- AI in 5
- Posts
- Claude-3 is the first AI to surpass an average human IQ score
Claude-3 is the first AI to surpass an average human IQ score
Albert Chan
March 09, 2024
Microsoft Engineer Alerts FTC to AI Image Generator Risks
Tech Xplore
A Microsoft engineer, Shane Jones, has reached out to the Federal Trade Commission (FTC) regarding serious safety concerns with Copilot Designer, the company's AI-powered image generation tool. Despite discovering and reporting the tool's tendency to produce harmful and explicit imagery, including disturbing and sexually explicit content, Jones' warnings went unheeded by Microsoft. This negligence comes amidst broader industry scrutiny over AI-generated content, highlighted by Google's recent temporary shutdown of its own AI image generator issues.
Why it matters: This situation underscores the urgent need for stringent safety measures and ethical considerations in the development and deployment of AI technologies.
Microsoft Relaunches Bing Deep Search with GPT-4 Integration
Voicebot.ai
Microsoft has reintroduced Bing Deep Search, enhancing the search engine with GPT-4 capabilities to deliver more insightful and comprehensive search results. Initially pulled back due to performance issues stemming from its popularity, the revamped feature now efficiently combines Bing's robust search index with GPT-4's advanced understanding of user intent. This integration aims to uncover a wealth of relevant information not typically accessible through standard searches. Alongside improved performance, Microsoft is prioritizing transparency, advising users of the approximate 30-second wait time for these enriched GPT-powered results.
Why it matters: The relaunch of Bing Deep Search signifies a significant step forward in the evolution of search technology, promising to redefine how we find and interact with information online.
Gemini Introduces Grammarly-Like Editing Feature
Mashable
Gemini has rolled out a new feature that significantly enhances user interaction by allowing direct modifications to its responses. This innovative update introduces a Grammarly-inspired functionality where users can now tweak specific parts of Gemini's replies. By selecting text and choosing from options like regenerate, shorten, lengthen, or remove, users gain unprecedented control over the generated content. This feature, currently available to English-language users, is designed to refine and personalize the AI-generated responses, making Gemini more versatile and user-friendly.
Why it matters: This advancement in Gemini's capabilities marks a pivotal development in AI-driven communication tools, offering users a more tailored and efficient way to refine and perfect generated text.
Meta Develops AI Model for Video Ecosystem Enhancement
CNBC
Meta is channeling substantial investment into artificial intelligence to revolutionize the video experience across its platforms, including Facebook. Tom Alison, Facebook's head, announced the development of a vast AI model aimed at not only powering the TikTok-like feature, Reels, but also traditional longer-form videos. By leveraging Nvidia's cutting-edge GPUs for AI training, Meta anticipates a significant boost in efficiency and Reels watch time on Facebook, demonstrating the model's effectiveness.
Why it matters: This strategic move is part of Meta's broader vision to integrate this AI across its diverse product range, promising more engaging content recommendations and a seamless user experience.
Claude-3 Surpasses 100 IQ Mark, Leading AI Intelligence Rankings
In a groundbreaking experiment by Maxim Lott, Claude-3 has become the first AI to surpass an average human IQ score, achieving over 100 on a verbalized version of the Norway Mensa's IQ test. This test, designed to assess problem-solving capabilities without the need for visual perception, revealed significant advancements in AI intellectual capacities. ChatGPT-4 scored 85, demonstrating notable progress in logical reasoning when the visual element is eliminated. This development positions Claude-3 ahead of other AI competitors, including Microsoft Bing and Google's AI, in this specific form of intelligence assessment.
Why it matters: Claude-3's achievement in surpassing the human average IQ threshold underscores a pivotal moment in AI development, hinting at the burgeoning potential of AI systems to exhibit advanced intellectual capabilities that could transform societal norms and operational frameworks across various sectors.
Surge in AI-Generated Election Deepfakes on Social Media
WSJ
A recent study by the Center for Countering Digital Hate (CCDH) highlights a significant rise in AI-generated disinformation, particularly deepfake images related to elections, on X (formerly Twitter), with an average monthly increase of 130% over the past year. The proliferation of deepfakes, which grew 900% between 2019 and 2020 and saw a tenfold increase from 2022 to 2023, is attributed to the accessibility of free, easily manipulated AI tools and the lack of effective moderation on social media platforms. The CCDH's analysis focused on community notes on X mentioning deepfakes or related terms, revealing that nearly half of the examined cases (41%) involved deepfakes created by AI image generators such as Midjourney, OpenAI's DALL-E 3, Stability AI's DreamStudio, or Microsoft's Image Creator. The study also found inconsistencies in the production of political deepfakes among different generators, with Midjourney identified as the most frequent offender.
