- The Rundown AI
- Posts
- An open-source GPT-4o competitor
An open-source GPT-4o competitor
PLUS: Salesforce's 'Tiny Giant' breakthrough
Rowan Cheung
July 04, 2024
Sign Up | Advertise | Tools | AI University
Welcome, AI enthusiasts.
Feeling down that OpenAI’s Advanced Voice Mode is still delayed and potentially months away? Have no fear — Moshi is here!
A new startup just launched its own real-time AI voice assistant, and the open-source model could be set to push the competition to new heights. Let’s investigate…
In today’s AI rundown:
French AI startup launches ‘Moshi’
Salesforce’s small model breakthrough
Turn thoughts into polished content
Perplexity gets major research upgrade
6 new AI tools & 4 new AI jobs
More AI & tech news
Read time: 4 minutes
LATEST DEVELOPMENTS
KYUTAI
Image source: Kyutai
The Rundown: French startup Kyutai just introduced Moshi, a new ‘real-time’ AI voice assistant capable of responding in a range of emotions and styles in a similar fashion to OpenAI’s delayed Voice Mode feature.
The details:
Moshi is capable of listening and speaking simultaneously, with 70 different emotions and speaking styles ranging from whispers to accented speech.
Kyutai claims Moshi is the first ‘real-time voice AI assistant’ released, with a 160ms latency that potentially outpaces OpenAI's offering.
The nonprofit group plans to open-source the research and model in the coming weeks, with Moshi currently available to try via Hugging Face.
The startup launched in 2023 with $324M in funding, with a team of 8 researchers developing Moshi in just four months.
Why it matters: Moshi looks to be a massive win for the French AI landscape, and another eye-opening rival that chips away at OpenAI’s perceived moat on the rest of the field. Plus, with that uniquely French accent, there certainly won’t be any ScarJo concerns about this model rollout.
TOGETHER WITH ARTISAN AI
The Rundown: Ready to bring your outbound into the age of AI? Meet Ava — an AI BDR that automates 80% of your outbound sales workflow and self-improves over time.
Ava operates on the Artisan Sales platform, which consolidates:
300M+ high-quality B2B contacts
Automatic lead research & enrichment with 10s of data sources
Deliverability tools, including email warmup
AI playbooks that research, write & send emails for you
Hire Ava and supercharge your pipeline today.
SALESFORCE
Image source: Salesforce
The Rundown: Salesforce just published new research on APIGen, an automated system that generates optimal datasets for AI training on function calling tasks — enabling the company’s xLAM model to outperform much larger rivals.
The details:
APIGen is designed to help models train on datasets that better reflect the real-world complexity of API usage.
Salesforce trained a both 7B and 1B parameter version of xLAM using APIGen, testing them against key function calling benchmarks.
xLAM’s 7B parameter model ranked 6th out of 46 models, matching or surpassing rivals 10x its size — including GPT-4.
xLAM’s 1B ‘Tiny Giant’ outperformed models like Claude Haiku and GPT-3.5, with CEO Mark Benioff calling it the best ‘micro-model’ for function calling.
Why it matters: While the AI race has been focused on building ever-larger models, Salesforce’s approach suggests that smarter data curation can lead to more efficient systems. The research is also a major step towards better on-device, agentic AI — packing the power of large models into a tiny frame.
AI TRAINING
The Rundown: ChatGPT's voice mode feature now allows you to convert your spoken ideas into well-written text, summaries, and action items, boosting your creativity and productivity.
Step-by-step:
Enable “Background Conversations” in the ChatGPT app settings.
Start a new chat with the prompt shown in the image above (it was too long for this email).
Speak your thoughts freely, pausing as needed, and say "I'm done" when you've expressed all your ideas.
Review the AI-generated text, summary, and action items, and save them to your notes.
Pro tip: Try going on a long walk and rambling any ideas to ChatGPT using this trick — you’ll be amazed by the summary you get at the end.
THE RUNDOWN AI UNIVERSITY
The Rundown: Runway’s Gen 3 AI video generator just dropped, and to showcase its powerful capabilities, we’re hosting a live workshop on how to create a fully AI-generated commercial using Gen 3, ElevenLabs, and Midjourney.
Join us on Friday at 4 PM EST to:
Understand the text-to-video capabilities and top prompt tips for Runway Gen-3.
Learn the industry-standard workflow for producing AI videos, from visual concepts to animated commercials using Midjourney and Runway.
Elevate your video with ElevenLabs sound effects and voiceovers.
If you’re a member of The Rundown University you can RSVP in the Upcoming Workshops space.
If you’re not a member yet, you can still join the workshop with a 14-day free trial to The Rundown University.
PERPLEXITY
Image source: Perplexity
The Rundown: Perplexity just announced new upgrades to its ‘Pro Search’ feature, enhancing capabilities for complex queries, multi-step reasoning, integration of Wolfram Alpha for math improvement, and more.
The details:
Pro Search can now tackle complex queries using multi-step reasoning, chaining together multiple searches to find more comprehensive answers.
A new integration with Wolfram Alpha allows for solving advanced mathematical problems, alongside upgraded code execution abilities.
Free users get 5 Pro Searches every four hours, while subscribers to the $20/month plan get 600 per day.
The upgrade comes amid recent controversy over Perplexity's data scraping and attribution practices.
Why it matters: Given Google’s struggles with AI overviews, Perplexity’s upgrades will continue the push towards ‘answer engines’ that take the heavy lifting out of the user’s hand. But the recent accusations aren’t going away — and could cloud the whole AI-powered search sector until precedent is set.
NEW TOOLS & JOBS
🎵 Suno iOS - Generate high-quality songs on your phone
🚀 Ariglad - Auto maintain and update support tickets
⭐ Google Gemma 2 - Gogle’s most efficient and cost-saving AI model, available in Google AI studio
🤖 Inline GPT - Chat with ChatGPT in any apps
👩🏫 AI Tutor Pro - An AI tutor to grow your knowledge on any topic
🚀 BIK - AI marketing automation for eCommerce brands
👩💼 DeepMind - AI Ethics & Safety Manager
📝 Parloa - Product Instructional Designer
💼 Glean - Enterprise Account Executive
🤖 Sanctuary AI - Field Technician - Robotics
QUICK HITS
Cloudflare released a free tool to detect and block AI bots circumventing website scraping protections, aiming to address concerns over unauthorized data collection for AI training.
App Store chief Phil Schiller is joining OpenAI’s board in an observer role, representing Apple as part of the recently announced AI partnership.
Shanghai AI Lab introduced InternLM 2.5-7B, a model with a 1M context window and the ability to use tools that surged up the Open LLM Leaderboard upon release.
Magic is set to raise over $200M at a $1.5B valuation, despite having no product or revenue yet — as the company continues to develop its coding-specialized models that can handle large context windows.
Citadel CEO Ken Griffin told the company’s new class of interns that he is ‘not convinced’ AI will achieve breakthroughs that automate human jobs in the next three years.
ElevenLabs launched Voice Isolator, a new feature designed to help users remove background noise from recordings and create studio-quality audio.
THAT’S A WRAP
SPONSOR US
Get your product in front of over 600k+ AI enthusiasts
Our newsletter is read by thousands of tech professionals, investors, engineers, managers, and business owners around the world. Get in touch today.
FEEDBACK
How would you rate today's newsletter?Vote below to help us improve the newsletter for you. |
If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.
Join the conversation