BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
17 September 2026
Speaking at the All-In Summit, Elon Musk proposed that the major AI competitors run each other's models through their own security test harnesses, and said it should happen "as soon as possible, if not immediately." His argument is a liability one: once a rival has formally graded your model and flagged a risk, ignoring that warning becomes legally and reputationally expensive in a way that internal red teaming never is. He caveated that any scheme has to be one China will accept, otherwise it simply handicaps the West, and warned that oversight is easy to escalate and very hard to dial back. No lab has publicly taken him up on it.
Source: All-In, 15 September 2026, Elon Musk and Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Merger
Jensen Huang was mid-session on stage when a call came through from the President, and put him on speaker for the room. Trump said robots would not take over, argued that slowing US AI development would mainly benefit China, and described data centres as "the oil of the next 25 years," worth trillions in inbound investment. Huang's own line back was that Nvidia is "not going to let" a slowdown happen. Note that Huang is the single most financially exposed participant in this debate, and that Trump followed the call with seven Truth Social posts naming Anthropic and Dario Amodei directly.
Source: All-In, 14 September 2026, Jensen Huang: The Doomer Hoax, Superintelligence Is Here, and The Future of AI
Musk told the summit that SpaceX and Tesla either build their own fab or fail to scale, tying the roughly 25 billion dollar Austin-area chip project explicitly to Taiwan supply risk and to edge compute demand from robots, vehicles and Mars missions. The venture is joint between Tesla, SpaceX and xAI, with a first phase in Grimes County, Texas, worth more than 16.8 billion dollars. He also volunteered that the idea "sort of did come to me in a dream" before offering the supply chain rationale. The strategic framing is new; the project is not.
Source: All-In, 15 September 2026, Elon Musk and Gwynne Shotwell on AI Risks and Peer Review, Starship, Terafab, SpaceX/Tesla Merger
The signatories object to how AI companies are entering academic mathematics, arguing that treating famous unsolved problems as benchmarks damages the science. Their case is that solving a problem is only a proxy for conceptual understanding, and that mass-producing true and false statements at speed could exhaust fertile ground rather than create new ideas. Tao is himself a recent convert who concedes AI is extraordinarily useful to working mathematicians, and frames his field as a microcosm of a broader threat to intellectual work. Reaction split, with Eric Weinstein and Nassim Taleb calling the signatories an unwanted priesthood and Steven Strogatz arguing that inevitable change is still worth mourning.
Source: The AI Daily Brief, 15 September 2026, Trump Rails Against AI Slowdown Hoax
The working paper, "Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors," analysed hiring and salary data from 2016 to 2024 covering about 29 percent of US bachelor's degrees and found a clear break starting with ChatGPT's release in November 2022. Graduates entering highly exposed fields saw a five percentage point drop in employment and a 13 percent reduction in earnings, partly from taking lower paid roles outside their field. The authors compare the decline to graduating into a large recession. The caveat worth carrying: causation is much less clear than correlation, given the concurrent tech sector correction and studies that find work from home an equally good predictor.
Source: The AI Daily Brief, 15 September 2026, Trump Rails Against AI Slowdown Hoax
The Chinese open-source model company announced the raise on Sunday, split between new stock and convertible bonds, with 60 percent of proceeds going to next-generation model training including what it calls a "fully self-training" loop. That makes Z.ai the first Chinese lab to publicly commit to building RSI, at the exact moment Western labs are arguing about whether to slow down. In the same week, 33 ByteDance-backed researchers published a theoretical roadmap titled "The Last AI Built by Humans," which concludes that AI can automate parts of the training process but finds no evidence yet of it improving that process itself.
Source: The AI Daily Brief, 15 September 2026, Trump Rails Against AI Slowdown Hoax
Rune Kvist, Anthropic's first product hire, announced AIUC's Series A on-mic, led by Ribbit Capital and First Harmonic, with Cursor, Harvey, Lovable and ElevenLabs as named customers. When the company raised its seed, "risk holds down adoption" was a hypothesis; Kvist now calls it fact, saying risk is literally what limits deployment of current frontier models. The product is AIUC-1, a certification standard for agents backed by real insurance capacity, where agents are audited and stress-tested for jailbreaks, hallucinations and data leakage. His complaint is that most AI companies optimise the happy path and never seriously test the adversarial one.
Source: Latent Space, 16 September 2026, Underwriting Superintelligence: Backing Agents you can Sue, Rune Kvist, AIUC
The same episode works through how Lloyd's can underwrite AI systems, and what an AI policy actually covers. The central hypothetical is the liability question nobody has answered: when a 20 dollar Cursor subscription contributes to a 200 million dollar plane crash, who is responsible. The Air Canada chatbot case is treated as the first real data point on where legal liability lands, and copyright is named as the hardest AI risk to insure because of adverse selection. Kvist's closing argument is that even after AGI the labs can never credibly serve as their own watchdogs.
