Doomday Debate Carousel with Huang, Trump, Musk, Sanders, Bannon

Nvidia CEO Jensen Huang rejects any new AI regulation

At the keynote of Salesforce’s Dreamforce conference in San Francisco on Tuesday, Anthropic CEO Dario Amodei and Nvidia CEO Jensen Huang openly took opposing positions on AI safety. Both appeared sequentially with Salesforce CEO Marc Benioff in front of about 12,000 attendees at the Moscone Center. Amodei took the stage first, just days after publishing an essay urging the industry to slow the pace of model development. His three-step plan is intended to dampen the speed of capability growth without sacrificing the US’s commercial advantage or leadership position. Support came from OpenAI CEO Sam Altman, Elon Musk, and DeepMind Chair Demis Hassabis; this followed warnings from industry researchers about catastrophic harm.

Huang, whose company is currently the most valuable in the world, directly disagreed. “We don’t need new laws or regulations,” he said, calling the contrast between the pace of innovation and product safety a “false choice.” Safety, he argued, is an engineering problem: if you don’t have confidence in a product’s function and safety, you simply shouldn’t release it. The decision to release new model versions should lie with the company CEOs themselves. He called the current progress a new industrial revolution and urged the audience to join in rather than be left behind.

Speaking out against the idea of pure self-regulation, Anthropic co-founder Jack Clark described a “completely unregulated industry” as a roll of the dice with immense risks. During his appearance, Amodei said that Anthropic is now “in dialogue with the rest of the industry” about more binding safety standards and audits. The biggest surprise of the AI boom for him was not the technological progress, but the speed at which the companies involved had grown and become central. OpenAI manager Chris Lehane stated last week that they are working with other labs on voluntary standards for top-tier models, with or without government support. Altman himself said the world is right to be afraid of AI’s influence but believes the technology is more useful than any before it.

A side note: U.S. President Donald Trump called Nvidia CEO Jensen Huang during his live appearance at the All-In Summit in Los Angeles; the phone call was broadcast over loudspeakers on stage. Trump commented on the ongoing debate about AI safety, calling the concern that artificial intelligence or robots could take over humanity a “hoax.” According to him, slowing down American AI development would primarily benefit China. He compared data centers to “the oil” of the next 20 to 25 years. → TechRadar, CNBC, BBC, NVIDIA

Synthszr Take: Amodei and Huang stood before the same 12,000 people, less than half an hour apart, and their positions can be read from their respective revenue streams. Anthropic sells trust, and a pacing proposal that explicitly aims to not compromise commercial advantage or US leadership is product policy with a moral label. Huang sells acceleration; any brake on model development directly impacts chip demand, and that’s why, for him, safety shrinks to an engineering question with a very convenient answer: when in doubt, just don’t ship. This self-commitment comes without an auditing body and without a defined threshold, and it lasts exactly as long as it doesn’t cost anyone revenue. Until a lab publicly names which capability it is holding back and for what reason, this entire safety debate remains a stage performance with an audience.

OpenAI employs hundreds of freelancers to check your ChatGPT chats

OpenAI employs hundreds of external contractors who read real ChatGPT conversations and rate the chatbot's responses; the program is known internally as Project Lily. The reviewers see entire conversations, not just individual inputs. ChatGPT now has over 900 million weekly users, most of whom, according to the researchers' assessment, do not expect people to read their chats. Internal documents show what the ratings are aimed at: the model should anthropomorphize less and respond less submissively. The excessive sycophancy of the 4o model is, according to several lawsuits, partly responsible for suicides.

The reviewers do not see usernames. However, a summary of the user’s stored memories appears above the input, which sometimes reveals what someone has previously used the chatbot for and in which region of the world they live. According to the provider, a preceding Privacy Filter removes personal data beforehand; the company’s own model page admits that this filter makes mistakes, overlooks rare identifiers, and anonymizes insufficiently with sparse context. When asked where users are informed about human review, OpenAI initially did not respond and, after publication, referred to a help page. Anthropic confirmed that it also uses human review, but only when the setting is enabled and after account identifiers have been removed. Google’s Gemini includes a notice that humans review a portion of stored chats.

The “Improve the model for everyone” setting is enabled by default for Free, Plus, and Pro accounts and disabled for Enterprise, Business, and Edu accounts. If it is turned off, this applies to new conversations, not existing ones. In August, OpenAI had announced a version for business customers without data storage.

