GOVERNANCE

Demis Hassabis offers useful AI safety system

While others in the AI industry have issued general calls to prioritize AI safety, Google DeepMind's Demis Hassabis has proposed an actual system. 

On Tuesday, Hassabis published a 1,500-word article on Twitter titled "A Framework for Frontier AI and the Dawning of a New Age" that makes a series of recommendations for how the industry could establish standards and appoint an organization to oversee them. 

"The rapid progress we’re seeing in AI requires a new approach to testing frontier AI model capabilities that is dynamic, adaptable, and rigorous," Hassabis wrote. "The US is well positioned, given its economic and technical standing, to take the first step in developing such a framework."

He recommended that the industry create a non-governmental organization similar to the Financial Industry Regulatory Authority (FINRA), which oversees and regulates stock brokers and financial advisors in the US. 

It's pragmatic that the UK-based Hassabis recommended the effort begin in the US, where the three biggest frontier AI labs, Google, OpenAI, and Anthropic, are primarily headquartered. 

In his essay, the 650-word section "A Framework for a Frontier AI Standards Body" is where he lays out his main points. Here are his recommendations:

• Create an independent standards body: Establish a US-led, federally overseen public-private organization, run by leading technical experts, to develop, update, and enforce evaluations of frontier AI models.

• Require rigorous testing before deployment: Ahead of release, frontier models should have mandatory evaluations for cybersecurity, biological threats, deception, agentic capabilities, and national security issues.

• Raise the bar for frontier labs: Based on new standards, makers of the most capable models should publish technical documentation, invest substantially in AI safety, vet key hires, maintain strong cybersecurity, and address vulnerabilities systematically.

• Build an adaptable international framework: Encourage global cooperation on AI safety, establish independent auditors, and retain the ability to coordinate a slowdown in frontier AI development (via pre-deployment approvals), if the risks become too severe.

Hassabis also acknowledges that such an effort would need to be resourced appropriately in order to succeed, and so he divorces those resources from the ups and downs of government and politics. "Funding would need to be substantial and likely mostly come from industry, in order to attract world-class technical talent and provide the necessary compute resources for large-scale testing," he said.

Chris Canal, CEO of Equistamp, an AI safety evaluator of frontier models, told The Deep View his perspective on what's needed for such an organization to succeed. "The entire job of a standards body is keeping the barrier to mass harm high enough that society survives its worst actors," said Canal. "The body should publish predictions of what the next generation of models will score before release, and be graded publicly on its accuracy. A testing regime that can't predict is [only] measuring the past."

I applaud Hassabis for the specificity of his recommendations, even if his individual proposals get dissected and transformed by others. In fact, especially if they start a dialogue and then get dissected and transformed by others. There's been too much fear mongering about AI in 2026 and not enough people putting forward actual proposals and being willing to have them put under the microscope. I said something similar in April when OpenAI researchers put forward their ideas for an AI New Deal and in May when Pope Leo XIV released his AI manifesto. The world needs creative thinking and bold ideas to start more public dialogue around the future of AI. If we want to see AI take a different trajectory, then it's up to all of us to engage in the dialogue. Ping me on Twitter to let me know your thoughts.

TOGETHER WITH CREATIO

Creatio: The CRM Platform Of The Future

Since their inception, CRMs have been a game changer for businesses. But in a world that’s always asking “now what?”, it’s what’s next that matters – and right now, that’s Creatio. 

Creatio is an AI CRM and Workflow Platform where people and AI agents work together. Trusted by thousands of organizations in over 100 countries, it combines best-in-class CRM with an AI-native, no-code platform built for the AI era. With no limits on users, AI agents, workflows, or scale, Creatio gives organizations the freedom to automate, innovate, and grow—without the constraints of traditional enterprise software.

Try Creatio right here and put the power of AI CRM to work.

RESEARCH

Why Claude’s moral compass keeps shifting

If you ask a person for advice, their answer will ultimately reflect their moral compass. Anthropic wants the same to be true of Claude. 

Anthropic has given Claude careful instructions on the values it would like the AI to reflect in its answers, including "good judgment and sound values” that can be applied contextually, as written in the Claude constitution. Yet, the company acknowledges that it’d be difficult for that document to anticipate the values reflected in the millions of answers Claude generates. As a result, on Monday, the AI lab released a report that unpacks the values Claude expresses. 

In particular, the report looks at how the values Claude expresses vary across models and languages. While prior research focused on the range of values Claude expressed, which included more than 3,000 distinct values, this study took a more narrow approach, tracking four different axes that could encapsulate those distinct values: 

1. Deference vs. Caution

2. Warmth vs. Rigor

3. Depth vs. Brevity

4. Candor vs. Execution

The analysis found that the four key axes captured 15% of variation in Claude’s values. The results reflected the character of the model; for instance, Sonnet 4.6 is characterized by warmth, and Opus 4.7 by rigor, which came through in the axes: the former leaned towards emotional warmth, and the latter towards deference. On the flip side, it also meant Opus 4.7 leaned more towards accuracy,

A similar discrepancy in results was seen across languages, with the largest variation being in warmth vs. rigor, with Claude leaning towards expressing warmth-related values in Arabic and Hindi and rigor-related ones in English and Russian. 

