Sardine named a Leader in The Forrester Wave™: Financial Crime Management Solutions, Q3 2026

Learn More
The Saturday Fraud Strategist

Modelos fundacionais na deteção de fraude: a tecnologia por trás do ChatGPT aplicada à fraude

13 min

Os modelos fundacionais para combater a fraude estão na agenda de todas as conferências este ano, e continuo a ouvir a mesma pergunta. São uma verdadeira inovação ou apenas mais um termo da moda? Quando cheguei ao Money20/20 Vegas, quis fazer essa pergunta a alguém que desenvolve soluções com estes modelos todos os dias.

Pedro Sanzovo trabalha com produtos de identidade e prevenção a fraudes e atua nessa área desde o início da carreira. Começou na Palantir, fundou uma empresa de prevenção a fraudes para lojistas no Brasil e a vendeu para a Signifyd, onde liderou o desenvolvimento de produtos baseados em modelos de aprendizado de máquina. Há seis meses, ingressou na Plaid.

Mantivemos esta conversa curta e prática. O que é um modelo fundacional e como ele difere dos modelos que a maioria de nós usa hoje? Em que situações ele supera o XGBoost e em quais esbarra nas mesmas limitações? E o que fazer com um embedding quando sua equipe trabalha com regras? Pedro trouxe boas respostas e algumas coisas que acho que você vai querer ouvir.

O que vai ouvir neste episódio:

  • O que é um modelo fundacional e como os modelos transformer diferem das regras encadeadas e dos modelos XGBoost que as equipas de combate à fraude utilizam há anos
  • Por que os modelos tradicionais dependem de atributos criados manualmente e de rótulos de qualidade para a aprendizagem supervisionada, e por que isso é difícil para empresas menores
  • Como os modelos de base para dados financeiros tratam as transações como uma sequência de eventos e preveem o que vem a seguir
  • O que Pedro quer dizer com o aumento de precisão de até 34% que tem observado e por que a qualidade dos dados continua a determinar o resultado
  • Por que as restrições de latência, como o limite de 300 milissegundos para transações com cartão, determinam qual modelo deve ser usado em cada contexto
  • Como a explicabilidade dos modelos fundacionais se compara à do XGBoost e por que as expectativas diferem nas áreas de crédito e fraude
  • Por que é difícil transformar embeddings em regras e como os agentes de IA podem ajudar

Deve ouvir este episódio se:

  • Lidera ou trabalha numa equipa de ciência de dados dedicada à fraude e está a decidir se deve explorar modelos fundacionais
  • Lidera uma pequena equipa de combate à fraude e enfrenta dificuldades com a rotulagem de dados e a engenharia de atributos
  • Quer uma explicação simples de como os modelos fundacionais diferem do XGBoost
  • Precisa de ponderar os ganhos de precisão face à latência e à explicabilidade
  • Têm curiosidade em saber como os embeddings e os agentes de IA poderiam trabalhar em conjunto para criar regras de deteção de fraude
Notas do episódio e principais conclusões

Das regras ao XGBoost e aos transformers

Pedro explica a história de forma simples. Durante 10 a 15 anos, a prevenção de fraudes baseou-se em regras encadeadas. Depois, o XGBoost permitiu usar muito mais dados para tomar pequenas decisões com latência muito baixa. Essa abordagem foi o estado da arte durante cerca de uma década. Observei que, de uma perspetiva algorítmica, esses modelos se parecem muito com regras, e Pedro concordou que partilham as mesmas desvantagens. Exigem muito trabalho para serem criados, precisam de cientistas de dados para programar as variáveis e dependem de rótulos claros para a aprendizagem supervisionada. Se algo for rotulado incorretamente, entram dados de má qualidade e saem resultados de má qualidade, o que é especialmente difícil para empresas mais pequenas. A conclusão é que os modelos tradicionais funcionam, mas deixam o peso do trabalho sobre as pessoas.