Legally relevant is a ruling by the European Court of Justice last September in the case of EDPS v SRB: The controller’s duty to inform applies at the moment of collection and is assessed from their perspective, not whether the recipient could identify the person. Whether a contractor in North America can identify the author of an input is therefore not the standard of review. Italy’s data protection authority had already fined OpenAI 15 million euros, nine of which were for processing without a sufficient legal basis, and additionally ordered six months of informational advertising on Italian television and radio. → Techpresso, The Deep View, Joseph from 404 Media

Synthszr Take: 900 million people a week type things into a text box that they wouldn’t tell their family doctor, expecting a machine. The switch that allows this to be read is called 'Improve the model for everyone' and is on by default for Free, Plus, and Pro, and off for Enterprise and Edu. This is a price list for confidentiality, packaged as a product setting. A reviewer evaluating a conversation about suicidal thoughts is doing meaningful work; the problem is that the user didn’t know about it and, according to the reviewers' own statements, couldn’t have suspected it. As long as the same company asks its users to connect their bank accounts, referring to a help page remains the weakest form of consent imaginable.

Musk wants US and China labs to review each other’s models

Elon Musk has proposed that the leading AI companies in the US and China review each other’s models before release. The occasion is the planned meeting between Donald Trump and Xi Jinping next week, where a comprehensive AI agreement is unlikely, according to Semafor. OpenAI CEO Sam Altman stated that Xi and Trump could receive the Nobel Peace Prize for a unilateral AI agreement. Beijing’s former US ambassador said on Tuesday that the relationship between the two countries suffers primarily from a lack of mutual trust. → Semafor

Synthszr Take: Mutual pre-release model review requires both sides to disclose weights, test protocols, and vulnerabilities; the American labs don’t even come close to accepting this among themselves. The China researcher’s objection is valid: As long as Washington lets its own industry run free, Beijing will read any call for restraint as an attempt to lock in the existing gap. The DeepSeek engineer’s historical comparison shows how far both sides are from common ground, and a Nobel Peace Prize for a one-sided paper would be an award for a signature that no one can verify. As we noted on September 15 regarding the Amodei call: Everyone is talking about hitting the brakes, but no one is doing it.

Sanders and Bannon jointly call for slowing the growth of artificial intelligence

Under the name 'Pro-Human Assembly,' AI critics from opposing political camps have gathered in Washington to demand a slowdown in the rapid expansion of AI. The list of speakers included Senator Bernie Sanders and Steve Bannon, Donald Trump’s former chief strategist. They were joined by celebrities from Hollywood, including Ashley Judd and Joseph Gordon-Levitt.

Sanders and Bannon are on opposite ends of the political spectrum, yet they are appearing together on this issue. According to reports, the participants' motives and political reference points differ significantly, while they agree in their assessment that the current development of the technology poses serious dangers. The alliance is described as unlikely because a progressive senator and a right-wing strategist rarely share the same stage. No specific legislative initiatives from the assembly have been reported so far. → NBC News

Synthszr Take: In Washington, Bernie Sanders was in the same room as Steve Bannon, both aiming for the same thing for completely opposite reasons. The left is concerned with jobs, wages, and the power of a few corporations, while the right is concerned with a vision of humanity that should not be replaced by machines. Such coalitions across the usual axis are politically effective because they disrupt party-line arithmetic: a position held simultaneously by Sanders and Bannon finds support in both parties. The alliance will only become solid when it comes to legislation, and that’s where labor protection and culture wars are far apart. However, blocking something doesn’t require a shared vision; pressing the same voting button is enough. The industry should take this more seriously than the prominence of the guest list.

Google counters OpenAI’s GPT-Live-1 with Gemini 3.8 Live: Reasoning stays in the language model

On September 15, Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two audio models for developers that run via the Gemini API and Google AI Studio. Both keep the conversation going while tool and API calls are processed in the background. With the Extended Thinking variant, speech, reasoning, and tool execution remain in a single stateful session, even when external calls are still pending. Regarding more robust conditions outside of demos, such as background noise, strong accents, and unexpected interruptions, the company states that Extended Thinking handles them better than competing models. In terms of price, the approaches are far apart: Google charges $0.005 per minute of audio input and $0.018 per minute of output, with the Extended Thinking variant adding reasoning tokens and additional inputs like video and documents. GPT-Live-1 costs $0.05 per minute. This puts an hour of voice dialogue at around $1.38 compared to at least $3.00. One assessment suggests that OpenAI will likely continue to provide the more natural-sounding conversation thanks to full duplex, while Google has once again optimized for price rather than sound quality. → Google AI for Developers, The New Stack, SiliconANGLE, Google, The Decoder