Source: Latent Space, 16 September 2026, Underwriting Superintelligence: Backing Agents you can Sue, Rune Kvist, AIUC
Ming-Yu Liu, who leads Cosmos research at Nvidia, describes a single model that captions video, generates video and audio, and emits robot actions. A vision-language model reasons autoregressively, and its weights then initialise a bidirectional diffusion generator, with a shared temporal scheme aligning signals running at different rates. Open weights are out in three sizes, Super, Nano and Edge, with Edge targeting Jetson Thor, Jetson Orin and DGX Spark. Liu's practical argument is that a neural simulator does not need accurate absolute success rates, only the ability to rank policy A above policy B the way the real world would, which is enough to triage which checkpoints deserve expensive real-world trials. The episode is a paid partnership with Nvidia, so treat the claims as vendor framed.
Source: Machine Learning Street Talk, 17 September 2026, How Physical AI Learns Across Language, Video and Action, Ming-Yu Liu
Pavan Muddireddy, who leads audio research at Mistral, says customers running voice agents across millions of sessions describe scaffolding rather than a solved problem, with quality dropping sharply outside the top few languages. Cascades of specialised models survive because each component stays separately adaptable, observable and constrainable. Mistral's own Voxtral stack feeds a 3B Ministral trunk with continuous audio embeddings passed straight into the decoder rather than via cross-attention as Whisper does, so emotion, timing and speaker identity are not discarded by an intermediate transcript. Named failure modes are streaming diarisation and compounding hallucination, where one out-of-distribution error loops or skips whole segments.
Source: Machine Learning Street Talk, 16 September 2026, Speech Recognition Is Not a Solved Problem, Pavan Muddireddy
Anton Leicht of the Carnegie Endowment sets out how a middle power gets dependable frontier model access without building a competing lab: host the data centres in exchange for model access, align security practice with the US, and hold chokepoint assets such as ASML lithography and ZEISS optics so that a coercive cutoff is unattractive. Australia is named alongside Norway, Singapore and the UAE on the strength of its infrastructure and security relationships. His view is that the binding constraint is policymakers not understanding the stakes, not construction capacity. Note that the show's published notes are AI-generated from the transcript by the publisher's own admission.
Source: The Cognitive Revolution, 15 September 2026, The Balance of AI Power: Anton Leicht on Politics, Pacing Deals, and Muddling Through Well
Gwynne Shotwell used her summit slot to position SpaceX as an AI infrastructure company, with a dedicated segment on orbital data centres alongside Starlink direct-to-cell and rocket retirement, tying launch, Starlink, terrestrial compute and xAI into one platform thesis. Anton Leicht makes the strategic counterpoint on The Cognitive Revolution: if most marginal compute eventually goes to orbit, launch capacity becomes the bottleneck, terrestrial host countries lose their bargaining power entirely, and anti-satellite capability enters the deterrence conversation. Debris and cascading collisions complicate all of it. Musk and Shotwell are talking their own book across xAI, SpaceX, Tesla and Terafab simultaneously.
Source: All-In, 15 September 2026, Elon Musk and Gwynne Shotwell; and The Cognitive Revolution, 15 September 2026, Anton Leicht
The Vice President argued that a lab claiming to have built something dangerous while calling for global AI governance has the sequence backwards. If the model is genuinely dangerous, the first step is to stop building it, and if the capability is already released, the priority is defensive tooling rather than rulemaking. He named Anthropic directly and elsewhere called the industry's request for regulation "a bit of a Trojan horse." In the same session he said H-1B reform will come through executive authority rather than legislation, on the reasoning that Congress will not pass a bill.
Source: All-In, 15 September 2026, JD Vance on AI, Entitlement Fraud, Iran War, Israel, H-1B Abuse and the Midterms
Newsom signed the package on 10 September, described by his office as the strongest child safety chatbot and social media laws in the nation. AB 1709 bars platforms from serving behaviourally addictive features such as infinite scroll and autoplay to under-16s. SB 1119 regulates chatbots with parental controls, notification when a child disables safety settings, and a crisis protocol for suicide-related queries, while a separate measure bans toys containing AI companion chatbots and another extends CSAM sanctions to AI-generated and digitally altered material. AB 2 sets financial penalties in suits against platforms accused of harming children.
Source: Pivot, 15 September 2026, AI Panic: Dario's Warning, Trump's Dismissal, and OpenAI's IPO Delay
The most systematically hostile read of the current AI safety discourse among this week's shows. Zitron, host of Better Offline, joined Ed Elson to ask why the safety story has attracted so much attention relative to demonstrated harm, and what that attention is doing for the people generating it. It is a useful counterweight to the All-In framing, which treats the same debate as an obstacle to be cleared. The episode closes on the 30-year US mortgage rate crossing 7 percent for the first time in 15 months, which is the cost of capital backdrop the AI capex argument sits inside.
Source: Prof G Markets, 16 September 2026, AI Insiders Keep Saying We're In Danger, Where's The Evidence?