BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
19 September 2026
Hard Fork frames the defining shift of the past fortnight: frontier AI companies are now asking to be regulated, and the Trump administration is refusing. The hosts walk through reporting on lab leaders calling for a slowdown, Trump's personal intervention against one, and the China argument that keeps the race running. The novelty is the inversion. Industry lobbying for rules and government declining to write them is the opposite of the usual pattern.
Source: Hard Fork, 18 September 2026, A.I. Safety Goes Mainstream + a Hard Fork Exit AMA
The same episode is billed as an exit AMA, with Kevin Roose and Casey Newton answering listener questions to close out what the show notes call this chapter of Hard Fork. The Pocket Casts feed currently lists only three episodes, which points to a feed reset or restructure. The notes stop short of saying the show is ending, so treat it as a change of format rather than a cancellation until the hosts say otherwise.
Source: Hard Fork, 18 September 2026, A.I. Safety Goes Mainstream + a Hard Fork Exit AMA
The Federal Reserve raised rates unanimously, its first increase since 2023, with two more projected by the end of 2027 and one more possible this year. The point for tech is the financing mix underneath the data centre boom, which has shifted heavily to debt, with Moody's projecting around 240 billion dollars of hyperscaler bond issuance this year. Rates are now high enough to push the 30-year mortgage back above 7 per cent but not high enough to slow hyperscaler borrowing. Taming inflation from here means pressure on equities, on AI capex, or on both.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents; and Prof G Markets, 17 September 2026, Fed Hikes Rates For First Time In 3 Years
DeepMind has launched the DeepMind Institute to fund interdisciplinary research on how advanced AI reshapes institutions, what humans will value, and how communities of agents should be governed. It funds academic work inside DeepMind and Google and at outside institutions, led by AGI chief scientist Shane Legg with chairman Demis Hassabis. Legg's own framing is notably unhurried. Today's systems still fail basic tasks and lack the consistency and creativity for full AGI, but he expects those gaps to close soon.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents
Reporting from The Information and Bloomberg has Apple developing two server configurations built on its M8 chips, a twin M8 Ultra and a four-chip version, targeted at around 2029. The market is AI developers, businesses and governments rather than hyperscaler racks. The detail that matters is the use of NVIDIA's NVLink Fusion networking, the first substantive Apple and NVIDIA collaboration in decades after the falling out of the Jobs era. New chief executive John Ternus reportedly approved the line a year ago while still running engineering.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents
A model codenamed Union Alpha is live for free testing on OpenRouter and OpenCode, scoring 74 per cent on DeepSWE against 72.7 per cent for GPT-5.6 Sol, at a cost per task closer to the cheaper tiers. If that holds, the interesting claim is not raw capability but the cost and capability frontier moving. Its origin is unknown and one theory is a blended model aggregating several providers. Unusually for a stealth test, OpenRouter says it is not training on user data.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents
Anthropic has collapsed its Cowork and chat surfaces into one product, saying users were spending effort deciding where a task belonged and that work begun in one did not carry into the other. Boris Cherny framed the merge as inevitable, arguing Claude Code proved developers could hand off a feature and Cowork proved knowledge workers could hand off a brief, so the destination was always one assistant carrying context across both. The brief registers one reservation. Power users who ran different settings for work and personal tasks trade a mode switch for a model switch and lose some fine-grained control.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents
The Information reports Instinct is raising at a 10 billion dollar valuation, which would be a remarkable price for a company in a category that barely existed a year ago. The brief sets two reactions against each other. Cisco president Jeetu Patel says no product has changed his life the way Instinct has since ChatGPT and that done right it could be a trillion-dollar company, while OpenCode's Dax says everything about the company smells weird. Both quotes are worth holding onto as the personal agent market prices itself.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents
David Pawlan's assistantbenchmark.com scores personal AI assistants across roughly 17 dimensions, including travel booking, purchasing, email, proactive behaviour, permissions and privacy, memory, phone calls and chained tasks, each rated out of 10 and averaged. It grew from three assistants to 108 by 15 September. Meta's Muse leads on 9.1 with Instinct on 8.4. The caveat is large. Averages cover different numbers of scored dimensions, seven against eleven in those two cases, and the whole project is scored manually.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents
CHOP has built a cardiac modelling system on NVIDIA's open-source MONAI framework that turns CT, MRI and ultrasound imaging into anatomically accurate 3D models of children's hearts. Congenital heart defects affect roughly 1 per cent of children born each year and every case is anatomically distinct, so pre-operative modelling is the bottleneck. Work that took a skilled researcher four hours now takes seconds, which puts it within reach of urgent care rather than scheduled surgery. Because the stack is open source, other paediatric centres can replicate it.
Source: The AI Daily Brief, 17 September 2026, Why Everyone Is Getting Excited About Personal AI Agents
Prof G Markets gives Bostrom a full hour on existential risk, and his central claim is that the warnings now coming from AI researchers are genuine rather than positioning. He sets out best and worst case superintelligence scenarios alongside his read on how fast capabilities are really moving. The timing places him directly inside the slowdown argument now running through Washington and the labs. The show notes do not carry the probability figure he gives, so it is not quoted here.
Source: Prof G Markets, 18 September 2026, He Warned AI Could Destroy Us. Now The Industry Is Listening, ft. Nick Bostrom
Stanford professor and Inception co-founder Stefano Ermon argues autoregressive language models are running into hardware and latency limits, and that diffusion architectures applied to discrete text and code, not just images and video, scale inference better through parallel token generation and higher GPU utilisation. He discusses Inception's Mercury models and voice agent deployments where latency is the product. His wider claim is that the next phase of competition is decided by efficiency rather than raw capability. This rests on the episode description rather than a transcript.
Source: No Priors, 18 September 2026, Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon
ChinaTalk's second instalment on Taiwanese drones maps where the red-free supply chain strategy breaks. Flight controllers, brushless motors, sensors and gimbals, and batteries are all choke points, with Chinese flight control parts costing a quarter to two thirds of US alternatives, roughly 90 per cent of neodymium production Chinese, and US camera sensors carrying a 200 to 300 per cent premium. Taiwan has moved from 24th to third largest supplier of drone batteries to Ukraine and cut Chinese cell reliance from 70 per cent to 50 per cent, while China still controls 65 to 97 per cent of upstream cathode, precursor and graphite material. In August Taiwan prosecuted two drone firm employees for passing Chinese flight control chips off as red-free on an army contract.
Source: ChinaTalk newsletter, 16 September 2026, Can Taiwan Build Drones? Part 2
Pivot covers Universal's reluctance to release the Gibney film, for which it holds international rights and paid a significant sum, and which still has no release date. Reporting in The Hollywood Reporter has some inside the company expecting it to be shelved. The film drew a five-minute ovation at Venice, Bleecker Street still plans a US theatrical release on 9 October, and HBO holds North American television and streaming rights. Musk has called it a hit piece.
Source: Pivot, 18 September 2026, Zuck's AI Safety Pitch, Canada Turns to Europe
Jared Isaacman takes All-In through the agency's direction, covering a Moon base and the new space industries around it, what would make Mars realistic, recruiting, and a nuclear NASA moonshot with new vehicles. He argues for keeping humans in space and for shifting capital allocation towards deep-space research, with competition from China and Russia as the framing. Worth noting the surrounding interests. All-In's hosts are active tech investors, the episode carries AI infrastructure sponsors, and Isaacman came from commercial spaceflight to run the agency that buys from it.
Source: All-In, 18 September 2026, Jared Isaacman: A New Era for NASA and American Space Exploration