CULTURE

AI has a copyright problem

Copyright and intellectual property remain major hurdles in embedding AI in creative spaces. Adobe might be looking to fix it. 

The company on Monday launched AI Foundry, a program that works with enterprises to build custom AI models trained specifically on their branding and IP. The models are capable of producing text, images, video, 3D renderings and more. 

Adobe launched Firefly, its family of image models, in 2023, with the models’ biggest selling point being that they are trained on licensed data and safe for commercial use. 

With Foundry, offering the ability to fine-tune these models stands to make Firefly’s offering a lot more useful. Hannah Elsakr, VP of generative AI new business ventures at Adobe, told TechCrunch that the Foundry offering is “elevating a lot of the capabilities we already had.”

Adobe’s offering seeks to reckon with the massive copyright issue plaguing developers of creative AI and those seeking to use it. Several model developers have already faced legal run-ins as a result, with major firms like Anthropic, OpenAI, Meta and Perplexity among them.

An offering like Foundry may also particularly strike a chord as AI video tools like OpenAI’s Sora, Google Veo, and xAI Imagine gain rapid popularity – and face backlash from the entertainment industry due to the copyright complications at stake.

Copyright might not be the only problem that crops up with AI in creative spaces. Adobe is widely known as an industry standard for things like photo and video editing, and its dedication to AI might signal where the creative and entertainment industry are headed. However, a lot of people within creative industries have ethical qualms about AI’s use to create art generally. While the pitch for AI Foundry as it stands is targeted towards enterprises, it represents a push to embed generative AI into a process that has long been essentially human: the creation of art.

TOGETHER WITH QA WOLF

✂️ Cut your QA cycles down from hours to minutes

If slow QA processes and flaky tests are a bottleneck for your engineering team, you need QA Wolf.

QA Wolf's AI-native service supports both web and mobile apps, delivering 80% automated test coverage in weeks and helping teams ship 5x faster by reducing QA cycles to minutes.

With QA Wolf, you get:

✅ Unlimited parallel test runs

✅ 15-min QA cycles

✅ 24-hour maintenance and on-demand test creation

✅ Zero-flake guarantee

The result? Drata’s team of 80+ engineers saw 4x more test cases and 86% faster QA cycles.

No flakes, no delays, just better QA — that’s QA Wolf.

Schedule a demo to learn more

RESEARCH

Anthropic’s takes on drug discovery with Claude

Anthropic is going all-in on the $2 trillion biotech market.

On Monday, the AI safety and research startup launched Claude for Life Sciences, a new capability designed to support scientists working on drug discovery: one of the most complex and costly processes in research, CNBC reported.

Part of the Claude model family, the tool is built to automate time-consuming tasks like hypothesis generation, literature review and regulatory drafting. The goal is to reduce the busywork that slows down early-stage research and development and, in turn, ship potentially life saving treatments to market faster. 

It’s a timely, and arguably Anthropic’s most practical use of AI yet. Developing a single drug can take 10 to 15 years and cost $2.6 billion. Claude isn’t running lab experiments or clinical trials, but with its new feature, Anthropic believes it can streamline the paperwork and data handling that often bog down progress.

“We want a meaningful percentage of all life-science work in the world to run on Claude,” Eric Kauderer-Abrams, head of Anthropic’s biology and life sciences division, told CNBC. 

The launch comes two months after Anthropic hired Kauderer-Abrams to lead its biology and life science division. Since then, Anthropic has partnered with platforms like Benchling, which manages lab data for over 200,000 researchers, and 10x Genomics, known for its genomic sequencing tools, to integrate their tools into Claude. 

In a demo shared with CNBC, a scientist used Claude to query experimental data in Benchling, generate summary tables, and compile a report — cutting a multi-day process down to minutes.

Anthropic isn’t alone in targeting the drug discovery space. Last week, Google DeepMind released Cell2Sentence-Scale 27B, a model developed with Yale to help uncover cancer cell behaviors. OpenAI, as well as startups like Recursion and Cradle, are also using AI to analyze proteins and extract insights from lab data.

