BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
11 September 2026
A researcher left Anthropic this week with a public thread arguing that the leading labs are racing towards self improving superintelligence and gambling with everyone's lives. The post reached roughly 100 million views inside 24 hours. What made it land was the response: Anthropic's alignment science lead publicly said the departing researcher was correct, put the odds of AI killing all humans within the decade above 10 per cent, and conceded the company has no plan for aligning superintelligence.
Source: Prof G Markets, 10 September 2026, Anthropic Whistleblower Says AI Could Kill Us All; also The AI Daily Brief, 10 September 2026; AI:AM, 10 September 2026; and Latent Space, 9 September 2026
The Big Technology host pushed back on the doom scenario, arguing that AI systems depend on a human built economy for chips, data centres and power, so removing humans would be self defeating. His concern is narrower and nearer: models capable of attacking financial institutions and hacking companies. He argued the extinction framing actively muffles the safety message, though he conceded he takes the risk more seriously than he used to, given the labs called early scaling correctly.
Source: Prof G Markets, 10 September 2026, Anthropic Whistleblower Says AI Could Kill Us All
Elson's own view is that extinction talk is a massive exaggeration, and he listed a string of walked back predictions from Altman, Stuart Russell and Dario Amodei to make the point. But his argument is that once a serving alignment lead endorses the claim publicly, it becomes a corporate communication and material public information for investors. He called for congressional testimony and subpoenas, and said that if the claims turn out to be hyperbole from companies preparing to list, that is a securities question.
Source: Prof G Markets, 10 September 2026, Anthropic Whistleblower Says AI Could Kill Us All
Ramp's lead economist Ara Kharazian described a level of concentration risk unseen in any other software category the firm tracks, noting that non AI software only reaches comparable concentration across its top 10 to 20 per cent of customers. The compounding problem is correlation. That top 1 per cent is overwhelmingly high growth technology companies and other AI startups, so when spending tightens they pull back together rather than in sequence.
Source: Prof G Markets, 10 September 2026, Anthropic Whistleblower Says AI Could Kill Us All
Per employee AI spend among the top 1 per cent of businesses dropped from roughly 8,000 dollars to 7,205 dollars month on month. Kharazian rejected summer seasonality as a full explanation because token volume and model usage rose over the same period. The Ramp index adds that the effective price per million tokens has fallen 41 per cent to 68 cents, down from a March peak of 1.15 dollars, so buyers are getting more for less rather than doing less.
Source: Prof G Markets, 10 September 2026, Anthropic Whistleblower Says AI Could Kill Us All
Frontier models fell to 45 per cent of token share, down from a 53 per cent peak in August, as firms decide that standard and light tiers deliver adequate return for most work. Kharazian was explicit that this is not Chinese or open source substitution: only 6.4 per cent of AI spending businesses use those models, and that figure is likely an overcount. Anthropic now leads on paid business adoption at 43.8 per cent of US businesses against OpenAI's 39.8 per cent, and both have confidentially filed with the SEC to go public.
Source: Prof G Markets, 10 September 2026, Anthropic Whistleblower Says AI Could Kill Us All
Apple unveiled the iPhone Duo on 9 September, about the size of a passport, with a display 80 per cent larger than the most recent iPhone, alongside the iPhone 18 and 18 Pro, AirPods 5 and a new Watch. Shares closed roughly flat. Patrick McGee, author of Apple in China, called it a marvel of industrial engineering but does not expect volume, comparing it to the iPhone Air: a halo product answering the charge that Apple no longer innovates rather than a new platform.
Source: Prof G Markets, 10 September 2026, Anthropic Whistleblower Says AI Could Kill Us All
The suit, filed on behalf of Max plan subscribers, alleges deceptive marketing and opaque fine print, arguing the 100 dollar and 200 dollar tiers do not deliver the advertised five times and twenty times the 20 dollar Pro plan's usage. The mechanism at issue is the interaction between the five hour rolling window and the weekly cap, which plaintiffs say pushes real usable capacity well below the headline multiples.
Source: The AI Daily Brief, 9 September 2026, AI Model Month Is Off to a Blistering Start
Ethan Tandowsky, most recently chief financial officer of Dutch fintech Adyen, joins to prepare the company for an IPO. ElevenLabs says it is on track for 600 million dollars in annualised revenue by year end, up from 350 million at the end of last year, has reached profitability, and now takes more than half its revenue from large enterprise customers. That last number is the one that matters for a voice company that started as a consumer novelty.
Source: The AI Daily Brief, 9 September 2026, AI Model Month Is Off to a Blistering Start
Google's third updated Flash model in six weeks scores 73.7 per cent on DeepSWE, just under Opus 5, but collapses from 89.4 per cent on Terminal Bench 2.1 to 19.1 per cent on version 4.0. Meta's Muse Spark 1.3 tied Opus 5 for first on the Artificial Analysis coding agent index at 55 cents per task, roughly a quarter of the cost. Alexandr Wang's reply conceded Spark 1.3 is not claimed to match Astra or Fable 5.1, only to be far cheaper for the same output.
Source: The AI Daily Brief, 9 September 2026, AI Model Month Is Off to a Blistering Start
ChatGPT Images 2.5 arrives with Flare for fast iteration and Sunburst for professional workflows that need edit to edit control, plus a new Sketch tool and roughly 50 per cent lower latency than Images 2.0. Arena placed the release first and second across text to image, image edit and multi image edit leaderboards, with the largest gains in multi image editing. ChatGPT is still generating more than three billion images a week.
Source: The AI Daily Brief, 9 September 2026, AI Model Month Is Off to a Blistering Start
The Series D was disclosed the same day as OpenAI's Navier-Stokes announcement and was buried by it. Latent Space flags it alongside Cognition's round as a raise that would have been the headline in any other week. It is a reasonable measure of how much capital is now moving through the sector that a European lab can close a round at that scale and lose the news cycle to a fluid dynamics result.
Source: Latent Space, 9 September 2026, AINews on the OpenAI Navier-Stokes report, newsletter post
Christiano led OpenAI's alignment work from 2017 to 2021 and pioneered reinforcement learning from human feedback, then left to found the Alignment Research Center and is currently senior technical adviser at the US Commerce Department's CAISI. He joins the OpenAI Foundation board and the Safety and Security Committee chaired by Zico Kolter, plus a non voting observer seat on the OpenAI Group PBC board. He is among the more prominent public voices on catastrophic risk.
Source: AI:AM with Nathan Labenz, 10 September 2026, Building Systems You Can Keep
Raffi Krikorian argues that as agents spread across the web, the browser is the natural enforcement point for human control and permissioning, because it already sits between the user and every site an agent touches. He is Mozilla's CTO, so the position is self interested, but the underlying problem is real: agentic systems currently have no consistent boundary layer. He also covered Project Glasswing, which gives open source maintainers free access to an Anthropic model to scan critical infrastructure code, and flagged the unresolved question of who pays for continuous scanning at scale.
Source: AI:AM with Nathan Labenz, 10 September 2026, Building Systems You Can Keep
Qwen-Drive-1.0-4B is derived from an unchanged Qwen3.5 4B vision language model, with a full BF16 checkpoint around 9B, adding separate planner supervised fine tuning, planner reinforcement learning and perception modules. It includes an external bird's eye view perception head for 3D detection, semantic occupancy and map segmentation, plus a planning expert for trajectory generation. It reports competitive results on WOD-E2E and NAVSIM while largely preserving general multimodal ability.
Source: Latent Space, 9 September 2026, AINews on the OpenAI Navier-Stokes report, newsletter post