Welcome back. Microsoft and OpenAI are locked in a bitter contract dispute over artificial general intelligence—a milestone that "everyone laughed at" in 2019 but now threatens their $10 billion partnership. OpenAI's Sam Altman claims they're close to AGI (which would cut off Microsoft's access), while Microsoft's Satya Nadella calls it "nonsensical benchmark hacking." The companies are now considering moving the goalpost to "artificial superintelligence" instead, proving that when you're fighting over world-changing AI, even the impossible becomes negotiable.

In today’s newsletter:

• 🧬 AI for Good: AlphaGenome reads DNA like a scientist-in-a-box

• 📚 Meta wins AI copyright case, following Anthropic’s victory

• 🈸 Claude apps now let anyone build and share AI tools instantly

🧬 AI for Good: AlphaGenome reads DNA like a scientist-in-a-box

Google DeepMind has unveiled AlphaGenome, an AI model that predicts how small changes in DNA affect molecular processes. The system helps researchers anticipate shifts in gene activity from single-letter genetic variations, a task that usually requires expensive, time-consuming lab work.

What researchers built: AlphaGenome uses transformer architecture — the same underlying system found in large language models — but trained on genomic data from public scientific projects. Unlike previous models that analyzed short DNA fragments, AlphaGenome can process sequences up to one million DNA letters and make thousands of predictions about biological properties.

• The model achieved state-of-the-art performance across genomic prediction benchmarks, outperforming existing tools in 22 out of 24 sequence prediction tests.

• In one case study, researchers applied AlphaGenome to mutations in leukemia patients and accurately predicted that non-coding mutations indirectly activated a nearby cancer-driving gene.

• Training the entire model took just four hours using Google's custom processors — half the computational budget of previous models.

DeepMind is making AlphaGenome available for non-commercial research through an API, with plans to explore commercial licensing for biotech companies.

Why this matters: Most people with rare diseases never learn what's causing them. Even when a genome is fully sequenced, doctors often don't know which mutation to focus on. AlphaGenome could help narrow that search by virtually testing thousands of genetic variants, potentially speeding diagnosis and drug discovery without requiring physical lab experiments.

Integrate the leading open source models: Llama, DeepSeek, OpenAI, and more

SambaNova Cloud lets you build with the best open source models from Llama, DeepSeek, OpenAI and more, all optimized with the fastest inference speed, powered by their new chip the RDU.

They have integrations to the most popular frameworks and tools (like Hugging Face, CrewAI, and Cline), APIs for building, and a playground to test things out. 

Get started with SambaNova Cloud today and unlock the full potential of the best open source models (also available through the AWS Marketplace).

📚 Meta wins AI copyright case, following Anthropic’s victory

A federal judge dismissed a copyright lawsuit against Meta's Llama AI training, but the victory came with significant caveats. Unlike this week's Anthropic ruling that embraced fair use, Judge Vince Chhabria ruled that the 13 authors simply "made the wrong arguments" rather than endorsing Meta's practices.

The details: Judge Chhabria ruled that plaintiffs — including Sarah Silverman, Ta-Nehisi Coates and Jacqueline Woodson — failed to present compelling evidence of market harm from Meta's training methods. The authors had argued that Meta used their books from pirated online repositories without permission to train Llama.

• Chhabria explicitly stated that "This ruling does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful," distinguishing it from broader validation of AI training practices.

• The judge criticized this week's Anthropic decision, arguing that Judge Alsup "focused heavily on the transformative nature of generative AI while brushing aside concerns about the harm it can inflict on the market."

• The ruling only affects these 13 authors, not "the countless others whose works Meta used to train its models," Chhabria noted.

Unlike the Anthropic case, which addressed fair use doctrine directly, Meta's victory was procedural. Chhabria said that while he had "no choice" but to grant Meta's summary judgment, the consequences are limited since this wasn't a class action.

A separate claim alleging that Meta illegally distributed copyrighted works via torrenting remains pending. The judge also suggested that stronger market harm arguments could succeed in future cases.

Why it matters: While Meta avoided immediate liability, this ruling provides far less legal protection than Anthropic received. Chhabria explicitly left the door open for other authors to bring similar lawsuits with better legal strategies. As the judge noted, the decision doesn't validate Meta's training methods — it simply found that these particular plaintiffs failed to make their case effectively.

🐋 DeepSeek R1-0528 + 🎨 FLUX.1 Kontext now on Together!

Build faster with the latest open models—DeepSeek for smarter reasoning & coding, and FLUX.1 for image editing via natural language.

🔌 OpenAI-compatible API, no setup headaches.

