• AI Time to Impact
  • Posts
  • . . AI: Startup Phind Reportedly Outperforms GPT-4 Turbo (2.23.24)

. . AI: Startup Phind Reportedly Outperforms GPT-4 Turbo (2.23.24)

Perplexity, NVIDIA and more

Friends, there are 4 hot stories in AI today but one big one: YC startup Phind has released a new model it says beats GPT-4 Turbo. Comparable in quality, 4X as fast, it's impressive. It's easy to test and it is really nice.

By the way, of our 4 stories today, 2 have not been covered at all yet by the biggest AI newsletters on the market or Techmeme. And one, the top story about Phind, got one sentence yesterday in two of the top AI newsletters. Here’s some more context and a little actual testing.

If you find this newsletter useful, I hope you’ll share it with a friend. Want to know what the most-clicked link of the week was among our readers? #1 was last week’s top link again, Universal Time Series Forecasting, #2 Stable Cascade image generation, #3 the fractals found in training data!

Now here’s today’s news.

Marshall Kirkpatrick, Editor

First impacted: AI developers, Data scientists
Time to impact: Short

YC company Phind has launched Phind-70B, a model they report outperforms GPT-4 Turbo with an 82.3% score on the HumanEval benchmark. The company says that Phind-70B, derived from the CodeLlama-70B model and fine-tuned with an additional 50 billion tokens, operates four times faster than GPT-4 Turbo through the use of NVIDIA's TensorRT-LLM library on H100 GPUs. The company says it melted one of the H100s while training the system. You can test 10 free prompts per day, or sign up for $10 and $20 monthly plans. In our initial testing, it works very well, very fast. [Introducing Phind-70B – closing the code quality gap with GPT-4 Turbo while running 4x faster] Explore more of our coverage of: AI Models, Performance Evaluation, NVIDIA Technologies.

First impacted: AI Researchers, Robotic Engineers
Time to impact: Medium

NVIDIA's new GEAR team, co-founded by Dr. Jim Fan and Prof. Yuke Zhu, says it's working on developing foundational models for AI agents that can operate in both virtual and real-world environments.The group says it aims to concentrate on multimodal base models, general-purpose robots, and simulation and synthetic data, with a goal to build large action models that can autonomously explore and continuously bootstrap their capabilities across different games and simulations. That means autonomous robots running around the real world, learning things. See Fan's thread on X highlighting some of his team's relevant work so far. https://twitter.com/DrJimFan/status/1761052023821369639 [NVIDIA Deep Imagination Research Homepage] Explore more of our coverage of: NVIDIA GEAR, AI Agents, Simulation and Synthetic Data.

First impacted: Investors, Robotics Engineers
Time to impact: Medium

WIRED's profile of Jensen Huang: NVIDIA has seen a 231% rise in stock price over the past year, pushing the company towards a $2 trillion valuation. Despite facing challenges due to the US-China tech conflict, Huang remains hopeful about the future of AI, noting its potential in future robotics and an undisclosed data center project Nvidia refers to as an "AI factory", while also stating that their tech has enhanced AI about a million-fold in the last decade. [Nvidia CEO Jensen Huang Is Powering the AI Revolution] Explore more of our coverage of: Nvidia Stock Growth, AI Development, US-China Tech Conflict.

First impacted: Podcast listeners, Perplexity AI and ElevenLabs customers
Time to impact: Medium

Perplexity AI has collaborated with voice company ElevenLabs to launch a podcast named "Discover Daily." The first episode produced by this combination of technology from two companies that have raied a combined $200M is a 3 minute overview of top news stories narrated by a robot David Attenborough knock-off voice. [Discover Daily by Perplexity] Explore more of our coverage of: AI Podcasts, Perplexity AI, ElevenLabs Collaboration.

That’s it! More AI news on Monday!