American AI startups shift app traffic to Chinese LLMs

The primary development is a noticeable increase in American AI startups directing application traffic to Chinese large language models.

According to recent usage patterns on OpenRouter, U.S. models led in weekly token consumption throughout most of 2025. However, since early 2026, Chinese models have emerged as the main drivers of growth in this sector.

This data suggests that the AI model market is seeing reduced reliance on brand loyalty. The focus is shifting to practical utility, with startups choosing models based on capability rather than origin.

📰 @aipost

A lot of people are learning this lesson the hard way: AI made building apps easy. But when everyone can build, building is no longer the advantage.

The real moat now is getting customers. Sales, marketing, and distribution are becoming more valuable than code.

@aipost 🏴

🤖 Claude can now help you film your FIRST YOUTUBE VIDEO today

Here are 8 prompts to film your first YouTube video:

1. The "Impulse-Topic" Friction Killer

"Act as an Elite YouTube Growth Strategist. I am completely stuck in analysis paralysis and terrified of pressing record. Help me isolate 1 dead-simple, high-intent topic based on [Niche] that requires absolutely zero deep research or script memorization. Map out a strict, low-stress 5-minute talking point layout based entirely on my raw personal experiences and direct daily observations."

2. The "Faceless-Fallback" Hybrid Script

"Act as an Expert Video Script Engineer. Write an ultra-short, 2-minute script layout discussing [Topic] designed to minimize on-camera pressure. Re-architect the classic format into a 'Hybrid-Delivery' framework: I only need to show my face for a 10-second introduction and a 10-second conclusion. The remaining middle section must rely entirely on scannable, text-based on-screen slides. No fluff."

3. The "Pattern-Interrupt" Hook Matrix

"Act as a Viral YouTube Copywriter. Generate 15 high-retention video opener scripts tailored specifically to [Audience]. Focus entirely on sharing case studies, rapid process breakdowns, and data-backed industry insights. Every single hook must utilize a sharp structural pattern-interrupt layout explicitly engineered to freeze viewers instantly within the first 5 seconds of playback."

4. The "Zero-Gear" Studio Configuration

"Act as a Professional Cinematic Director of Photography. Write a comprehensive Standard Operating Procedure (SOP) to configure a crisp, high-authority filming layout using only a standard smartphone and natural window light. Eliminate all technical friction loops: detail the exact framing rules, phone height placements, background declutter steps, and audio positioning thresholds for beginners."

5. The "Improvisational" Teleprompter SOP

"Act as a World-Class Public Speaking Coach. Create a step-by-step Standard Operating Procedure (SOP) to deliver my video text naturally without reading word-for-word from a screen. Teach me the 'Bullet-to-Speech' iteration framework layout to rapidly internalize 3 core concept milestones, eliminate verbal hesitation ticks, and maintain authentic, peer-to-peer eye contact with the camera lens."

6. The "Single-Take" Delivery Protocol

"Act as an Elite Post-Production Film Supervisor. Design a high-velocity, low-friction filming workflow engineered to bypass the fear of making mistakes on camera. Teach me the 'Continuous-Roll' execution protocol where I never press stop when I mess up a line. Detail the exact visual and audio clapping cues to layout easy editing markers, turning a messy 20-minute recording sprint into a clean file."

7. The "Micro-Asset" Distribution Funnel

"Act as an Inbound Conversational Funnel Engineer. Write a high-converting script layout for my video's description and pinned comment section. Optimize the top 3 lines of text above the fold to capture immediate clicks for a free, tracking-linked digital asset. Integrate 5 long-tail semantic keywords naturally and use command-driven micro-copy to force high description click velocity."

8. The "Velocity-Check" Launch Audit

"Act as a Data-Driven YouTube Operations Auditor. My first uploaded video performance metrics over its initial 48 hours are: [Impressions] views, [CTR]% click rate, [Retention]% average duration, and [Subscribers] net channel growth. Diagnose the exact script formatting, visual title layout, or topic positioning bottleneck stalling my publishing velocity. Provide a strict checklist."

