Khe Hy says Fable one-shot several tasks that Opus likely could have done only with heavy steering, after he fed it about 200 call transcripts and 50 decks.
Brett Caughran plans a real-world test of 100 earnings previews as earnings season starts, a test he expects Fable to fail.
Khe Hy describes a Karpathy-style self-updating knowledge wiki as the fix for clients whose search pulls only about 17 of 500 documents into context.
Fable impresses in vibe checks, but investors still need real evals
Fable looks like a real step up from Opus when tested on messy, real business material, but the hosts only have vibe checks and are moving toward structured evals. The harder problem is no longer building AI prototypes but organizing knowledge, choosing the right tools and deploying them to improve investment decisions. Client skepticism and token costs show why that gap matters.
Takeaways
Khe Hy tests new models with his own business files. Lacking formal evals, he feeds each model about 200 call transcripts, 50 decks, email and survey data, then asks for a website, strategy document and deck. Fable's strategy document was most insightful, surfacing one minute from roughly ten thousand minutes of calls.
Fable needs less steering than Opus. Fable one-shot tasks Opus could likely only do with heavy steering. Brett Caughran saw it build a coherent pattern recognition engine for timing shorts. Khe's client tool generates about 20 Claude skills per investor, though his cost guess was nearer $500 than $30.
Brett will run a large real-world eval. He plans 100 earnings previews on 100 companies, forecasting revenue, EBITDA, EPS and stock direction. He cites a Thinking Machines report where a model scored high 40s on financial judgment and mid 70s with a judgment layer added.
A self-updating wiki beats pointing models at raw files. Keyword-style search may pull only about 17 of 500 documents into context. A wiki converts files to markdown, summarizes with tags and links ideas. Khe advises starting with five names, and keeping index documents so agents can find models.
Skills work best at moderate depth and single purpose. Brett turned a 120 page modeling manual into skills, arguing a four to six page prompt is right, since one page is thin and eight loses attention. Subagents use separate context windows. Excel modeling still disappoints, so firms are trying dozens of subagents.
Claude Cowork and Claude Code are converging. Cowork offers live artifacts, easier scheduled tasks and a disposable virtual machine that IT prefers. Claude Code has root access and suits developers using GitHub scrapers. Both hosts say the gap is compressing, and Cowork will soon run tasks with the laptop closed.
Client skepticism is rising while deployment lags. Khe says client skepticism rose about 15 percent last quarter, citing token costs and invisible checking work, and he could not tell Opus 4.5, 4.6 and 4.8 apart. Brett says prototypes are easy but deploying AI for better decisions is hard.
Tap a name to go there · tap a line to see why they are linked