The Information Machine
The edition

Wednesday September 16, 2026

In this edition

  1. New today
  2. NewThe open-weight versus closed AI model gapMozilla Puts Open-Weight AI Gap at 4.4 Months Behind Closed Frontier
  3. NewGoogle DeepMind Gemini 3.8 Live voice modelsGoogle DeepMind Launches Gemini 3.8 Live Real-Time Voice Models
  4. NewPerplexity's local Windows AI agentPerplexity Launches Local AI Agent for Windows PCs
  5. NewUK Artificial Superintelligence BillUK ASI ban bill tabled; Altman says safety standards lag capabilities
  6. Updates
  7. Day 19AI data center buildout and community resistanceAI Infrastructure Spending Reshapes Energy, Memory, and Labor Markets
  8. Day 6Anthropic's Claude misuse threat intelligence reportAnthropic Report Exposes Chinese AI Distillation at Scale, Draws Diplomatic Threats
  9. Day 2OpenAI's GPT-6 Astra modelGPT-6 Astra Leads Fable 5.1 Across Benchmarks in Same-Week Launch
  10. Day 5Senate AI safety bill and competing legislationTrump Calls AI Existential Risk a Hoax as Sanders Pushes Pause Treaty
  11. Day 1OpenAI ChatGPT for Financial ServicesOpenAI Launches ChatGPT for Financial Services on GPT-6 Astra

New today

01
New

The open-weight versus closed AI model gap

  • Nvidia CEO Jensen Huang on September 16 disputed the cost argument in Mozilla's finding that open-weight models trail closed frontier models by 4.4 months, saying self-hosting makes closed models cheaper for most users and that open models' main value is control.
  • The Institute of Foundation Models released K2 Horizon, six models from 0.9B to 375B parameters with training code, data recipes, and production infrastructure, and US lawmakers probed DoorDash, Airbnb, Cursor, and Apple over their use of Chinese AI models, which have led the open-weight field since roughly 2024.
The gist

The narrowing gap means organizations can access near-frontier AI performance at substantially lower cost using open models. The choice between open and closed models involves contested tradeoffs around cost, control, and workload type that organizations are actively navigating.

Continue this story →4 sources
02
New

Google DeepMind Gemini 3.8 Live voice models

  • Google DeepMind released two speech-to-speech models on September 15: Gemini 3.8 Live, framed as cost-efficient and scalable, and Gemini 3.8 Live Extended Thinking, which reasons and speaks simultaneously while executing tools in the background.
  • Google DeepMind states Extended Thinking ranked first on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6 and leads agentic task completion benchmarks, while Gemini 3.8 Live placed second in the Speech Agent Arena.
  • Both models process real-time visual input, detect language across 97 languages mid-conversation, and watermark all audio output via SynthID.
  • Simon Willison noted the models are comparable to OpenAI's GPT-Live models and built a browser-based testing UI from the documentation.
The gist

The release adds two production-grade real-time voice models to the developer market, with one claiming a top benchmark ranking for speech quality and agentic task completion. Audio watermarking via SynthID is built into all output to keep AI-generated audio detectable.

Continue this story →2 sources
03
New

Perplexity's local Windows AI agent

  • Perplexity released a local AI agent for Windows on September 15, running on-device across local files, Microsoft 365, and the web on NVIDIA GeForce RTX or RTX PRO GPUs with 24GB or more of VRAM.
  • The agent ships pre-configured with a post-trained local model such as Qwen 3.8 27B and connects to Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub.
  • Perplexity stated that sensitive information stays on device, and when cloud reasoning is needed, the agent requests user permission before sending data off-device.
The gist

On-device execution means sensitive data does not leave the PC, and locally completed tasks do not consume cloud credits. The integration with major productivity services positions the agent as a cross-platform desktop tool for everyday work.

Continue this story →2 sources
04
New

UK Artificial Superintelligence Bill

  • Labour MP Alex Sobel tabled the Artificial Superintelligence Bill on September 8, which would criminalize ASI development and deployment in the UK and require the government to pursue a global ban, with cross-party backing from over 70 MPs and peers including 15 former ministers urging Prime Minister Andy Burnham to support it.
  • In a Fortune interview, OpenAI CEO Sam Altman said current safety standards are 'not at a place' to advance capabilities further, acknowledged uncontrolled AI as 'absolutely' possible, and alluded to a pact among lab leaders to slow development while safety and alignment catch up.
The gist

A cross-party legislative push in the UK to criminalize ASI development, combined with a public call from a sitting US AI lab CEO that safety standards cannot support further capability advances, puts concrete pressure on both regulatory and industry timelines for how advanced AI systems are built and governed.

