OpenAI News & Insights: Security Alerts, Model Upgrades, and AI Milestones

OpenAI Hit By Two Big Security Issues

OpenAI is facing dual security concerns. The first involves its Mac ChatGPT app, where engineer Pedro José Pereira Vieito discovered that user conversations were stored in plain text rather than encrypted. Since the app is only available on OpenAI's website, it bypasses Apple's sandboxing requirements. After Vieito's findings were publicized, OpenAI updated the app to include encryption. Sandboxing is a security measure that prevents vulnerabilities from spreading across applications, and storing data in plain text poses risks of exposure to other apps or malware.

The second issue dates back to 2023 when a hacker accessed OpenAI's internal messaging systems, raising concerns about internal vulnerabilities. OpenAI's technical program manager, Leopold Aschenbrenner, who highlighted these security flaws, claims he was fired for whistleblowing. OpenAI disputes this, stating his departure was unrelated. These incidents raise concerns about OpenAI's ability to manage data security effectively amidst its rapid adoption and internal challenges.

OpenAI's new, lightweight GPT-4o mini model promises an improved ChatGPT experience

OpenAI has introduced GPT-4o mini, a smaller and more affordable version of its flagship language model, offering developers a 60% cost reduction compared to GPT-3.5 Turbo. This new model will replace GPT-3.5 Turbo for free ChatGPT users, enhancing the baseline experience. GPT-4o mini scored 82% on the MMLU benchmark, slightly lower than GPT-4o's 88.7% but higher than GPT-3.5 Turbo's 70%. While AI experts caution against over-relying on benchmarks, they remain a standard measure of performance.

Smaller models like GPT-4o mini provide developers with flexibility and cost-efficiency, similar to Google's Gemini 1.5 Flash and other compact models from AI companies like Anthropic. GPT-4o mini currently supports text and image processing, with future capabilities planned for audio and video. Although GPT-3.5 Turbo will no longer be available on ChatGPT, developers can still access it via OpenAI’s API until it is phased out.

Former OpenAI researcher’s new company will teach you how to build an LLM

Former OpenAI researcher Andrej Karpathy has launched Eureka Labs, an AI learning platform aiming to create an "AI native" educational experience. Its first course, LLM101n, targets undergraduates, teaching them to build a "Storyteller AI Large Language Model" to create and illustrate stories. The platform pairs expert-designed materials with an AI-powered teaching assistant for personalized guidance.

Karpathy, a prominent figure in AI, announced the venture to build publicly. Eureka Labs will initially offer courses online, with future plans for in-person groups. Karpathy's background includes a PhD from Stanford, founding membership at OpenAI, and senior AI director at Tesla. He has also produced popular AI tutorials on YouTube.

Eureka Labs aspires to expand beyond its initial AI course to offer a broad curriculum, aiming to leverage AI to enhance human potential. Karpathy emphasized his longstanding passion for both AI and education as the driving force behind this venture.

OpenAI reportedly nears breakthrough with “reasoning” AI, reveals progress framework

OpenAI has introduced a five-tier system to evaluate its progress toward developing artificial general intelligence (AGI). Unveiled to employees during an all-hands meeting, this system aims to provide a clear framework for AI advancement, though it describes hypothetical technology and may serve as a marketing strategy to attract investors. OpenAI's primary goal is achieving AGI, which refers to AI capable of performing novel tasks like humans without specialized training. CEO Sam Altman believes AGI could be reached within this decade.

The five levels range from current conversational AI (Level 1) to systems managing entire organizations (Level 5). OpenAI’s technology, such as GPT-4o, is at Level 1, with executives suggesting they're nearing Level 2, "Reasoners," capable of human-level problem-solving. Higher levels include autonomous agents (Level 3), innovative AI (Level 4), and organization-managing AI (Level 5).

This system, still under refinement, aims to communicate milestones and garner feedback from employees, investors, and board members. Similar classification attempts by other AI labs highlight the challenge of quantifying AI progress and the potential for overpromising capabilities.

OpenAI explores custom AI chip with this NVIDIA rival

OpenAI, currently using NVIDIA GPUs for its AI models, is exploring the development of its own AI-specific chip. CEO Sam Altman is in discussions with Broadcom and other chip manufacturers to reduce reliance on NVIDIA. This aligns with Altman's broader vision to enhance infrastructure, including power supplies and data centers, to support powerful AI models. OpenAI is also recruiting former Google employees experienced with Google’s tensor processing units (TPUs).

A potential partnership with Broadcom is in the works, leveraging Broadcom's expertise in custom AI accelerators and previous collaborations on Google’s TPU project. Broadcom, a fabless chip designer, offers comprehensive silicon solutions essential for data centers, including networking components, PCIe controllers, SSD controllers, and custom ASICs. OpenAI could utilize Broadcom's full vertical stack of products to meet its data center needs, benefiting from Broadcom's expertise in communication technologies crucial for AI infrastructure.