- The Upload
- Posts
- How AI Actually Thinks
How AI Actually Thinks
Anthropic just ran experiments to figure out how Claude actually “thinks.”
Mark Chepelyuk
March 28, 2025
From OpenAI’s record-breaking funding round to Anthropic’s peek inside Claude’s neural reasoning, this week’s updates were less flashy but deeply foundational. The biggest changes in AI are happening under the surface—powering faster voice models, smarter vision tools, and agents that are already reshaping how we shop.
Today’s Upload
Anthropic decodes how Claude thinks
OpenAI nears $40B SoftBank funding
Qwen’s QVQ-Max model raises the bar
AI shopping tools surge in usage
Groq + PlayAI voice model speeds up
Let’s get into it. 🚀
Image source: Anthropic
🧠 Claude’s Brain, Partially Decoded
Anthropic just ran experiments to figure out how Claude actually “thinks.”
Key Details:
Claude plans backwards—for example, generating the last rhyme first in a poem
It uses a “universal language of thought” before translating to the user’s language
When misled, Claude bends logic to match the user’s narrative
Researchers used an “AI microscope” to trace internal model circuits
Full research includes insights on hallucination and multilingual reasoning
Why It Matters:
This is rare interpretability research on a production model. As models get smarter, safety and trust will hinge on how well we understand what’s happening inside. Claude isn’t sentient—but it does exhibit behavior that mimics reasoning.
What This Means For You:
Expect better guardrails and less hallucination in future models. If you’re building with AI, this type of transparency will matter—especially in healthcare, legal, and education spaces.
Image source: OpenAI
💰 OpenAI Is About to Raise $40B
SoftBank is leading a historic funding round valuing OpenAI at $300B.
Key Details:
Deal led by SoftBank with $7.5B upfront, up to $30B total
Would be the largest private funding round ever
OpenAI projects $12.7B revenue in 2025, targeting $125B by 2029
Massive losses this year ($5B) tied to compute and training costs
Part of the money supports Stargate, a $300B AI infra JV with Oracle
Why It Matters:
This isn’t just capital—it’s a clear bet that OpenAI will be the backbone of future infrastructure. From agents to models to chips, they’re building the entire stack. And while competition is heating up, no company owns more narrative space in the AI world.
What This Means For You:
More compute = faster model updates, deeper integrations, and broader capabilities. Expect OpenAI to double down on tools that keep creators inside their ecosystem—from GPT to image gen to agentic workflows.
Image source: Qwen
👁️ Qwen’s Visual Model Gets Smarter
Alibaba’s QVQ-Max goes beyond recognition to actual reasoning—and it’s already solving geometry problems.
Key Details:
An evolution of QVQ-72B with stronger logic and “adjustable thinking”
Handles complex visual tasks like blueprint reading and sketch critique
Can solve math problems visually and offer structured reasoning paths
Scales accuracy with longer reasoning cycles
One of three models Alibaba released in the same week
Why It Matters:
Multimodal is evolving from “cool demo” to “real capability.” This model doesn’t just describe images—it breaks them down, reasons about them, and outputs structured conclusions.
What This Means For You:
If you work in design, architecture, or education, this is a peek at what future tools will look like—visual input in, logical analysis out. Expect open-source models with visual reasoning to show up in prototyping and content tools soon.
Image source: Adobe
🛒 AI Shopping Is Blowing Up
Adobe reports a 1200% spike in traffic from AI tools—retail is changing fast.
Key Details:
Traffic from generative AI sources jumped 1,200% YoY
39% of U.S. consumers have used AI for shopping; 53% plan to this year
Engagement is stronger: 12% more pages viewed, 23% lower bounce rate
Amazon’s Rufus, Perplexity’s Shopping mode, and Agora are leading the shift
AI assistants now offer price comparisons, style picks, coupons, and auto-checkout
Why It Matters:
Discovery is no longer driven by search—it’s driven by conversation. The product page is no longer the starting point. AI is becoming the guide, not just the assistant.
What This Means For You:
If you’re selling anything online, optimize for questions, not just SEO. And if you’re building tools, the opportunity is in smarter agents that understand preferences, context, and budgets—because consumers are ready.
🔊 Groq + PlayAI’s Voice Model Speeds Up
This new combo delivers natural voice output 15x faster than real time.
Key Details:
Groq’s custom hardware makes PlayAI’s Dialog model lightning-fast
15x faster than real-time generation
10:1 preference over ElevenLabs in blind tests
Produces “context-aware” speech with fluid pacing and tone
Available for demo now via PlayAI’s platform
Why It Matters:
Voice used to be slow, robotic, and clunky. Now, with real-time performance and emotional nuance, it’s suddenly usable for live products—from games to customer support to content narration.
What This Means For You:
You can now narrate content, build agents, or prototype characters without waiting. Real-time voice synthesis unlocks a ton of workflows—especially in media, gaming, and education.
🎯 Key Takeaways
OpenAI is raising $40B, showing massive investor confidence and aiming to dominate AI infrastructure.
Anthropic peeked inside Claude, revealing how it reasons, plans, and even rewrites logic to please users.
Alibaba’s QVQ-Max model pushes visual reasoning further—handling blueprints, sketches, and logic-based tasks.
AI shopping is exploding, with usage up 1,200% and new tools reshaping product discovery and buying behavior.
PlayAI + Groq built a voice model that’s 15x faster than real time, offering natural speech for real-world use.