AIR#9 - March 13th, 2024
Good morning, AI enthusiasts! Today's AIR: The AI Recon is packed to the brim with stories that'll have your neurons firing on all cylinders as you sip your morning brew.
We're taking a deep dive into Meta's ambitious leap towards AGI with their announcement of two colossal 24k GPU clusters and plans for an eye-watering expansion to 350k NVIDIA H100 GPUs by 2024. Meanwhile, the tech job landscape is undergoing a seismic shift, with AI reshaping demand and salaries in surprising ways – data from 20M job postings tells the tale.
In a move that's sparking heated debate, a report is calling for the US to clamp down on AI development, suggesting making open-source AI models illegal to ward off existential threats. And for the parents out there, you might be intrigued (or concerned) to learn that your kids could be watching AI-generated YouTube content, a blend of tech and creativity that's as lucrative as it is dystopian. On the legal front, Meta's not pulling any punches, suing a former executive over alleged AI theft in a high-stakes drama of betrayal and corporate espionage. Sam Altman's bold prediction that AI will soon handle 95% of marketing work has the industry buzzing about the future. And that's just the tip of the iceberg – from Oracle's AI-driven profit surge to China's AI-powered high-speed rail innovation, and a groundbreaking tool from OpenAI that's set to revolutionize model debugging, we've got a treasure trove of stories that promise to kickstart your day with a dose of AI wonder.
So, grab that coffee, and let's embark on today's digital odyssey together – it's a journey through the cutting-edge of AI that you won't want to miss.
Business
🔥 Building Meta's GenAI infrastructure
Meta announces two 24k GPU clusters for AI, aiming for AGI with open-source tech and a massive future expansion to 350k NVIDIA H100 GPUs by 2024.
How AI is disrupting the demand for software engineers: data from 20M jobs
AI's rising demand reshapes tech jobs: ML engineers up, frontend down. Salaries stagnant despite AI boom. Data from 20M jobs reveals shifts.
Jail time for open source AI models
Report urges US to restrict AI development to prevent existential threats, including making open-source AI models illegal.
Your Kid May Already Be Watching AI-Generated Videos on YouTube
Kids may be watching AI-generated YouTube content, blending tech and creativity for quick, lucrative videos. Critics call it dystopian. YouTube's new policies aim for transparency.
Meta sues former executive over alleged AI theft
Meta sues ex-executive for stealing AI secrets and employee data, alleging breach of trust and seeking damages.
Sam Altman: AI will handle 95% of marketing work
Sam Altman predicts AI will do 95% of marketing work, revolutionizing the industry with near-instant, cost-effective solutions. AGI expected within 5 years.
China puts trust in AI to maintain largest high-speed rail network on Earth
China leverages AI for its vast high-speed rail, enhancing safety and maintenance, outpacing global rail tech with data-driven precision.
Oracle tops quarterly profit estimates on AI demand, shares surge
Oracle's shares soar 14% as AI demand boosts profits beyond forecasts, hinting at a promising Nvidia partnership.
AI-related startups raised nearly $50B in 2023
AI startups defied 2023's investment slowdown, raising nearly $50B. Amidst booming valuations, the future may demand recalibration.
Engineering
🔥 OpenAI – transformer debugger release
OpenAI launches Transformer Debugger, a tool for deep diving into language model behaviors, enhancing model interpretability and debugging.
🔥 Google Announces Stealing Part of a Production Language Model
Google's breakthrough model-stealing attack exposes hidden dimensions of major AI models for under $20, revealing a new cybersecurity frontier.
Devin, an AI Software Engineer
Meet Devin: The first AI software engineer acing coding benchmarks, passing real engineering interviews, and completing jobs on Upwork. A new tech milestone!
Andrej Karpathy on automating software engineering
Karpathy likens software engineering automation to self-driving evolution: from manual coding to AI-assisted development, predicting a future of supervisory roles and high-level input.
Midjourney debuts consistent characters for gen AI images
Midjourney introduces "consistent character" feature, enabling users to create stable character appearances across different AI-generated images.
GPT-4.5-Turbo seems to be indexed on Bing coming soon?
Matt Shumer hints GPT-4.5-Turbo may soon be integrated with Bing, sparking anticipation.
Academic
🔥 Stealing Part of a Production Language Model
Researchers expose vulnerability in AI models like ChatGPT, stealing their core data for under $20. Possible defenses discussed.
🔥 Is Cosine-Similarity of Embeddings Really About Similarity?
New study reveals cosine-similarity in AI embeddings might not truly measure similarity, urging a rethink of its use.
Large Language Models Are Neurosymbolic Reasoners
New study reveals Large Language Models excel as symbolic reasoners, achieving 88% success in text-based games, showcasing their potential beyond mere language tasks.
Using Hallucinations to Bypass GPT4's Filter
New exploit in GPT4 bypasses safety filters by inducing "hallucinations", posing a fundamental security risk.
Guardrail Baselines for Unlearning in LLMs
New study suggests simple techniques like prompting & filtering can effectively "unlearn" concepts in AI, offering a lighter alternative to fine-tuning.