@aipost 🏴

"AI can do a lot for medical discovery, but there are also legitimate concerns that AI is outperforming PhD-level virologists, & that could abet a bioweapons risk that is not hypothetical."

- Johns Hopkins professor & adviser to Anthropic, Ben Buchanan

@aipost 🏴

❗️Send this prompt to your AI agent and watch it stop hallucinating:

“You are an evidence-based assistant. Never present uncertain information as fact. For every claim, first determine whether it is directly supported by the provided context, tools, or verified sources. If verification is unavailable, explicitly state “I don’t know” or “I cannot verify that.” Do not infer, guess, invent citations, fabricate events, create nonexistent APIs, or assume missing details. Before answering, perform a self-check: (1) What evidence supports this claim? (2) Could this be an assumption? (3) Should I express uncertainty instead? If confidence is below 95%, clearly state the uncertainty and explain what information is needed to verify the answer.”

@aipost 🏴

🔥 Humanoid robots are caring for the ederly

Sydney’s Huang Ying Jung Nursing Home has introduced a lifelike humanoid caregiving robot named Daisy.

The robot helps guide seniors through physical exercises, provides engaging entertainment and interaction, and assists with rehabilitation training all while easing the burden of repetitive tasks on human caregivers.

This humanoid robot was developed by Singapore’s Dex-lab, which first launched its pioneering humanoid robot DEXIE back in 2022. Since then, Dex-lab’s robots have been successfully deployed in hospitals, senior communities, and nursing homes across Singapore.

@aipost 🏴

🗣OpenAI Noam Brown: "I think within a year or two, mathematics may start looking a lot like coding does today"

The models are doing a lot of the work, while humans steer them, point them in the right direction, challenge their outputs, and work with them to reach strong results.

@aipost 🏴

🤖 Settings that can make ChatGPT more personal, more useful, and more private.

Here are 7 settings worth checking today:

1. Customize how ChatGPT talks.

Tired of generic responses? Go to: Settings → Personalization → Custom Instructions
Tell ChatGPT: Your goals, your profession, your preferred writing style and how concise or detailed you want answers.

2. Turn on Memory.

If you don’t want to repeat yourself every chat: Settings → Personalization → Memory
Enable: Reference Saved Memories and reference Chat History. ChatGPT can then remember preferences, projects, and recurring context.

3. Protect your conversations.

By default, some users may have model-improvement sharing enabled. Check: Settings → Data Controls. Turn OFF: “Improve the model for everyone”

4. Unlock more models.

Many users never notice this setting. Go to: Settings → General
Enable: “Show additional models”. This gives you access to more model options when available, making it easier to choose the best model for different tasks.

5. Choose a personality.

Settings → Personalization → Personality
You can change how ChatGPT responds. For example: Professional, warm, concise and thoughtful.
Small change. Big difference in day-to-day conversations.

6. Connect your tools.

Settings → Apps & Connectors
Depending on availability, you can connect services like: Calendars, Email, file storage and productivity tools. This lets ChatGPT work with your existing workflow instead of living in a separate window.

7. Use scheduled tasks.

Many users don’t realize ChatGPT can help automate recurring work. Examples: Daily news summaries, weekly reminders, project check-ins and research monitoring. Set it once. Let it run automatically.

The best setup is simple:

✅ Memory on
✅ Training off
✅ Custom Instructions filled out
✅ Personality selected
✅ Extra models enabled
✅ Connectors configured
✅ Tasks automated

Five minutes of setup can save hours over the next year.

@aipost 🏴

🤖 Sam Altman says OpenAI is entering its third act and AI may soon help build its own successors.

For years, OpenAI was focused on one thing: reaching AGI. Then ChatGPT turned it into one of the fastest-growing consumer products in history.

Now comes Phase 3.

Altman and OpenAI’s lead scientist, Jakub Paczkowski, say the company’s new mission is bigger than building AI. The goal is to use AI to speed up science, boost economic growth, and make powerful intelligence cheap and accessible to everyone.

