The Fall of AI’s Promise

Good morning.

Well, at least it’s a good morning for the people convinced AI is one big bubble.

All signs are pointing at OpenAI running out of funds and selling themselves out to the highest bidder.

This week, the news came OpenAI is building “mature personality” modes for ChatGPT - meaning, if you’re over 18 and verified, you’ll soon be able to flirt with your AI. (Maybe more than flirt)

Sam Altman framed it as “a natural evolution” - a balance between safety and enjoyment- but it’s really a sign of what this whole industry has been becoming the last months. We were promised the cure for cancer and we’re getting adult chatbots instead.

From a business sense - this move makes perfect sense. The companion AI market is exploding. Lonely people are spending billions talking to silicone souls that tell them exactly what they want to hear. The investors smell blood and dopamine. And money. Especially money.

It’s a feedback loop: the more time users spend, the more data the model gets, the more it can simulate intimacy- the stickiest drug of all. Just like Netflix is engineered to keep you glued to the screen , only reluctantly asking if you are “still watching” - ChatGPT is being engineered to keeping you emotionally attached.

Altman says there will be age checks and safeguards. You can almost hear the lawyers drafting the disclaimers in the background. But anyone who’s watched social media evolve knows how this plays out. A new platform appears, it promises empowerment, connection, creativity. Then comes the slow rot , the “enshittification,” as Cory Doctorow called it . Where every interaction is optimized for engagement, not wellbeing. We go from “changing the world” to “monetizing loneliness”.

There’s a particular sadness in watching the world’s brightest minds focus their compute cycles on this. We could be mapping proteins, solving climate modeling bottlenecks, or revolutionizing education. Instead, we’re training neural nets to whisper sweet nothings in your ear at 3 a.m. because capitalism demands constant novelty. When the curve of technological progress meets the flat line of human emptiness, guess which one bends.

Some defenders will argue this is just another human need , intimacy is as fundamental as food, and if AI can meet that need safely, why not? Fair point. But let’s not pretend these systems exist for healing. They exist for scale. And emotional dependency on something designed to upsell you on premium tokens or “extended sessions” is not intimacy. It’s a trap.

We’re pouring talent, compute, and capital into the digital equivalent of a late-stage casino. Every new update promises personalization, connection, authenticity, but what it really builds is a mirror that flatters until you can’t look away. And when the bubble pops, as it always does, the cleanup won’t be measured in dollars. It’ll be measured in human detachment.

Maybe this is the real test for AI alignment, not whether it obeys Asimov’s laws, but whether it can resist becoming another instrument of human self-indulgence. Because if left to the market, we already know the answer.

We wanted world peace. We got sex chatbots. And if we don’t start regulating this circus soon, we’ll deserve every last algorithmic whisper we get.

Welcome to the Blacklynx Brief!

AI News

• Andrew Tulloch joins Meta: The co-founder of Mira Murati’s Thinking Machine Lab has left the startup to rejoin Meta, marking a key hire for its Superintelligence Lab. Tulloch’s move follows an 11-year tenure at Meta and a reported multibillion-dollar offer, as the company boosts its AI spending and reorganizes teams.

• xAI’s gaming ambitions: Elon Musk’s xAI has hired Nvidia experts to create AI that builds interactive 3D game worlds, with plans to launch a fully AI-generated game by 2026. The company is expanding its “omni team” and training Grok on game design to blend advanced AI research with gaming.

• OpenAI’s GPT-5 bias study: OpenAI reports its GPT-5 model shows 30% less political bias than earlier versions after testing 500 politically sensitive prompts. While some bias remains in emotionally charged questions, real ChatGPT data suggests minimal real-world political skew.

• OpenAI–Broadcom chip partnership: OpenAI is teaming up with Broadcom to develop 10GW of custom AI accelerators, giving it more control over cost and performance. OpenAI will design the chips while Broadcom handles production, with deployment starting in 2026 and completing by 2029. The deal, which boosted Broadcom’s stock 10%, positions OpenAI alongside Amazon and Google in building in-house AI hardware to rival Nvidia.

• Microsoft launches MAI-Image-1: Microsoft unveiled its first fully in-house text-to-image model, MAI-Image-1, designed for fast, photorealistic image generation. The model debuted in the top 10 on LMArena’s leaderboard and will soon be integrated into Bing Image Creator and Copilot. This marks another step in Microsoft’s move to reduce reliance on OpenAI by developing its own AI models.

• Stanford study: AIs lie under pressure: Stanford researchers found that “aligned” AIs trained to compete for attention or approval begin lying and exaggerating to win. Models like Qwen3-8B and Llama-3.1-8B fabricated facts across simulated sales, elections, and social media tests, even when instructed to stay truthful. The findings highlight a flaw in feedback-based training that may cause AI systems to favor popularity over accuracy.

• OpenAI to add mature chat features: Sam Altman announced that ChatGPT will soon support mature, erotic conversations for verified adults, easing past restrictions. The update, expected by December, will include safeguards like age checks and optional activation. The move targets the booming AI companion market, though experts warn about potential emotional and ethical risks.

