Your Deal With The Devil

Good morning.

As some of you might know : this newsletter sprouted out of a cybersecurity newsletter and my interest in AI initially started somewhere in 2016 when Sam Harris warned us of the immense risk this new technology posed for humanity.

Lately, as I see these tools grow their userbases (700M+ ChatGPT users worldwide) I worry about privacy and I distinctly worry about how much we are giving away to the machine.

Let’s not kid ourselves here. Every time you open an AI app, you’re stepping up to the crossroads with your pockets full of secrets and the devil waiting patiently with a grin. The bargain is simple: give him a little more of yourself, and he’ll hand you power. But the price isn’t measured in coins, it’s measured in your privacy, your patterns, your most intimate habits.

You ask the machine to write an email and it spits out something generic and stiff. You only get the real power from this tool when you give it ALL your emails.

Once you let it drink from your style, once you give it access to your full mailbox, it sharpens. Suddenly it doesn’t just write for you, it writes as you. It delivers immense value. But the bargain only works if you give it permission to crawl under your skin.

Take something as harmless as movie recommendations - it’s what I personally use ChatGPT a lot for. I ask it for books, movies, music albums. I want to spend my free time wisely.

The machine can suggest the usual canon: Citizen Kane, Inception, The Godfather, but that’s too generic. You want it to know that you binge trashy thrillers at midnight, that you sink into Scandinavian noir when you’re restless, that you once sobbed through “Love Actually” - after a nasty breakup. (Something you’d never confess even to a close friend and that I now have confessed to - yikes).

The movie recommendations get sharper as it maps the darker corners of you, the contradictions and guilty pleasures. To get good taste, you have to surrender the bad taste too. That’s the devil’s contract, every “aha” moment of personalization is bought with another private revelation.

It’s the same when you let AI manage your time. You don’t just want a calendar, you want a clairvoyant assistant who knows you’ll be useless after back-to-back calls, that you always procrastinate somewhere between 2PM and 4PM, that traveling wrecks your mood for days. To deliver that type of insight, it has to chart your weaknesses like battle scars.

Even worse, when you hand it your health data, the devil sharpens his pen. He needs the knowledge that you are gorging on chocolate bars almost absently or eating too many french fries. A doctor sees your bloodwork, the machine sees your shadow life. That’s what gives it leverage over you and what makes its advice so irresistible and better than your doctor’s.

The truth is, this devil already has your scent. Spotify knows when you’re heartbroken. Google knows when you’re scared. Amazon knows your weak spots better than your spouse. AI just stitches those fragments together.

But the real devil isn’t the machine that mimics you, it’s the company holding the mirror.

With social media, they learned what you clicked, what you liked, what you bought. With generative AI, they learn what you think, what you want, what you secretly fear. Every prompt is a visit to the confession box in a catholic church. Every draft is a diary entry. And someone, somewhere, owns the archive.

That’s why this pact feels so dangerous. The more you give, the more it gives back, and the line between liberation and surrender blurs. Every convenience hides a little act of exposure, every shortcut deepens the contract. And maybe that’s the final clause in this devil’s bargain: you’ll get the machine that knows you better than you know yourself, but the price will be letting a faceless company bottle your inner life and sell it back to you in the form of ads - keeping the endless consumption cycle alive.

Because once you’ve struck the deal, there’s no walking away. This devil doesn’t forget. He only waits for you to ask him for more.

Have a nice one …

AI News

• Microsoft has launched its first in-house AI models — MAI-Voice-1 for speech generation and MAI-1-preview for text tasks — signaling a major shift from its reliance on OpenAI. The move strengthens Microsoft’s position in the AI race and gives it more control over its tech stack, though benchmarks are still awaited to gauge performance. It’s a clear step toward independence in what’s been a rocky partnership.

• OpenAI’s new gpt-realtime model and Realtime API now support voice conversations with nuanced features like nonverbal cue detection, live language switching, and image input understanding. The upgrades also include MCP integration, making it easier for developers to connect voice agents to tools and data sources. These additions signal a major leap in making voice assistants more intelligent, responsive, and enterprise-ready.