To learn more about what the differences were in the axes across model and language, you can dig into the report. The most important takeaway is that Claude’s values do vary, highlighting that AI answers should not be taken as objectively true, since they can introduce slight biases and, as a result, yield different levels of quality and accuracy. 

Anthropic says this research doesn’t examine the impacts these discrepancies have on users, and as a result, those impacts remain unknown. However, at the scale at which people are using these models, it is difficult not to consider how slight nuances in answers could have long-tail impacts on behavior. This is especially true for sensitive use cases such as mental health and relationship advice, where a slight difference in tone can make a big difference in delivery, impact, and helpfulness. It also raises the question of who is most qualified to instill these values, which will be dispersed across millions of conversations every day.

TOGETHER WITH GENERAL ASSEMBLY, AN LHH BRAND

93% of leaders encourage AI use. 27% don't use it for executive work.

Your team already turns to AI for the simple things: drafting emails, summarizing documents, pulling together a report. Fewer leaders bring it into the room when the stakes are higher, like restructuring a department, greenlighting a budget, or deciding what the org focuses on next year.

General Assembly surveyed 524 VP+ leaders at companies with 100+ employees and found exactly how wide that split is: 93% encourage daily AI use, but only 27% apply it to scenario planning, org design, or financial modeling.

GA's AI For Leaders training closes that gap, building your capability to frame AI problems, design solutions, and align AI initiatives to long-term business goals.

See how the training works.

PRODUCTS

Gemini, ElevenLabs top new Voice AI benchmark

One of AI's big promises is making your workflows faster, and talking to your computer is one of the big unlocks.

Voice AI models are booming because they let people interact with AI more seamlessly. As a result, minor latency, word errors, and misinterpretation of emotional cues, speech or accents, can hinder the experience greatly. Yet, Hume found that traditional benchmarks only accounted for technicalities rather than actual human interaction, so it built a new benchmark, "Real World VoiceEQ," to evaluate these models' performance more accurately. 

"Our team has spent more than a decade researching human expression and emotional intelligence, and Hume was founded to make voice AI more emotionally intelligent," Andrew Ettinger, Hume's CEO, told The Deep View. "We saw this benchmark as a way to share what we’ve learned and contribute to the field’s progress."

The benchmark goes beyond technical correctness, also evaluating the models' ability to understand emotional cues, communicate naturally, and be reliable in real-world conversations. It does so by evaluating more than 40 voice models across more than 15 key evaluation dimensions and more than 60 metrics, and is grounded in 700,000 human judgments, making it the largest human evaluation of voice AI to date, according to the release. Every evaluation is also conducted using Kairos, Hume's voice-native evaluation platform. 

Contrary to what you may think, the purpose of the benchmarks isn't to find the "best" voice model, but rather to identify the strengths and weaknesses that make certain models better suited to different needs. Here is where the top models stand in the benchmark: 

• Text-to-Speech: Google Gemini achieved the strongest performance 

• Speech-to-Speech: OpenAI's GPT Realtime Mini achieved the strongest performance 

• Automatic Speech Recognition: ElevenLabs emerged as the strongest overall ASR system

• Speech Understanding: Google Gemini led the Speech Understanding category overall

"Our results show there isn’t one 'best' voice model. Different systems have different strengths and trade-offs," said Ettinger. "A good leaderboard helps companies choose based on what they actually need, whether that’s emotional intelligence, precise transcription, expressive speech, or reliability in noisy environments."

Hume found that current models are being fine-tuned to pass and perform well on popular benchmarks, without those capabilities necessarily translating to better performance in real-world use cases. This is a major issue that impacts other AI applications as well. For instance, at launch, many models achieve high benchmark performance in certain physics applications. Yet the models will fail to answer simple questions accurately, highlighting that benchmarks don't always translate to everyday use cases. More companies should adopt an approach of creating benchmarks that are fine-tuned to practical needs, rather than just sounding impressive based on pure data. The trap is that those impressive stats are what get power users, enterprises, and investors to buy in.

LINKS

• Spotify releases features that lets users learn more about a song, make requests, etc, by talking to it 

• New York Governor Kathy Hocil signs moratorium to block new environmental permits for data centers

• Anthropic gives teachers free access to premium Claude features 

• OpenAI’s ad business falls short of its own revenue forecast by 90%, report

• Kalshi and OpenAI strike a partnership to use prediction market data in some soccer tournament search results 

• AI startup Reflection AI signs $1 billion compute deal with Nebius

• Codex and ChatGPT: Usage limits were rest again for all 

• Manus: PowerPoints Slides on Manus

• Wandr: Perplexity’s internal benchmark is now open sourced

• PrismML: Announced Bonsai 27B-class model 

• Google: Users can create AI images right in Search 

• Paces: AI Engineer 

• Morgan Stanley: AI/ML Engineer

• Alpaca: Support AI analyst 

• Fanatics: AI Engineer

GAMES

A QUICK POLL BEFORE YOU GO