CULTURE

New tech is coming to tackle the AI slop crisis

It's getting harder to discern authentic content from stuff that's AI-generated, but there may be hope. 

Some tech companies are starting to draw a line in the sand. One is Anthropic, which has updated its Claude models to watermark all generated text to comply with transparency regulations in the EU AI Act. The act, which requires AI companies to identify AI-generated and edited content, has also received commitment from companies such OpenAI, Microsoft, Google, Cohere, Meta and Synthesia. 

As such, Anthropic noted in an update to its support page that all models released after August 2 will automatically watermark both text and files that they generate or edit, with files using the C2PA open standard. 

Eventually, the company will expand the watermarking to older models and add in a system that enables watermarks to follow users when they copy and paste text. The standard will apply to Anthropic's suite of products, including Claude Code, Claude Cowork, and Claude Tag.

And it's not the only company that's getting serious about identifying machine-made content.

• Spotify announced on Monday that it's rolling out "AI Personas," or badges that identify when an artist "does not represent a real person," both for artists that self-disclose this information and for "relevant profiles" that do not. Additionally, Spotify will not recommend these artists in user algorithms unless users follow them.  

• AI music platform Suno, meanwhile, is implementing a new "watermarking and fingerprinting technology" as part of a suite of transparency tools, as well as changing its download policy to fight spam and fraud. 

• And newsletter distributor Substack has integrated AI detection tool Pangram into its platform, allowing users to see how much of a blog post has been written with AI. 

This shift towards transparency comes at a time when an increasing number of people are unable to detect when content is AI-generated. A study from music streamer Deezer found that 97% of listeners aren't able to tell the difference between authentic and AI-generated content. And the lines are blurry with images, too: On average, around 50% of readers of The Deep View are unable to spot AI-generated images in our daily "AI or Not" game at the bottom of the newsletter.

Whether it's writing, photography, art, or music, AI has presented a massive existential question mark for creatives. Watermarking and transparency measures represent a significant step in the right direction for protecting the sanctity of creation, both in the consumer's right to know when content wasn't created by hand and for those trying to make a living as creators. Still, the problem at hand is similar to the one facing AI-powered cybersecurity: When presented with a 10-foot wall, bad actors will always come prepared with a 12-foot ladder, whether they're climbing over cyber defenses or watermarking tools. Additionally, a bigger question remains. Though consumers currently are generally against AI-generated content, especially when that content purports to be human-made, will consumers eventually stop caring? 

TOGETHER WITH RAMP

Choose Your AI Tool in Excel: Claude Vs. ChatGPT vs. Copilot

Claude, Copilot, and ChatGPT are the three leading AI tools that are reshaping finance work.

But how do you know which tool you should be using for certain tasks in Excel, PDFs, and other Microsoft Office applications?

As part of the Fast Track for AI Summer Series, Kenji Farre, Co-founder of Career Principles and creator of “Kenji Explains” - a YouTube channel with over 950k+ subscribers, is hosting a live, hands-on session on August 17th, comparing the three tools inside Excel across the same set of everyday finance scenarios, so you’ll know which tool to reach for the job in front of you.

Join him live

ENTERPRISE

Why enterprise AI is shifting from rent to own

As intelligence becomes your organization's most valuable asset, the question is whether it makes sense to rent it from another company.

Oumi, a startup run by a former leader on the Google Gemini team, has launched tech that lets you build and own your AI models, deploy them within hours, and use agentic AI to keep them updated and compound their value over time. 

On Tuesday, Oumi released new capabilities that let enterprises deploy specialized AI models, continuously improve them using production data, and manage the full lifecycle of your models from one console. 

• With a single click, enterprises can now deploy their own custom models with production inference that can automatically scale to only use the GPU power they require, saving money and energy

• Organizations can now automatically retrain their custom models using live production data captured from the internal environment 

• A publicly accessible CLI can now be operated by agents and provides developers with full access to training, evaluation, and workflows 

One way to think about this is that the appeal of open models for enterprises is that you can download them and then customize them to exactly what your organization requires. For example, at the Ai4 conference last week, I bumped into Joel Hron, CTO of Thomson Reuters, who explained why the company has built its own foundation model after concluding that it needed to own its own intelligence and not outsource it to frontier labs. The result has been a model that's also much smaller and cheaper to run.

However, that requires expensive AI expertise and time. Even if you can afford to hire AI engineers and convince them to come work for your company instead of a hot AI startup, it can take months or longer to develop your own custom models. 