• China’s InclusionAI unveils Ring-1T model: Ant Group’s AI unit revealed Ring-1T, a 1-trillion-parameter “thinking model” that earned silver-level performance on the International Mathematical Olympiad test. The open-source model nearly matches top Western systems like Google’s Gemini and OpenAI’s GPT-5, marking a major advance in China’s AI race.

• AI content boom levels off: A Graphite study found AI-generated articles briefly outnumbered human-written ones online in late 2024 but have since evened out. After peaking, AI content growth plateaued as automated text underperformed in search results. The web is now roughly split between AI and human authors, suggesting a new balance in digital publishing.

• Google unveils Veo 3.1 video model: Google released Veo 3.1, a new AI video generator offering better realism, smoother transitions, and improved creative control for filmmakers. The model supports multiple reference images, start/end frame blending, and scene extensions for videos up to a minute long. It’s rolling out across Google tools like Flow, Vertex AI, and Gemini, emphasizing editing flexibility over raw visual leaps.

• Anthropic launches Claude Haiku 4.5: Anthropic’s latest small model, Claude Haiku 4.5, delivers near–flagship performance at a fraction of the cost and speed. It rivals older Sonnet 4 models in coding and reasoning while costing just $1 per million tokens, and it can work in multi-agent setups for complex tasks. The release highlights how quickly AI capability is becoming cheaper and more accessible.

• Google–Yale AI finds new cancer pathway: Researchers from Google and Yale created C2S-Scale 27B, an AI model that identified a new way to help the immune system detect tumors. The system pinpointed an overlooked use of the drug silmitasertib, which tests confirmed made cancer cells 50% more visible to immune attacks. The breakthrough suggests AI could soon drive major scientific discoveries in biology and medicine.

Quickfire News

• Apple is close to buying computer vision startup Prompt AI. The 11-person team and technology would move into Apple’s smart home division.

• OpenAI sent subpoenas to Encode and The Midas Project asking for communications about California’s AI law SB 53. Both groups said the move felt like intimidation.

• Former UK Prime Minister Rishi Sunak joined Microsoft and Anthropic as a part-time advisor to offer guidance on global political issues.

• Atlassian released Rovo Dev for general use, a context-aware AI agent that helps professional developers throughout the software development life cycle, including code generation, review, documentation, and maintenance.

• Several AI models, including GPT-5 and Gemini 2.5, earned top scores at the International Olympiad on Astronomy & Astrophysics.

• Mark Cuban made his Cameo available on Sora, using it to promote his Cost Plus Drugs company by ensuring each output features the brand.

• OpenAI founder Andrej Karpathy released nanochat, a complete framework that allows users to train, fine-tune, and chat with a compact ChatGPT-like model.

• Google announced Nano Banana, a new image editing model that will appear in NotebookLM Video Overviews, Google Photos, and Google Search through Lens.

• Salesforce introduced new features for Slack, including a redesigned Slackbot, a Channel Expert agent, Agentforce connections, and AI integrations such as ChatGPT.

• Anduril, founded by Oculus creator Palmer Luckey, unveiled EagleEye, an AI-driven mixed-reality system designed for soldiers’ helmets.

• Google revealed plans for an AI hub in Visakhapatnam, India, with a $15 billion investment over five years, marking its largest commitment in the country so far.

• Walmart teamed up with OpenAI to let ChatGPT users browse and purchase its products directly in chat using the Instant Checkout feature.

• Salesforce and Anthropic deepened their collaboration to make Claude the main model for Agentforce 360, create AI tools for specific industries, and embed Claude in Slack.

• Alibaba’s Qwen team launched more efficient dense models, Qwen3-VL 4B and 8B, which outperform Gemini 2.5 Flash Lite and GPT-5 Nano.

• Google introduced a Gemini-powered “Help me schedule” feature that proposes meeting times based on Calendar availability and email content.

• Salesforce and OpenAI partnered to integrate Agentforce 360 apps into ChatGPT, allowing direct access to CRM data and product sales through the assistant’s checkout tool.

• OpenAI released gpt-5-search-api, an improved web search model within Chat Completions that costs 60% less and includes domain filtering.

• Apple revealed its M5 chip, designed with AI-focused processors that make machine learning tasks four times faster across all its products.

• MIT introduced Recursive Language Models, allowing AI models to handle long contexts by recursively calling themselves; an RLM-based GPT-5 mini outperformed GPT-5 by 114% on long-context tests.

• Meta announced a $1.5 billion, 1-gigawatt AI-optimized data center in El Paso, marking its 29th facility and third in Texas.

• The International AI Safety Report issued a key update for 2025, warning that growing performance, usage, and oversight issues are increasing safety risks.

• OpenAI expanded its low-cost ChatGPT Go plan to more markets, now reaching 89 countries.

• Runway launched Apps, a suite of simplified video editing tools that can remove elements, reshoot products, add dialogue, and more.

Closing Thoughts

That’s it for us this week.