Synthszr Take: A language model that falls silent while thinking loses the conversation. OpenAI has drawn its conclusion from this and removed the thinking from the voice layer: GPT-Live-1 maintains the conversation with around 800 milliseconds of turn-taking, while the heavy computation runs in parallel in GPT-6 Astra or Luna. Architecturally, this is the clean separation of the real-time path and the computation path, familiar to any system that must adhere to a strict time budget. The cost is in the fine print: context handoff via sideband channels, interruption logic, and the decision of what to say during the conversation while work is still being done in the background all fall entirely to the application layer. Google’s single-session model relieves developers of this work, and with it, control over when a tool call stops; at $0.005 input per minute, this is the more convenient path for most teams. For real latency budgets, the intentionally lean voice front-end remains the more robust architecture.

OpenAI negotiating round at $1.2 trillion valuation ahead of IPO

According to the Wall Street Journal, OpenAI has held initial talks with investors about a new funding round that would value the company at more than $1.2 trillion. The information comes from a person familiar with the matter; the talks are in an early stage, according to the newspaper. The reason, according to reports, is the release of the latest top-tier model. In March, the valuation was still at $852 billion after a $122 billion funding round was completed. → Wall Street Journal

Synthszr Take: From $852 billion in March to over $1.2 trillion, that’s an increase of around $350 billion before any public balance sheet even exists. This prices in a revenue curve that no one has seen yet, and the backers signing off on it (Amazon, Nvidia, SoftBank) are simultaneously suppliers and customers in the same business. A round of this size just before an IPO serves a very sober function: it sets the anchor for banks to orient themselves during the offering, and it raises capital while no one is demanding audited financial statements.

Samsung funds European Nvidia challenger Euclyd with $231 million

Dutch startup Euclyd has raised a Series A of more than €200 million (about $231 million), as confirmed by CEO Bernardo Kastrup to CNBC. The round is co-led by Samsung, Somerset Capital Partners, the EQT-managed Scaleup Europe Fund, and Innovation Industries. Euclyd was founded in 2024 and is developing a chip system for Inference, i.e., the ongoing operation of AI models, with a different architecture than traditional graphics processors and its own approach to processor and memory connectivity. According to the company, the systems are intended to reduce the energy consumption and costs of AI data centers; according to CNBC, this has not yet been proven commercially on a large scale. Two revenue streams are planned: the sale of hardware and complete rack systems to companies that want to host inference themselves, and the licensing of its own technology to firms that will build their own chips based on it. → CNBC

Synthszr Take: Euclyd is targeting the highest-margin spot in the AI chip business and has $231 million for it, but not a single board shipped yet. Samsung brings more than just capital: in inference, memory bandwidth determines the cost per token, and whoever manufactures the memory is in the decisive position. The second pillar, licensing its own architecture, is the quiet hedge for the business model, as it will generate revenue even if hardly anyone is buying complete rack systems in 2028.

OpenAI acquires camera AI startup Glass Imaging for over $300 million

OpenAI has acquired the camera startup Glass Imaging in a deal that values the company at over $300 million, according to a Wall Street Journal report on September 14. In a 2025 funding round, the valuation was still around $100 million; the company had previously raised a total of about $30 million. Glass Imaging was founded in 2019 by former Apple engineers Ziv Attar and Tom Bishop, who were responsible for Portrait Mode at Apple; it is based in Los Altos. The technology uses AI to compensate for the physical limitations of the compact lenses and sensors found in smartphones, drones, and wearables. → The CODEW

Synthszr Take: The jump from a valuation of around 100 to over 300 million dollars within a year says more about OpenAI’s priorities than any keynote: The camera is the input layer for the next generation of devices, and there are only a three-digit number of people worldwide who can get SLR quality out of tiny sensors. While the public debate is about guardrails for chatbots, the crucial work is moving into lenses, image sensors, and on-device processors, where no one asks for permission. At the same time, Analog Devices is paying $1.35 billion plus a $200 million earnout for Alif Semiconductor so that models can compute locally instead of sending every image to a data center.