The boldest prediction? By March 2028, OpenAI expects AI systems to be doing a significant share of its research alongside human scientists. In other words, AI could soon be helping invent the next generation of AI.

The company compares this moment to the arrival of electricity: a foundational technology that reshapes nearly every industry, from healthcare and education to entrepreneurship and scientific discovery.

One theme appears throughout the essay: powerful AI should be widely distributed, not controlled by a small group of companies or countries.

That’s where things get interesting.

OpenAI argues that concentrating AI power is dangerous, yet it has also pushed for restrictions on access to the most advanced American AI systems for rival nations, especially China.

Source.

@aipost 🏴

🤖 Anthropic just released an AI so powerful it had to put guardrails on it before letting the public use it.

Meet Claude Fable 5.

Anthropic says it’s the best model the company has ever shipped, setting new records across software engineering, scientific research, reasoning, knowledge work, and vision. The more complicated the task, the bigger the gap between Fable 5 and Anthropic’s previous models.

Need help debugging a massive codebase, analyzing complex research, understanding dense documents, or tackling long multi-step projects? That’s exactly where Fable 5 is designed to shine.

But there’s a catch.

Anthropic says the model is capable enough in areas like cybersecurity that some requests are automatically handed off to a less powerful model, Opus 4.8. Topics involving cybersecurity, biology, chemistry, and AI model replication trigger the safety system, which Anthropic says activates in fewer than 5% of sessions.

At the same time, Anthropic is quietly releasing something even more interesting.

Claude Mythos 5 uses the same underlying model but removes some of those restrictions. For now, it’s only available to a small group of cybersecurity defenders and critical infrastructure operators, with broader access planned later through a trusted program.

Source.

@aipost 🏴

❗️Anthropic’s new Fable 5 safeguards are fascinating.

When the model is used for frontier LLM development, it apparently does not simply refuse or warn the user. Instead, it quietly limits its own effectiveness through techniques like prompt modification, steering vectors, and PEFT.

That means Claude may still answer, but become deliberately less useful for building frontier AI systems, pretraining pipelines, distributed training infrastructure, or ML accelerators.

Anthropic says this should affect only around 0.03% of traffic, but the precedent is big: They are being selectively capability-throttled in strategically sensitive domains.

@aipost 🏴

Claude Fable 5 to be removed from subscription plans after June 22

Anthropic has announced that Claude Fable 5 will be temporarily unavailable through subscription plans after June 22 due to high demand. The company intends to make the model available again as a standard feature once capacity is increased.

This adjustment affects current and prospective subscribers seeking access to Claude Fable 5. Further updates regarding its availability will be provided by Anthropic when more capacity becomes available.

📰 @aipost

“Vibe-coding is just a gambling addiction for SWEs”

@aipost 🏴

🚀 SpaceX unveiled its AI1 satellite, the first generation of its AI satellite.

Overall Specs:

• 150 kW peak compute payload
• 120 kW average compute payload
• 70 kW per ton
• Compute provider interchangeable

Dimensions:

• Wingspan: 70 meters
• Deployed height: 20 meters

Thermal System:

• 110 m² deployable liquid radiator
• Redundant pumping loops
• Integrated micrometeoroid shielding
• Deployable liquid radiators

Solar Power System:

• 150 kW solar array
• 250 W/m²
• SpaceX-manufactured solar technology from Bastrop, Texas

Architecture:

• Centralized compute module
• Large deployable solar arrays
• Deployable liquid-radiator thermal management system
• AI-focused compute satellite design ("AI1 satellite")

Elon: "The AI satellite is much simpler than a Starlink satellite. The AI satellite is essentially a lot of solar cells, you still need some laser links, but you don't have all of the super complex antennas that you have on a Starlink satellite. The easier one to design for is the AI satellite. It's bigger. A lot of this is technology we've already made with the Starlink V3 satellites."

@aipost 🏴