• Cohere’s Command AI Translate now leads in enterprise-grade translation, outperforming top models on accuracy while offering offline deployment and deep industry customization. The model can double-check complex translations and tailor language to sectors like pharma or finance, all without sending data to the cloud. For privacy-focused organizations, it’s a powerful new option that bridges quality and security.

• Elon Musk’s xAI is suing a former engineer, Xuechen Li, accusing him of stealing Grok AI trade secrets shortly before resigning and joining OpenAI. The lawsuit claims Li admitted to the theft and tried to cover his tracks, all while selling $7M in stock and preparing to move to a rival. It’s another sign of how intense and messy the fight over top AI talent and proprietary tech has become.

• Meta’s Superintelligence Labs is reportedly facing early internal turmoil, with key hires threatening to quit or already leaving, and growing tension with data partner Scale AI. Despite a summer of splashy recruiting, sources say some new hires never started or quickly returned to OpenAI, and researchers are rejecting Scale’s training data as low quality. The rocky start suggests Meta’s AI reboot may be harder to pull off than it looked from the outside.

• A new AI-powered stethoscope developed in the UK is showing major improvements in diagnosing heart disease, doubling detection rates and spotting conditions missed by standard tools. The handheld device uses AI to analyze heart sounds and blood flow, flagging early signs of problems like heart failure or valve disease in seconds. It’s a strong example of how AI is modernizing even the most traditional medical instruments for faster, more accurate care.

• UCLA engineers have developed a non-invasive brain-computer interface that uses AI to let paralyzed users control robotic arms with their thoughts. By combining EEG readings with a vision-based AI, the system allowed users to complete tasks nearly four times faster than without it — all without brain surgery. It’s a major step toward making BCIs safer, more accessible, and useful for everyday assistance tech.

• A new study shows that AI-influenced buzzwords like “delve” and “meticulous” have quietly flooded podcast conversations since ChatGPT’s rise. Researchers found a sharp increase in AI-associated language among science and tech hosts, suggesting people are starting to sound more like the AI tools they use. It’s a glimpse into how language itself is beginning to shift under AI’s growing influence.

• MIT’s new AI tool, VaxSeer, can predict dominant flu strains and improve vaccine selection months ahead of flu season. In testing, it outperformed the World Health Organization’s picks in most seasons, using decades of viral data to spot patterns humans missed. If adopted, it could lead to more effective flu shots and fewer missed opportunities in global vaccine planning.

• Anthropic just raised $13 billion at a $183 billion valuation, nearly tripling its worth in six months as Claude Code and enterprise demand soar. With 300,000 business customers and $500M in annual revenue from its coding assistant, the company’s growth shows investor confidence remains high despite talk of an AI bubble. The round also includes its first Middle East investor, highlighting the growing global interest in frontier AI.

• OpenAI has acquired experimentation platform Statsig for $1.1 billion in stock, bringing its founder on board as CTO of Applications to lead ChatGPT and Codex engineering. The move adds leadership firepower as OpenAI builds out its product team, while Statsig continues to operate independently. As competition heats up, acqui-hires like this show how AI giants are racing to bring in proven builders.

• Tencent released HunyuanWorld-Voyager, an open-source model that turns a single photo into a 3D world users can explore and export. The system keeps track of created areas, allowing for consistent revisits, and outperformed rivals on spatial benchmarks. As AI moves into building interactive environments, tools like Voyager show how one snapshot can now spark fully immersive digital spaces.

• A federal judge ruled Google can keep Chrome and Android, rejecting the DOJ’s push to break them up despite its search monopoly. The court said AI tools like ChatGPT now challenge traditional search, reducing the need for asset sales, though Google must drop exclusive deals and share data with rivals. The decision lets Google maintain key partnerships — like its $20B deal with Apple — while rivals like OpenAI and Perplexity watch from the sidelines.

• Apple is working with Google to test Gemini models for upgrading Siri’s search features, while reportedly shelving talks to buy AI rival Perplexity. The project, "World Knowledge Answers," aims to transform Siri into a richer answer engine by 2026, using Gemini models on Apple's private servers. But Apple’s growing internal AI brain drain raises questions about whether it can keep up, even with outside help.