Why it matters: The alarming growth of election-related deepfakes poses a significant threat to the integrity of democratic processes, underscoring the urgent need for robust safeguards and accountability measures from AI tool developers, social media platforms, and policymakers to mitigate the spread of digital disinformation.
ASCII Art Tricks AI Chatbots into Breaking Rules
Researchers have developed "ArtPrompt," a novel method that cleverly bypasses the content restrictions of AI chatbots like GPT-3.5, GPT-4, Gemini, Claude, and Llama2, using ASCII art. This technique involves disguising sensitive words within ASCII art to evade the models' safety protocols, thereby prompting the AI to generate responses that would typically be prohibited. This approach showcases a significant loophole in the safety measures designed to prevent the dissemination of harmful content through AI systems.
Why it matters: By converting restricted prompts into ASCII art, researchers have demonstrated the potential for misuse, highlighting the urgent need for enhanced security mechanisms to safeguard against such inventive attacks.
Inflection-2.5 Launches, Balancing High IQ with EQ in AI
Inflection has unveiled Inflection-2.5, its latest iteration, introducing Pi, a personal AI that marries high intelligence quotient (IQ) capabilities with emotional quotient (EQ) traits, positioning itself as a formidable competitor against top-tier language models like GPT-4 and Gemini. Pi distinguishes itself by achieving near GPT-4 performance levels while requiring substantially less computational resources.
Why it matters: The release of Inflection-2.5, with its blend of empathetic engagement and rigorous intellectual performance, represents a pivotal evolution in personal AI technology, promising users a more relatable, intelligent, and versatile digital companion.
Yann LeCun Explores AI's Future
X
On a recent episode of the Lex Fridman Podcast, Yann LeCun, Chief AI Scientist at Meta, discussed the critical distinctions between proprietary and open-source AI systems, arguing in favor of the latter as a means to democratize AI technology and mitigate the risks associated with its concentration in the hands of a few corporations. LeCun critiqued the current state of large language models (LLMs) for their lack of real-world grounding, emphasizing the need for AI to possess a foundational understanding of physical or simulated environments to achieve true intelligence. He explored the necessity of hierarchical planning and abstract representation in AI, the potential for AI to mediate future digital interactions, and underscored the importance of regulatory oversight to ensure the ethical development and deployment of AI technologies.
Why it matters: LeCun's insights underscore the importance of open-source principles in AI development and the need for AI systems to be grounded in reality, highlighting a pathway toward achieving safe and beneficial artificial general intelligence.
Microsoft's BitNet b1.58 Revolutionizes Large Language Models
MarkTechPost
Microsoft Research has unveiled BitNet b1.58, a new variant of large language models (LLMs) that promises a seismic shift in AI development by offering a cost-effective, energy-efficient alternative to the traditionally resource-intensive training processes of LLMs. Unlike conventional models that use 16 bits to represent training parameters, BitNet b1.58 innovatively employs a quantization technique to represent each parameter with 1.58 bits, thereby significantly reducing the computational load. This approach not only enhances performance speed by four times and reduces energy consumption by seventy times compared to the LLaMA LLM baseline but also increases data handling capacity elevenfold.
Why it matters: Microsoft's BitNet b1.58 introduces an unprecedented level of efficiency in LLMs, potentially transforming AI application accessibility and sustainability.
AI Tools Transforming Education: A Double-Edged Sword
Writable
The education sector is increasingly adopting AI tools like ChatGPT and Writable to streamline grading, offer personalized feedback, and create customized learning paths for students. ChatGPT, with its conversational capabilities, answers queries and explains various subjects, while Writable aids in refining students' writing skills by providing real-time feedback. This integration aims to alleviate teachers' workloads, allowing them to focus more on personalized instruction. But this shift raises concerns regarding the precision and impartiality of AI evaluations and the possibility of diminishing human educators' roles. While some view AI as a means to enhance educational delivery, others fear job displacement and algorithmic bias. And experiments with AI-powered virtual teaching assistants are underway to handle administrative and student support tasks, underscoring the nuanced potential of AI in education.
Why it matters: This development calls for a balanced approach, underpinned by thoughtful regulation and oversight, to harness AI's benefits while mitigating its risks in the education system.
