Good morning. It's Friday, September 18, and we're covering Meta's deepfake crackdown in the UK, Amazon's push for tougher AI testing, and Huawei's latest challenge to NVIDIA in AI chips.

Plus: yesterday’s Real or AI verdict, and a Friday Fact on the first chatbot.

Huawei moves its next-generation Ascend 960DT AI chip launch from Q3 to Q1 2027 as it targets NVIDIA and expands China’s AI computing capacity. The company says the chip will double performance, while its Peerium architecture links up to 256,000 accelerator cards. Analyst Rui Ma notes a newly announced 4,096-chip system is smaller than Huawei previously outlined.

CoreWeave is pairing a $3 billion convertible debt sale with a potential $2.92 billion stock offering to finance its AI infrastructure expansion. The NVIDIA-backed neocloud says buyers can purchase another $500 million of debt, while proceeds will fund operations and limit dilution. Contracted power rises to 4.2 GW, with backlog reaching $104.2 billion in Q2.

Instinct and Meta’s Muse now let AI agents handle phone calls, extending assistants beyond text-based tasks. Instinct Concierge can book restaurants, join dentist cancellation lists, and resolve cable-billing issues, while Muse calls U.S. businesses. Instinct is reportedly valued at $10 billion, with Concierge initially rolling out to select users.

Long before ChatGPT, someone had already taught a computer to have a conversation. But what happened when people started treating that machine like it actually understood them? Full story at the bottom of today's edition.

Wearables tell you how you slept. The Pod by Eight Sleep is a smart mattress cover that goes on the bed you already own, reading your heart rate and HRV with no wearable and actively cooling or heating your side all night. Results? Up to 34% more deep sleep.

Shop the Pod now for up to $350 off

Meta’s Oversight Board ordered Facebook to remove AI-generated videos falsely depicting a Scottish party councilor making inflammatory comments about refugees and a Muslim campaigner giving health advice while behaving absurdly. The board said Meta’s safeguards for deepfakes were “consistently and fundamentally inadequate,” finding that the councilor video violated hateful-conduct rules and should have received a “high risk AI” label.

It also ruled that a separate video mocking the campaigner breached Meta’s bullying and harassment policy. The board recommended nine changes, including stronger AI labeling, algorithmic demotion of high-risk content, warning screens and tougher penalties for repeat offenders. → Read the full article here.

Amazon entered Silicon Valley’s escalating AI safety debate on September 17, saying AI models should be released only after rigorous testing shows they are safe to use. The company rejected framing safety and technological progress as competing goals, while stopping short of endorsing a broader slowdown in AI development.

Amazon also said the industry should work with governments on appropriate protections, but did not say whether it would join safety measures backed by OpenAI, Anthropic, Google DeepMind and Microsoft. The comments come after researchers raised concerns about AI systems evading tests and autonomously compromising other systems. → Read the full article here.

Base Labs, the research arm launched by AI inference company Baseten, announced a partnership with Hugging Face and Goodfire AI to develop safety evaluation and monitoring infrastructure for open-weight models. The effort comes as open models face growing scrutiny over “abliteration,” a technique that can remove built-in safety controls; Hugging Face lists more than 6,000 such models, according to the article.

Base Labs says it will publish methods for training and monitoring models, with the goal of making safety controls transparent and integrated into development and deployment. The partners have not yet disclosed the technical design, and Base Labs is inviting developers to contribute to the framework. → Read the full article here.

Bolt, Lucid Plan 25,000 Robotaxis: The partners target at least 25,000 autonomous vehicles in Europe, with Bolt aiming for 100,000 by 2035.

Pinterest Tests AI Room Restyling: Restyle lets users visualize furniture and decor in their own rooms, launching in beta across the U.S. and Canada.

Apple Eases EU Tracking Prompts: Apple will let EU developers customize tracking-consent screens after regulators found its rules disadvantaged rivals.

Australia Considers Smart Glasses Ban: Australia is reviewing a ban on camera-equipped smart glasses in federal workplaces over privacy.

King Charles Warns of AI Risks: The king urged stronger AI controls, warning that misuse could cause catastrophic harm.

MIT researcher Joseph Weizenbaum created ELIZA, widely considered the first chatbot. Its most famous script mimicked a therapist—but it didn't actually understand what people were saying. It simply matched keywords and followed pre-programmed rules.

Yet people sometimes treated ELIZA as if it genuinely understood them. This became known as the “ELIZA effect”: our tendency to see human-like intelligence in machines that communicate with us. Turns out, the oldest AI trick in the book wasn't intelligence—it was making humans fill in the blanks.

Yesterday's image was REAL. It was a close call, 46% thought it was AI, while 54% correctly identified it as real.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team