DeepSeek Crushes; Runway Video Amazes; Morphing Bot
Chinese Frontier Models That Rewrite the Price Curve
DeepSeek released V3.2 and V3.2 Speciale, two open-source reasoning models it claims are matching or beating GPT 5, Gemini 3 Pro, and Sonnet 4.5 on math, coding and tool-use benchmarks. Speciale even hit medal-level performance at the 2025 International Math Olympiad and Informatics Olympiad.
AI Secret surfaced the hidden advantage. DeepSeek Sparse Attention quietly cuts long-context inference costs by about 70 percent, enabling 300-page documents to run on mid-tier GPUs.
Why It Matters
DeepSeek’s move shifts the competition away from massive training budgets and toward compute efficiency. Anyone with consumer hardware can now run near-frontier reasoning locally, outside U.S. API control.
The Deets
- Two 685B parameter models
- MIT license with full weights on Hugging Face
- $0.28 input and $0.42 output per 1M tokens
- 70 percent inference savings on long contexts
Key Takeaway
Frontier AI is no longer expensive. It is downloadable.
🧩 Jargon Buster - Sparse Attention: A method that reduces computation by skipping unnecessary token interactions while preserving quality.
More: The Rundown AI | AI Secret
🏛 Power Plays
Microsoft Sharpens Its “Positive Sum” AI Thesis
Satya Nadella reiterated a shift in Microsoft’s AI philosophy: focus on economic uplift, not dominance. The company is doubling down on AI super-factories, per-agent pricing and an expanded right to OpenAI’s IP under the reworked partnership.
Why It Matters
Microsoft wants to be the infrastructure layer everyone builds on, including rivals. It is less apex predator, more neutral backbone.
The Deets
- Per-agent pricing aligns with autonomous system economics
- Microsoft gains long-term access to OpenAI’s model IP
- AI super-factories framed as global accelerators
Key Takeaway
Microsoft’s strategy is to lift the whole ecosystem and quietly own the rails.
🧩 Jargon Buster - Per Agent Pricing: Paying for AI agents instead of human users.
More: LinkedIn
🧰 Tools and Products
Runway Gen 4.5 Takes the Video Crown
Runway’s new Gen 4.5 dethroned competitors on Artificial Analysis’s text-to-video leaderboard. Its realism, physics handling, and hair and fabric coherence set a new bar.
Why It Matters
This is the strongest push yet toward true cinematic-grade AI video. The jump in realism feels like a preview of Hollywood pipelines in the near future.
The Deets
- Fluid dynamics and motion consistency improved
- Handles realism and cinematic styles especially well
- Originally codenamed Whisper Thunder and David
Key Takeaway
Gen 4.5 compresses years of VFX capability into a prompt box.
🧩 Jargon Buster - Fluid Dynamics Simulation: Mathematical modeling of how liquids and gases move to produce realistic visuals.
More: The Rundown AI
Kling O1 Unifies Video Generation and Editing
Kuaishou’s Kling O1 can take up to seven multimodal inputs and generate or edit three to ten second clips with fine grained control.
Why It Matters
It merges creation and editing into one interface, letting users restyle scenes, remove elements, shift lighting and change motion without touching traditional software.
The Deets
- Supports video references, camera motions, actions, subjects
- Edits existing footage through text commands
- Outperforms Veo 3.1 and Runway Aleph in internal tests
Key Takeaway
Kling O1 brings generative and editorial video under one roof.
🧩 Jargon Buster - Multimodal Input: Feeding multiple input types like text, images, and video into the same model.
More: The Rundown AI
Pomelli Auto-Builds Your Brand’s Entire Identity
Google Labs and DeepMind’s Pomelli scans your website, extracts brand signals, and generates campaign-ready creative assets.
It is like handing your brand to a designer, strategist, and art director in one tool that runs in seconds.
The Deets
- Detects fonts, colors, tone, keywords
- “Business DNA” fully editable
- Auto-propagates creatives across formats
Key Takeaway
Pomelli gives small teams a full creative department.
