OpenAI's three core focus areas for ChatGPT in financial services
OpenAI focused on three core areas when building ChatGPT for Financial Services: financial data, an interface built for finance, and the firm's format.
36,915 results · Most relevant
OpenAI focused on three core areas when building ChatGPT for Financial Services: financial data, an interface built for finance, and the firm's format.
The speaker saw the value of the outbound follow-up queue not in centralizing all agent-sent messages — agents send so many that would be data overload — but in using agents to help with follow-up: what happens when a prospect responds to outbound. Previously the team logged into each platform, responded, and hoped the prospect booked a meeting.
None stated.
Wells Fargo's downstream task for the foundation model was fraud classification: given a transaction, predict whether it will be fraudulent.
The same foundation model is now used across different organizations within Wells Fargo — for fraud detection, credit line increase, anti-money laundering, and marketing.
The goal of the transaction foundation model is to learn the relationship between transactions over time and transactions across customers. A transaction foundation model can be used for downstream tasks like credit line increase and fraud detection, flagging bad actors from an anti-money-laundering perspective.
Wells Fargo pre-trained their transformer blocks at scale from scratch using their own transaction data. They used the open-source architecture (Llama-style releases) but pre-trained it from scratch on their own GPU cluster. The foundation model was built on billions of transactions from Wells Fargo customers, and they did not pre-train from any existing open-source model — they used an existing architecture but pre-trained from zero on their own data.
The Wells Fargo foundation model beats XGBoost from a latency perspective and delivers much better reduction of false positives and a better fraud detection rate.
Kurtz reports that shadow AI has become a huge issue for many companies: agents like Claude Code and Codex are popping up all over enterprise environments, and organizations struggle to understand where their AI is taking place, who's doing what, and what identity agents are operating under. Nobody knew where rogue and shadow AI was. He cites a stat that an employee can have 90 different AI agents on average — which he thinks is a low estimate — and says this represents a much bigger opportunity for CrowdStrike than just EDR.
stat is publisher-reported; Kurtz calls it a low estimate
Kurtz argues the only way to fight AI today is with better AI. Defenders have advantages: they know their own systems, have the architectural diagram, and have the code. Putting that defender knowledge into AI dramatically increases the probability of outsmarting offensive AI. The hosts add that knowing your environment is one of the only advantages a defender has in classic cybersecurity.
Kurtz's strategic argument
Kurtz claims enterprises have 2-3 magnitudes more AI agents operating in their environments than they are aware of, and that an employee can have around 90 different agents — which he believes is a low estimate. He warns against assuming someone is reviewing the production code agents generate: nobody is watching what agents are doing, and a 'give it a task and come back in a few days' approach is problematic.
estimates attributed to Kurtz's belief
When a place lacks its own typographic identity, dig into its history to find why — colonialism often breaks the connection between spoken language and its visual form. Foley checked whether anyone had researched Irish typography with cultural context (often only chronologies exist) and looked back through Irish history asking 'what if Ireland had had a typographic evolution of its own?'
Evan Conrad argues it is very hard to solve this problem unless you are operating both the market and the clusters at the same time, because GPU compute isn't sufficiently standardized or reliable for the market operator to be divorced from the physical infrastructure.
Greg Brockman reports that OpenAI published results resolving almost 400 very important open math problems, and describes this week's math results as a real milestone in AI's ability to massively accelerate scientific discovery.
A16Z is announcing the expansion of the ACC crypto fund into A16Z Frontier. The firm historically created a vertical fund behind each new area after Chris identified it, but its vertical strategy, while working well, still leaves new things falling outside the verticals, and the process of turning new areas into a practice has historically been somewhat undefined. If the frontier investing motion works, it may become a whole fund at some point.
The speaker argues that hardware has historically been terra incognita for software people — it has been very easy to blow up companies going into hardware. But basically any form of hardware can now be taken over and reprogrammed using AI-powered reverse engineering tools. Anybody who understands software may automatically be able to do many things that were historically difficult in hardware, in a way never conceivable before. AI is entering a renaissance where software understanding unlocks historically difficult hardware problems.
Marc Andreessen says the gap between what early adopters understand is possible with AI and what normal people understand is the biggest delta he has ever seen in his career. AI-native early-stage startups have a completely different view on what they can do with AI than typical incumbents in almost any industry, and AI capability is racing ahead week by week, with leading-edge entrepreneurs getting steadily more advanced than incumbents by the week.
asserted
The guest frames spec-driven development as a call to action for designers, design-minded people, and design psychologists to insert themselves into deciding what products should be, stepping upstream in the decision process — reasserting their value by taking their design intent and design skills into the upstream decision step. Shaping product outcomes so they are user-centered and follow sound design principles has always been the designer's goal.
The guest endorses that the shift from building to deciding is a big opportunity for UX people and design psychologists to reevaluate where they can add value.
The research model must find the right reports, compile the same reporting periods, keep different businesses separate, and explain what the numbers mean. The analyst receiving the report needs to open the source behind an important claim, check a calculation, and see what evidence could change the conclusion. A useful report gives the analyst something they can check, and that is the standard for a finished research job.
Geico began strategically investing in Geico-managed AI agents on a variety of harnesses and models to offload patch and minor-version work all the way from spec to deployment, so engineers can stay focused on major work.
Choose sources → to get a library of your own and a daily brief.