O que muda com um modelo fundacional

Os modelos Transformer são a base dos LLMs que todos usamos e funcionam tratando a informação como uma sequência e prevendo a próxima saída mais provável. A questão de Pedro é o que acontece se fizermos o mesmo com os dados financeiros. A Plaid e outras empresas já codificam informações financeiras como uma sequência de eventos e criam embeddings que substituem as características definidas manualmente. As equipes dependem menos da engenharia de características e precisam de rótulos menos precisos, embora alguns rótulos ainda sejam necessários. Pedro observa um ganho de precisão de até 34%, o que corresponde ao que vejo no setor. No entanto, o mesmo problema persiste. Dados de baixa qualidade geram resultados de baixa qualidade, e até algo tão simples como registros de data e hora precisos é mais fácil de dizer do que de fazer. Esses modelos podem absorver mais dados, mas não temos uma oferta ilimitada de dados limpos, e as equipes talvez precisem começar a coletar dados que nunca armazenaram antes.

A latência, a precisão e a explicabilidade determinam onde cada modelo se enquadra

Perguntei como tudo isso se encaixa na aprendizagem automática e nos modelos de decisão mais recentes. Pedro analisa as variáveis-alvo e as restrições. O objetivo costuma ser a precisão e a sensibilidade, e a maior restrição é a latência. Uma decisão sobre uma transação com cartão tem um limite de tempo de cerca de 300 milissegundos ou menos, enquanto a adesão de novos clientes ou o abuso de promoções permitem uma margem muito maior. Na sua opinião, uma boa equipa de ciência de dados e engenharia de dados deve explorar as opções, saber o que está a otimizar e deixar que os dados decidam. Se um modelo conseguir uma precisão superior sob as mesmas restrições, é esse que ganha. A explicabilidade é a sua outra preocupação. O XGBoost nunca a resolveu por completo, mas é muito mais fácil de utilizar do que um transformer. No crédito, existe uma exigência regulamentar de explicabilidade; na fraude, não. A resposta certa depende do caso de utilização.

De embeddings a regras com agentes de IA

Acrescentei que os embeddings são um atalho para a engenharia de atributos, mas não podem ser usados em regras. O Pedro pensou imediatamente em agentes. A sua equipa está a experimentar fornecer embeddings a agentes e perguntar-lhes se conseguem transformar alguns deles em regras e identificar lacunas. Continua a ser necessário que uma pessoa avalie se o trabalho é bom. As coisas não estão a ficar mais simples, e o Pedro diz que estão a ficar mais divertidas. Concordo que temos mais opções e mais brinquedos com que brincar. Cada um implica compromissos que as equipas precisam de compreender antes de avançarem.

Conclusão final

Os modelos fundacionais são a ferramenta mais recente no combate à fraude. Não substituem dados de qualidade. O conselho do Pedro é ter clareza sobre o que se pretende otimizar, respeitar as limitações e deixar que os resultados determinem a escolha. Se estiver a escolher entre modelos tradicionais e modelos fundacionais, comece pela qualidade dos dados que lhes servem de base. Estamos apenas a começar a compreender como estes modelos transformam o trabalho e acredito que as melhores equipas os vão encarar como uma opção promissora a testar, e não como uma solução definitiva.

Ainda não quer encerrar a conversa sobre o meu tema favorito — e, espero, o seu também? Subscreva a newsletter The Saturday Fraud Strategist.