Oumi is using AI to do what Thomson Reuters and other advanced enterprises have done and make it more accessible to organizations that don't have the same level of in-house expertise. The promise is also that it can happen a lot faster. Using Oumi's web console, you can prompt the tool to build your organization's custom model in a way that's very similar to prompting Claude Code to build you an app. 

"It would [previously] take you weeks and months, and you can do it now literally often in minutes," Manos Koukoumidas, CEO of Oumi and former Google Gemini lead, told The Deep View. "[It's] hours on the clock... but it could be minutes of human effort."

Over the past couple years, frontier labs Anthropic, Google, and OpenAI have convinced a lot of enterprises that AI is going to revolutionize their future. But right now, one of the natural next conclusions is that many enterprises don't want to rent their intelligence from these companies. Instead, they'd rather control it so that they can customize it, better manage on-going costs, and have full control over their most proprietary data. That's a big part of what's driving the rising interest and momentum around open models. But it's also expensive and time-consuming up-front to customize open models for your enterprise. And then there's the on-going cost of retraining your custom model to keep it up to date. That's where Oumi has stepped up to offer a platform that makes building, deploying, and updating your own custom models far more streamlined. 

TOGETHER WITH QUIQ

Most CX teams aren't struggling because the model is too weak

Most CX teams aren't struggling to build AI agents because the model is too weak. They're struggling because customer inquiries are messy, policies have edge cases, and every answer needs the right mix of speed, judgment, and control. That's usually where a strong demo starts to quietly fall apart.

In Quiq's new guide, "CX Agent Design: How to Build One You Can Read and Trust," it breaks down the two dead ends most teams hit when building AI for customer service, the building blocks (Guides, Skills, Tools, and governed boundaries) that replace them, and why a CX agent needs more than good code to actually hold up in production.

Get your free copy of the guide here.

MARKETS

Databricks acquisition unlocks AI coding on mobile

Running a team of agents to build stuff and carry out tasks for you while you touch grass just got a lot better.

On Tuesday, Databricks announced that it acquired Electric, a company that runs a lightweight database that has exploded in popularity because of its usefulness in the era of AI agents.

Technologically, Electric runs PGlite, which is a lightweight version of the popular open-source Postgres database that can run directly on a laptop, tablet, phone, browser, or AI agent sandbox using WebAssembly (WASM). This lets software developers and vibe coders do much faster and more powerful stuff straight from their mobile devices. 

"You can run it in very low‑powered environments," Databricks VP of engineering Nikita Shamgunov told The Deep View. "You can run it on Raspberry Pi. You can run it on your phone. You can run it on your laptop."

PGlite enables two important things:

• By running locally, it can reduce latency, work when you have unreliable connectivity, lower cloud costs, and make agent loops faster by not needing to make a network request every time data is needed; as a result, agents can constantly read information, update their context, call tools, coordinate with other agents, and decide what to do next

• Simultaneously, Databricks and Electric pair PGlite with a real-time synchronization system that keeps those small local databases connected to Databricks' central Lakebase Postgres database; that enables developers and companies to potentially scale up the best apps and deploy them more broadly across the organization

"Before, we were bottlenecked on people typing code, and therefore the rest of the infrastructure had to go at human speed," said Shamgunov. "But now they go at machine speed… If you don't embrace that as company, your competitors will and you'll be left in the dust."

The paradigm shift here is companies being able to safely enable software developers and team members who want to vibe code to work on much more sophisticated projects straight from their phones or tablets. At the same time, their data can be synced to an enterprise database so that the work can be centralized and have the option for wider deployment as well as dissemination of learning across the organization. 

LINKS

• Nvidia reportedly developing Nemotron 4 with 1 trillion parameters

• Google CEO Sundar Pichai said Gemini app now reaches 1 billion users

• IBM, Together AI partner on $240 million deal for Nvidia inference clusters

• Manus is reportedly close to reversing its acquisition by Meta

• OpenAI special projects lead Brad Lightcap to depart for new venture

• AI startup Trajectory raises $40 million at $300 million valuation

• Nemotron 3.5 Lightning: Nvidia's latest open source model sitting at a lightweight 30 billion parameters and can run on a single GPU or laptop. 

• Grok Bot: SpaceXAI has developed an iPhone and Mac app that it calls an "AI teammate." 

• LTX-2.5: The latest world model from LTX, featuring higher pixel fidelity, multishot scenes, and a pretrained foundation built to be finetuned across domains.

• ChatGPT: Now in preview, OpenAI has unveiled the ChatGPT desktop for Linux. 

• Retell: Founders Initiatives, AI Agents

• Gamma: AI Engineer

• Artificial Analysis: Member of Technical Staff

• Walden: AI Engineer

GAMES

A QUICK POLL BEFORE YOU GO