Paralyzed individuals regain speech through innovative neurotech

Pivot 5: 5 stories. 5 minutes a day. 5 days a week.

1. Paralyzed individuals regain speech through innovative neurotech

Two research groups—one at Stanford University and the other at the University of California, San Francisco—have developed brain-computer interfaces that have enabled two paralyzed individuals to speak again. These BCIs translate brain activity into speech and can even digitally convey facial expressions, offering a glimpse into the future of assistive technology.

The BCIs use algorithms trained to recognize phonemes, the basic building blocks of speech, from neural activity. The Stanford participant, a 68-year-old woman with ALS, communicated at an average of 62 words per minute with a 24% error rate. The UCSF participant, a 47-year-old woman paralyzed due to a brainstem stroke, communicated at 78 words per minute with a 25% error rate.

While the technology is promising, it still has room for improvement, particularly in decoding accuracy. The UCSF team even created a digital avatar capable of expressing facial emotions based on the participant's neural activity. As BCIs continue to evolve, they hold the potential to dramatically improve the quality of life for individuals with paralysis and other conditions that impair speech.

Read the full story here 

2. Nvidia's Q2 earnings skyrocket, driven by Data Center growth

Nvidia has set new benchmarks in the tech industry, reporting a record Q2 revenue of $13.51 billion, a staggering 101% increase year-over-year. The data center segment has been the linchpin, growing 171% annually to $10.32 billion and accounting for 76% of the total revenue. "The new Nvidia computing era has begun," said Jensen Huang, Nvidia's CEO, emphasizing the company's focus on large language models.

The company's gross margin soared to 71.2%, up from 45.9% a year ago. This increased profitability is being channeled back into research and development. Nvidia aims to launch a new product or platform every six months, focusing on generative LLMs.

Nvidia's commitment to innovation is evident in its R&D spending, which rose 10% to $2.04 billion in Q2 FY2024. This investment is geared toward next-generation platforms like its Hopper architecture, designed to maintain Nvidia's lead in workload accelerated computing. With its focus on complete infrastructure, Nvidia is well-positioned to meet the rising enterprise demand for LLM solutions.

Read the full story here 

3. YouTube experiments with Hum-to-Search song identification on Android

YouTube is experimenting with a new feature that allows Android users to identify songs by humming, singing, or recording them. This search-by-song capability is currently available to a select group of Android users and aims to enhance the music discovery experience on the platform.

Building on a similar capability launched by parent company Google in 2020, YouTube's feature requires users to hum, sing, or record a song for just three seconds. The platform then identifies the tune and directs users to relevant YouTube videos, whether it be the official music video, user-generated content, or Shorts.

Both YouTube's and Google's features are built on machine learning models that can match a person's hum to a song's "fingerprint" or signature melody. While other apps like SoundHound and MusixMatch offer similar capabilities, they aren't as popular as YouTube and Google. As YouTube continues to test this feature, it could become a significant tool for song discovery.

Read the full story here

4. OpenAI unveils fine-tuning for GPT-3.5 Turbo, offering customers to train ChatGPT on their own data

OpenAI has rolled out a fine-tuning feature for its GPT-3.5 Turbo model, allowing users to train the model with custom data via its API. This new capability aims to enhance the model's performance in specific tasks and can even rival GPT-4 in certain use-cases. "Fine-tuning allows developers to create unique and differentiated experiences for their users," OpenAI stated.

The fine-tuning process not only improves the model's performance but also adds functionalities like improved steerability, reliable output formatting, and custom tone. "Early tests have shown a fine-tuned version of GPT-3.5 Turbo can match, or even outperform, base GPT-4-level capabilities on certain narrow tasks," OpenAI reported.

However, these advancements come at a price. Training costs are set at $0.008 per 1,000 tokens, and usage costs are $0.012 per 1,000 tokens for text input and $0.016 for text output. Despite the higher costs, OpenAI ensures data privacy and uses GPT-4 for moderation purposes.

Read the full story here 

5. U.S. faces calls for stronger AI regulation

The United States is grappling with the need for stronger regulation of artificial intelligence, particularly in the wake of the viral success of generative platforms like ChatGPT. Concerns about job displacement and the unlawful exploitation of intellectual property have led to a series of legal actions. "The U.S. is lagging in tech regulation, and it's time to catch up," said a policy analyst.

Europe has already taken significant steps in this direction with the passage of the AI Act, considered the world's first comprehensive tech law. The act has received mixed reviews, with over 150 executives signing an open letter to the European Commission, expressing concerns about the legislation's aggressive policies.

In response to growing concerns, the White House has been urging companies to pledge to develop tech responsibly. Key figures in the industry, including Microsoft, Google, and OpenAI, were scheduled to meet at the White House to discuss protective measures. As tech continues to gain attention on Wall Street and among global leaders, the push for responsible and comprehensive regulation is becoming increasingly urgent.

Read the full story here 

Listen to the daily Pivot 5 podcast on Spotify, Apple Podcasts, and more