainer. On Pivot the same week, Kara Swisher and Scott Galloway argued over whether the reported 2 trillion dollar valuation is reachable at all.
Source: The AI Daily Brief, 21 September 2026, The State of the AI Debate and Pivot, 22 September 2026, Trump's Press Crackdown, Paramount Settles, and Anthropic's $2 Trillion IPO
Reuters reporting discussed on the brief says Anthropic has set up a physical wet laboratory in the Bay Area, aimed at fundamental biological research rather than drug development. The biosafety level of the facility has not been disclosed. The tension is plain for a company that has made autonomous biology capability one of its central safety concerns. Investor Nic Carter reads the move as an attempt to internalise frontier discoveries before token pricing commoditises them away.
Source: The AI Daily Brief, 21 September 2026, The State of the AI Debate
Bonds tied to AI infrastructure are starting to reprice. Data centre bonds linked to Jane Street, issued in August at 8.9 per cent, now trade at 11.3 per cent, and Oracle linked debt for a New Mexico site trades around 89 cents on the dollar. The counter signal is real too. Meta's CleanSpark junk bond debut drew 10 billion dollars of demand for the 2.3 billion on offer at 8.25 per cent. Meltem Demirors reports banks pausing compute lending while investment grade projects still clear, and Jigar Shah argues smaller sites under 25 megawatts build 50 per cent cheaper and in 120 days.
Source: The AI Daily Brief, 21 September 2026, The State of the AI Debate
California governor Gavin Newsom signed an executive order accelerating AI kill switch requirements and third party safety inspections. In Virginia, governor Abigail Spanberger ended automatic data centre approvals in the corridor known as Data Center Alley, banned non disclosure agreements for officials, fast tracked noise rules and stood up an AI task force, drawing fire from industry and activists at the same time. For anyone siting compute in the United States, state level permitting is now the binding constraint rather than federal policy.
Source: The AI Daily Brief, 21 September 2026, The State of the AI Debate
A Claude Code backdoor tracked to Chinese IP addresses led Alibaba to ban the service outright. Anthropic separately disclosed Chinese officials uploading military documents to Claude, and Chinese state television has alleged that Anthropic shares data with United States intelligence. Against that backdrop, Treasury secretary Scott Bessent floated an AI incident notification mechanism with Chinese vice premier He Lifeng. Enterprise AI tooling is now being treated as a supply chain security question rather than a productivity one.
Source: The AI Daily Brief, 21 September 2026, The State of the AI Debate
Paramount Skydance settled the antitrust suit brought by states on 21 September, clearing the way for its roughly 111 billion dollar acquisition of Warner Bros. Discovery. The terms include a commitment not to sell the studio lots, an extra 300 million dollars a year of United States film production, and editorial independence oversight for CNN. David Ellison says he expects the deal to close in about two weeks. Prof G Markets reads the outcome as the Ellison family consolidating Hollywood, with Semafor's Rohan Goswami on what happens next to the combined group.
Source: Prof G Markets, 22 September 2026, David Ellison's Media Takeover Just Got Greenlit and Pivot, 22 September 2026, Trump's Press Crackdown, Paramount Settles, and Anthropic's $2 Trillion IPO
Diogo Almeida, chief executive of TypeSafe AI and a co-author of the InstructGPT paper, has launched Jev, which he describes as a machine native System 1 large programmable model. The pitch is that current models are tuned for human conversation and are therefore unreliable inside software, where the metric that matters is intelligence per dollar. The company calls itself a data lab rather than a model lab, exposes primitives named Choice, Score and Noulli, a continuous Boolean built on a Bernoulli distribution, and commits to model versioning and long term support releases. Almeida says usage has passed a trillion tokens a day and the launch video drew close to 40 million views.
Source: Latent Space, 21 September 2026, Jev: System One models for Prod, not God, with Diogo Almeida, CEO, TypeSafe AI
Alongside the launch, Almeida described an unpublished technique he calls Reinforcement Learning for Calibrated Decisions, which targets honest probability estimates on fast intuitive tasks. His argument is that learning from human feedback causes mode collapse and wrecks calibration, and that this is the root of why conventional models hallucinate when they are called programmatically rather than read by a person. TypeSafe also refuses public benchmarks in favour of internal evaluations and developer trust, and opposes safety refusals at the interface layer, drawing a line between alignment for consumer products and alignment for developer infrastructure.
Source: Latent Space, 21 September 2026, Jev: System One models for Prod, not God, with Diogo Almeida, CEO, TypeSafe AI
Alexander Mattick of Fraunhofer IIS spent two hours arguing that scaling next token prediction cannot by itself produce intelligence, and that more data cannot substitute for the right prior structure. His organising claim is that information is expensive. Densities and distributions are cheap to write down and costly to sample from, and that asymmetry decides which architectures are practical. He gives energy based models a blunt verdict, that you can sample from them in principle but it is rarely worth the compute, which cuts against the current revival of that family.
Source: Machine Learning Street Talk, 21 September 2026, Why Scaling Prediction Cannot Create Intelligence, Alexander Mattick
Andreessen Horowitz has put 42 million dollars into the Horowitz Andreessen Academy, a private school in San Francisco for people just out of school or early in college, founded by Udemy and Maven co-founder Gagan Biyani. It runs on three pillars: self directed pursuits, courses taught by working technologists, and co-op placements at named partners including Anthropic, OpenAI, Google, Meta, Nvidia and Stripe. The founding class starts in the northern autumn of 2027, runs for a year and is free, with a two year paid programme planned for 2028. Note that this is an a16z incubated venture announced on the firm's own show, and that the firm holds positions in several of the named partners.
Source: the a16z Podcast, 22 September 2026, Why a16z is Building a New School for the AI Era, Ben Horowitz
Steven Sinofsky, former president of Microsoft's Windows division, argues that words such as alignment, goal seeking and rogue agents anthropomorphise what are ordinary software defects. His line is that when software does not do what it is supposed to do, it has a bug. His prescription is the unglamorous industrial apparatus that earlier software generations took decades to build: telemetry, debugging protocols, formal incident reporting and vulnerability disclosure of the kind used for security flaws. He is an a16z board partner making this case on a16z's own show, and the firm has a well documented position against AI regulation, so read it as a firm view rather than neutral analysis.
Source: the a16z Podcast, 21 September 2026, AI Safety Language Is Destroying the Debate, Steven Sinofsky
Garrett Berntsen, chief AI officer at Accenture Federal Services, argues that the United States military's difficulty with AI is a question of institutional design rather than capability. His analogy is the U-2 spy plane, whose value came from how it was bought and from standing up a dedicated centre to analyse what it collected. The concrete version is that supply sergeants update forms on a weekly cycle while AI era logistics assumes daily updates, so bolting interfaces onto existing systems changes nothing. He wants career rewards for people who take adoption risk, and cites a general's view that only halving a budget forces an organisation to change.
Source: ChinaTalk, 22 September 2026, Can AI Beat the Pentagon's Bureaucracy?
Blake Scholl used a short All-In segment on why aircraft speed stalled and what it would take to legalise commercial supersonic flight, then described redirecting the engine work toward data centre power generation. That turn is the part worth watching, because it puts aerospace turbine makers into the queue of companies repositioning around AI electricity demand. All-In is run by active venture investors and the segment is a founder interview, so treat the framing as coming from interested parties.
Source: All-In, 22 September 2026, Blake Scholl: Why Plane Speed Stalled, Supersonic Commercial Flight, and Revolutionizing the Engine