WEEKLY AI NEWS: RESEARCH, NEWS, RESOURCES, AND PERSPECTIVES
AI & ML news: Week 9–15 September
OpenAI’s new model, Apple to Unveil iPhone 16 and ‘Apple Intelligence’ AI Features, California AI Regulation Bill Nears Approval, and much more
21 min readSep 16, 2024
The most interesting news, repository, articles, and resources of the week
Check and star this repository where the news will be collected and indexed:
You will find the news first in GitHub. All the Weekly News stories are also collected here:
Research
- De novo design of high-affinity protein binders with AlphaProteo. demonstrates a family of machine learning models that have been trained for protein design; reports 3-to 300-fold improvements in binding affinities and higher experimental success rates when compared to other methods on seven target proteins; demonstrates that AlphaProteo’s performance is similar to the seven targets when tested on hundreds of target proteins from the PDB.
- In Defense of RAG in the Era of Long-Context Language Models. reports that one of the main problems that an RAG system addresses (i.e., uses more relevant information) is that longer-context LLMs suffer from a diminished focus on relevant information. They suggest an order-preserving RAG mechanism that enhances performance on long-context question answering, but it’s not perfect — in fact, the quality of responses increases and then declines as retrieved chunks increase. They also mention a sweet spot where it can achieve better quality with a lot fewer tokens than long-context LLMs.
- Strategic Chain-of-Thought: Guiding Accurate Reasoning in LLMs through Strategy Elicitation. a technique to improve LLM performance by adding strategic information before the intermediate CoT reasoning phases; the strategy for addressing problems aids in directing the creation of the CoT paths and solutions; promises to use the Llama3–8b model to get a 21.05% gain on the GSM8K datasets.
- The Effects of Generative AI on High Skilled Work: Evidence from Three Field Experiments with Software Developers. Examines the effects of generative AI on software developers, highlighting a 26.08% rise in completed tasks among developers utilizing AI tools such as GitHub Copilot. Additionally, it indicates that less experienced developers are more inclined to adopt AI tools and experience significant productivity improvements.
- LongCite: Enabling LLMs to Generate Fine-grained Citations in Long-context QA. Creates a large-scale supervised fine-tuning (SFT) dataset using off-the-shelf large language models (LLMs) to enhance long-context question answering with citations. The training focuses on 8B and 9B parameter models, improving their ability to generate citations from extended contexts while enhancing response accuracy. It claims to outperform GPT-4o on its proposed LongBench-Cite benchmark.
- MemLong: Memory-Augmented Retrieval for Long Text Modeling. Employs an external retriever to gather historical information, enhancing the performance of long-context large language models (LLMs). It consistently surpasses other state-of-the-art LLMs on long-context benchmarks and can extend context length from 4k to 80k on a single 3090 GPU.
- Pandora’s Box or Aladdin’s Lamp: A Comprehensive Analysis Revealing the Role of RAG Noise in Large Language Models. Introduces a benchmark, NoiserBench, to assess how various types of noisy information impact the performance of retrieval-augmented generation (RAG) models. The study reveals that, among different beneficial noise types (e.g., semantic, datatype, and illegal sentence), illegal sentence noise leads to the greatest performance improvement across models and datasets.
- Beyond Preferences in AI Alignment. Critiques the prevailing AI alignment method of human preference tuning, highlighting how it fails to grasp the rich, nuanced content of human values. The argument is made that AI alignment requires reframing, suggesting that instead of aligning with individual human preferences, AI systems should align with normative standards relevant to their societal roles.
- Planning In Natural Language Improves LLM Search For Code Generation. Obtaining a variety of candidate solutions is one of the difficulties in code creation. Even repeated sampling frequently falls short of producing enough originality to address an issue. But if you start with a natural language plan and generate ideas for potential solution paths, the resulting generation is much more varied and diverse, which leads to better solutions for code creation.
