Neoteric AI News Digest 10: Small Models, Big Moves & New Standards

Neoteric AI News Digest 10: Small Models, Big Moves & New Standards

Phew! August passed in the blink of an eye, and somehow, we’re already closing out the first week of September! Time flies—and so does the AI world—so we’ve got plenty of hot news to catch up on!

In this issue, we’ll explore Microsoft’s Phi-3.5, a compact model outperforming big names, Meta and Spotify's concerns over EU regulations stifling AI innovation, and Anthropic’s push for AI transparency. But that’s not all! Have you heard about io.net and FLock.io launching the first "Proof of AI" consensus? We’re covering that too! Plus, we dive into healthcare innovations with Piramidal’s brainwave model and DeepMind’s GenRM, which lets models verify their own answers.

Ready? Let’s go!

Credit: Microsoft

Microsoft’s Phi-3.5: A Small Model Making Big Waves

Looks like Microsoft is taking on the AI heavyweights with its latest release, Phi-3.5. Despite being a “small” language model, Phi-3.5 outperforms some big names like Google’s Gemini and OpenAI’s GPT-4o on several key benchmarks. Coming in at 3.8 billion, 4.15 billion, and 41.9 billion parameters, Phi-3.5 is available for free download and can even run on local devices. It's not just about size; this model boasts a vision version that understands images, not just text, and features a mixture of expert models for more efficient processing.

What's really cool is how Phi-3.5 opens up possibilities beyond just cloud computing. Imagine running an AI like this on a smart doorbell for facial recognition without sending data to the cloud—pretty futuristic, right? However, real-world performance does show some quirks. It shines in STEM and social sciences but can fumble on simple language tasks. Still, Phi-3.5 proves that you don’t need to go big to make an impact in the AI world.

Curious about Phi-3.5's potential? Dive into the full scoop here.


Meta and Spotify Slam EU's AI Regulations as Innovation Stiflers

Mark Zuckerberg and Daniel Ek shared their concerns about the EU's new AI regulations. In joint statements, Meta and Spotify's CEOs took aim at what they see as overly restrictive AI regulations, particularly regarding open-source development. They argue that these rules are stifling innovation, making it tough for European companies to compete globally.

Zuckerberg emphasized that Meta's AI models are hampered by EU data privacy laws, preventing them from training on public data from platforms like Facebook and Instagram. Similarly, Ek pointed out that AI has been crucial to Spotify’s success in personalizing user experiences but fears EU regulations could slow the pace of open-source AI growth, affecting the creative ecosystem.

Both CEOs call for a balanced approach that allows innovation while protecting user data. But for now, it seems Europe might miss out on the latest AI advancements as companies like Meta hold back their next-gen models due to regulatory uncertainties.

For more, check out the full articles on Cointelegraph and TechCrunch.

Anthropic’s Bold Step Toward Transparency — System Prompts Release

Did you hear about Anthropic’s latest move? It seems like the company is setting a new standard for transparency in the AI industry, as they are the first ones ever to release the system prompts for its models — specifically, the Claude family.?

These prompts serve as the “operating instructions” for models, guiding their behavior, responses, and even how they handle controversial topics. Anthropic has shared details for Claude 3.5 Sonnet, Claude 3 Haiku, and Claude 3 Opus, highlighting their unique capabilities and knowledge cut-offs.

This move has won praise for helping to demystify the often-criticized “black box” nature of AI, though it doesn’t make Claude fully open source. While the actual source code and training data remain proprietary, Anthropic’s approach offers a glimpse into AI's inner workings, encouraging other companies to follow suit in enhancing transparency.

You can read the full article on VentureBeat.

io.net and FLock.io Launch World's First Proof-of-AI Consensus

io.net and FLock.io Launch First-Ever Proof-of-AI Consensus

Speaking of “the first ones” — io.net has teamed up with FLock.io to roll out a brand-new consensus mechanism called "Proof of AI" (PoAI). This isn't just a fancy new term—PoAI is designed to make decentralized AI networks more trustworthy and scalable by ensuring that only those who genuinely contribute computing power get rewarded, shutting down anyone just trying to game the system.

According to io.net's CMO, Milan Amin, PoAI is all about making decentralized networks a solid alternative to traditional cloud providers. The idea is that these networks will perform better, scale easier, and ultimately be more cost-effective.

This mechanism checks if nodes are actually pulling their weight, making sure that honest participants are fairly compensated while keeping freeloaders at bay. It’s a move that not only boosts trust but also supports the long-term growth of decentralized AI platforms.

Wanna dive deeper? Read the full article on Cointelegraph.

Piramidal’s Foundation Model for Brain Waves Could Supercharge EEGs

It's always exciting to see AI stepping up to support medical professionals. This time, a startup called Piramidal is making waves (literally) with a new foundational model designed to analyze brain scan data from EEGs.

Founders Dimitris Sakellariou and Kris Pahuja noticed a huge gap in how EEGs are used across different hospitals—varying machines, inconsistent data, and a need for specialized knowledge to interpret the results. Piramidal’s new model aims to streamline this process by automatically flagging critical patterns, like those signaling a stroke or seizure, which could improve outcomes for patients and ease the burden on overworked nurses and doctors.

The idea here is similar to how foundational language models work; think of it like Meta’s LLaMA models but for brainwaves instead of text. Piramidal’s model promises to read and interpret EEG signals from any setup, regardless of the equipment or patient. They’ve already built and tested their model and are now gearing up to deploy it in hospitals early next year. They’re planning four pilot programs in ICUs, where the model will be put through its paces in real-world conditions.

Piramidal has raised $6 million in seed funding to help scale this technology, with plans to integrate a vast amount of open-source and new patient data to refine their model further.

This tech is set to be the biggest EEG model yet, potentially identifying patterns that even the most trained human eyes might miss. With a foundation like this, the potential to revolutionize neurological care is on the horizon, and we’re eager to see where Piramidal takes it next!

You can read all about it on TechCrunch.

DeepMind’s GenRM: LLMs That Verify Their Own Work

It’s no secret that large language models often make mistakes (duh), especially with complex tasks. DeepMind's new approach, called GenRM, aims to fix this by having models verify their own outputs, using their generative abilities to double-check answers.

Unlike traditional verifiers that score responses, GenRM leverages LLMs’ strength in text generation. It prompts the model to create a reasoning process before deciding if an answer is correct—think of it as the model grading its own test with an extra step of “thinking out loud.” This method consistently outperformed existing verification techniques in tests, even surpassing well-known models like GPT-4 in certain tasks.

GenRM is designed to be adaptable, improving accuracy without needing separate verification models. By combining answer generation and self-checking, it opens up new possibilities for making LLMs more reliable and accurate across various applications. DeepMind’s approach could be a big step forward in reducing errors and boosting the credibility of AI outputs.

Wanna know more? Read a full piece on VentureBeat.


That’s a wrap for this issue! As AI continues to evolve at lightning speed, we’ll be here to keep you updated on all the latest breakthroughs, challenges, and insights. Thanks for reading, don't hesitate to share your thoughts in the comment section, and stay tuned for more cutting-edge AI news in our next edition. Until then, keep exploring and stay curious!

P.S. Looking for a trusted tech partner for your AI-powered software development project? We’ve been building AI projects since 2017 ?? See how we can help you!

要查看或添加评论,请登录

Neoteric的更多文章

社区洞察

其他会员也浏览了