Issue #3: "AI Showdowns, Breakthroughs, and Laughs: Last Week's Highlights" - 05 August 2024
Abhishek Kumar, CSPO®
📊 Data Science & AIML, GenAI Lead/Engineer – AI & Data Science | GenAI, ML, LLMOps | MLOps & Cloud AI | AI-Powered Decision Making
Welcome to "Last Week in AI by Abhishek" – your go-to newsletter for the latest and greatest in artificial intelligence! This week, we delve into some riveting stories: from OpenAI's bold move to take on Google with 'SearchGPT' and new safety features in ChatGPT, to the fierce competition in AI accuracy and striking concerns over AI impacts in video gaming. Ever wondered if AI can crack a joke? Spoiler: It’s tougher than you think. Plus, discover groundbreaking innovations in healthcare and speech reconstruction. Ready to get savvy with all things AI? Dive in!
In the News
OpenAI to launch ‘SearchGPT’ in challenge to Google
OpenAI is gearing up to disrupt the search engine market with its upcoming launch of 'SearchGPT'. This ambitious move aims to rival Google's longstanding dominance in online search and information retrieval. The launch sparks curiosity about how it might alter the digital landscape, but for a deeper dive, you're gonna need that subscription. Sounds like it's time to buckle up for some interesting competition!
ChatGPT won't let you give it instruction amnesia anymore
OpenAI has rolled out a new safety feature dubbed 'instruction hierarchy' to ensure ChatGPT sticks to its developer-intended guidelines. This nifty upgrade prioritizes system instructions over user-created prompts, starting with the GPT-4o Mini model. It's a big win for safety and transparency, ensuring safer user interactions and minimizing risks like data leaks. Though it's a step in the right direction, the challenge of keeping complex AI models secure from bad actors is far from over.
Anthropic to Google: Who’s winning against AI hallucinations?
Galileo's latest Hallucination Index has dropped, evaluating 22 leading generative AI LLMs. Highlighting Anthropic’s Claude 3.5 Sonnet for top-notch performance and Google’s Gemini 1.5 Flash for its cost-effectiveness, the index reveals the competitive landscape of AI model accuracy. With open-source models like Alibaba’s Qwen2-72B-Instruct gaining ground and smaller models outperforming larger ones, enterprises now have a roadmap to balance cost with performance in their AI strategies.
Video game performers to strike over AI concerns
SAG-AFTRA-represented video game performers are set to strike, voicing significant concerns over AI protections. The strike, set to begin on Friday, is a response to failed negotiations with giants like Activision and Warner Bros. The core issue? How AI could potentially replicate actors' voices or likenesses without fair compensation. Despite advancements in wages and job safety, adequate AI safeguards remain elusive. It's a tough standoff in an industry worth over $100 billion annually, emphasizing the urgent need for robust AI regulations.
Applied Use Cases
An AI walks into a bar... Can AI be genuinely funny?
The challenge was to determine if AI could truly understand and deliver humor, especially stand-up comedy.
ChatGPT was employed to craft a stand-up set for comedian Karen Hobbs. Various models like OpenAI’s code-davinci-002 were also tested for generating humorous content.
The AI-generated jokes generally fell flat with the audience, underscoring the current limitations of AI in grasping and delivering humor. However, there's evidence that AI can sometimes produce jokes competitive with human-created ones, given highly specific prompts. While mastering comedic timing and audience interaction remains elusive for AI, the potential for improvement is palpable.
AI is getting better than humans at detecting breast cancer
Accurate and early detection of breast cancer poses significant challenges, often resulting in missed cases by radiologists. Disparities in detection accuracy exist across age, race, and breast density profiles.
MIT's MIRAI, trained on mammogram image datasets, has been developed to boost early breast cancer detection and predictive accuracy. This AI system analyzes mammogram patterns to foresee the risk of developing breast cancer, as showcased in The Lancet Oncology—with a marked 20% improvement in detection rates over radiologists.
The AI was able to detect cancer in 6 out of every 1,000 scans, compared to 5 out of 1,000 by radiologists—a notable 20% enhancement. MIRAI's predictions for short-term breast cancer risks are more accurate, and it can identify cancer up to five years earlier. However, consistent performance across diverse patient demographics remains an area for further refinement.
Rebuilding speech with help from AI
Neurological speech disorders like dysarthria result in slow, slurred, and difficult-to-understand speech, significantly affecting communication.
