Visual TL;DR
Agents Hit InvestingDriver
From the articleFrank Long told AI Engineer that agents will hit investing harder than coding and he demoed how Coatue is wiring that bet into its hedge fund workflow.
Fund As FunctionContext
From the article 2 mentionsHe described a fund as a function that ingests external context like news, prices and filings and internal context like positions, liquidity and past notes, then outputs selection, size and timing as a portfolio time series.
Korean Media PipelineEffect
From the articleUsing Korean media on semis as his case study, Long broke gathering into collect, dedupe, filter, translate, extract and cross check, treating chunking, image handling and model choice as hyperparameters to hill climb.
No Single PromptDriver
no prompt can compress all context into selection size timing tuple
From the articleHe was blunt that no single prompt can compress all context into the tuple, that many steps still want deterministic code or a person in the meeting, and that LLMs still fail at mapping news to portfolio impact without human judgment plus search tools.
LLMs Fail MappingOutcome
From the articleHe was blunt that no single prompt can compress all context into the tuple, that many steps still want deterministic code or a person in the meeting, and that LLMs still fail at mapping news to portfolio impact without human judgment plus search tools.
Agents Hit InvestingDriver
From the articleFrank Long told AI Engineer that agents will hit investing harder than coding and he demoed how Coatue is wiring that bet into its hedge fund workflow.
Fund As FunctionContext
From the article 2 mentionsHe described a fund as a function that ingests external context like news, prices and filings and internal context like positions, liquidity and past notes, then outputs selection, size and timing as a portfolio time series.
Every Pixel Via APICore
From the articleLong said every pixel a human sees should be reachable by an agent via API, and he credited a decade of KKR led data discipline at Coatue for filing every trade note to make that exposure possible.
Korean Media PipelineEffect
From the articleUsing Korean media on semis as his case study, Long broke gathering into collect, dedupe, filter, translate, extract and cross check, treating chunking, image handling and model choice as hyperparameters to hill climb.
Hyperparameters Hill ClimbContext
From the articleUsing Korean media on semis as his case study, Long broke gathering into collect, dedupe, filter, translate, extract and cross check, treating chunking, image handling and model choice as hyperparameters to hill climb.
No Single PromptDriver
no prompt can compress all context into selection size timing tuple
From the articleHe was blunt that no single prompt can compress all context into the tuple, that many steps still want deterministic code or a person in the meeting, and that LLMs still fail at mapping news to portfolio impact without human judgment plus search tools.
Human Plus CodeEffect
From the article 2 mentionsHe was blunt that no single prompt can compress all context into the tuple, that many steps still want deterministic code or a person in the meeting, and that LLMs still fail at mapping news to portfolio impact without human judgment plus search tools.
LLMs Fail MappingOutcome
From the articleHe was blunt that no single prompt can compress all context into the tuple, that many steps still want deterministic code or a person in the meeting, and that LLMs still fail at mapping news to portfolio impact without human judgment plus search tools.