- Imitating Language via Scalable Inverse Reinforcement Learning. Modern language modeling can largely be viewed as a specialized form of imitation learning, which benefits from extensive research in the broader field. This paper investigates the application of inverse reinforcement learning to mimic entire sequences rather than individual tokens. The findings are encouraging and suggest that reinforcement learning could play an increasingly important role in the training pipelines of language models moving forward.
- Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers. This longitudinal study evaluated the abilities of 100 NLP researchers to generate and review novel ideas. The findings revealed that while LLMs were able to produce more innovative ideas, these ideas were slightly less practical compared to those created by human researchers.
- Superhuman Automated Forecasting. The Safe AI Institute has published research on a system capable of surpassing human experts in forecasting accuracy.
- The AdEMAMix Optimizer: Better, Faster, Older. This paper from Apple introduces an alternative to the traditional exponential moving average optimization method, incorporating contributions from older gradients to significantly enhance learning convergence.
- DiverGen: Improving Instance Segmentation by Learning Wider Data Distribution with More Diverse Generative Data. DiverGen is an innovative approach for generating datasets to improve instance segmentation models. Instead of relying on expensive manual annotations, it leverages generative models to create diverse data, helping to mitigate overfitting and boost model performance.
- Policy Filtration in RLHF to Fine-Tune LLM for Code Generation. Policy Filtration for Proximal Policy Optimization (PF-PPO) is a technique aimed at enhancing the precision of reinforcement learning from human feedback (RLHF), specifically in the context of code generation tasks.
- Data Augmentation via Latent Diffusion for Saliency Prediction. Researchers have introduced a novel data augmentation technique to enhance saliency prediction models, which have historically struggled due to the scarcity of labeled data.
News
- Google using anti-competitive tactics in UK ad market, claims watchdog. CMA says tech company has ‘abused its dominant position’ to the detriment of publishers and advertisers
- Apple to unveil iPhone 16 and ‘Apple Intelligence’ AI features. Apple watchers also expect new colors for the iPhone at the annual launch event, this year titled ‘It’s Glow time’
- TSMC’s $65 billion Arizona facility can now match Taiwan production yields according to early trials.the US is committed to establishing semiconductor manufacturing within its borders, and perhaps no effort is more crucial to this goal than TSMC’s three-fab facility in Arizona. The government is pouring billions into the development, alongside TSMC’s $65 billion investment.
- AI Firm’s Misconfigured Server Exposed 5.3 TB of Mental Health Records. A misconfigured server from a US-based AI healthcare firm Confidant Health exposed 5.3 TB of sensitive mental health records, including personal details, assessments, and medical information, posing serious privacy risks for patients.
- California’s big AI regulation bill is headed to Gavin Newsom. A California bill requiring makers of large AI systems to test them for potential harm cleared the Legislature today. It could still face a veto by Gov. Gavin Newsom.
- Google search monopoly US case remedies to come by December. The U.S. Department of Justice plans to issue an outline by December on what Alphabet’s, must do to restore competition after a judge earlier found the company illegally monopolized the market for online search, prosecutors said at a court hearing in Washington on Friday.
- Intel reveals first Lunar Lake laptop CPUs: everything you need to know. Previously known as Lunar Lake, Intel has introduced its Core Ultra 200V portfolio, which features competitive integrated GPUs for tiny notebooks, fast CPUs, and enhanced AI capabilities. The CPUs have 32GB RAM capacity, eight CPU cores, integrated memory, and improved efficiency. Prominent producers such as Acer, Asus, Dell, and HP will introduce laptops equipped with these novel CPUs. Reviews to support Intel’s assertions are still pending.
- OpenAI, Still Haunted by Its Chaotic Past, Is Trying to Grow Up. To draw in significant investors such as Microsoft, Apple, and Nvidia, OpenAI is reorganizing its management and organization intending to reach a $100 billion valuation. Internal disagreements within the organization regarding its safety procedures and objectives have resulted in a high employee turnover rate, with important researchers leaving to work for competitors such as Anthropic. OpenAI struggles to strike a balance between business goals and moral considerations while developing AI technology, despite increasing income and user base growth.