Continue this story →2 sources

Updates

05
Day 19

AI data center buildout and community resistance

  • In a September 15 forecast, BloombergNEF projected US data centers will consume 20% of national power by 2035, modeling a 19 GW supply shortfall even after record grid additions, with at least 15 coal plants postponing retirements since January 2025 to meet AI demand.
  • Emerald AI and NVIDIA demonstrated a flexible-load program with Silicon Valley Power that throttles low-priority AI workloads on real-time grid signals while keeping critical jobs running, adding a demand-side management dimension to the energy constraint picture.
  • Qualcomm and Amazon also announced a multi-generation agreement to co-develop custom AI inference chips and high-speed connectivity hardware for AWS data centers.
The gist

The AI buildout is simultaneously straining global memory supply, electrical labor, energy infrastructure, and capital markets, creating interlocking physical constraints that determine how fast AI capacity can actually grow. The concentration of compute among a small number of actors and the structural shift of memory toward long-term contracts signal that near-term supply cannot flex quickly to meet demand.

Continue this story →9 sources
06
Day 6

Anthropic's Claude misuse threat intelligence report

  • The quantified scale of Chinese lab distillation in Anthropic's September 10 report emerged September 15: Alibaba's Qwen ran over 151 million exchanges through fraudulent accounts, DeepSeek relayed 12.1 million in two weeks, and Moonshot forwarded 300,000 Kimi user requests to Claude.
  • China's foreign ministry threatened resolute countermeasures, and U.S.
  • House committees requested testimony from Anthropic, Google, and Quantum Xchange following Anthropic's assessment that a PRC-backed actor executed an autonomous AI cyberattack.
  • Analyst Ryan Fedasiuk called it a world-historic Chinese counterintelligence failure.
The gist

Anthropic's acknowledgment that newer Claude models may meaningfully assist dangerous biological research marks a capability threshold the company previously said had not been reached. The Chinese distillation cases, at the scale documented, and China's diplomatic response have added a geopolitical dimension that extends the story beyond AI safety into state-level competition and data privacy.

Continue this story →6 sources
07
Day 2

OpenAI's GPT-6 Astra model

  • Competitive programmer Psyho found on September 15 that GPT-6 Astra, launched September 3, solved 100% of logic puzzles tested, including ones impossible via backtracking, against Sol 5.6's 20-30%.
  • In gaming, Astra completed Portal in one shot and reached seven Factorio Space Age science packs while Fable 5.1 reached two.
  • Researchers described Astra resolving overnight a Seymour Conjecture theorem that had been a month-long target, with one writing the result showed 'genuine creativity in methodology'.
The gist

Astra's cybersecurity classification as 'Critical' marks the first time OpenAI has applied that tier to any model, and the model's benchmark leads over Fable 5.1 across agentic, mathematical, and gaming tasks represent a concrete shift in the competitive standing between the two labs' flagship models. The same-week launch created a direct public comparison across a wide range of independent evaluations.

Continue this story →18 sources
08
Day 5

Senate AI safety bill and competing legislation

  • Trump declared AI existential risk a hoax on September 15, comparing it to what he calls the 'Russia hoax' and climate change, resolving prior ambiguity that commentators had been debating.
  • Senator Bernie Sanders separately called on Trump and Chinese President Xi Jinping to negotiate a treaty that would pause AI development broadly and ban superintelligence outright.
The gist

Trump's public dismissal of AI existential risk as a hoax sets a clear executive posture against safety-oriented regulation. Meanwhile, competing pressures from a Democratic PAC, a Senate push for a US-China pause treaty, and intra-Republican conflict show AI policy becoming a live fault line across party lines ahead of midterms.

Continue this story →4 sources
09
Concluded today

OpenAI ChatGPT for Financial Services

  • September 15 reporting covered OpenAI's September 10 launch of ChatGPT for Financial Services, built on GPT-6 Astra with more than 50 connectors to providers including Datasite and FactSet, co-developed with Morgan Stanley and Evercore.
  • Follow-up coverage added that GPT-6 Astra's API is priced at $10 per million input tokens and $50 per million output tokens, with consumer plans from free to $200 per month for Pro, while noting OpenAI has not disclosed pricing or eligibility criteria for the vertical product.
The gist

Financial institutions can access a curated set of licensed data sources through a single platform without managing separate MCP connectors or credentials. U.S. banking regulators are scrutinizing AI adoption in areas including data access, governance, third-party risk, and human oversight, according to VentureBeat.

Continue this story →2 sources
← September 15, 2026

What is moving now · Every edition · Every story

The daily email

Want this in your inbox?

I send one email each morning with the stories that moved. If you would rather just read here, that works too.

Subscribe free