Sign In

Carlos Ξ

crypto / econ / data / research @ flipsidecrypto

Carlos Mercado is a data scientist at Flipside Crypto, focusing on the intersections of data.1 He has experience working at Flipside and received his education from Johns Hopkins University.1 Carlos is also active on GitHub under the username CR-Mercado, where he has 54 followers and 3 following.2 Additionally, he has been featured on The Harpreet Podcast, where he shared his journey into data science and discussed the importance of building a personal brand.3 Carlos can be found on LinkedIn with the username crmercado.12

Highlights

Mar 8 · linkedin

How to think about building #ai models. Download to read (LinkedIn 🤝 blurry compression). Here's the summary:


How do LLMs work? 🏭

Users give [Instructions] w/ [Context] attached to a [System Prompt] (optionally passed through a [Tuning filter]) so a [Foundational Model] can generate a matching set of words as a response using its token weights.

Instruction: summarize the 5 main points of this argument with quotes

Context: the writer's argument + user preferences on how to summarize

System Prompt: rules, behavior, censorship, capabilities (e.g., web browse, image generation, code execution)

Tuning Filter: Custom training layer

Foundational Model: GPT 3.5; Claude; Gemini; Llama-2


How do you scope a custom LLM? 📃

The more ambitious the goal, the more examples/prep you'll need to do.

SEARCH: Often solved w/o "AI", using vectorization, term frequency, cosine similarity & other "basic" algorithms. Finding already-created content for a user request.

SUMMARIZE: Often solved by "framing" requests. Attaching already-created content to requests and letting the foundation model do most of work but and guiding its output to follow a defined format.

Mixture of Expert models horizontally scale by adding new experts in specific domains to intelligent route summarizations (and often summarize those summaries).

GENERATE: The holy grail. Teach a foundational model the essence of a human idea. Something learned thru experience not just being told what to do.

Emergent Understanding - a foundational model, with tuning (given 100+ perfect input/output pairs) can outperform the sum of its parts.

"Bigger the better" - a vertically scaled model knows so much about so many topics that an "A-Ha!" moment happens, finding latent patterns across examples without being told the pattern explicitly.


What are the tradeoffs between popular Foundational Models? 💱

Cost Controls, Enterprise Support, Hosting, Tuning, A/B Prompt Testing, Automatic Scaling all have to be balanced against the risk of getting the rug pulled from under you.

LLMs are chaotic systems- the smallest change in a system prompt could cascade into your model losing quality- or worse- refusing to perform at all!

Version control & temperature setting (i.e. forcing deterministic behavior) are huge opportunity spaces in Open Source foundation models.


Is Context enough? 📓

For narrow scopes, it's easier to attach relevant information to every user request. For a car problem AI, you attach 100s of car manuals and let the foundation model do the work.

When is tuning worth it? ✍

For wider scopes (or pricy APIs), you may need emergent understanding. Instead of attaching 100s of car manuals. You train the model on 100s of the most common questions & answers about cars. As the model understands car parts and which problems correlate to which parts, it can better help the user identify which mechanic service they need.

#artificialintelligence #llms

Nov 16 · linkedin

📢 Proud of the Flipside Crypto team for launching the #web3 Active Directory.

One of the major mistakes #DAOs have made over the last 2-3 years is weak exclusion. Anyone can join the discord. Anyone can spam gm. Anyone can hold the token.

Large, successful organizations don't operate like this. Open organizations are good. But decentralized doesn't mean direct democracy.

It's wildly inefficient to have every decision go to a vote.

In fact, the 1944 CIA "Simple Sabotage Field Manual" details exactly how to exploit democratic decision-making to ruin an organization. How many have you experienced in DAOs?

  1. Everything has a channel - no shortcuts.
  2. Talk a lot, often.
  3. Bring up irrelevant issues constantly - wen moon??
  4. Pass tasks around endless committees.
  5. Haggle over wording.
  6. Re-open conversations already decided on.
  7. Emphasize "caution" and "reasonableness" to slow things down.
  8. Question whether decisions are proper, in-scope, or conflict with a defined morality.

How many DAOs have you been in where lack of getting sht done* meant all the most talented people left?

The future of DAOs is segmented trust for autonomous decision making. Anyone of X rank can make X decisions. Anyone of Y rank can make Y decisions.

In traditional orgs, they use services like Microsoft Active Directory to control which users have what privileges and access to systems.

We've released Badger to make this trivial for web3 orgs too.

Badger uses NFTs to assign and manage roles within an organization.

Combining this with out of the box integrations to Guild, Discord, Telegram, Twitter, etc. to segment real power across these roles.

Allowing DAOs to decentralize the autonomous decision-making of its members.

Check out more on Badger here: http://trybadger.com

#crypto #technology #economics #blockchain #nfts #ethereum

Related Questions

What projects has Carlos Mercado worked on at Flipside Crypto?
What is Carlos Mercado's educational background?
How did Carlos Mercado transition into a career in data science?
What are some notable achievements of Carlos Mercado in his field?
How does Carlos Mercado contribute to the data science community?
Carlos Ξ
Carlos Ξ, photo 1
Carlos Ξ, photo 2
Add to my network

Location

Washington, District of Columbia, United States