Fable Is Here, But Is It Actually Better? | Invest with AI Vibe Check
In this episode of Invest with AI, Brett Caughran and Khe Hy run a focused vibe check, a round-the-horn read on where AI actually stands for investing right now with important context for the people who are in the seat.
Brett and Khe are daily power users that work with buyside investors to integrate these tools into real workflows. Two people who run these models daily still won't tell you they've gotten meaningfully better over the last six months. Nobody noticed much of a gap between Opus 4.5, 4.6, and 4.8, then Fable shipped from Anthropic as its most capable release yet. The question that matters for a PM is whether the job has actually changed, and where the alpha sits when most of the street is misusing these tools the same way.
We get into:
Why two daily users still won't say the models got meaningfully better in six months, and why that's a fair question to evaluate
Khe's informal eval on every new model: dump everything the business has ever produced and watch what it surfaces (Fable found the one salient minute buried in 10,000 minutes of calls)
Brett's short-signal test, where Fable one-shotted a pattern-recognition engine that used to take two or three rounds of steering
The 100-company earnings-preview eval Brett is running live this season: forecast revenue, EBITDA, and EPS, plus whether the stock trades up or down on the print
The Bridgewater / Thinking Machines result, where bolting a judgment layer onto an LLM moved financial judgment from the high-40s to the mid-70s
The Karpathy "knowledge wiki": turning a 500-document deal folder into an ontology an agent can actually traverse, instead of Cowork running command-F across 17 files
Why making your file system legible to agents (markdown, metadata, naming conventions) beats pointing a model at an unorganized folder
Thin vs. thick skills, and why single-purpose composable subagents that ring-fence their own context beat one bloated skill
Why complex Excel modeling still breaks, and the 30-subagent approach one firm is building to fix it
Claude Code vs. Claude Cowork: why the line is collapsing for knowledge work, plus the infosec case for a disposable VM over giving an agent root access
The buy-side vibe check: skepticism up ~15% this quarter, "bot-sitting" (Glean's term for the 30% of AI time spent babysitting the model), and the cost question
We're not coming at this as "experts" with all the answers. We're in it every day, testing, breaking things, and trying to understand where this is going. The goal of the podcast is simple: bring you along as we learn, and give you a clearer view of how AI is actually being used in investing. If you work in equity research, at a hedge fund, or on the buy-side and you're trying to make sense of AI, this is a good place to start.
Chapters (Timestamps)
[00:00] Intro
[00:57] — First Impressions of Fable (and Why the Ban Was a Relief)
[03:05] — The Eval That Found the Thing He Was Avoiding
[05:14] — Why Evals Are Hard, and the One-Shot Short Signal
[06:56] — The 100-Company Earnings Preview Test
[07:46] — The Bridgewater Judgment Layer (High-40s → Mid-70s)
[10:48] — Braintrust and Scoring MCP Vendors
[12:30] — The Karpathy Knowledge Wiki for 500-Doc Folders
[18:20] — Making Your File System Legible to Agents
[21:11] — Turning a 120-Page Manual into Skills
[22:04] — Thin vs. Thick Skills and Composable Subagents
[25:36] — Why Complex Excel Still Breaks (and the Subagent Fix)
[27:24] — Claude Code vs. Cowork: The Line Is Collapsing
[29:57] — The Infosec Case for Cowork (VM vs. Root Access)
[33:50] — Vibe Check: Skepticism Is Up, and "Bot-Sitting"
[36:26] — Digital Twins and Where the Buy-Side Actually Is
Want to actually build these workflows yourself?
The AI Accelerator is Fundamental Edge's 6-month cohort for investors who want repeatable AI workflows. Learn More below:
https://www.fundamentedge.com/ai-accelerator
Follow Invest with AI on:
Spotify: https://open.spotify.com/show/033xcEEovVViS7hIYwNuGZ
Apple Podcasts: https://podcasts.apple.com/us/podcast/invest-with-ai/id1896918892