💬 Try them free at chat.together.ai

• Silicon Valley’s ‘tiny team’ era is here

• TSMC is issuing $10B in stock to handle currency swings

• Gemini gets a command-line interface for terminal users

• Nvidia steps into cloud, shaking up other tech giants

• Ring camera alerts are getting a major AI upgrade

• Rubrik is buying AI startup Predibase for $100M+

• The little-known startup that has surged past Scale AI—without any investors

• Dell: AI Solutions Product Manager

• Hugging Face: Machine Learning Engineer, WebML

• Krisp: Removes background noise, echoes and voices from video and audio calls

• Papercup: Dubs videos into other languages using AI voice translation

• Greptile: Automated AI code reviews to catch bugs, anti-patterns and security issues

🈸 Claude apps now let anyone build and share AI tools instantly

Anthropic just launched a feature that transforms Claude into a platform for creating, hosting and sharing AI-powered apps, but without any coding, hosting, or deployment required. Users can now build interactive applications that embed Claude's intelligence directly, then share them instantly with a simple publish button.

Since launching Artifacts last year, users have created over 500 million artifacts — from productivity tools to educational games. Now Anthropic has added a dedicated artifacts space accessible via the sidebar, plus the ability to embed AI capabilities directly into creations.

The details: The new system removes traditional barriers to AI app development. Instead of managing API keys, hosting infrastructure, or usage costs, developers can build functional AI applications entirely within Claude.

• Zero-friction sharing: Apps are published with a single click and shared via link. Anyone with a Claude account can access them immediately.

• Built-in AI capabilities: Unlike static artifacts, these apps can answer questions, generate content, and adapt responses based on user input through Claude's API.

• Collaborative remixing: When users access shared apps, they can modify them by talking to Claude, creating their own variations without affecting the original.

Early applications include AI-powered games with NPCs that remember conversations, learning tools that adjust to individual skill levels, and data analysis apps where users upload files and ask follow-up questions in natural language.

The feature is available to Free, Pro, and Max users.

Examples in the wild: The breadth of user creations suggests the platform's potential. Users have built everything from personalized tutoring systems to a one-page Product Requirements Document maker. Music producer Rick Rubin even created a poetic introduction to visual coding that users can modify through prompts.

The system handles complex orchestration behind the scenes. Claude writes production-ready code that manages AI functionality, user authentication, and API calls — all visible and modifiable by users. The economics work because API usage is tied to end users' Claude plans, not the app creator's costs.

The platform includes sophisticated debugging capabilities. When something breaks, users can either click "Fix with Claude" or describe problems in plain language. Version control is automatic — every iteration creates a new artifact version, and users can experiment with forking by editing previous messages to try different approaches without losing original work.

Anthropic just turned AI development into a conversation. By removing technical barriers and hosting costs, they've created the first truly accessible AI app platform, one where anyone can build and deploy intelligent software by simply describing what they want.

Since users authenticate with their own Claude accounts, creators don't face scaling costs that have historically limited AI app sharing. This could enable viral AI applications without traditional infrastructure concerns.

But this is also Anthropic's platform play. Apps are completely tethered to Claude's ecosystem — no exports, no migrations, no escape hatches. Creators get convenience in exchange for total dependency. It's the classic build-versus-buy trade-off, except Anthropic made "buy" so frictionless that most probably won't realize they're locked in until they want to leave.

Current limitations around persistent storage and external APIs restrict enterprise use cases. But the velocity and simplicity here represent something bigger: a preview of software development where the distinction between using AI and programming with it disappears entirely.

🤔 Your thought process:

Selected Image 1 (Left):

• “Image one looks like a science fiction landscape. The way the moon looks in this image won it for me. Tiny and blurry - totally genuine and a fine detail I would be surprised if AI wouldn't try to ‘perfect’.”

• “Nothing stood out as obviously fake to me in the other photo, but this stood out as obviously real in a way that (I don’t think) AI has mastered yet. The way each blade on the wind turbines sat at different angles and the way the edge of the half moon slightly fades into the sky shows a kind simple delicacy that I believe AI still struggles with. I’m sure that will change soon, though.”

Selected Image 2 (Right):

• “AI did better with shadows than mother nature on this one. Had me fooled”

• “Should have trusted my gut. I thought it looked a bit more washed out.”

💭 Poll results

Here’s your view on “Should AI companies be allowed to train on copyrighted books only if they obtain the texts legally?”:

Yes — fair use depends on legal copies (55%):

• “If I become a best selling author that success would be due to the books I've read. Some feel that if A.I. does the same thing, that's a violation of copyrights? Doesn't seem logical, so "training" AI software by "reading" a lot of books very fast is okay when AI buys books legally and reads them to learn.”

• “An answer “yes“ to this question is intuitively obvious.”

No — any training needs licensing regardless (28%):

• “I think there needs to be some middle option that I'm not aware of. If big companies like Disney or McDonald's can sue if you use their products/ licensing without permission in a way that they have not agreed to, but parody, satire, copyright expiration into public domain, and creative original works based on their material is allowed, then surely the same protections should be given to a regular writer or artist.”

• “For profit use should demand licensing as this is the author's "product"”

Other (17%):

• “Large language models perform better if the data is validated and continuously tested for accuracy. ”

The Deep View is written by Faris Kojok, Chris Bibey and The Deep View crew. Please reply with any feedback.

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!

P.P.S. 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.