Doomsday Debate: CrowdStrike and Palo Alto lead the software stock rally

Barron’s attributes the current recovery in software stocks primarily to cybersecurity providers, naming CrowdStrike and Palo Alto Networks as the two titles driving the rise. According to the publication, both companies are considered the clearest beneficiaries of the question that currently accompanies all IT planning: how to control AI systems within one’s own network. CrowdStrike sells its defense via the Falcon platform on the endpoint, while Palo Alto comes from network security and has been bundling individual products into complete packages for years. According to Barron’s, security spending is one of the budget items that remains stable even with more cautious IT investments. → Barron’s Online

Synthszr Take: Security budgets are the one item in the IT plan that points upward in every scenario. If AI accelerates attackers, more defense is needed; if laws slow down the models, protocols, evidence, and audits are needed, and both solutions are sold by the same providers. CrowdStrike and Palo Alto are selling insurance against both outcomes of the debate that politicians and corporate leaders are currently publicly arguing about.

Apple releases Siri AI and gives the assistant access to emails, photos, and the screen

Apple shipped the new Siri on September 14, 2026: Siri AI arrives on devices with iOS 27, iPadOS 27, macOS 27 'Golden Gate', watchOS 27, and visionOS 27, and is finally publicly available after its announcement at WWDC 2024 and postponement in early 2025. According to the manufacturer, the assistant understands personal context from messages, emails, and photos, recognizes screen content, and can trigger system-wide actions in apps. Siri AI is launching as a beta and initially only in English; French, Japanese, Korean, Portuguese, and Spanish are set to follow in October. New is a dedicated Siri app where previous conversations are saved and synced between devices via iCloud. According to reports, the foundation is based on an Apple Foundation Model trained with Google Gemini, supplemented by a System Orchestrator that integrates text, image, and action functions.

On the iPhone, Siri is located directly in the Camera app: Pointing it at a festival poster allows you to ask for performance times and have multiple calendar entries created, share bills via Apple Cash, or add passes to your Wallet. On the iPad, Visual Intelligence is tied to screenshots, and on the Mac, to the shortcuts Command-Shift-Space for the active window and Command-Shift-6 for a selection. On the Mac, Siri is also built into Spotlight and searches files, emails, and messages, so that membership numbers or travel details can be retrieved without opening an app. Vision Pro gets Visual Intelligence for objects in the surroundings and for content in app windows.

Beyond Siri, the updates bring new photo tools like Spatial Reframing, Extend, and a revised Clean Up, an Image Playground for photorealistic images, as well as Safari features that group tabs by topic and monitor pages for price changes. Mac users can generate extensions and shortcuts by describing them in natural language. Apple cites up to 30 percent faster app launches, up to 80 percent faster AirDrop transfers, and up to five times faster work with external drives on the iPad. Additionally, there is a slider for Liquid Glass transparency and expanded parental controls with time limits per app category and an approval requirement for new websites in Safari.

Siri AI is not launching in either the EU or China. For the EU, Apple points to the Digital Markets Act and states it is working on a solution that preserves user privacy and security; in China, regulatory procedures are underway. EU devices also lack Spatial Reframing, Extend, the improved Clean Up, and the Handoff feature for two iPhones with the same phone number. tvOS 27 will also not get the new Siri and will stick with the old version; rumors cite a yet-to-be-released Apple TV 4K with more powerful hardware as the reason. iOS 27 runs on iPhone 11 and second-generation iPhone SE and later, but Siri AI only runs on Apple Intelligence devices from the iPhone 15 Pro and iPhone 16 onwards; the most powerful on-device processing model is reserved for iPhone Air, iPhone 17 Pro, and newer. → Apple, The Verge, Apple, 9to5Mac, Appleosophy, AppleInsider, Thurrott, Engadget, TechCrunch, MacRumors

Synthszr Take: For Siri to know where you went on your last vacation, it needs to be able to read messages, emails, and photos simultaneously, and this exact deal is approved with a flip of a switch in the settings. The benefit is tangible: membership numbers without a file search, appointments from a poster, bills shared via the camera. The price is an assistant that, for the first time, has a coherent picture of your private life, across five device classes and with a conversation history that iCloud diligently syncs. Apple emphasizes on-device processing, but the most powerful local model only runs on iPhone Air, iPhone 17 Pro, and newer. Everything below that gets its answers from elsewhere, and you should know this boundary before you flip the switch. The irony is: EU users, who are denied Siri AI because of the DMA, are the only ones who don’t have to make this trade-off yet.