• AI Emergence
  • Posts
  • Claude 3.5 Sonnet - New, Better, Consistent LLM ?

Claude 3.5 Sonnet - New, Better, Consistent LLM ?

Along with: OpenAI acquires multiple companies to build for the future enterprises

Hey there, 

When OpenAI released GPT4 last year, I thought they would now accelerate to bring better intelligence with a laser-sharp focus. Little did I know that they were thinking differently! 

OpenAI still hasn’t released a GPT 5-level model yet and has acquired 2 companies this week which indicate they want to be the builders as well as an Enterprise Intelligence solution at the same time.

They acquired a company focusing on collaboration and another one that focuses on building data infrastructure for Enterprises. Looks like they want to focus on offering a better “Teams plan” than the next level of intelligence!

What it does is give Anthropic (and possibly others) to bring out competitive models and make a claim to the throne! Something that Anthropic attempted last week with Claude 3.5 Sonnet, and it has impressed many!

One thing is for sure - the second half of 2024 will see some exc-AI-ting battles!

Before we dive in - Don’t miss the Final Sales offer at DataHack Summit, India’s largest GenAI event, running until June 28th. You can save up to INR 11,200 on tickets!

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

Please note: This is an abbreviated version of our newsletter due to email length restrictions. For the complete experience, visit our website and enjoy the full, unabridged edition.

Table of Contents

This week, Anthropic has introduced Claude 3.5 Sonnet, the mid-tier model from the upcoming Claude 3 family. Remarkably, it outperforms the higher-tier Claude 3 Opus while maintaining the speed and cost efficiency of the earlier mid-tier model, Claude 3 Sonnet.

It is available free on Claude.ai and the Claude iOS app and accessible via Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI.

Performance:

  1. Operates at twice the speed of Claude 3 Opus

  2. Outperforms in GPQA, MMLU, and HumanEval benchmarks against GPT 4o and Gemini 1.5

New Feature:

Artifacts for real-time interaction with AI-generated content, enhancing Claude’s role as a collaborative tool. For example, you can create a game using prompts and play it on the right console using Artifacts

Capabilities:

  1. Superior at writing, translating code, handling multi-step workflows, interpreting charts and graphs, and transcribing text from images

  2. Improved understanding of nuance, humor, and complex instructions
    (source)

Remember the exciting voice capability update demoed by OpenAI during the Spring update? It’s running on a delay. The Advanced Voice Mode for ChatGPT, initially planned for June, is now set for an alpha release in July to a select group of ChatGPT Plus users.

In a post on X, OpenAI cited the delay as improving the model’s ability to detect and refuse certain content, enhancing user experience, and preparing infrastructure to scale to millions while maintaining real-time responses. (source)

Amidst trouble, Stability AI found some “stability” and has made big headlines this week due to two reasons.

Stability AI has named Prem Akkaraju as its new CEO. He was the former CEO of the famous digital effects company Weta Digital. This comes months after the exit of founder and CEO Emad Mostaque left the cash-strapped company.

The firm has raised about $80 million in funding, according to a person familiar with the matter, from a group of investors including Coatue Management, Lightspeed Venture Partners, and former Google CEO Eric Schmidt. Stability AI has named former Facebook President Sean Parker the executive board chairman. (source)

Amazon is reportedly developing a consumer AI chatbot named "Metis," aiming to rival OpenAI's ChatGPT. 

This new chatbot, accessible via a web browser, is part of Amazon's broader strategy to enhance its presence in the AI space, alongside its enterprise AI assistant and AI shopping tool.

The company is integrating its AI technology across different products, including a new paid tier for its Alexa voice assistant, which could significantly boost its revenue. (source)

Synthesia, backed by Nvidia has enhanced its video production suite with new updates including the creation of AI avatars using webcams or phones.

These updates transform Synthesia into a comprehensive video production tool for large companies, moving beyond just AI-generated avatars.

Notable additions include

  • Full-body avatars, 

  • A screen recording tool, and 

  • The ability to clone voices in multiple languages. 

This democratization of video creation aims to lower the barriers for those without large budgets, making sophisticated video production accessible to more users. (source)

ByteDance, the parent company of TikTok, is collaborating with U.S. chipmaker Broadcom to develop a custom 5-nanometer AI chip, despite the tight U.S. export controls on advanced semiconductor technology.

This chip, designed to comply with U.S. regulations, will be manufactured by Taiwan's TSMC. This strategic move by ByteDance aims to secure a stable supply of high-performance chips for its expanding AI initiatives, amid intense global competition for semiconductor resources.

This partnership highlights ByteDance's ongoing efforts to enhance its technological capabilities in the face of international supply chain challenges. (source)

Days after The Wall Street Journal reported talks between Apple and Meta to integrate the latter’s AI models, Bloomberg reported that Apple has no such plans. 

According to Bloomberg, Apple briefly considered a partnership with Meta in March while exploring AI integrations with various companies. However, privacy concerns led Apple to shelve the idea, as integrating Meta’s AI models was seen as potentially harmful to Apple's image due to its ongoing criticism of Meta's privacy practices. (source)

Sony Music, Universal Music Group, and Warner Records have filed lawsuits against AI companies Suno and Udio, alleging massive copyright infringements for using their music to train AI systems that produce competitive music.