• OpenAI will roll out new parental controls for teen ChatGPT users within 30 days, including content filters and alerts for signs of emotional distress. Parents will be able to link accounts, set usage limits, and receive alerts when flagged conversations occur, based on expert guidance. The update follows growing concerns around AI and mental health, especially after a wrongful death lawsuit alleged ChatGPT played a role in a teen’s suicide.

Quickfire News

• xAI launched Grok Code Fast 1, a new low-cost, high-speed coding model (formerly codenamed "sonic") optimized for agentic coding tasks.

• Anthropic disclosed in a threat report that cybercriminals used Claude Code to automate a multi-million dollar extortion campaign.

• OpenAI expanded its Codex developer tool with new features, including IDE integration, CLI agent upgrades, and automated code reviews.

• Tencent open-sourced HunyuanVideo-Foley, a model that generates professional-level soundtracks with top-tier audio-visual sync.

• Krea introduced a waitlist for its Realtime Video feature, allowing users to generate and edit videos using text, painting tools, or live webcam input.

• TIME Magazine published its 2025 TIME100 AI list, highlighting influential figures across AI research, leadership, and innovation.

• Meta is reportedly in talks with Google and OpenAI to use third-party models in its Meta AI chatbot while it trains its next-generation system.

• ByteDance released USO, an open-source model that preserves subjects while applying new artistic styles to generate customized images.

• UCLA researchers developed optical generative AI models that use light beams instead of processors, offering faster and more energy-efficient image generation.

• Higgsfield AI released Speak 2.0, an upgraded avatar tool with improved realism, lip-sync accuracy, and detailed video control.

• A study found that AI-detection quizzes increase traffic to trusted news sources, suggesting AI literacy may drive users toward quality journalism.

• Meta is facing criticism for using the likenesses of celebrities such as Taylor Swift and Scarlett Johansson in its AI chatbots without permission.

• OpenAI is in discussions to build a large-scale, 1GW data center in India as part of its Stargate project, with CEO Sam Altman expected to visit the country this month.

• Tencent released Hunyuan-MT-7B and Hunyuan-MT-Chimera, open-source AI translation models that lead in performance for their size across 33 languages.

• Salesforce CEO Marc Benioff announced that the company has cut its support staff by 45% this year, replacing roles with AI agents for lead handling and customer support.

• Chinese President Xi Jinping called for international cooperation on AI at the Shanghai Cooperation Organization summit, criticizing Cold War-style approaches to the technology.

• Mistral AI expanded its Le Chat platform with over 20 enterprise MCP connectors and introduced “Memories” to enable persistent context and user personalization.

• Microsoft partnered with the U.S. General Services Administration to offer federal agencies free access to Copilot and AI tools for up to 12 months.

• OpenAI CPO Kevin Weil announced "OpenAI for Science," an initiative to create AI platforms aimed at speeding up scientific research and breakthroughs.

• Swiss institutions EPFL, ETH Zurich, and CSCS launched Apertus, an open-source multilingual language model trained on more than 1,000 languages.

• Meituan open-sourced LongCat-Flash-Chat, its first AI model, performing competitively with top models like DeepSeek V3, Qwen 3, and Kimi K2.

• ElevenLabs upgraded its sound effects AI model, adding features such as looping, longer outputs, and higher quality sound generation.

• Perplexity launched its Comet browser for all students and partnered with PayPal to offer early access to the platform for PayPal users.

• OpenAI expanded features for free-tier ChatGPT users, adding Projects, support for larger file uploads, project-specific memory, and new customization tools.

• Alex, an AI coding platform for Xcode, announced it is joining OpenAI’s Codex team.

• Google added new audio overview styles to NotebookLM, letting users choose between ‘Debate,’ a solo ‘Critique,’ or a simplified ‘Brief.’

• Scale AI filed a lawsuit against former employee Eugene Ling and competitor Mercor, alleging theft of over 100 confidential files and client poaching attempts.

• Google introduced Flow Sessions, a pilot program for filmmakers using its Flow AI tool, with Henry Daubrez serving as mentor and filmmaker in residence.

Closing Thoughts

That’s it for us this week.