Good Morning! OpenAI's dropping their latest o3-mini model in GitHub Copilot, promising faster reasoning with the same brainpower we loved in o1. DeepSeek's making waves for all the wrong reasons as security researchers found their AI safeguards crumbling under even the most basic jailbreak attempts. And in the world of fintech, banks are finally catching up to modern tech stacks by embracing cloud computing and APIs, transforming their sluggish legacy systems into something that actually moves at internet speed.

— Forrest Knight & Meghanadh Vasireddy

OpenAI's o3-mini Hits GitHub: Faster Reasoning, Same Brain Power

OpenAI just dropped their latest reasoning model, o3-mini, and it's doing well in the dev ecosystem. GitHub has wasted no time integrating it into Copilot and GitHub Models, marking a significant leap in AI-assisted development.

What's New: The o3-mini model matches o1's performance in STEM tasks while delivering responses 24% faster (averaging 7.7s vs 10.16s). For devs using GitHub Copilot Pro, Business, or Enterprise, you'll find it in VS Code and github.com chat, with JetBrains support coming soon.

Performance improvements:

• Matches o1 on competition coding (Codeforces)

• 48.9% accuracy on SWE-bench Verified tasks

• 39% reduction in major errors on complex problems

• Supports function calling, structured outputs, and developer messages

• Configurable reasoning effort levels (low/medium/high)

The Practical Stuff: Paid Copilot subscribers get 50 messages per 12-hour period. Enterprise admins can manage access through org settings. Plus, the GitHub Models playground lets you experiment with o3-mini alongside models from Cohere, DeepSeek, Meta, and Mistral – perfect for comparing performance and finding the right fit for your workflow.

Read More Here

DeepSeek's AI Safety Issues

While most AI companies are beefing up their safety measures, security researchers at Adversa just uncovered some serious vulnerabilities in DeepSeek's safeguards. And we're not talking about sophisticated exploits here – even the most basic jailbreak techniques worked like a charm.

The Testing: Adversa put DeepSeek through its paces with 50 different jailbreak attempts. The results? A perfect 100% success rate – and not in a good way. The model failed every single safeguard test, from simple role-playing tricks to more complex token manipulation techniques.

Vulnerability types discovered:

• Basic linguistic jailbreaks (including UCAR variants)

• Programming jailbreaks via SQL query manipulation

• Token chain exploitation (similar to the infamous "anatomcalifwmg" case)

• Character-based attacks like DAN variations

• Adversarial approaches targeting token representations

This revelation raises serious concerns about AI safety implementation. While other major players have patched similar vulnerabilities, DeepSeek's comprehensive failure to implement basic safety measures shows we've still got a long way to go in securing AI systems. For developers and security professionals, it's a stark reminder of the importance of robust safety testing in AI development.

Read More Here

Cloud & APIs: Banking's Tech Evolution Gets Real

Context: The banking sector is finally shedding its monolithic legacy systems for something more our speed - cloud computing and APIs. This shift isn't just about modernizing; it's about transforming banks from slow-moving giants into agile service providers that can actually keep up with today's tech demands.

Gone are the days of month-long deployment cycles: Banks are now leveraging cloud platforms to roll out services in days, while APIs are turning them into proper platform players. The real game-changer? The combination of both technologies enabling real-time processing and advanced analytics at scale.

Key improvements in banking infrastructure:

• Shift from CapEx to OpEx models

• Multi-cloud architecture support

• Real-time data processing capabilities

• API-driven service integration

• Enhanced security through cloud-native tools

• Blockchain and IoT integration potential

The Bottom Line: While regulatory hurdles and legacy system integration remain challenging, the cloud-API combo is proving to be a solid foundation for modern banking. It's not just about cost savings anymore - we're talking about a complete overhaul of how financial services are built and delivered. Banks that nail this transition are positioning themselves to handle whatever fintech throws at them next.

🔥 More Notes

• Tana snaps up $25M as its AI-powered knowledge graph for work racks up a 160k+ waitlist

• Keir Starmer scrapped email account in 2022 after Russian hacking, says report

• SoftBank and OpenAI set up joint company to push artificial intelligence services

📹 Youtube Spotlight

Was this forwarded to you? Sign Up Here