The LLMcorns: 4 New Billion Dollar Gen AI Valuations in One Week

LLM providers are still commanding remarkable valuations in this fundraising climate.

Created by DALL-E

Next Week in The Sequence:

Edge 365: We review a new LLM reasoning method: Reflexion and we dive into the original paper that proposed that technique. We also review Flowise, a visual platform for building LLM apps.
Edge 366: We discuss Anthropic’s Sleeper Agents research about LLM security vulnerability.

You can subscribe below!

📝 Editorial: The LLMcorns: 4 New Billion Dollar Gen AI Valuations in One Week

Eye-popping fundraising events are nothing new in the generative AI market, but last week was particularly special. Valuations in the generative AI space have been a hot topic in venture capital circles for the past couple of years. Questions arise: Where will value accrue in the gen AI market? Are we approaching a correction point in gen AI startup valuations? Will LLMs become commoditized? Such questions typically signal a shift towards a more conservative fundraising climate. However, if last week is anything to go by, that shift is not imminent. Last week was incredible from a venture fundraising perspective in gen AI, with four new companies surpassing $1 billion in valuation:

Brett Taylor, former Co-CEO of Salesforce, is raising $85 million for Sierra, a new enterprise AI company, at a rumored $1 billion valuation.
Kutrim, an Indian AI platform started by Bhavish Aggarwal, founder of Ola, raised $50 million at over a $1 billion valuation.
Voice AI startup ElevenLabs raised $80 million at a $1.1 billion valuation.
Elon Musk’s xAI is raising $1 billion at a $6 billion valuation. Details about Taylor’s Sierra are not yet public, but the common denominator among these companies is that they are all building their own foundation models. These are capital-intensive problems requiring significant compute and research, and can yield asymmetric results. Whether the area of foundation models in the generative AI stack will become commoditized over time is an intriguing intellectual debate. For now, foundation models are still commanding remarkable valuations relative to their development stage and market adoption. What a remarkable week!”

🔎 ML Research

Transformer for Graph Data

Google Research published a paper detailing Exphormer, a transformer model optimized for graph datasets. The method uses sparse attention and leverages spectral graph theory to achieve strong performance in graph data —> Read more.

Vision Language Models with Multiple Inputs

Amazon Science published a paper proposing a technique for processing multiple inputs in vision-language models. In essence, the method produces an aggregated embedding for multiple input images which leads to a richer representation —> Read more.

Heuristics Performance

Microsoft Research published a paper proposing MetaOpt, a method to evalute the performance of heuristic algorithms. MetaOpt is able to compare different heuristics and evaluate its performance in different settings —> Read more.

LLM Hallucination Detection

Researchers from University of Washington, Allen AI and Carnegie Mellon University published a paper proposing a task and benchmark for detecting hallucinations in LLMs. The benchmark includes human judgments outputs on different LLM outputs across different domains —> Read more.

ChatQA

NVIDIA Research published a paper introducing ChatQA, a series of conversational models that achieve ChatGPT type performance across different QA benchmarks. ChatQA is based on a dense retriever fine-tuned on a multi-turn QA dataset —> Read more.