- BP extends the use of AI in a five-year deal with spy tech firm Palantir. Oil and gas company to use artificial intelligence to speed up decision-making by engineers
- Google’s second antitrust suit brought by US begins, over online ads. DoJ accused tech giant of more monopolistic behavior a month after a judge found it illegally cornered online search
- What is Apple Intelligence, when is it coming and who will get it? At WWDC 2024, Apple unveiled Apple Intelligence, a platform designed to integrate AI capabilities into existing applications like Mail, Messages, and Siri. Utilizing large language models, it supports functions such as text summarization and image generation, all aimed at enhancing the user experience. A beta version will be available in the U.S. starting this October, with plans to expand globally in 2025.
- New open source AI leader Reflection 70B’s performance questioned, accused of ‘fraud’. HyperWrite’s Reflection 70B, a variant of Meta’s Llama 3.1 LLM, is under scrutiny after independent evaluators were unable to reproduce its advertised performance. The problems were traced back to corrupted model weights during the upload to Hugging Face, causing inconsistencies. The AI community is now awaiting further clarifications and updates to better understand the model’s true capabilities.
- The new Shortwave AI Assistant. Shortwave has substantially enhanced its AI Assistant, equipping it to handle complex, multi-step tasks like advanced searches, calendar lookups, and in-depth email analysis, making it more versatile and powerful in managing user tasks.
- OpenAI might use Apple’s TSMC for chips. OpenAI could greatly lower operational costs by adopting more efficient chips, which would be particularly beneficial as its user base continues to expand, allowing for better scalability and resource management.
- Apple takes direct aim at Microsoft’s Copilot+ PCs in new AI-focused Mac promos. Apple is actively marketing the Mac as the “best AI PC,” positioning it as a direct competitor to Microsoft’s Copilot+ PCs. This strategic push highlights Apple’s focus on integrating AI capabilities into its devices, aiming to challenge Microsoft’s AI-driven offerings in the PC market.
- GPT-fabricated scientific papers on Google Scholar: Key features, spread, and implications for preempting evidence manipulation. Generative AI tools, such as ChatGPT, are increasingly generating fraudulent research papers that are finding their way into databases like Google Scholar, mixing with legitimate studies. These papers, frequently addressing sensitive topics like health and the environment, threaten the integrity of science and public trust. Strengthened oversight and improved filtering mechanisms in academic search engines are crucial to addressing this rising concern.
- Apple announces its new A18 and A18 Pro iPhone chips. At its “Glowtime” event, Apple introduced the A18 and A18 Pro chips, highlighting substantial CPU and GPU upgrades compared to the A16 Bionic. The A18 Pro offers increased memory bandwidth and improved image processing. Both chips come equipped with advanced AI capabilities, with the A18 Pro specifically enhancing on-device model performance and thermal design for a superior gaming experience.
- AMD announces unified UDNA GPU architecture — bringing RDNA and CDNA together to take on Nvidia’s CUDA ecosystem. At IFA 2024, AMD revealed plans to merge its RDNA and CDNA architectures into a unified UDNA microarchitecture, positioning itself to compete more effectively with Nvidia’s CUDA ecosystem. This strategic shift is aimed at simplifying development and strengthening AMD’s foothold in the AI and high-performance computing (HPC) markets. The move to UDNA marks a significant transition, with full-scale adoption anticipated after the release of the RDNA 4 generation.
- Waymo Giving 100,000 Robotaxi Rides Per Week But Not Making Any Money. Waymo is now delivering over 100,000 paid autonomous rides per week in San Francisco, Phoenix, and Los Angeles, a figure that has doubled since May. Despite this growth, the company remains unprofitable, with Google’s experimental division facing a $2 billion operating loss. The high costs of vehicles and city mapping, along with ongoing public hesitation, continue to hinder Waymo’s journey to profitability.
- iOS 18.1 with Apple Intelligence launches in October, and more languages rolling out over time. Apple announced that Apple Intelligence will launch in beta with iOS 18.1 in October, initially available exclusively for US English users.
