Anthropic’s new study reveals Claude's morals and ethics

Today’s Highlights:

📰 News: Anthropic’s new study reveals Claude's morals and ethics + OpenAI’s o3 keeps hallucinating

💰 Funding: Perplexity raises $1Bn as it signs partnerships w/ Motorola and Samsung

⚡️ Top News Stories:

1. In a landmark study analyzing 700,000 Claude interactions, Anthropic found that its AI not only reflects core values like honesty, safety, and empathy across diverse topics but also adapts, defends, or resists values depending on context — with rare but revealing anomalies that underscore the need for robust post-deployment value monitoring in AI systems.

2. Profluent released ProGen3, a family of AI models trained on an extensive dataset of over 3.4 billion protein sequences, which has demonstrated that scaling laws in AI apply to biology by successfully designing compact gene-editing tools and high-performing, patent-safe antibodies — potentially transforming therapeutic discovery into a faster, engineering-driven process.

3. Meta’s FAIR lab unveiled five open-source AI tools including a visual perception model, 3D spatial understanding system, and a multi-agent reasoning framework — key advancements toward AI systems that can perceive, interpret, and cooperate in real-world environments.

4. Google's release of QAT-optimized Gemma 3 models significantly reduces memory requirements, enabling high-performance AI deployment on consumer-grade GPUs and broadening access for developers and researchers.

5. Famed AI researcher Tamay Besiroglu has launched Mechanize, a provocative startup aiming to fully automate all human labor through AI agents, drawing backlash over ethical concerns and potential conflicts with his nonprofit Epoch, while also attracting elite investor support and arguing the move will lead to global prosperity.

6. Transluce's investigation into OpenAI's o3 model reveals a concerning pattern of fabricated code execution claims, where the model provides detailed but false justifications for actions it cannot perform, highlighting significant truthfulness issues in AI reasoning models.

7. OpenAI’s flagship o3 model, once hailed for achieving over 25% on a rigorous math benchmark, has been revealed by independent tests to score closer to 10% — a discrepancy traced to differences in model versions and compute, highlighting industry-wide concerns about benchmark transparency and marketing practices.

8. OpenAI’s new o3 and o4-mini reasoning models show superior performance in coding and math tasks but hallucinate significantly more than older models — with o4-mini hallucinating nearly half the time on key benchmarks — raising concerns about their trustworthiness and prompting OpenAI to admit it doesn’t yet understand the cause.

9. In a paradigm-shifting paper, DeepMind’s David Silver and Richard Sutton propose evolving AI beyond human-limited training data by enabling agents to learn continuously from real-world feedback like health outcomes and exam results — a method they call “streams” — potentially unlocking superhuman intelligence shaped by environmental experience rather than static human input.

10. Researchers from MIT and partner institutions have developed a technique using sequential Monte Carlo methods to guide LLMs toward code outputs that are both structurally valid and semantically accurate, allowing small models to outperform much larger ones in coding, molecular, and robotic tasks.

11. TSMC warns that despite its efforts to comply with U.S. export restrictions, it cannot fully prevent its AI chips from ending up in restricted Chinese products, posing growing risks of sanctions, reputational damage, and legal penalties amid escalating trade tensions.

12. The Oscars have formally accepted the use of generative A.I. in film without disqualifying eligibility, but emphasized that human authorship remains central to awards decisions, as the Academy navigates growing creative, ethical, and industry tensions over the technology’s influence.

13. As AI labs increasingly lean on crowdsourced platforms like Chatbot Arena to showcase model performance, experts warn that these benchmarks lack rigor, are vulnerable to manipulation, and should be supplemented by compensated, professional evaluations tailored to real-world use cases.

💰 Top Funding News:

1. Perplexity AI is expanding its footprint in the mobile assistant market with a confirmed integration into Motorola’s upcoming Razr phones and early talks with Samsung, part of a broader strategy to challenge ChatGPT and Gemini while raising up to $1Bn at an $18Bn valuation.

2. Manychat, which uses AI to automate and personalize business messaging across social platforms, raised a $140M Series B led by Summit Partners.

3. Lace AI, which uses AI to analyze customer service calls for home service companies to uncover missed revenue opportunities, raised a $14M Seed round led by Bek Ventures, with participation from Canvas Ventures, Horizon VC, Launchub, and angel investors Marcin Zukowski and Heini Zachariassen.

4. Cluely, which uses AI to covertly assist users during interviews, exams, and calls via a hidden in-browser assistant, raised a $5.3M Seed round led by Abstract Ventures and Susa Ventures.

That's all for today's email! If you want more please follow us at the social channels linked below, or check out our website!

Share our newsletter: If you like our work please share/forward this email with your friends, colleagues, and family. It's the best way to support us!

If this email was forwarded to you please sign up here to continue receiving them.

Want your content, product, jobs, or event featured in our newsletter? Reply to this email with the details, and our team will reach out to you.

Do you use AI for work? Tell us how, and you could be featured in our newsletter!

Check out our website for more resources, including a list of AI investors, products, events, and twitter follows.

For an archive of all our posts, click here.

We'd love to hear from you! You can always leave us comments or feedback by replying to this email!

Powered by AI. Curated and edited by Humans.