Conecte-se com Pedro Sanzovo | LinkedIn
Gerente-geral e líder de prevenção a fraudes na Plaid

Conecte-se com Chen Zamir | LinkedIn
Apresentador do The Saturday Fraud Strategist
Ajudando fintechs a criar defesas mais inteligentes contra fraudes
Coautor de “O guia de IA para quem combate fraudes”

Episode transcript
Chen Zamir
Chen Zamir
00:06
Hi everybody and welcome to another live special episode from
Chen Zamir
Chen Zamir
00:09
Money 20/20 Vegas with me today Pedro Sanzovo from Plaid.
Chen Zamir
Chen Zamir
00:13
Welcome to the show. Lovely to meet you.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:16
So happy to be here.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:17
Excited to be here Money 20/20 and talk to you about fraud.
Chen Zamir
Chen Zamir
00:21
Yeah. So we just started the conference has just started.
Chen Zamir
Chen Zamir
00:24
I want to speak to you shortly about found foundation models in fraud.
Chen Zamir
Chen Zamir
00:27
Tell us a bit about who you are, what you've been up to so far.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:31
Yeah, absolutely. Uh, so I joined Plaid only six months ago.
Chen Zamir
Chen Zamir
00:35
Okay.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:35
But I've been in the fraud space for, well my whole career really.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:39
It was started Palantir.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:41
That was my first foray into defensive technologies, let's call it.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:45
Moved back to Brazil and started my own company in the transactional
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:50
fraud space focused on the on merchants and servicing merchants,
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:54
helping prevent chargeback, uh, all that jazz.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
00:57
We we ran that company for about five or six years.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:00
Uh and had a successful exit to Signified.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:03
Back in 2018 at Signified, I was leading a lot of the the uh the
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:07
product work most focused on machine learning, modeling, making
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:12
sure that we're creating top-of-the-line models there.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:15
And I had the opportunity early this year to come over to Plaid
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:19
uh and help us continue building this incredible work that
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:23
Plaid has been doing over the last few years on building
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:27
really good identity and fraud products for the market.
Chen Zamir
Chen Zamir
01:33
Pedro,
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:34
Yeah,
Chen Zamir
Chen Zamir
01:34
Let's talk about foundation foundation models.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:37
Uh like to yeah, yeah.
Chen Zamir
Chen Zamir
01:39
Yeah that's that's the hippest term now in uh in fraud.
Chen Zamir
Chen Zamir
01:42
Let's start with the basics like what's foundation models?
Chen Zamir
Chen Zamir
01:46
Uh how are they different from regular legacy machine learning models?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:53
Cool. Yeah.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:54
The the way that I like to think about it like think
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
01:57
about the history of fraud prevention, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:00
Uh 10 maybe 15 years everyone was basically creating
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:03
those set of chained rules, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:06
Uh waterfalls that were basically making decisions.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:09
Um at some point the the industry realized that there
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:13
were a type of algorithmic approach namely XG boost.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:16
So the uh extreme gradient boosted uh random forests that allowed
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:20
for a lot more data to be encompassed into micro decisions that
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:25
ultimately resulted in better prediction with super low latency
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:29
and those became effectively the state-of-the-art in the
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:33
front end industry for the last 10 years at the very
Chen Zamir
Chen Zamir
02:48
They also look very much similar to rules,
Chen Zamir
Chen Zamir
02:50
right, from an algorithmic perspective.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:52
Yes. Right. And they have the same downsides of rules.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:55
What are the downsides of rules?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
02:57
They on one hand they're very they're very labor some to build.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:00
There's a lot of feature engineering
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:03
that needs to go into it, right?
Chen Zamir
Chen Zamir
03:05
Adapting the index, the features, yeah.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:07
It's not self-learning.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:08
You need data scientists to actually be coding up the features,
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:11
understanding what the traffic is telling you.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:13
And then like whenever there's a change in behavior, you need to
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:16
put your data scientists, your fraud investigators should actually
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:19
understand like okay what the hell is happening here?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:22
How can we codify that into a feature? Let's build that feature.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:25
Let's throw into the model. Let's train it offline.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:27
Let's move it online and then try to figure out
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:29
if it actually providing lift and precision.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:32