- Bringing generative AI to video with Adobe Firefly Video Model. Adobe’s Firefly Video Model introduces AI-driven tools to video editing programs such as Premiere Pro. Set to launch in beta later this year, the model provides editors with improved workflows, enabling them to experiment with creative concepts, fill gaps in timelines, and incorporate new elements into their videos.
- Mistral releases Pixtral 12B, its first multimodal model. French AI startup Mistral has introduced Pixtral 12B, a multimodal model with 12 billion parameters designed to handle both images and text. The model, accessible through GitHub and Hugging Face, can be fine-tuned and is available under the Apache 2.0 license. This release comes after Mistral secured $645 million in funding, strengthening its role as a key player in Europe’s AI industry.
- Elon Musk says Tesla has ‘no need’ to license xAI models. Elon Musk has refuted claims that Tesla will share revenue with his AI startup xAI in exchange for using its AI models. He explained that while Tesla has gained from xAI engineers’ expertise, it doesn’t need to license xAI’s models. Musk also noted that xAI’s large models are incompatible with Tesla’s vehicle computers.
- Apple is thinking about a rival to Meta Ray-Ban glasses. Apple might be developing non-AR smart glasses, positioning them as potential competitors to Meta’s $299 Ray-Ban glasses, which also lack AR functionality. Meta’s glasses come equipped with features like a camera and an AI chatbot. By excluding AR capabilities, Apple’s glasses could be more affordable, lighter, and have improved battery life due to reduced complexity.
- OpenAI in talks to raise funds at $150B valuation, Bloomberg says. OpenAI is in talks to raise $6.5B from investors at a valuation of $150B, people familiar with the matter told Bloomberg
- Meta fed its AI on almost everything you’ve posted publicly since 2007. Unless you’re in the EU, there’s no ability to opt out of AI training settings that keep Facebook or Instagram posts public.
- Google is using AI to make fake podcasts from your notes. Google’s NotebookLM app can now generate ‘lively’ audio discussions with two AI hosts about the documents you’ve given it.
- Introducing OpenAI o1-preview. OpenAI has launched its latest model, designed to think carefully before responding. It was trained using reasoning processes, allowing it to take time to deliberate before providing an answer. This approach has resulted in superhuman performance in certain areas. Initially, users will be limited to around 30 queries per week, though OpenAI plans to remove this restriction shortly.
- Google is now rolling out Gemini Live to free users on Android. Google is launching Gemini Live, its conversational AI tool, to all free Android users following a month of early access for advanced users. With this feature, users can interrupt responses to provide new information and receive text transcripts of their conversations. While extensions like Gmail are not yet supported, Gemini Live introduces ten new voice options, with additional features expected to be added soon.
- Sergey Brin says he’s working on AI at Google ‘pretty much every day’. Google co-founder and ex-Alphabet president Sergey Brin said he’s back working at Google “pretty much every day” because he hasn’t seen anything as exciting as the recent progress in AI — and doesn’t want to miss out.
- Amazon starts testing ads in its Rufus chatbot. Amazon’s shopping chatbot, Rufus, will soon incorporate sponsored ads, displaying them based on the user’s search queries and the context of their conversations.
Resources
- OLMoE: Open Mixture-of-Experts Language Models. Presents a fully open large language model (LLM) that utilizes a sparse Mixture-of-Experts approach. OLMoE is a 7B parameter model with 1B active parameter per input token. An instruction-tuned version is also available, which reportedly surpasses the performance of Llama-2–13B-Chat and DeepSeekMoE 16B.
- Large Language Model-Based Agents for Software Engineering: A Survey. A survey paper on large language model (LLM)-based agents in software engineering, offering insights across various areas such as requirements engineering, test generation, and software maintenance.
- DepthCrafter: Generating Consistent Long Depth Sequences for Open-world Videos. Researchers were able to produce very accurate depth information without requiring any camera posture or optical flow information by using Stable Diffusion video as a prior model.
- SmileyLlama: Modifying Large Language Models for Directed Chemical Space Exploration. Using DPO-style data and supervised fine-tuning on open-source language models, LLMs can be trained to produce compounds with intriguing features for potential medicinal development.
