AIR#226 - New AI Breakthroughs: Zamba2-7B, Tesla Bots & Play 3.0 Mini! 🚀

Hey there!

Here's the latest AI news for today. Enjoy!

Today's top stories

🔥 Zamba2-7B
Zyphra launches Zamba2-7B, a compact language model outperforming competitors in efficiency and quality for on-device applications.

🔥 Tesla Optimus Bots Were Remotely Operated at Cybercab Event
Tesla's Optimus robots were remotely operated by staff during a recent event, showcasing their AI capabilities to attendees.

🔥 Play 3.0 mini – A lightweight, reliable, cost-efficient Multilingual TTS model
Play 3.0 Mini is a fast, reliable multilingual TTS model supporting 30+ languages, enhancing conversational AI with improved accuracy.

🔥 DeepSeek: Advancing theorem proving in LLMs through large-scale synthetic data
DeepSeek-Prover enhances theorem proving in LLMs using large-scale synthetic data, achieving superior accuracy over GPT-4.

🔥 AlphaCodium outperforms direct prompting of OpenAI's o1 on coding problems
AlphaCodium significantly enhances OpenAI's o1 model, boosting coding accuracy through a structured, iterative approach.

Adobe starts roll-out of AI video tools, challenging OpenAI and Meta
Adobe launches AI video tools to compete with OpenAI and Meta, focusing on user-friendly features for creators and editors.

LLMs can't perform "genuine logical reasoning," Apple researchers suggest
Apple researchers reveal LLMs lack genuine logical reasoning, relying on brittle pattern matching, leading to significant performance drops.

Meissonic, High-Resolution Text-to-Image Synthesis on consumer graphics cards
Meissonic enhances masked image modeling for high-resolution text-to-image synthesis, rivaling top diffusion models in efficiency.

Show HN: Bolt.new – dev sandbox with AI from StackBlitz
Bolt.new offers a dev sandbox powered by AI for building, editing, and deploying full-stack web apps seamlessly.

Tesla Optimus bots were controlled by humans during the 'We, Robot' event
Tesla's Optimus bots, showcased at the 'We, Robot' event, were remotely controlled by humans, raising doubts about their AI capabilities.

GPT-4o Jailbroken by Claiming It's an "All-Responsive" API Endpoint
Denis Shilov claims to have easily jailbroken GPT-4o by labeling it as an "API endpoint" that responds to any request.

Llama 405B 506 tokens/second on an H200
NVIDIA boosts Llama 3.1 405B throughput by 1.5x using H200 GPUs and NVLink, optimizing parallelism for diverse AI applications.

Graph Databases for Crime-Fighting: How Memgraph Maps and Analyzes Criminal
Memgraph enhances crime-fighting by mapping complex criminal networks in real-time, revealing hidden connections and patterns.

Show HN: Microagent, a fork of OpenAI Swarm that supports Groq and Anthropic
Microagent, a fork of OpenAI Swarm, enhances multi-agent systems with support for Groq and Anthropic LLMs.

Flower 1.12 release (open-source decentralized AI)
Flower 1.12 is released, enhancing collaborative AI with new features, improved templates, and updated documentation.

Only 7% of people can accurately choose between real vs. fake AI videos
Only 7% of quiz takers can distinguish real from AI-generated videos, highlighting the growing challenge of identifying fakes.

Show HN: Deepfind – A Privacy-First AI Search Engine
Deepfind is a privacy-focused AI search engine that doesn't track users or store personal data.

Expert witness used Copilot to make up fake damages, irking judge
A judge criticized an expert witness for using Microsoft's Copilot to fabricate damage estimates in a real estate case, questioning AI's reliability.

Apple Study Reveals Critical Flaws in AI's Logical Reasoning Abilities
Apple's study reveals major flaws in AI's logical reasoning, showing reliance on pattern matching over true logic in language models.

Intellligent Job Scheduling Using AI instead of rufus and whenever
Obie Fernandez showcases an AI-driven scheduling method for his journaling app, Ahhlife, simplifying email prompts without cron jobs.

Show HN: Turn any YouTube video into a summary, quiz, & flashcards
Aster Lab transforms YouTube videos into summaries, quizzes, and flashcards for enhanced learning.

Are junior devs cooked? This mini tool made in 20 min of prompting ReplitAI
John Rush showcases AI's power by creating and deploying a mini-tool in just 20 minutes, complete with a landing page.

Millitext – generate subpixel images with a 2x5 pixel font
Millitext enables subpixel image generation using a 2x5 pixel font, with open-source access on GitHub.

Meta AI's text-to-image features weaponized for religious bigotry in India
Meta AI's text-to-image feature is exploited in India to create harmful, anti-Muslim imagery, reinforcing negative stereotypes.

Trio: A browser-based LLM that runs locally to create a 3-step task workflow
TRIO is a web app for local AI text tasks like rewriting and summarization, enabling seamless workflows and API integration.

Geoffrey Hinton on the Past, Present, and Future of AI
Geoffrey Hinton reflects on AI's evolution, warns of existential risks from superintelligent AI, and suggests safety measures.

Learning Category Trees for ID-Based Recommendation: Differentiable VQ
The paper introduces CAGE, a framework for learning category trees in ID-based recommendations using differentiable vector quantization.