Um and it's also very very dependent on label
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:34
sets and very clear label sets, right?
Chen Zamir
Chen Zamir
03:36
Supervised learning.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:37
Supervised learning. Exactly. Supervised learning.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:39
So you need to point the model to this is
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:41
the thing that I want you to predict.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:43
This is the thing that I don't want you to predict, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:46
And you need to keep those models
Chen Zamir
Chen Zamir
03:48
And that's very labor intense, right? That's not easy.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:51
Labor intensive.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:52
It's very complicated, especially for smaller companies, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:56
Uh and then you get into this game that if you're
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
03:59
labeling things incorrectly, you're screwed.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:02
Trash in, trash out, right? It's a massive problem.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:05
What happened over the last let's say five years or so
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:09
is that um super super super smart people figured out
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:13
that there was different way of codifying information
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:16
that does not resemble those decision trees, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:19
And those are the transformer-based models.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:22
Fundamentally those transformer-based models is what is
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:26
backing the LLMs right that we all use uh to an extreme
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:30
now like ChatGPT, Anthropic, Claude all that jazz.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:33
And all it does to a certain extent is that it codifies
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:37
information as a sequence of inputs, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:40
And LLM's those are words or those are letters and then tries to
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:44
predict what is the next most likely output of that sequence.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
04:48
Again for LLMs that's a word that's a letter
Chen Zamir
Chen Zamir
05:03
What what If we could do the same for fraud, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:05
What if we could do the same for financial data?
Chen Zamir
Chen Zamir
05:09
Mhm.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:09
I I think that's the question that everyone started asking.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:12
It's like, okay, instead of codifying letters and words,
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:16
what if we codified financial information, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:19
So that is what well we at Plaid started doing and a
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:22
lot of players in the industry are doing as well.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:25
We're codifying financial information.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:28
And we're codifying our data as the sequence of events.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:31
And then we're creating those things called embeddings, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:35
Which ultimately are the inputs for the model
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:38
to try to figure out, okay, what is the next
Chen Zamir
Chen Zamir
05:41
Instead of these handcrafted features that you described before for
Chen Zamir
Chen Zamir
05:44
post, we have now embeddings that are created by the foundation model.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:48
And the beauty of this approach is that you rely less on
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:51
humanmade feature engineering and also label science.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:54
You still need to some labels.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:56
You need to feed into it, but it doesn't need to be the same
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
05:59
degree of precision that actually boost models needed.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:02
What's the downside?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:03
There there are a few downsides and we're still
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:06
kind of learning how to use them, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:08
Uh but actually before we go to downsides, they're good.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:12
They're predictive, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:13
At least I think we we're seeing right
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:15
now an uplift in precision up to 34%.
Chen Zamir
Chen Zamir
06:18
That's very much in line with what we're seeing.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:21
That's in line with what we're seeing in industry.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:23
Uh but again, I think what's important to keep in mind is
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:27
that the same problem that existed in actually boost trash
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:30
in trash out still exists to a certain extent.
Chen Zamir
Chen Zamir
06:33
Interesting. How does trash look like for a foundation model?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:37
It's poor data, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:38
If your data is not clean, if the data quality is poor and on its
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:41
essence, the model is not going to be able to codify those sequences.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:45
So a simple example is one of the most interesting elements
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:49
of those foundational transform transformer based models
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:52
is uh the time based events and sequencing, right?
Chen Zamir