- Running a LLM on the ESP32. This code demonstrates how to execute a small language model on an Arduino board, showcasing the process of deploying and running AI models on resource-constrained hardware.
- DocAI. This is another example of effectively leveraging existing models to extract structured information from documents, demonstrating the innovative use of pre-trained AI models to automate data extraction tasks efficiently.
- FluxMusic. Text-to-music generation using a rectified flow transformer involves converting text inputs into musical compositions by utilizing a model that combines transformer architectures with rectified flow techniques. This approach enhances the model’s ability to generate coherent and diverse music sequences based on textual descriptions.
- iText2KG: Incremental Knowledge Graphs Construction Using Large Language Models. iText2KG is a Python package that leverages large language models to extract entities and relationships from text, progressively constructing consistent knowledge graphs. This tool automates the process of transforming unstructured text into structured knowledge, allowing for the incremental growth of comprehensive knowledge graphs.
- Multimodal RAG using ColPali (with Byaldi) and Qwen2-VL. Merve has created a great resource for using language and vision models to improve retrieval.
- Awesome-Text2X-Resources. This is an open collection of state-of-the-art (SOTA) and novel Text-to-X methods (where X can represent any output, such as images, audio, or 3D models). The collection includes papers, code, and datasets, aimed at staying up-to-date with the expected surge in research developments in this area over the coming months.
- Qihoo-T2X: An Efficiency-Focused Diffusion Transformer via Proxy Tokens for Text-to-Any-Task. The Proxy Token Diffusion Transformer optimizes diffusion transformers by minimizing redundant computations, employing a reduced set of representative tokens for attention processing. This approach enhances efficiency while maintaining model performance.
- UniDet3D: Multi-dataset Indoor 3D Object Detection. UniDet3D is a robust 3D object detection model designed to operate across multiple indoor datasets, delivering strong performance in identifying and detecting objects in three-dimensional spaces.
- Starst3r. This innovative tool leverages Mast3r along with smart optimizations to efficiently reconstruct 3D scenes from just a few 2D images, offering impressive results with minimal input.
- simple_tma. Image processing and cropping that can be run on the GPU.
- Lexicon3D. In a recent study comparing seven visual encoding models for 3D scene understanding, researchers found that the most effective model varied based on the specific task. DINOv2 emerged as the top performer overall, while video models excelled in object-level tasks, and diffusion models outperformed others in geometric tasks. Surprisingly, models pre-trained on language showed notable limitations in this context.
- One-DM:One-Shot Diffusion Mimicker for Handwritten Text Generation. The One-DM model generates handwritten text that can imitate any style using only a single sample as a reference. This approach allows for highly personalized handwriting generation with minimal input data.
- optillm. Optillm assists in optimizing prompts by utilizing various well-established research algorithms, including Monte Carlo Tree Search, Z3 solvers, and Self Consistency, to improve performance.
- Train Till You Drop: Towards Stable and Robust Source-free Unsupervised 3D Domain Adaptation. Researchers tackled the challenge of source-free unsupervised domain adaptation for 3D semantic segmentation by implementing regularization techniques and proposing a new criterion to improve adaptation performance.
- Memory-Efficient Optical Flow. HCVFlow is a newly developed memory-efficient optical flow method designed to address the high computational demands of all-pairs cost volumes in high-resolution images.
- Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models. Concept Sliders offer a powerful mechanism for controlling the output of diffusion models. Recent efforts have been made to integrate them with the new Flux suite of models, enhancing their functionality and adaptability.
- Minifying HTML for GPT-4o: Remove all the HTML Tags. Converting HTML to plain text can significantly reduce costs with minimal performance loss in GPT-4o for data extraction tasks. Tests on the Mercury Prize dataset demonstrated that GPT-4o performs effectively even without the HTML structure, and GPT-4o mini offers a cost-efficient solution for handling unstructured questions. For structured extraction tasks, it’s advisable to test both versions to find the right balance between cost and accuracy.
