OpenAI releases new o3-mini model along w/ new Deep Research agent for complex research tasks + Anthropic releases system to block jailbreaks

Today’s Highlights:

?? News: OpenAI releases new o3-mini model along w/ new Deep Research agent for complex research tasks + Anthropic releases system to block jailbreaks

?? Funding: AI-powered knowledge graph platform raises $25M

?? Top News Stories:

1. OpenAI has launched o3-mini, a cost-efficient reasoning model optimized for STEM applications, offering faster, more accurate performance while significantly reducing operational costs.

• For the first time, free-tier users can experience AI reasoning capabilities, while ChatGPT Plus, Team, and Pro subscribers receive higher rate limits, with Pro users gaining unlimited access to o3-mini and o3-mini-high.
• o3-mini outperforms the previous o1 model, responding 24% faster, and includes adjustable reasoning effort settings (low, medium, high) to balance speed and accuracy.
• The model operates at 63% lower costs than its predecessor, reducing API pricing to $1.10 per million input tokens, making advanced AI reasoning more affordable.
• This release positions OpenAI competitively against recent AI advancements, while setting the stage for the full o3 model launch in the coming months.

2. OpenAI has launched Deep Research, an advanced ChatGPT agent designed for professionals in finance, science, policy, and engineering, capable of autonomously conducting multi-step research, analyzing online sources, text, images, and PDFs, and generating comprehensive reports.

• Its powered by a specialized version of the upcoming o3 model, optimized for web browsing, data analysis, and multimodal processing, allowing it to synthesize data within 5 to 30 minutes with citations and in-depth analysis, with future plans to integrate graphs, embedded images, and connections to specialized data sources.
• Deep Research is currently available to ChatGPT Pro users (limited to 100 queries per month), with support for Plus, Team, and Enterprise users coming soon.

3. OpenAI’s leadership held a Reddit AMA, where they hinted at potential openness to open-source AI, teased the o3 release timeline, and acknowledged DeepSeek as a rising competitor.

• Altman admitted that OpenAI has been “on the wrong side of history” regarding open-source AI, suggesting a need for a new strategy but noting that not everyone at OpenAI shares this view, and it’s not a top priority.
• Altman acknowledged that DeepSeek has reduced OpenAI’s competitive lead, but he remains confident OpenAI will continue producing better models even as the gap narrows.
• CPO Kevin Weil hinted at OpenAI potentially open-sourcing older AI models and introducing new transparency features to reveal reasoning processes, similar to DeepSeek’s R1 model.
• Weil confirmed OpenAI is working on a successor to DALL-E 3, stating it will be “worth the wait,” while Altman estimated that the full o3 model would arrive in “more than a few weeks, less than a few months.”
• Altman suggested that recursive self-improvement (AI improving itself without human input) is more plausible than he previously believed, but reiterated that OpenAI remains cautious about its advancements.

4. Anthropic has introduced "Constitutional Classifiers," a system designed to block nearly all AI jailbreak attempts by filtering harmful prompts and responses in real-time using a constitution of natural language rules.

• The system relies on a "constitution" of adaptable rules that define what is permitted and restricted, enabling rapid adjustments to counter evolving jailbreak techniques.
• The classifiers filter malicious prompts using synthetically generated data, significantly reducing jailbreak success rates (from 86% to near-zero in automated tests) without major computational overhead.
• A months-long bug bounty program with 183 security experts spending over 3,000 hours resulted in only a 50% success rate on a set of 10 forbidden prompts, demonstrating the system’s robustness.
• Anthropic is inviting the public to test the system until February 10, offering a challenge to break through its safeguards while promising rapid adaptations to defend against new attack methods.

5. OpenAI’s new trademark filing suggests future ventures into AI-powered humanoid robots, smart devices, quantum computing, and custom AI chips.

6. OpenAI and ソフトバンク have launched "Cristal Intelligence," an enterprise AI assistant for Japanese businesses that learns and adapts to company operations, strengthening OpenAI’s global expansion while reducing its reliance on Microsoft, with SoftBank Group companies gaining first access to OpenAI’s latest AI models in Japan.

7. The EU’s AI Act has reached its first compliance deadline (Feb. 2), banning AI systems deemed an "unacceptable risk", including social scoring, real-time biometric surveillance, and manipulative AI, with violators facing fines up to €35M ($36M) or 7% of global revenue, though enforcement begins in Aug ‘25.

8. Meta releases "Frontier AI Framework" that aims to balance innovation with AI risk mitigation, focusing on cybersecurity threats and potential misuse in chemical and biological weapons, while introducing threat modeling, risk thresholds, and expert collaboration to prevent catastrophic AI risks.

9. DeepSeek’s AI has been banned by multiple governments, including Italy, Taiwan, the U.S. Navy, and NASA, over privacy, security, and data sovereignty concerns.

10. Thomas Shedd, former Tesla engineer and current Technology Transformation Services (TTS) director, told General Services Administration (GSA) workers that the federal government would pursue an "AI-first strategy," emphasizing automation, centralized data collection, and AI coding agents across federal agencies, despite concerns over security risks and government-specific challenges.

11. Microsoft AI is launching the Advanced Planning Unit (APU) to analyze AI’s effects on society, health, and work, operating under CEO Mustafa Suleyman and hiring experts across economics, psychology, quantum, and nuclear fields.

12. MLCommons and Hugging Face have launched the "Unsupervised People’s Speech" dataset, featuring over a million hours of audio in 89+ languages to advance AI speech research, improve accent recognition, and enhance low-resource language models.

?? Top Funding News:

1. Tana , an AI-powered knowledge graph platform automating task management and workflows, raised a $25M funding round led by Tola Capital , w/? Lightspeed , Northzone , Alliance VC , and notable angels.

2. Jump - Advisor AI , an AI-driven assistant automating financial advisor workflows, raised a $20M Series A led by Battery Ventures , with participation from 花旗 Ventures, Sorenson Capital , and Pelion Venture Partners .

3. Cinamon , an AI-driven 3D animation platform streamlining video production for creators, raised an $8.5M Series B led by Altos Ventures and Saehan Venture Capital.