AIR#4 - March 8th, 2024
Dive into today's edition of AIR: The AI Recon, where we unpack a treasure trove of AI marvels and misadventures that are shaping our digital and real worlds in ways we're just beginning to grasp.
Imagine a world where uttering "AI, no ads please" could topple a trillion-dollar industry, or where personal AIs like Inflection-2.5 charm us with empathy while being light on our processors. We've got the inside scoop on the drama at OpenAI that nearly saw Sam Altman ousted, and how Reddit's AI bouncer is setting the stage for a safer online community.
Meta's bold move to build a giant AI model could change how we consume videos, while Elon Musk's cheeky challenge to OpenAI adds a twist to the tale. From accusations of data scraping between AI titans to the AI industry's flirtation with nuclear power for a sustainable future, we're covering the ground that matters.
And for those who've ever been spooked by a scam call, the rise of voice cloning scams is a chilling reminder of AI's darker potential. With nearly 40% of us unable to distinguish between real and AI-generated faces, the stakes for digital literacy have never been higher. Meanwhile, the developer and scholarly communities are buzzing with breakthroughs like GaLore's memory-efficient training methods and Coalescence's speed boost for LLMs, not to mention the ethical and existential ponderings over AGI's elusive horizon.
So, whether you're a business mogul, a code wizard, or an AI ethicist, today's AIR: The AI Recon is your guide to the fascinating, fast-evolving world of artificial intelligence. Sip that coffee and let's unravel the future together!
Business
🔥 "AI, no ads please": 4 words to wipe out $1T
"AI's potential to decimate $1T ad industry explored. 'AI, no ads please' could redefine content consumption, challenging tech giants."
Inflection-2.5: meet the best personal AI
Inflection-2.5, blending IQ with EQ, rivals GPT-4 using 40% less compute, enhancing Pi's user interaction with empathy and advanced capabilities.
Key OpenAI executive played a pivotal role in Sam Altman's ouster
OpenAI's Sam Altman briefly ousted after CTO Murati and co-founder Sutskever raise concerns about his leadership, sparking company turmoil.
Reddit rolling out AI bouncer to halt harrassment
Reddit introduces AI-powered harassment filter to help moderators, enhancing safety and preparing for IPO with a focus on AI integration.
Meta is building a giant AI model to power its 'entire video ecosystem'
Meta invests billions in AI to unify video recommendations across platforms, aiming for efficiency and engagement boosts by 2026.
Musk to OpenAI: Change your name to ClosedAI and I will drop the lawsuit
Elon Musk challenges OpenAI to rebrand as ClosedAI to avoid lawsuit.
Midjourney accuses StabilityAI of scraping their image data and bans employees
Midjourney accuses StabilityAI of data scraping, causing service disruption; bans their employees immediately.
The AI industry is pushing a nuclear power revival – partly to fuel itself
AI's energy hunger drives tech giants to invest in nuclear power, aiming for a sustainable future but facing regulatory hurdles.
The Terrifying A.I. Scam That Uses Your Loved One's Voice
Beware: AI scams now use cloned voices of loved ones to demand money. Always verify before acting. #VoiceCloningScam
Nearly 40% fooled by AI-generated faces
Nearly 40% can't spot AI-generated faces, revealing a gap in our digital literacy and raising alarms over disinformation tools.
Andreessen Horowitz Is Planning to Raise Two AI-Focused Funds
Andreessen Horowitz eyes AI future, plans to raise $6.9 billion for two AI-focused funds.
India announces $1.2B investment in AI projects
India greenlights $1.25B for AI projects to boost computing infrastructure, support startups, and develop public sector applications.
Engineering
🔥 The Pile is a 825 GiB diverse, open-source language modelling data set (2020)
The Pile, an 825 GiB open-source dataset, boosts AI's cross-domain knowledge and generalization, proving a robust benchmark for language models.