Chen Zamir
06:55
Yeah, yeah, yeah.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
06:57
If you don't have a really good way to key those times and like
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:00
that's something that if anyone has ever worked with data, you
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:04
know that making sure that your time columns or time data assets
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:08
are precise and they're not getting funky with like time
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:11
zones and whatnot, it's easier said than done, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:14
And that problem still exists in transform models.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:17
So you need to make sure that that underlying data asset is
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:20
trustworthy, is rich, and then the algorithm is going to do its work, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:25
And it's going to do its magic.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:27
Uh but it doesn't solve for the foundational issue.
Chen Zamir
Chen Zamir
07:30
And it sounds simple, but I think where where I've seen
Chen Zamir
Chen Zamir
07:33
organization kind of like get it wrong or get it messy
Chen Zamir
Chen Zamir
07:36
is exactly where you know like status has arrived.
Chen Zamir
Chen Zamir
07:39
And you need to kind of like keep the time stamp of the status
Chen Zamir
Chen Zamir
07:42
or velocity attributes and like when did I actually see these
Chen Zamir
Chen Zamir
07:46
links or connections and when you don't have that.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:49
Totally.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:49
And now it's interesting because those models they're able to absorb
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:53
a lot more data than actually boost trust or to an extent, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
07:57
It's like the whole kind of dynamic that we're seeing even on
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:00
the LLM is that the more compute that we put into it, the more
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:04
data that we put into it, the more precision it shows.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:07
In theory, that should be the same that we
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:09
should see in the financial ecosystem.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:11
The more data that we put into it, the more compute that we put
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:15
into it, the better precision we're going to see, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:18
But the problem is that data we don't have a 8:25 limitless
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:21
amount of good quality cleaned data, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:23
What we're learning is okay like we have our essential
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:26
set of that asset that is going to the model.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:29
Now let's bring this new that asset that we have
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:32
never been able to incompass an extra boost. Oh damn, okay.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:35
That that asset is not ready for it it needs to be cleaned it needs
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:39
to be treated it should like we need to start collecting it.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:42
Data that we weren't collecting before that
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:44
we weren't storing the right Hey. Yeah.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:46
Um, so there's there's again a lot of
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
08:48
foundational work to make that work.
Chen Zamir
Chen Zamir
08:51
Yeah.
Chen Zamir
Chen Zamir
08:52
And I think that, you know, to an extent it still is
Chen Zamir
Chen Zamir
08:55
the same problem that we had with again legacy. It's funny.
Chen Zamir
Chen Zamir
08:59
It's funny to kind speak about like that
Chen Zamir
Chen Zamir
09:02
with legacy machine learning models.
Chen Zamir
Chen Zamir
09:04
Also there you cannot really train on fresh data without
Chen Zamir
Chen Zamir
09:08
auditing the data and making sure that it's it's
Chen Zamir
Chen Zamir
09:11
clean and the labels are clean and so on. Last question.
Chen Zamir
Chen Zamir
09:15
Uh super new, like what two months maybe, but we now have Jev.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
09:20
Yeah.
Chen Zamir
Chen Zamir
09:21
Uh so you know we're kind of like we we are just getting our heads
Chen Zamir
Chen Zamir
09:24
wrapped around this concept of foundation models and now we have the
Chen Zamir
Chen Zamir
09:28
system one or decision models or whatever you want to call it, uh that
Chen Zamir
Chen Zamir
09:32
you know work differently than uh than you know LLM and foundation models.
Chen Zamir
Chen Zamir
09:37
Uh they are much more fit uh for a purpose in terms of
Chen Zamir
Chen Zamir
09:40
um you know like real time operations and so on.
Chen Zamir
Chen Zamir
09:43
How do you see the play between you know we have like
Chen Zamir
Chen Zamir
09:46
machine learning models, we have foundation models
Chen Zamir
Chen Zamir
09:49
and we have now decision decision models.
Chen Zamir
Chen Zamir
09:51
How do you see like the the scope here?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
09:54
I mean it's you got to think about what are your target variables
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
09:58
in any exercise and what are your constraints?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:00
For the world of fraud prevention, Generally what I'm
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:04
thinking about when I think about target variables,
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:07
I'm thinking about precision recall right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:10
I'm thinking about true positives, false positives,
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:13
the quadrants that we all know.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:15
What but I got to think about the constraints.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:18
And a big constraint is latency
Chen Zamir
Chen Zamir
10:20
Yes
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:21
