BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
24 September 2026
Meta's personal AI agent Muse has reached number one on the United States App Store, ahead of ChatGPT, the first agent of its kind to break out with mainstream consumers. Users report it buying groceries, services and subscriptions on their behalf within hours of activation, and the surge pulled semiconductor stocks up with it. On Prof G Markets the week before, Scott Galloway and Ed Elson made the case that Meta could win the AI race without ever being seen to lead it, on distribution across billions of users rather than frontier model bragging rights. The brief frames the shift as commerce moving to a business to agent model.
Source: The AI Daily Brief, 23 September 2026, Agent Wars! and Prof G Markets, 21 September 2026, How Meta Could Quietly Win The AI Race
Shopify has given Muse backend access and switched on agentic checkout through Shop Pay across its stores, the opposite of the blocking posture other large retailers have taken. The asymmetry is commercial. Amazon earns roughly 76 billion dollars a year from advertising that agents route around, while Shopify has no equivalent ad business to defend. Revenue sharing talks are the obvious next step, with the risk that high per agent fees leave agentic commerce open only to the largest players.
Source: The AI Daily Brief, 23 September 2026, Agent Wars!
Reports cited on the brief say OpenAI is working on a competing personal agent under the codename Aeon, possibly for launch at Dev Day. This is reporting rather than an OpenAI announcement and should be treated as unconfirmed. If it lands, the consumer agent market goes from one breakout product to a two horse race inside a month of Muse shipping.
Source: The AI Daily Brief, 23 September 2026, Agent Wars!
The Information reports that OpenAI and Anthropic had contracts close to final for reciprocal safety testing of each other's models, then abandoned the negotiations. Elon Musk has separately floated a similar bilateral arrangement, arguing that distillation would show up in testing logs and that shipping an unsafe model after such a test would create liability. Cross lab evaluation was one of the few voluntary mechanisms the industry pointed to as an alternative to regulation, so its collapse narrows what the labs can put in front of policymakers.
Source: The AI Daily Brief, 23 September 2026, Agent Wars!
Treasury secretary Scott Bessent has publicly rejected proposals to give AI labs legal protection, saying responsibility sits with management rather than with automated systems, and naming OpenAI management over past incidents. He signalled the administration will not act as a backstop for AI companies. That points to a liability led rather than a regulation led approach in the United States, which changes how frontier labs have to price risk into what they ship.
Source: The AI Daily Brief, 23 September 2026, Agent Wars!
Kara Swisher and Scott Galloway work through Trump's announcement of an AI Force, modelled on the Space Force he created in his first term, and an AI czar to oversee the sector. Trump dismissed industry calls to slow development as politically motivated and predicted AI could eventually account for as much as a quarter of United States gross domestic product, arguing existing criminal and civil law is sufficient to handle AI misconduct. The czar post is open because David Sacks left the AI and crypto role to co-chair the President's science and technology council. It is the clearest signal yet that the White House treats AI as industrial policy rather than a regulatory target.
Source: Pivot, 22 September 2026, Trump's Press Crackdown, Paramount Settles, and Anthropic's $2 Trillion IPO
SpaceXAI has released Grok 4.7 and claims seventh place on the Artificial Analysis Intelligence Index, with gains over version 4.6. Independent testing was less flattering. Critics measured the model as 30 to 80 per cent less token efficient than rivals such as Astra, which means a higher effective cost despite marketing built on affordability. Benchmark rank and sticker price are pulling apart from real inference economics, and that gap is increasingly what buyers actually decide on.
Source: The AI Daily Brief, 23 September 2026, Agent Wars!
John Platt of Google described ERA, short for Empirical Research Assistance, a system that automates scientific problems which can be framed as optimisation tasks, pairing Gemini with Monte Carlo tree search across experimental notebooks. Platt reports a step change in capability between Gemini 2.0 and 2.5, and says ERA has now produced at least ten peer reviewed papers solving open scientific problems. It is one of the more concrete claims so far that an agent plus a frontier model yields publishable original science rather than literature summaries.
Source: Latent Space, 22 September 2026, An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science
Condensation trails from aircraft account for roughly 1 per cent of human induced warming, and contrail avoidance is one of the cheapest climate levers available. Platt says ERA resolved a modelling challenge around counterfactual warming effects that had been open for two years. He also pointed to FireSat, a Google satellite constellation aimed at cutting wildfire detection latency. In both cases the value sits in operational deployment rather than in the model itself.
Source: Latent Space, 22 September 2026, An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science
Naveen Rao, formerly AI chief at Databricks and founder of MosaicML, argues that energy rather than chips or data is the binding constraint on AI, and that today's matrix mathematics is running into physical and thermodynamic limits. He makes the case for a different substrate built on non linear dynamical systems, and frames the real unit economics of AI as the cost and power draw of a token, against a human brain that runs on about 20 watts. Rao is a venture backed founder pitching a thesis that favours his own company, on a show hosted by active venture investors, so read the framing as an interested argument rather than neutral analysis.
Source: All-In, 21 September 2026, Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology
Julian Gewirtz, author and former China director on the National Security Council, told ChinaTalk that for all the speeches coming out of Beijing, the question that matters most remains unanswered, namely how China will regulate open weight models proliferating globally. His read is that Beijing itself has not decided, because the field is moving quickly and the trade offs are large. Western policy debate routinely assumes a settled Chinese open weights strategy, and this is a direct argument that no such strategy exists yet.
Source: ChinaTalk, 22 September 2026, Xi Descends on DC, Julian Gewirtz on the Trump Xi Summit and AI
Gewirtz describes a reversal in United States and China technology diplomacy ahead of Xi's first White House visit since 2015. Selling China advanced accelerators such as the H200 used to be the concession, something Washington withheld. The framing now runs the other way, with Beijing agreeing to buy them treated as a concession to Trump. That inverts a decade of export control logic and weakens the leverage those controls were built to create.
Source: ChinaTalk, 22 September 2026, Xi Descends on DC, Julian Gewirtz on the Trump Xi Summit and AI
Gewirtz draws a structural asymmetry in how AI risk gets discussed. An American lab chief executive can talk at length about risks they have no solution for, while a Chinese minister cannot, because naming a risk publicly implies the Party sees it and will handle it. The absence of open ended worry from Chinese officials is therefore not evidence that they are unconcerned. Jordan Schneider adds that some United States actors lean hard on the China argument against domestic regulation, which tangles two separate debates together.
Source: ChinaTalk, 22 September 2026, Xi Descends on DC, Julian Gewirtz on the Trump Xi Summit and AI
Replit founder and chief executive Amjad Masad argues that the found a company young script is itself a trap, and that starting a company can be a form of premature optimisation. He makes the case for project based learning and curiosity over grade optimisation, tracing his own route from chess to AI research. Read the framing with care. Replit is an a16z portfolio company, his co-guest Gagan Biyani runs the firm's own 42 million dollar academy, and this is the third consecutive episode on the a16z feed promoting that school.
Source: the a16z Podcast, 23 September 2026, Amjad Masad on Rethinking College for the AI Era