🧩 Jargon Buster: Business DNA - A structured summary of a brand’s visual and tonal identity.
More: The Rundown AI
ChatGPT’s Code Leak Reveals Ads Are Coming
Ad modules inside the ChatGPT Android beta show formats like search ads and carousel ads. With OpenAI’s compute burn hitting billions per quarter, monetization moves from optional to necessary.
Expect “friendly” sponsored answers woven into conversations. This is the business model shift that follows runaway compute demand.
The Deets
- Fully formed ad surfaces discovered
- HSBC projects $100B cumulative losses by 2029
- Ads likely embedded directly in conversations
Key Takeaway
ChatGPT is not courting advertisers. It is signaling distress.
🧩 Jargon Buster - Carousel Module: A horizontal card based ad surface.
More: AI Secret
🧪 Research and Models
NVIDIA Launches Alpamayo R1 for Autonomous Reasoning
NVIDIA’s Alpamayo R1 is an open reasoning model for autonomous driving that unifies vision, language and path planning.
Essentially NVIDIA is giving smaller automakers a head start on “driver minds” without massive proprietary datasets.
The Deets
- Open sourced on GitHub and Hugging Face
- Runs on NVIDIA hardware
- Useful for logistics, mining, agriculture, and robotics
Key Takeaway
NVIDIA is open sourcing autonomy but keeping the silicon moat.
🧩 Jargon Buster: Path Planning - How a robot decides safe navigation from point A to point B.
More: AI Secret
DSTAR Reinvents Mobility Through Morphing Design
Ben Gurion University’s DSTAR morphs its shape to navigate tight spaces, climb obstacles, and cross unstable terrain.
Why It Matters
Perfect for disaster zones, underground tunnels, and planetary exploration where conventional robots cannot fit.
The Deets
- Fits through 10 centimeter openings
- Climbs 20 centimeter planks
- Navigates soil, slopes, and grass
Key Takeaway
In robotics, flexibility is becoming the new strength.
🧩 Jargon Buster - Four Bar Linkage: A mechanical configuration that lets bodies change shape while staying stable.
More: Robotics Herald
⚡ Quick Hits
Accenture and OpenAI Expand Their Enterprise Alliance
Accenture is equipping tens of thousands of consultants with ChatGPT Enterprise and launching a program to help clients deploy agents across workflows. More: The Rundown AI
Nvidia Invests $2B in Synopsys
The investment accelerates chip design automation by pairing AI with advanced semiconductor engineering. More: The Rundown AI
Epic CEO Pushes Back on “Made With AI” Labels
Tim Sweeney says the tag will soon apply to everything and therefore means nothing, urging marketplaces like Steam to drop it. More: The Rundown AI
MIT Spinout Liquid AI Releases LFM2
The blueprint aims to make small models viable as on-device control layers for enterprise autonomy. More: AI Secret
Zig Foundation Leaves GitHub
Citing unaddressed GitHub Actions bugs and Microsoft’s AI-first shift, the foundation is moving to Codeberg. More: AI Secret
Telegram Launches Cocoon GPU Network
A decentralized platform connecting GPU owners, developers, and users for lower cost private AI compute. More: AI Secret
Changan Auto Invests RMB 225M in Its Robotics Arm
Another major automaker joins the humanoid race.
More: Robotics Herald
FieldAI Deploys Boston Dynamics Spot for Mega-Site Scans
Autonomous scans cover more than 500,000 square feet, pushing construction automation deeper into daily workflows. More: Robotics Herald
🧵 Tools of the Day
Artbreeder for collaborative image creation more…
Wobo for automated job applications more…
Powtoon for AI explainers more…
Reflect for backlink-based knowledge mapping more…
Colossyan for training videos built from PDFs more…
Lingo Champion for news-driven language learning more…
Wave for real time meeting transcription more…
Nebius Token Factory for fast open source inference more…
Today’s Sources: AI Secret | The Rundown AI | Robotics Herald