Good morning! This is your daily ☕️ Techpresso.
On this day, 57 years ago, the spacecraft Surveyor 3 was launched from Cape Kennedy, Florida, marking the second US soft landing on the Moon.
In today's Techpresso:
✈️ US Air Force confirms first successful AI dogfight
🏆 Mistral's latest model sets new records for open source LLMs
🧠 Intel unveils the world's largest neuromorphic computer
🏴☠️ Billions of public Discord messages may be sold through a scraping servic
🎭 Microsoft's new AI model creates hyper-realistic video using static image
👁️ GPT-4 nearly matches expert doctors in eye assessments
🎁 + 9 other news you might like
🔮 + 4 handpicked research papers and tools
✈️ US Air Force confirms first successful AI dogfightLINK
The US Air Force, via DARPA, announced that an AI-controlled jet successfully engaged in an in-air dogfight against a human pilot for the first time, during tests at Edwards Air Force Base in California in September 2023.
DARPA has been working on AI for air combat through its Air Combat Evolution (ACE) program since December 2022, aiming to develop AI capable of autonomously flying fighter jets while adhering to safety protocols.
The AI was tested in a real aircraft, the experimental X-62A, against an F-16 flown by a human, achieving close maneuvers without the need for human pilots to intervene.
🏆 Mistral's latest model sets new records for open source LLMsLINK
French AI startup Mistral AI has released Mixtral 8x22B, claiming it to be the highest-performing and most efficient open-source language model, utilizing a sparse mixture-of-experts model with 39 billion of its 141 billion parameters active.
Mixtral 8x22B excels in multilingual support and possesses strong math and programming capabilities, despite having a smaller context window compared to leading commercial models like GPT-4 or Claude 3.
The model achieves top results on various comprehension and logic benchmarks and outperforms other models in its supported languages on specific tests.
🧠 Intel unveils the world's largest neuromorphic computerLINK
Intel Labs introduced its largest neuromorphic computer yet, the Hala Point, featuring 1.15 billion neurons, likened to the brain capacity of an owl, that aims to process information more efficiently by emulating the brain's neurons and synapses in silicon.
The Hala Point system, consuming 2,600 W, is designed to achieve deep neural network efficiencies up to 15 TOPS/W at 8-bit precision, significantly surpassing Nvidia's current and forthcoming systems in energy efficiency.
While showcasing remarkable potential for AI inference and optimization problems with significantly reduced power consumption, Intel's neuromorphic technology is not yet a universal solution for all AI workloads.
🏴☠️ Billions of public Discord messages may be sold through a scraping servicLINK
A service called Spy Pet is selling access to a database containing around 3 billion Discord messages from over 14,000 servers, tracking more than 600 million users.
Spy Pet allows customers to purchase "data credits" with cryptocurrencies to access information about users, including message history and connected accounts like GitHub.
Despite Discord's effort to make server messages less accessible to the public, Spy Pet has compiled extensive user data that could compromise privacy, although private messages are not mentioned in Spy Pet's offerings.
🎭 Microsoft's new AI model creates hyper-realistic video using static imageLINK
Microsoft introduced VASA-1, an AI model that produces hyper-realistic videos from a single photo and audio clip, featuring realistic lip syncs and facial movements.
The model can create 512x512p resolution videos at 40fps from one image, support modifications like eye gaze and emotional expressions, and even incorporate singing or non-English audio.
While Microsoft recognizes the AI's potential for misuse in creating deepfakes, it intends to use VASA-1 solely for developing virtual interactive characters and advancing forgery detection.
👁️ GPT-4 nearly matches expert doctors in eye assessmentsLINK
OpenAI's GPT-4 almost matched the performance of expert ophthalmologists in an eye assessment study, as reported by the Financial Times and conducted by the University of Cambridge's School of Clinical Medicine.
GPT-4 scored higher than trainee and junior doctors with 60 correct answers out of 87, closely following the expert doctors' average score of 66.4, in a test evaluating knowledge on various ophthalmology topics.
The study, highlighting both potential benefits and risks, indicates that while GPT-4 shows promise in medical assessments, concerns about inaccuracies and the model's tendency to "hallucinate" answers remain.
Other news you might like
Samsung shifts executives to six-day workweeks to ‘inject a sense of crisis’.LINK
Slack's new AI 'Recap' feature will send you a daily digest of important convos.LINK
Stability AI's text-to-image model now available to more developers.LINK
Airchat is the latest app trying to make 'social audio' cool again.LINK
The hidden story behind one of SpaceX’s wettest and wildest launches.LINK
Google cutting more jobs as it restructures teams as part of shift to AI.LINK
TikTok's answer to Instagram, Notes, is rolling out.LINK
Google will invest more than $100 billion in AI, DeepMind founder says.LINK
Apple cuts greenhouse emissions in half as it focuses on carbon neutral goal by 2030.LINK
Latest research and tools
Short Stack: the world’s smallest functional scale-model Nintendo Wii, featuring HDMI output for digital audio and video and a MicroSD card for game and save storage, encapsulated in a design roughly the size of a deck of playing cards.LINK
Collapse of self-trained language models: the paper explores how self-trained language models can deteriorate over time without continuous updates and oversight.LINK
RHttp: a tool designed for easy and fast manipulation of HTTP requests in a terminal with features like a built-in editor for JSON payloads, automatic response syntax highlighting, and JSON auto-formatting.LINK
Chinchilla Scaling: A Replication Attempt: the study fails to replicate the results of previous research on chinchilla growth patterns.LINK
Want to get the latest news differently? Find us on:
See you tomorrow for a new dose of ☕️ Techpresso!
-min.png){kind=small}
.png){kind=small}