- Prompt2Fashion: An automatically generated fashion dataset. This dataset, created with large language models, curates outfit recommendations for various occasions, styles, and body types, providing high-quality and relevant suggestions.
- Sources of Uncertainty in 3D Scene Reconstruction. Researchers are improving 3D scene reconstruction techniques such as Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (GS) by incorporating uncertainty estimation methods. Although these approaches produce high-quality renders, they face challenges in addressing uncertainties caused by noise, occlusions, and camera inaccuracies.
- 🦙🎧 LLaMA-Omni: Seamless Speech Interaction with Large Language Models. Llama Omni is a speech input-output model built on Llama 3.1 8B, designed to operate with extremely low latency while maintaining high-quality responses.
- AWS AI Stack. This ready-to-use, full-stack boilerplate project is designed for building serverless AI applications on AWS. It is ideal for developers looking for a reliable AWS foundation for AI apps and seamless access to powerful LLM models through Bedrock while ensuring your app’s data remains separate from model providers.
- Internet of Agents. The Internet of Agents (IoA) is a novel framework aimed at enhancing multi-agent collaboration by enabling more efficient integration of diverse third-party agents.
- ell: The Language Model Programming Library. Ell is a newly released package developed by a former OpenAI scientist, designed to manage prompts as code, streamlining the process of working with prompts in AI applications.
- EMO-Disentanger. This research employs a two-stage model to separate and analyze emotive elements in piano music generation, enabling more expressive and nuanced performances.
- Reader-LM: Small Language Models for Cleaning and Converting HTML to Markdown. Jina has unveiled two cutting-edge models capable of transforming noisy HTML into clean, structured Markdown, optimized for training and reasoning tasks.
- Agent Workflow Memory. Agent Workflow Memory (AWM) is a technique that enables language model-based agents to learn and retain reusable task workflows from previous experiences, allowing them to effectively manage complex, long-horizon tasks.
- Hi3D-Official. Hi3D is a novel model designed to improve the generation of multi-view consistent, high-resolution 3D images from a single input. By using a video diffusion technique, it addresses the limitations of traditional 2D methods that lack 3D awareness, leveraging temporal consistency from video models to enhance geometric coherence across different views.
- Fine Tuning Llama 3.1 405B with Axolotl on a Lambda 1-Click Cluster. Axolotal AI has collaborated with Lambda Labs to demonstrate how their one-click cluster can be used to fine-tune the Llama 3.1 405B model. Although the process requires 64 GPUs, the new tools make it possible with minimal infrastructure setup, streamlining the process significantly.
- super-benchmark. SUPER is a newly introduced benchmark aimed at evaluating how effectively large language models (LLMs) can replicate tasks sourced from research repositories.
- Using GPT-4o for web scraping. An AI-powered web scraper, utilizing OpenAI’s GPT-4o, is designed to extract structured data from HTML tables. While it performs well on simple tables, its results are mixed when dealing with more complex tables, such as those with merged rows or intricate structures.
Perspectives
- ‘If journalism is going up in smoke, I might as well get high off the fumes’: confessions of a chatbot helper. Journalists and other writers are employed to improve the quality of chatbot replies. The irony of working for an industry that may well make their craft redundant is not lost on them
- Will AI make us overconfident? Students are increasingly turning to AI tools like ChatGPT to tackle complex research challenges, surprising educators with their swift advancements. AI-powered development tools, particularly in coding, greatly enhance both ambition and productivity, though they also introduce risks of overconfidence and mistakes. Despite occasional inaccuracies, AI offers valuable interactive starting points for difficult tasks, potentially fostering more active learning and encouraging exploration across disciplines.
- LLMs struggle to explain themselves. An interactive demo was employed to evaluate large language models’ (LLMs) ability to recognize and explain number sequences produced by random programs. The findings revealed that although LLMs often correctly identified the sequences, their explanations of the underlying patterns were frequently inaccurate. This underscores the limitations of LLMs’ reasoning capabilities, despite their strong performance on standardized tests.