Gemini won't return C++ coding help if you're under 18 "to preserve your safety"
Gemini blocks C++ help for under 18s, citing safety. Yann LeCun: "Wat?"
Jetbrains unbundles AI Assistant and is now available as a separate plugin
JetBrains' IntelliJ IDEA 2024.1 Beta is out, featuring Java 22 support, enhanced dev workflows, and AI Assistant now a separate plugin for flexible use.
Neural Chess
AI enthusiast trains neural network to play chess, drawing inspiration from a video on AI playing Pokemon. Despite a broad AI background, practical challenges in chess rules and neural network training prove formidable. #NeuralChess
Thanks to AI, the coder is no longer king: All hail the QA engineer
AI shifts software development: Coders speed up with AI but QA engineers face more work as they tackle AI-generated bugs and flaws. The balance of power changes.
Researchers jailbreak AI chatbots with ASCII art
Researchers use ASCII art to bypass AI chatbot safety, tricking them into advising on illegal activities.
Evolving from Rule-Based Classifier: ML Powered Auto Remediation
Netflix's ML Auto Remediation cuts Spark job errors by 56% and costs by 50%, blending rule-based classification with ML for seamless, automated fixes.
LLMs On-Device with MediaPipe and TensorFlow Lite
Google's MediaPipe & TensorFlow Lite now enable LLMs to run on devices, offering cross-platform support and optimizing memory and compute use.
GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection
GaLore revolutionizes LLM training, slashing memory use by up to 82.5% without sacrificing performance, enabling 7B model training on consumer GPUs.
Coalescence: Making LLM inference 5x faster
Coalescence framework boosts LLM structured generation speed by 5x, optimizing deterministic output without sacrificing quality.
Building on quicksand – the state of modern AI app development
AI app development in 2024: Innovate fast but wisely, focus on configurability and prioritize experiments. Building stable on shifting sands.
Ix: Autonomous GPT-4 agent platform
Ix platform launches, revolutionizing task automation with GPT-4 agents. Code, analyze data, and more with no-code AI agents.
Prompt Decomposition and Reconstruction Makes Powerful LLM Jailbreakers
DrAttack boosts LLM jailbreak success to 78% with fewer queries by splitting harmful prompts into less detectable sub-prompts.
The Claude 3 system prompt, explained by Anthropic
Amanda Askell unveils Claude 3's system prompt, breaking it down for clarity. Dive in for a closer look!
Evaluating LLM Models at Scale
Mozilla.ai leads in LLM evaluation at NeurIPS 2023, focusing on fine-tuning models efficiently and developing robust evaluation methods.
Google's Genie game maker is what happens when AI watches 30K hrs of video games
Google's Genie transforms 30K hrs of game footage into playable environments, showcasing AI's leap in creating interactive worlds.
Academic
OpenAI's "Own Goal"
OpenAI faces criticism for straying from its original open-source mission, amid lawsuits and concerns over ethics and transparency.
What if AGI is not coming?
Exploring the possibility that AGI may never arrive, amid scaling limits and lack of breakthroughs. Will billions in AI investment yield more than sophisticated hallucinations?
Social learning: Collaborative learning with large language models
Google Research explores social learning for LLMs, enabling them to share knowledge without compromising privacy, mirroring human learning methods.
Gemini and Claudes are killing RAG?
Despite Gemini's prowess in long-context processing, RAG's unique blend of external data integration ensures it remains vital in AI's evolving landscape.
Yann LeCun: Meta AI, Open Source, Limits of LLMs, AGI and the Future of AI [video]
Yann LeCun explores Meta AI, open source dynamics, LLMs' limits, AGI, and AI's future on Lex Fridman Podcast #416.
SaulLM-7B: A Pioneering Large Language Model for Law
Meet SaulLM-7B, the first 7B parameter AI designed to revolutionize legal document processing and comprehension.
The First Global Generative AI Landscape 2024
AIport launches first truly global GenAI landscape, showcasing 35 countries' advancements beyond the usual tech giants.