LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A language model that can place stock orders will contradict itself unless it stores the original thesis, the positions, and the goals between sessions. The lecture builds that memory as files, connects it to a broker workstation for live data, and then spends the rest of the demo fighting read-only flags, dropped market-data sessions, and orders that never leave the queue. The architecture is a memory and a permission system. It is not evidence that an unattended model should run a portfolio.
Why it matters
The motivation in the talk is that people, including policymakers, are bad at forward decisions. The proposed assistant is a Python service between a broker workstation and a language-model API. Market data, news, positions, and order status arrive through the workstation's client and callback split, often called the client and the wrapper. The model can be swapped. The failure the design is trying to prevent is a blank session: yesterday's buy is forgotten, today's headline flips the book, and the headline was noise. Memory is a set of structured files. One block records trades. One records the thesis that justified each trade. One records portfolio state. Chat is compressed on restart. The model is told that a high-confidence thesis should survive a re-check unless the thesis itself broke.
Under that sits a position table. Selecting a row shows a narrative, a target, an entry, and a confidence. A chat panel can argue with that row. Goals such as aggressive growth or capital preservation are editable. Preset questions cover main risks, whether the book matches the goals, and a twelve-month alpha and beta. An automation switch can re-read the book on a cadence the user picks: one, five, ten, fifteen, or thirty units of minutes, hours, or days.
The idea in plain words
The model is given tools, not a blank prompt. A first smoke test in the demo listed available models. A second, on the paper port 7497, pulled headlines. The interface then filled with demo numbers, including a net asset value of 1.2 million dollars, before live data replaced them. Stale demo targets were still on screen: 160 for Nvidia and 245 for Apple. When the market snapshot returned, Apple was around 320, about 31 percent above the old target. The lecturer objected. They wanted either a higher target or an exit.
Later the model proposed selling 62 Apple shares to bring that name to 12 percent of net asset value. Market-data connections dropped and returned. Tool calls hung until the workstation calls were made non-blocking. A missing exchange field canceled an order. Client identifiers collided. A worker reused a socket and knocked out the data feed. Timeouts hit position sizing. That was the behavior the project wanted to show. It arrived only after a string of execution failures.
How to use it
If you study the design, separate memory from permission. Memory answers what the model believed and what it already did. Permission answers whether a tool may send an order, how large, and what happens when the socket dies mid-loop. Require a thesis record before a re-evaluation, which the demo itself forgot for AVGO. Treat a target that collapses to the last price as a failed update. Keep automation off until a single order works twice. The lecture's closing line, that you could leave live capital with the model, is the enthusiasm of the build. The same transcript is a list of canceled and stuck orders. Those failures are the lesson.
What usually goes wrong
Starting from an empty prompt so the model trades the latest headline. Leaving a read-only flag and a live flag in disagreement. Sharing one client identifier across threads. Trusting a demo thesis after the price has moved 31 percent past the target. And storing a credential in source because a smoke test was faster that way.
Questions people ask
What is the brain, in concrete terms?
Files that keep trades, the thesis and target for each name, timestamped goals, a short global summary, and compressed chat. Live positions still come from the broker. The files exist so a restart does not invent a new philosophy.
Did the assistant manage a 1.2 million dollar portfolio?
That net-asset figure was demo data on the first screen. The later session read a small live book of Nvidia, AVGO, and Apple, and struggled to modify it. Do not merge the two screens into one audited account.
Why did Apple's target become the current price?
The model treated a still-valid story as a reason to mark the target up to about 320, the print it had just fetched, rather than to a level that would justify continued risk. The lecturer's correction was to exit if the only target is the price you can already sell at.