- No more free pass: Regulation starts to crack down on social media platforms. The arrest of Telegram’s CEO in France and the closure of X in Brazil are two of the latest signs that times are changing, with networks beginning to be held more accountable
- Here’s how 7 news audience directors are thinking about Google’s AI Overviews. Google’s AI Overviews, which use the Gemini language model, received significant criticism for inaccuracies and potentially harmful recommendations following their launch in the U.S. Despite the negative feedback, Google extended the feature to six additional countries, sparking concerns among publishers about decreased web traffic and distorted content. AI experts and SEO specialists stress the importance of transparency and improved citation methods to preserve trust and ensure consistent traffic.
- Diffusion is spectral autoregression. Diffusion models and autoregressive models share a fundamental similarity, as both rely on iterative refinement processes. The author demonstrates, using Fourier transform techniques, that diffusion models function similarly to approximate autoregression in the frequency domain, especially for visual data. This insight suggests promising pathways for unifying generative modeling approaches across various data types.
- Why We Fear Diverse Intelligence Like AI. The emergence of AI and various forms of intelligence is blurring traditional distinctions between “real beings” and machines. Rather than centering discussions only on AI, it’s important to recognize and ethically interact with a broad range of cognitive systems, including bioengineered, robotic, and hybrid entities. By broadening our understanding of intelligence and fostering compassion, we can better navigate the ethical challenges posed by these rapidly evolving technologies.
- SGSeg: Enabling Text-free Inference in Language-guided Segmentation of Chest X-rays via Self-guidance. SGSeg is a segmentation framework for chest X-rays that incorporates language guidance during training but allows for text-free inference during the prediction phase.
- Are novelists who worry about the rise of AI really ‘classist and ableist’?An international writing organization appeared to greenlight the use of AI, prompting anger, the resignation of four board members, and an entire creative community to ask: ‘What?!’
- AI Chatbots Have a Political Bias That Could Unknowingly Influence Society. A new study has uncovered strong evidence that we can now add political bias to that list, further demonstrating the potential of the emerging technology to unwittingly and perhaps even nefariously influence society’s values and attitudes.
- How influencers and algorithms mobilize propaganda — and distort reality. The engagement-fuelled logic of social media has bequeathed us a world in which what’s trending is a yardstick for what’s true.
- Artificial intelligence can help to make animal research redundant. One alternative in its early stages is artificial intelligence (AI), whereby generative adversarial networks produce animal data. However, there remains a disconnect between AI-generated animal data and human safety data. Computer models that simulate complex human physiological processes could close this gap, with AI used to analyze the resulting data sets.
- Wikipedia is facing an existential crisis. Can gen Z save it? The world’s most important knowledge platform needs young editors to rescue it from chatbots — and its own tired practices
- AI-Generated Junk Science Is Flooding Google Scholar, Study Claims. New study claims to have uncovered a disturbing trend in the world of academic research: AI tools like ChatGPT being used to produce fake scientific papers that are infiltrating Google Scholar, one of the most widely used academic search engines.
- Will the “AI Scientist” Bring Anything to Science? Researchers have created an AI tool capable of automating scientific workflows, from generating hypotheses to executing experiments and drafting research papers. While its accuracy and coherence require further development, critics warn that AI’s role in simulations, such as in quantum computing and materials science, may lead to narrower research questions and less impactful findings. Supporters, however, see potential in using this AI to streamline the early stages of research, helping scientists conceptualize and define their projects more efficiently.
- Is AI Quietly Sabotaging Itself — And The Internet? Amid the growth of AI content online, a group of researchers at Cambridge and Oxford universities set out to see what happens when generative AI tools query content produced by AI. What they found was alarming.
Meme of the week
What do you think about it? Some news that captured your attention? Let me know in the comments
If you have found this interesting:
You can look for my other articles, and you can also connect or reach me on LinkedIn. Check this repository containing weekly updated ML & AI news. I am open to collaborations and projects and you can reach me on LinkedIn. You can also subscribe for free to get notified when I publish a new story.
Here is the link to my GitHub repository, where I am collecting code and many resources related to machine learning, artificial intelligence, and more.
or you may be interested in one of my recent articles:
