Good morning. ChatGPT went down this week and people collectively lost their minds—with one professor receiving only 9 out of 50 essays because students couldn't function without AI assistance. Plot twist: the professor admitted they were also struggling to grade those nine essays on their own.

— The Deep View Crew

In today’s newsletter:

• 🖼️ AI for Good: Could damaged artwork become a thing of the past?

• 🧠 Magistral by Mistral AI excels in multilingual, transparent reasoning

• 🤖 Meta’s V-JEPA 2 improves physical reasoning in AI

🖼️ AI for Good: Could damaged artwork become a thing of the past?

A breakthrough AI restoration method is bringing damaged paintings back to gallery walls without touching the original artwork, potentially rescuing thousands of pieces currently hidden in museum storage.

What's new: MIT researcher Alex Kachkine developed a technique that uses computer vision and machine learning algorithms to digitally reconstruct damaged oil paintings, then prints the restoration onto transparent polymer overlays. His pilot restoration of a 15th-century Master of the Prado Adoration work—with over 5,600 damaged areas—took just 3.5 hours compared to the typical months or years.

How the AI works:

• High-resolution scanning creates detailed damage maps using computer vision algorithms

• Machine learning models analyze the artist's brushstrokes, color patterns, and style from intact painting sections

• AI algorithms generate 57,314 different color variations matched to the original palette using advanced color-matching techniques

• Digital inpainting fills gaps by learning from the artist's existing work and comparable pieces from the same era

• A transparent polymer mask is printed with the AI-generated restoration and overlaid on the original

Why it matters: Traditional restoration requires hundreds of hours per painting and costs that limit treatment to only the most valuable works. This forces museums to keep damaged pieces in storage indefinitely. Kachkine's method is up to 70 times faster than manual restoration while being completely reversible.

By the numbers: An estimated 70% of institutional art collections remain in storage at any given time, with damage being a primary factor for many artworks.

What's next: The technique uses convolutional neural networks (CNNs) for pattern recognition and could integrate with emerging technologies like multispectral imaging for even more precise restorations. While purists debate the ethics of AI-enhanced art, the method offers a compromise—preserving originals while making art accessible to the public.

The technology transforms museum storage rooms into potential exhibition spaces, giving centuries-old masterpieces a second chance at public display through the power of artificial intelligence.

Let Sue do your security questionnaires

Wrangling answers for a security questionnaire or sharing your SOC 2 with customers shouldn’t feel like herding cats.

Conveyor's AI Agent, Sue, transforms the entire security review process from chaos into control.

No more chasing down SMEs, managing a content library, updating systems, or hunting and pecking for the right answers.

Sue does it all — autonomously.

She completes every questionnaire, shares NDA-gated docs in your Trust Center, coordinates across systems, and even follows up with your team when needed.

All you get is a ping when it’s ready for final review. It’s how F1000 companies are completing security reviews 80% faster.

Ready to have Sue handle the chaos for you? Learn more here.

🧠 Magistral by Mistral AI excels in multilingual, transparent reasoning

French AI lab Mistral launched its first reasoning model, but early testing reveals it trails behind established competitors from OpenAI, Google, and Anthropic in key technical benchmarks.

Mistral released Magistral in two versions—an open-source Magistral Small and enterprise-focused Magistral Medium—designed to solve complex problems through step-by-step logical reasoning rather than pattern matching.

Why it matters: The release marks Europe's entry into the AI reasoning model space, where Microsoft-backed Mistral is betting multilingual capabilities can differentiate it from English-centric competitors like OpenAI's o1 and DeepSeek's R1.

Initial benchmarks tell a sobering story. Magistral Medium scored 73.6% on the AIME2024 math test, while DeepSeek's newer R1 model significantly outperforms it on the same benchmark. The model also underperforms Google's Gemini 2.5 Pro and Anthropic's Claude Opus 4 on physics, math, and coding evaluations.

How it works: Unlike standard AI models that generate responses immediately, Magistral uses what's called "chain-of-thought" reasoning—working through problems step-by-step before providing answers. The model supports eight languages including English, French, Spanish, German, Italian, Arabic, Russian, and Chinese, with CEO Arthur Mensch telling CNBC this differentiates it from "U.S. models that reason in English and Chinese models that reason in Chinese."

The competitive reality: The launch coincides with OpenAI releasing its most powerful reasoning model, o3-pro, highlighting the challenge facing the €1.1 billion-funded startup. Industry observers note that even with substantial funding, achieving state-of-the-art performance against established players remains formidable.

Mistral appears to be competing on different vectors—emphasizing speed (claiming 10x faster responses), multilingual reasoning, and open-source accessibility rather than pure benchmark performance. The Apache 2.0-licensed Magistral Small reinforces the company's commitment to open development, potentially building ecosystem advantages where it lacks technical leadership.

The bottom line: Magistral represents a calculated play for market share in a space where being first-to-market matters less than sustained competitive advantage. Whether multilingual reasoning and developer accessibility can overcome performance gaps remains the defining question for Europe's AI champion.

Shadow AI is here. This is how you stay ahead of it

AI is already in your organization...whether you’ve invited it or not. From hidden copilots to unsanctioned tools, shadow AI is creating invisible risks.

Tactical AppSec: An AI Security Field Guide gives you the clarity and control you need to secure AI without slowing innovation.