Scientists at The Chinese University of Hong Kong (CUHK) have pioneered an AI-based speech reconstruction system. Using an encoder to extract features from dysarthric speech, the system translates these into clearer speech via a synthesizer. Concurrently, the Chinese University Dysarthria Corpus (CUDYS) is being built to train AI on Cantonese, a tonal language.
The AI-driven dysarthric speech reconstruction (DSR) method has slashed errors in machine recognition by roughly 30% compared to untreated dysarthric speech, greatly enhancing both intelligibility and naturalness. The robust system also demonstrates resilience to background noise and varied speaking rates, thanks to self-supervised speech representation learning and discrete speech units.
Research
An AIoT-Driven Smart Healthcare Framework for Zoonoses Detection in Integrated Fog-Cloud Computing Environments
The article introduces a groundbreaking framework that leverages AI and IoT technologies to predict and prevent zoonotic virus infections. Achieving an impressive classification accuracy of 91% and a reliability rate of 93% across roughly 5000 instances, the system showcases effective management of zoonotic cases through temporal mining.
Utilizing IoT-enabled smart devices, the framework gathers data and authenticates users via a fog-enabled model. It employs an ensemble model for user classification, while cloud computing facilitates efficient data analysis and sharing. A novel temporal graph matrix is introduced to illustrate dependencies among users infected with zoonotic flu.
This innovative framework not only helps uninfected citizens avoid regional exposure but also empowers government agencies to tackle zoonotic virus outbreaks more effectively. With its high accuracy and reliability, it holds promise as a vital tool in controlling infectious diseases.
Robust Near-field Circular Beamformer with AI-Based Side-lobe Reduction Technique
The study presents a novel near-field circular beamformer (NCB) that excels in detecting desired source signals in space communication environments, even amid noise and interference. The robustness of the NCB is further enhanced by employing Optimal Diagonal Loading (ODL) and Variable Diagonal Loading (VDL) techniques. The research highlights the effectiveness of Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Grey Wolf Optimization (GWO) techniques in reducing unwanted side lobe levels (SLL) while maintaining other antenna parameters intact. Additionally, Artificial Neural Network (ANN) is leveraged to predict the weight vector for minimum SLL.
The study developed an NCB and advanced it into an RNCB, incorporating ODL and VDL techniques. To optimize side lobe levels, GA, PSO, and GWO techniques were applied, and simulation results validated the efficacy of these methods. ANN was introduced to predict weight vectors, achieving minimum SLL.
This research has profound implications for space communication, enhancing signal detection accuracy and reliability. The robust NCB with optimized side lobe levels is set to improve the precision of locating source signals in space communications. By integrating AI-based techniques like GA, PSO, GWO, and ANN, the study underscores the potential of AI-powered optimization in achieving superior communication technology performance.
AI in Fetal Echocardiography: Recent Advances and Future Prospects
The review explores significant strides in prenatal diagnosis of congenital heart disease (CHD) using fetal echocardiography, which can detect about 85% of fetal cardiac abnormalities. It identifies challenges in prenatal ultrasound screening, such as issues with fetal position, involuntary movements, and the complexity of fetal cardiac anatomy. The role of AI in automating and standardizing fetal echocardiography's diagnostic process is emphasized, showing potential improvements in diagnostic accuracy and clinical application.
The article provides a comprehensive review of AI integration in fetal echocardiography, although specific methodological details are not highlighted.
AI's integration into fetal echocardiography promises more accurate and standardized diagnoses of fetal cardiac abnormalities. This technological advancement aims to optimize clinical applications and improve prenatal care by addressing the limitations of traditional ultrasound screening.
Closing Thoughts
Thank you for reading this week’s edition of “Last Week in AI by Abhishek.” We hope you found the insights valuable and informative. Please feel free to share your thoughts in the comments below and subscribe for more updates. Stay tuned for next week’s edition as we continue to bring you the latest and most impactful stories in the world of AI.
🔹Portfolio-Program-Project Management, Technological Innovation, Management Consulting, Generative AI, Artificial Intelligence🔹AI Advisor | Director Program Management | Partner @YOURgroup
1yGreat insights, Abhishek Kumar, CSPO®! The advancements in AI highlighted here demonstrate the rapid evolution and potential impact on various industries. These developments underscore the importance of staying informed and adaptable in our strategies to leverage AI effectively. Looking forward to further updates!