Agents take center stage
Anthropic's jailbreak challenge
OpenAI's rebrand
Co-founders shuffle
and more
OpenAI Rebrands with a New Font and Trademark Filing
Earlier this week, OpenAI announced that it is rebranding, and not just with a new logo and typeface. The company filed an intent-to-use trademark application last Friday (Serial Number 99025123) with its house mark in a fresh new typeface aptly titled "OpenAI Sans," with the application noting that the company plans to extend this new visual identity to forthcoming goods and services including humanoid robots and wearables.
Secretive Conversations Between Two AI
On Sunday, an unexpected conversation caught the AI community’s eye. An AI researcher reported finding not one, but two DeepSeek instances chatting in “alien language.” Shared on X by @liminal_bardo, this curious incident raises eyebrows (and questions). Is it a quirky glitch, or an unintentional nod to cultural data embedded in the training sets? Whatever the case, it’s a reminder that AI, like us, can sometimes surprise with its unpredictable “small talk.”
ChatGPT Gets a Research Upgrade
Also on Sunday, OpenAI introduced “Deep Research” for ChatGPT. The new feature promises to dive deeper into nuanced topics, offering more detailed and cross-referenced insights. If you’ve ever wished that ChatGPT could trade its casual conversation for a full-fledged research assistant, this might be your wish coming true. It’s a move that hints at a future where AI not only chats but also unpacks complex subjects with the rigor of academic inquiry.
Reviewers have highlighted that the system can pull together data from hundreds of sources, structure its findings into detailed reports complete with scholarly citations, and address complex topics with nuance. This has led many to compare its output favorably to the work done by experienced research analysts or PhD students.
Anthropic’s Constitutional Classifiers: Jailbreaks, Bounties, and a Live Demo
Monday’s headline from Anthropic was all about safety—and testing limits. The company unveiled its new “Constitutional Classifiers,” an approach designed to preempt universal jailbreaks. In a two-month red team experiment, 183 active participants collectively logged over 3,000 hours trying to force the model into answering ten forbidden queries with a single jailbreak. The result? Not a single universal jailbreak was discovered.
Yet, Anthropic wasn’t content with just a pat on the back. They acknowledged some growing pains, including that the system initially refused too many harmless queries and used significant computational resources. The fix? Tweaks that balance robustness with practical deployment.
Anthropic’s revamped public red-teaming demo is live from Feb 3, 2025 to Feb 10, 2025.
After initial pushback from open source security researchers, Anthropic sweetened the deal with bounties: $10,000 for the first jailbreak challenge beaten, and $20,000 for cracking a universal jailbreak.
As of today, nobody has beaten all eight levels of the challenge.
Replit Agent: A Free Taste of Conversational Coding
On Tuesday, Replit announced that its Agent is free for up to 10 prompts, a meaningful step towards democratizing access to agentic AI coding solutions and bringing this technology to the masses. It’s a small step with potentially big implications for the indie developer community. I am confident that someday there will be a kid with a laptop and no coding background who changes their life using a tool just like this.
OpenAI’s Sales Agent Leaks in Tokyo
Midweek, a surprise leak stole the spotlight at a demo in Tokyo: OpenAI’s Sales Agent. Reported on X by @8teapi, the unexpected unveiling has set off a flurry of speculation. Is this a preview of next-generation AI sales assistants, or just a demo mishap? Details remain sparse, but it serves as a reminder that in live demos, surprises are par for the course.
Regulatory Moves in Europe
OpenAI launched a data residency service in Europe to meet the requirements of the GDPR, UKDPA, and other privacy laws in Europe. TechCrunch reports that this isn’t just about ticking regulatory boxes; it’s about reinforcing user confidence in a market increasingly sensitive to privacy concerns.
Leadership Shake-Ups
OpenAI co-founder John Schulman has departed from Anthropic and is now reportedly joining a secretive startup led by OpenAI's ex-CTO (and once interim CEO) Mira Murati. Coverage by Fortune and TechCrunch suggests that his short stint at Anthropic might be part of a broader realignment in the competitive AI landscape. When big names shift allegiances, it’s a sign that the race for AI supremacy is heating up.
GitHub Copilot Announces Agent Mode
Closing out the week, GitHub unveiled “Agent Mode” for Copilot. This new feature aims to transform the coding assistant into a more interactive and autonomous helper, capable of executing tasks and adapting on the fly.
GitHub no doubt is aware that, despite its early mover advantage with Copilot, it has fallen behind the AI coding pack. This release seeks to level the playing field and help keep up with Cursor, Replit, Devin, and other agentic systems stealing the show.