Apr 21, 2024

• Meta says you can't turn off its new AI tool on Facebook, Instagram
  Meta has integrated AI across its platforms including Instagram, Facebook, WhatsApp, and Messenger in Canada, offering features like restaurant recommendations and study aids, but cannot be disabled by users.

• Programming Is Mostly Thinking (2014)
  Programming involves significantly more thinking and decision-making than actual coding, with the physical act of typing constituting a minor fraction of the overall effort.

• Why everything is becoming a game
  Gamification is increasingly used to influence behavior across various domains, from consumer purchases to employee productivity, leveraging Skinner's insights on behaviorism and the effectiveness of immediate, unpredictable, and conditioned rewards.

• Financial market applications of LLMs
  Quantitative traders explore using Large Language Models (LLMs) for financial predictions, leveraging their autoregressive learning capabilities to model sequences of market data, akin to how they predict sequences of words.

• How important is leetcode in ML?
  Leetcode's prominence in the hiring process for Applied Data Scientist roles at FAANG companies is highlighted by a 3:1 ratio of Leetcode to ML interviews, suggesting a significant emphasis on coding skills.

• Self-reasoning tokens: teaching models to think ahead
  Self-Reasoning Tokens are introduced to incentivize models like GPT to plan for future tokens, effectively teaching them to think ahead by producing a secondary token that is only influenced by future tokens, not the immediate next one.

• Penzai: JAX research toolkit for building, editing, and visualizing neural nets
  Penzai is a JAX library designed for post-training model manipulation, offering tools for reverse-engineering, model surgery, and more, enhancing research flexibility and model understanding.

• Amazon grows to over 750k robots, replacing 100k humans
  Amazon's rapid advancement in robotics has led to the deployment of over 750,000 robots, working alongside its 1.5 million employees, marking a decrease of over 100,000 human workers since 2021.

• Kaiming He's lecture on DL architecture for Representation Learning
  Kaiming He's lecture dives deep into DL architecture for Representation Learning, highlighting historical architecture advances with unparalleled clarity.

• Fast food is already automated?
  Kernel, a new fast-casual restaurant by Chipotle's founder, utilizes robots for food preparation, aiming to revolutionize fast-food service with speed, accuracy, and lower costs.

• How to fix America's aviation system (2023)
  Near misses in U.S. aviation have increased by 25% over the past decade, attributed to outdated FAA technology and a shortage of air traffic controllers.

• Lossless Acceleration of LLM via Adaptive N-Gram Parallel Decoding
  Adaptive N-gram Parallel Decoding (ANPD) introduces a lossless method to accelerate Large Language Models (LLMs) by generating multiple tokens simultaneously, preserving the model's original output integrity.

• That IACR Preprint
  Cryptographer Yilei Chen's preprint claimed a polynomial-time quantum algorithm for lattice problems, notably solving the GapSVP problem with an approximation ratio of ~n4.5, which if correct, would have significant implications for the security of lattice-based cryptosystems.

• GPT-4 can exploit vulnerabilities by reading CVEs
  OpenAI's GPT-4 has demonstrated the ability to autonomously exploit real-world security vulnerabilities by interpreting CVE advisories, achieving an 87% success rate in exploiting vulnerabilities described in a dataset of 15 one-day vulnerabilities, including critical severity ones, as detailed in a paper by University of Illinois Urbana-Champaign researchers.

• Natural language processing in Bash (2023)
  The Bash toolchain utilizes the n-gram language model to generate random prose that mirrors the style of a given text corpus, exemplified through training on Herman Melville's Moby-Dick.