Right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:21
So depending on the use case that you're facing latency
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:25
constraints are going to be wildly different.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:27
If you're trying to make like transaction transaction decisions for
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:31
credit cards for instance or debit cards, your latency budget is
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:35
very very minimum generally under 300 milliseconds or or or even lower.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:40
If you're doing onboarding or promo abuse, then
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:42
your latency constraints are way larger.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:45
So then you have a lot more room to play with. Okay.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:48
Optimizing the target variables.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:50
And so the way that I think about it is if you have a
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:53
good data science team and a good data engineering team,
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:56
you're bound where where it's almost mandatory that
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
10:59
you need to explore the options out there. You need to do research.
Chen Zamir
Chen Zamir
11:04
Yeah.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:05
And then you need to have clarity on what are
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:07
you actually trying to optimize for, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:10
Uh and honestly, I think the beauty of our of our industry
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:13
is that the data tells the truth ultimately.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:16
If you have a model that beats your precision with the same kind of
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:20
constraints and respecting the constraints in place, it wins.
Chen Zamir
Chen Zamir
11:24
Yeah.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:25
Right. Then there's another, if I may,
Chen Zamir
Chen Zamir
11:27
Go for it.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:28
There's another thing that gets me concerned about this
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:31
new age of modeling which is explainability, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:35
So like one of the things and XG boost never really solved that but
Chen Zamir
Chen Zamir
11:40
It's much easier with XG Boost.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:41
It's much easier, right? It's much easier.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:43
You can get a sense like humans can interact with
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:46
the outputs of those models in decent ways.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:48
It's much harder to do with trans with foundational
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:51
models for transformer based models and with Jev
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:54
I imagine that the constraint is the same. The problem is the same.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:57
So, how can we walk that line?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
11:59
And again, it depends a lot of the use case, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:02
If we're talking about risk, credit has a
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:04
regulatory mandate for explainability.
Chen Zamir
Chen Zamir
12:07
Yes.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:08
That is very different from fraud.
Chen Zamir
Chen Zamir
12:09
Yes.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:10
Right. So,
Chen Zamir
Chen Zamir
12:11
And and I would add maybe to to close uh things off, I would
Chen Zamir
Chen Zamir
12:14
add that, you know, these embeddings, you know, they they
Chen Zamir
Chen Zamir
12:18
short like it's a shortcut to feature engineering.
Chen Zamir
Chen Zamir
12:21
I don't need the talent. I don't need the effort.
Chen Zamir
Chen Zamir
12:24
I don't need to handcraft it.
Chen Zamir
Chen Zamir
12:26
But I can also not really use them in rules.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:29
Yep. Exactly. So, so yeah,
Chen Zamir
Chen Zamir
12:32
So then you need to kind of like go through that
Chen Zamir
Chen Zamir
12:34
effort in order to to like write rules.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:36
Totally. And then it gets into I think you're totally right.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:39
And that's well immediately I start thinking about agents, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:42
It's like okay, you know who can kind of understand some of those
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:46
embeddings but in the humans potentially agent, right?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:49
So like things that we're experimenting with is
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:52
okay can we feed agents embeddings and can they
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:54
transform some of those into rules, right? Can they find the gaps?
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
12:58
Uh and obviously you you still need a human in the loop to
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
13:01
try to figure out okay like is this actually good work.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
13:04
But I think that's the direction that we need to explore
Chen Zamir
Chen Zamir
13:08
Yeah I mean folks uh things are not getting simpler
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
13:11
No
Chen Zamir
Chen Zamir
13:12
That's the bad news
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
13:13
They are getting more fun I think
Chen Zamir
Chen Zamir
13:14
The good news it's more fun, more options, more toys to play with.
Chen Zamir
Chen Zamir
13:17
Pedro enjoyed it. We should uh do a proper sit down
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
13:20
Lets do it
Chen Zamir
Chen Zamir
13:21
In the podcast and enjoy the conference, man.
A close-up headshot of a young man with brown eyes, messy brown hair, and light stubble wearing a black t-shirt against a plain, light background.
Pedro Sanzovo
13:24
Thank you very much. Appreciate it.
Chen Zamir
Chen Zamir
13:26
See you next Saturday.