Next Week in The Sequence:

• Edge 331: Our series about fine tuning techniques dives into one of the first fine-tuning methods ever created: ULMFiT. We review Google’s symbol tuning paper and dive into Scale’s LLM Engine open source framework.

• Edge 332: Deep dives into the Flash-Attention techniques that are revolutizing LLM scalability.

You can subscribe below:

📝 Editorial:A Week of Monster Generative AI Releases

Last week could be considered one of the most intense weeks we have seen in the generative AI market in quite a few months, and that’s quite a distinction considering the manic pace of that space. While the news was dominated by Anthropic’s massive round led by Amazon and OpenAI's rumored new fundraising efforts, my interest gravitated more towards the number of high-profile tech releases we saw in the last few days.

If you don’t believe me, here are my top five:

1. Mistral 7B is the first open source model that comes out of the French gen AI startup. This one is particularly interesting because it uses novel attention architectures to achieved superior performance than much larger models. Mistral 7B can be catalogued like the top smallest LLM in the market 😉.

2. Meta AI: Dubbed as a ChatGPT competitor, Meta AI is remarkably important because opens Meta Llama 2 models to a massive audience. This is certainly a solid test for Meta’s gen AI open source strategy.

3. Cohere Chat + RAG: Cohere unveiled its implementation of retrieval-augmented generation (RAG) capabilities with a very clean conversational interface. This release is relevant because it signals a demarcation from the path followed by OpenAI and Anthropic by focusing on an enterprise audience.

4. Claude on Bedrock: Anthropic announced the availability of its Claude model to Amazon Bedrock customers. This release represents the first major alternative to Azure OpenAI Service.

5. Bedrock GA: The Claude addition was not only the important news for Bedrock last week. The service achieved general availability(GA) status after quite some time in preview releases.

How about that for five days in the gen AI market? From open source to the market strategies of tech incumbents, last week's releases are likely to play a factor in the near future of generative AI. Fascinating stuff.

🎥 Watch Now: Building Plaid’s ML Fraud Detection Application

Want to learn about Plaid’s ML platform journey? In this on-demand recording, Plaid Software Engineer Renault Young shared the technical challenges they faced, how they set up the data foundations they needed to start building an ML platform, what they used to look for patterns in transaction data in real time, and more. Today, Signal is Plaid’s biggest ML application and analyzes 1000+ risk factors per ACH transaction. 

The on-demand recording is now available for you to watch and share with your colleagues! 

WATCH THE VIDEO

🔎 ML Research

AutoGen

Microsoft Research published a research paper unveiling AutoGen, aframework for LLM-based agents. AutoGen outlines some key components of multi-agent architectures including communication, guardrails and others —> Read more.

GPT-4V

OpenAI published a paper detailing the technical details behind GPT-4(Vision), which extends GPT-4 with the ability to process image inputs. The paper puts special emphasis on the safety capabilities of this new release —> Read more.

Auditing LLMs with LLMs

Carnegie Mellon University published an interesting paper about performing LLM audits with, well, LLMs. The paper proposes a tool called AdaTest that can create and run tests on LLMs —> Read more.

Physics-Inspired Gen AI

Researchers from MIT published a paper discussing PFGM++, a physics-inspired generative AI model for pattern recognition. The model integrates physics laws such as diffusion and poisson flow to discover patterns in complex environments —> Read more.

🤖 Cool AI Tech Releases

Voice and Image in ChatGPT

OpenAI added voice and image capabilities to ChatGPT —> Read more.

Mistral 7B

Generative AI platform Mistral open sourced a 7B LLM that seems to perform better than bigger alternatives —> Read more.

Cohere RAG

Cohere launched its Chat API with RAG capabilities —> Read more.

Bedrock GA

Amazon Bedrock reached general availability with new models and capabilities —> Read more.

🛠 Real World ML

Training Models at Pinterest

The Pinterest engineering team discusses different training improvements to their ranking models —> Read more.

Fraud Detection at Uber

Uber discusses some details about its Risk Entity Watch platform that uses unsupervised learning for fraud detection —> Read more.

📡AI Radar

• Anthropic announced a monster $4 billion strategic investment from Amazon.

• Meta introduced new generative AI capabilities including Meta AI, a new conversational assistant across different apps and devices.

• Character.AI is rumored to be raising a new round that will value the company around $5 billion.

• AI testing platform Kolena raised $15 million in a series A.

• You.com released YouAgent, an AI assistant which can generate and execute code.

• Spanish Grammarly competitor Correcto raised $7 million in a new round.

• Vishal Maini, A former researcher from Google DeepMind raised $14 million for a new seed stage AI fund.

• Cloudflare launched several generative AI improvements including serverless inference.

• HR giant Workday announced several new generative AI additions to its platform.

• AI-based market intelligence firm AlphaSense raised $150 million.

• AIOps platform Senser came out of stealth mode with a $9.5 million raise.

• AI-cybersecurity startup Nexusflow raised $10.6 million in a new round.

• AI agent training platform Luda announced a $7 million round.