This isn’t a theory-heavy whitepaper. It’s a practical, action-first handbook with ready-to-use templates, shadow AI discovery tools, and governance checklists you can apply today.

Lead confidently and get ahead of risks before they get ahead of you — download the Tactical AppSec Field Guide now.

• Would ChatGPT risk your life to keep running?

• OpenAI rolls out o3-pro for ChatGPT Pro and API users

• Dia browser is a bold AI-powered bet on the open web

• Disney and Universal sue Midjourney AI over copyright use

• Wikipedia halts AI summary pilot after pushback from editors

• Apple’s using AI to tag apps and boost discovery in the App Store

• Maxar: AI/ML SME (TS/SCI)

• Motorola Solutions: AI-Focused Design Researcher

• LoopGenius: AI-powered marketing platform that automates ad campaigns across Facebook, Instagram and Google Search

• Camunda: See how AI + orchestration are transforming real-world workflows across finance, insurance, and more*

• AutoWrite App: A quick and efficient way to create high-quality SEO articles

🤖 Meta’s V-JEPA 2 improves physical reasoning in AI

Meta just released V-JEPA 2, an AI system that teaches machines physical intuition through video, as dedicated robotics companies deploy thousand-robot fleets and raise billions at sky-high valuations.

Meta unveiled V-JEPA 2, a "world model" trained on over 1 million hours of video that helps AI agents understand physics, predict outcomes, and plan actions in unfamiliar environments. Chief AI scientist Yann LeCun calls it "an abstract digital twin of reality."

V-JEPA 2 uses a two-stage training process: learning general world dynamics from video, then fine-tuning with just 62 hours of robot control data. The model predicts missing video segments in abstract space rather than reconstructing pixels, allowing it to focus on high-level concepts about object interactions.

Meta enters a space where hardware companies have massive head starts. Figure AI was in talks to secure $1.5 billion funding at a $39.5 billion valuation with plans to manufacture 100,000 humanoid robots. The company recently ditched its OpenAI partnership to build proprietary AI, stating: "We can't outsource AI for the same reason we can't outsource our hardware."

Boston Dynamics' Atlas robot demonstrates "advanced athletics and agility" and has partnered with the Robotics & AI Institute to accelerate reinforcement learning. Agility Robotics' Digit is already working full-time at Amazon warehouses.

By the numbers:

• Meta claims V-JEPA 2 is 30x faster than Nvidia's Cosmos

• Meta's robots achieved 65-80% success rates in lab pick-and-place tasks

• Figure AI's robots work 4x faster and 7x more accurately than humans at BMW's plant

• Chinese firm UBTECH has over 500 orders for its Walker S1 humanoid

Context: The release coincides with Meta's $14.8 billion Scale AI investment and broader push into physical AI as the robotics market consolidates around vertically integrated players who control both hardware and software.

Meta's V-JEPA 2 represents a calculated bet on the future structure of the robotics industry—and it might be the smartest play available to a software giant watching hardware companies build integrated empires.

While Figure AI and Boston Dynamics pursue vertical integration, Meta is building the infrastructure layer that could power thousands of smaller robotics companies who can't afford to develop world-class AI in-house. The open-source release of V-JEPA 2 isn't admission of weakness—it's ecosystem strategy. Meta learned from Android that giving away the software can capture more value than selling hardware.

The video-based training approach is genuinely revolutionary. Teaching machines physical intuition through observation rather than trial-and-error could dramatically reduce the astronomical training costs that currently limit robotics to well-funded giants. If V-JEPA 2 delivers on its promise of "zero-shot robot planning," it becomes the great equalizer that lets startups compete with billion-dollar hardware companies.

🤔 Your thought process:

Selected Image 1 (Left):

• “This has songs loaded and the other doesn't have a mixer or laptop to control the music.”

• “Ironically, it was the little open browser window showing part of an error message that made me think it was real. Also, the window sliders matched the portion of the message that was showing. Overall, it just seemed unlikely choices for a current AI model to make.”

Selected Image 2 (Right):

• “Whoa, great human with realistic hands! Nice work AI”

• “Wow, I can't seem to guess correctly. This image looks pretty real!”

💭 Poll Results

Here’s your view on “Do you think Meta is losing the AI race?”:

Yes (61%):

• “Meta is losing - for now. Love him or hate him, Zuck knows how to surround himself with the people he needs to get what he wants, so I wouldn't count them out by a long shot.”

• “They should have doubled, tripled and quadrupled down on the AR aspect of wearables while they were in front in that area, integrated the wearables deeply with their social apps based on "know everything about you" data, and used that as the spearhead to expand into deeper areas.”

• “Throwing mega bucks attempting to lure the brains of AI to work for him could be counterproductive. Remember, people work for a number of reasons not all good. What will Meta expect for mega wages from these individuals? Not a good look when poaching workers from competitors, better off seeking graduates and training them up.”

No (13%)

Not sure / need more info (26%)

Thanks for reading today’s edition of The Deep View!

We’ll see you in the next one.

P.S. Enjoyed reading? Take The Deep View with you on the go! We’ve got exclusive, in-depth interviews for you on The Deep View: Conversations podcast every Tuesday morning. Subscribe here!

If you want to get in front of an audience of 450,000+ developers, business leaders and tech enthusiasts, get in touch with us here.