The lawsuits, filed in New York and Massachusetts, claim these companies copied hundreds of songs to create AI-generated music mimicking renowned artists, potentially damaging human musicians' work.

Both companies argue their technology is transformative, aimed at creating new content rather than copying existing music. (source)

OpenAI has strategically expanded its capabilities with the acquisition of two innovative startups: Multi and Rockset.

Multi, known for its AI-enhanced video collaboration platform, aligns with OpenAI’s aim to extend its reach beyond browser-based tools and potentially develop an AI-driven operating system. (source)

Rockset, on the other hand, brings expertise in real-time analytics with its serverless search engine, enhancing OpenAI’s ability to handle and analyze large-scale data in real-time. (source)

These acquisitions not only enhance OpenAI’s technological prowess in handling complex AI applications but also bolster its presence in the enterprise sector.

Nvidia's stock price has significantly declined, falling over 5% to around $120, marking its lowest point since a major stock split on June 10.

This downturn continues a trend from the previous week, resulting in a total loss of around 12% from its recent peak.

This drop has wiped out approximately $400 billion in market capitalization, even without any major negative catalysts, suggesting a market correction.

Despite these losses, Nvidia still holds a substantial market value of $2.9 trillion, maintaining its position as the world’s third-largest company, just behind Apple and Microsoft. (source)

Apple has decided not to release its new "Apple Intelligence" AI product and other key features in the European Union in 2024 due to concerns with the EU's Digital Markets Act (DMA). 

The DMA's interoperability requirements could compromise the integrity, privacy, and security of Apple's products. The features affected include iPhone Mirroring and enhancements to SharePlay. 

Despite the setback, Apple remains committed to negotiating with the EU to find a resolution that maintains customer safety without sacrificing feature availability. 

This regulatory challenge affects a significant portion of Apple's market, as Europe accounted for nearly a quarter of its global net sales in 2023. (source)

YouTube has updated its privacy request process, now enabling users to request the removal of AI-generated content that uses their likeness. This includes videos that use synthetic alterations of a person's face or voice.

The decision to remove such content will consider factors like the authenticity of the content, identifiability of the individual, and the context of the content, such as parody. (source)

Toys"R" Us has launched its first AI-generated commercial using OpenAI's Sora video generator.

The advertisement narrates the brand's history through its mascot, Geoffrey the Giraffe, and highlights the vision of founder Charles Lazarus in the 1930s.

It’s not known how much post-production work has been done on this video but the result seems amazing. You could also see challenges around consistency of character and also SORA struggling at small places - but nonetheless, it is fascinating!  (source)

Tool: revid.ai

Getting social media attention for your brand is increasingly challenging with thousands of brands competing for visibility. Can you produce high-quality content in a fraction of the time it takes to create a traditional video? 

This is where revid.ai comes in. It helps you create videos for your social media

Let’s say that you want to create a reel from a LinkedIn post and cross-post it on Instagram. Let’s see how it works -

How to access the tool:

1. Sign up or log in to revid.ai.

2. Enter the text or link to the content you want to convert into a video.

3. Select the desired video style and any specific preferences for voiceovers or animations.

4. Configure the output settings to match your brand's visual identity.

5. Publish the video directly to your social media channels or download it for later use.

How’s the output?

Using AI tools for creating social media content can be tricky. The results often lack polish and offer limited flexibility for adjustments. Use them cautiously!

  • In the recent episode of Leading with Data, I had a brief conversation with Jaidev Janardhan, CEO of ZOPA Bank, about his work in redefining the fintech landscape at Zopa and his experience in managing credit risk during the financial crisis, and how his leadership and innovative approaches have significantly expanded Zopa’s offerings and impact in the financial sector.

  • If you're keen on advancing your knowledge in AI development and exploring the practical applications of AI agents, the course "AI Agents in LangGraph" offered by deeplearning.ai is ideal for you. This course is specifically designed to provide insights into the development and application of AI agents within language graph systems. Also how to manage complex tasks, utilizing large language models (LLMs) to boost decision-making and automation.

  • Karpathy creates LLM101n on GitHub, a course to build AI storytellers using LLMs. The course, when launched would help students to create, iterate, and improve storytelling entirely using AI.

  • Gates Foundation is backing Jeff Hawkins (author of The Thousand Brains) in the race towards Artificial General Intelligence (AGI).

This week I was wondering how far are we to have a reliable assistant to write complete code based on instructions, autocorrect the bugs, and have a production-ready code for deployment.

The current levels are probably best demonstrated by Claude in this tweet by Alex Albert

The day we have this - the amount of code (and hence apps) generated and deployed will shoot up like the junk content has shot up on the internet. I also don’t expect it to be a point in time - it would be a continuous change, but I think that point is probably 2-3 years away.

Would love to know, what you think.

How far are we to have a reliable assistant to write complete code based on instructions, auto-correct the bugs, and have a production-ready code for deployment?

Login or Subscribe to participate in polls.

Whenever it happens, it would be a great enabler for many people who fear coding or are not interested in coding, and the world we know today, would have changed forever!

See you next week!

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Join the conversation

or to participate.