Kauderer-Abrams is under no illusions that Claude will solve biology’s hardest problems overnight. But he sees it as a tool that could give scientists a faster path to the next big breakthrough.

TOGETHER WITH FIDDLER

📊 The 2025 Agentic Observability Guide

The Fiddler Agentic Observability Guide breaks down exactly what’s needed to monitor and control complex, multi-agent systems in production.

Learn how to:

• Pinpoint issues across the agentic hierarchy

• Visualize aggregated metrics and granular insights

• Gain context on agent decision paths and chains of logic

Unlock visibility and control for your AI agents — download the guide today. 

Enterprises are racing to harness the potential of AI agents to drive faster decisions and deepen automation. But as these systems move from proof-of-concept to production, a critical truth becomes clear: you can’t trust what you can’t see. If you’re ready to take control of your AI agents - Download the Agentic Observability Guide Today.

RESEARCH

DeepSeek OCR could make AI cheaper, more efficient

DeepSeek is giving its models more to do. 

The Chinese AI firm has developed a new OCR, or optical character recognition system, that’s capable of handling a lot more text with far fewer resources, according to a paper published by DeepSeek researchers this weekend. Both the code and the trained model of this proof-of-concept are currently publicly available.

DeepSeek’s model parses text as images, requiring less computing power than typical models that only handle digital text. The OCR is capable of handling much longer documents without running into memory limits, achieving more than 96% precision in keeping the original meaning and detail of the text when that text is compressed by up to 10 times. 

“A single image containing document text can represent rich information using substantially fewer tokens than the equivalent digital text,” the researchers noted. The paper notes that vision encoders can potentially enhance large language models’ efficiency in processing text. 

The model was trained on 30 million pages of PDFs. DeepSeek’s system can handle several kinds of documents, including diagrams and formulas, and supports 100 languages, and can process data extremely quickly. With just one GPU, the OCR can process 200,000 pages per day. A whole cluster, meanwhile, can handle up to 33 million documents a day. 

DeepSeek’s discovery matters for a few reasons. For one, it tackles the problem of context windows, or the amount of information that a model can process at once, Bob Rogers, chief product and technology officer of Oii.ai and co-founder of BeeKeeper AI, told The Deep View. The larger the context window, the more expensive the query. DeepSeek’s system reduces the size of that context window without reducing the amount of information processed. 

Additionally, the fact that this tech is open source is “just a great bonus,” Rogers noted. It signals another discovery in developing more cost-effective AI for the Chinese open source company, which sits as a competitor to AI giants like Google, OpenAI and Anthropic.

“It’s helping us stay out of AI services prison,” said Rogers. “We're starting to use these AI services, and we're getting pretty dependent on them. Having the ability to do real things at scale with more locally deployable models is actually like a pressure release valve for that.”

LINKS

• Unitree releases H2 Destiny Awakening, it’s latest humanoid robot 

• Anthropic reportedly spent $2.66 billion on AWS in the first three quarters of 202f

• Langchain raises $125 million in fresh funding at a $1.25 billion valuation

• Medical chatbot OpenEvidence raises $200 million 

• South Korea rolls back AI textbook program after backlash from teachers, parents

• Enterprises should take a more modest approach to AI, Asana cofounder says

• Hackers claim to have personal data of NSA, government officials

• Andrej Karpathy discusses timeline for AGI

• Claude Code on Web: Anthropic has released its AI coding assistant on browsers. 

• VibeOnly: A tool for finding and hiring candidates with “elite” AI skills.  

• Fish Audio S1: Expressive and emotionally-rich text to speech for realistic voice cloning.

• Playabl.ai: An AI-powered game development platform for in-game assets across titles. 

• Nora: An AI coding tool for developing Web3 applications.

• OpenAI: Research Engineer

• Anthropic: Privacy Research Engineer, Safeguards

• Meta: AI Research Scientist, CoreML - Monetization

• Zoom: AI Scientist - Agentic Verticals

GAMES

POLL RESULTS

What will power the next wave of AI inference?

• GPUs (18%)

• LPUs (like Groq's) (20%)

• Hybrid chips (23%)

• Cloud optimization (15%)

• Too early to tell (24%)