Sign In

Paul Golding

Chief Scientist at Meta

Paul Golding is a Senior Technical Consultant and Chief Scientist at Meta, where he focuses on leveraging artificial intelligence to drive business growth. He has a notable background in the wireless and mobile technology sectors, with over fifteen years of experience. Golding is recognized for his strategic insights into AI applications, which include corporate strategy, product innovation, and operational excellence.13

In addition to his role at Meta, Golding is the CEO of Golding & Co, Inc., where he has executed numerous projects globally for major companies such as IBM and Vodafone. He holds an impressive portfolio of over 30 patents and has been involved in hundreds of inventions related to AI and technology.13

Golding's expertise extends to consulting on AI strategy and acceleration, helping businesses integrate AI into their operations effectively. His approach emphasizes practical solutions tailored to specific business needs, avoiding jargon and focusing on delivering tangible value.12

Highlights

Apr 25 · linkedin

Joint Embeddings -- But Why?

One of the most interesting developments for me in embeddings was when the idea of combining different "domains" (distributions) into a single embedding space arrived. A fantastic early example was the Starspace approach (via Facebook) which combined hash-tag embeddings with document embeddings.

I adapted it (during the XtoVec era) to build a model for next-best-action in digital marketing, mostly so that marketing folks could label touch points with hash-tags, which turns out to be the easiest way to maintain a "folksonomy" with little regard for formalisms etc.

This technique continues to deliver, more recently with the SANTA approach - Structure Aware DeNse ReTrievAl. This combines docs with structured text (e.g. JSON) -- https://lnkd.in/g_5VUwsE

Why do these approaches work? After all, they are taking what seem to be two different distributions and compressing them into the same latent space.

Well, hash-tags and JSON still use natural language whereby naming conventions are human interpretable.

These techniques are still valuable in chatbots where one is still interested in policy-directed dialogs as opposed to free-form. Indeed, I think we need to differentiate task-bots from oracle-bots.

Most use cases are going to be task-bots, which originally were solved using intent-based recognition. Early on, the intent logic still sat outside of the prompt sense-making logic, but innovators like Rasa figured out that both can be combined into a single embedding space (using something like Starspace).

It is no surprise then that ServiceNow has adopted a similar approach in their LLM efforts to combine NL and policy data, namely workflows.

However, it seems clear to me that at some point we will return to a more formal approach via the introduction of (Domain Specific Languages) DSLs, albeit possibly generated via LLMs, to specify workflows in ways more reliable than merely "hoping for the best" that disjoint distributions will align because structured representations happen to use appropriate naming conventions. Why not insist upon it?

In a way, this is what I was attempting with a hashtag folksonomy. There is a sense in which adding hashtags goes with a mental model of "indexing", which is kind of what we're hoping the tags to do in the embedding space. Over time, users will add tags that slowly evolve to resemble a "domain specific language" because users' mental models figure out that the tags are actually instructions to some downstream AI.

Nov 14 · linkedin

Evaluating LLMs: Possibly the most important topic in GenAI

Arvind and Sayash carry out various useful analyses of AI from a critical-thinking and research perspective. This contribution is a useful orientation into the world of evaluating LLMs. I previously posted a full-length survey paper which I have been analyzing and summarizing for a later post, especially contextualized to the very much open question of how to evaluate LLMs within the enterprise where a whole set of unique problems apply.

Related Questions

What are some of Paul Golding's most notable projects at Meta?
How has Paul Golding contributed to the development of AI at Meta?
What is the role of Golding & Co, Inc. in accelerating AI adoption?
Can you provide examples of Paul Golding's work with major tech companies like Google and Microsoft?
What are the key strategies Paul Golding uses to integrate AI into businesses?
Paul Golding
Paul Golding, photo 1
Paul Golding, photo 2
Add to my network

Location

San Francisco Bay Area