D.A.D.: Researchers Hand Stock-Picking Entirely to AI — and Report Returns Few Hedge Funds Match — 5/18
The Daily AI Digest
Your daily briefing on AI
May 18, 2026 · 7 items · ~4 min read
From: Hacker News, NBER, arXiv
D.A.D. Joke of the Day
My AI keeps giving me the same answer no matter how I ask. Guess it's not wrong, I'm just not ready to hear it.
What's New
AI developments from the last 24 hours
Counterpoint: AI Coding Tools Won't Help If Requirements Are Your Real Bottleneck
A provocative blog post argues that AI coding tools won't actually speed up software development—because coding was never the bottleneck. Drawing on process optimization classics like 'The Toyota Way' and 'The Goal,' the author contends that the real constraint is upstream: translating vague requirements into clear solutions. AI can generate code faster, but if teams are stuck waiting on stakeholder decisions or unclear specs, faster typing changes nothing. The post offers no data, but frames a counterintuitive challenge to AI productivity assumptions.
Why it matters: As companies invest in AI coding assistants expecting efficiency gains, this argument suggests the payoff depends entirely on whether coding—versus requirements gathering or decision-making—is actually your team's constraint.
Gruber Dismisses AI-Agent Hype: Apple Sells Experiences, Not Technologies
Tech commentator John Gruber pushed back on Steven Levy's Wired argument that Apple's next CEO must deliver a 'killer AI product.' Gruber contends AI is a technology enabler, not a product category—defending Apple executive John Ternus's view that Apple ships experiences, not technologies. He dismisses predictions that AI agents will disrupt the iPhone ecosystem by decade's end as 'pure fever dream high-on-the-hype fantasy.'
Why it matters: The debate reflects a genuine strategic divide: whether AI demands new product categories or will be absorbed into existing ones—a question every company building AI roadmaps is navigating.
Cost Analysis: Running AI Locally on MacBooks May Be 3x Pricier Than Cloud
A cost analysis argues that running AI models locally on a high-end MacBook Pro costs roughly three times more than cloud inference through services like OpenRouter. The calculation: amortizing a $4,299 M5 Max over five years plus electricity yields about $1.50 per million tokens, while cloud alternatives offer comparable models at $0.38-0.50 per million tokens with faster speeds (60-70 tokens/second versus 10-40 locally). Community reaction pushed back, noting the analysis charges the full laptop cost rather than incremental inference costs, and ignores the value of data privacy and avoiding future price increases.
Why it matters: For teams evaluating local versus cloud AI, this frames the tradeoff: cloud is cheaper per token today, but the math changes significantly if you're already buying the hardware—and privacy requirements may override cost considerations entirely.
What's Innovative
Clever new use cases for AI
Quiet day in what's innovative.
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Quiet day in what's controversial.
What's in the Lab
New announcements from major AI labs
Quiet day in what's in the lab.
What's in Academe
New papers on AI and its effects from researchers
Researchers Hand Stock-Picking Entirely to AI — and Report Returns Few Hedge Funds Match
Most AI in finance tries to predict what individual stocks will do, then leaves a human (or another model) to decide what to actually buy. AlphaPortfolio skips that middle step: the AI builds the portfolio itself, trained to maximize returns relative to the risk it takes on. In backtests on U.S. stocks with monthly rebalancing, the system reports beating a comparably risky benchmark by roughly 13 percentage points a year — a margin that, if it held up in live markets, would outperform almost any active fund manager. The researchers say the edge persists even after excluding small, hard-to-trade stocks. The team — Lin William Cong (NTU Singapore President's Chair Professor, on leave from his Cornell finance chair), with Ke Tang and Jingyuan Wang — describes it as among the first 'large' generative AI models in finance built for problems other than text.
Why it matters: If these results survive live trading — a significant if — this is a step beyond AI as forecasting tool: it's AI making the buy and sell calls itself, with claimed returns that would put it in elite hedge-fund company. That would reset what quantitative asset managers compete on.
University of Toronto Economist: Standard Methods for Measuring AI's Wage Impact May Get It Backwards
University of Toronto economist Joshua Gans argues that standard methods for measuring AI's effect on wages may be fundamentally flawed. Typical studies, he contends, hold job tasks fixed—but when AI automates some work, companies rebundle the remaining tasks into redesigned roles, and that reorganization determines which skills get rewarded. The paper's core claim: current wage regression analyses can actually get the direction wrong, showing wage gains where there are losses or vice versa, because they ignore how jobs reshape around automation.
Why it matters: If the analysis holds, much of the empirical research reassuring workers about AI's wage effects—or alarming them—may be measuring the wrong thing, complicating workforce planning and policy debates.
LLM Analysis of 114,000 Claims Shows Media Systematically Amplifies Climate Severity
A large-scale study using LLMs to analyze roughly 114,000 matched claim pairs found that climate information systematically shifts toward more severe interpretations as it moves from technical IPCC reports to policymaker summaries to newspaper coverage. The severity increase comes mainly from emphasizing higher-impact numbers within scientific ranges—not from dropping uncertainty language or cherry-picking worst-case scenarios. The pattern held across both left- and right-leaning outlets, and claims stayed within accepted scientific bounds. The University of Maryland's Sebastian Galiani, with Franco Mettola La Giglia and Raul A. Sosa of Argentina's Universidad de San Andrés, examined all six IPCC Assessment Reports (1990-2023) and coverage from ten major US and UK newspapers.
Why it matters: This offers empirical evidence for a long-suspected pattern in science communication—and demonstrates how LLMs can analyze information translation at scale, potentially useful for organizations tracking how their own technical content gets simplified downstream.
Researchers Give AI Companions Emotional Memory That Persists Across Conversations
Researchers developed Cross-Temporal Emotion Modeling (CTEM), a framework that gives AI companions persistent emotional memory—linking an agent's past interactions with users to its current emotional responses. Unlike standard chatbots that treat each conversation as isolated, CTEM-powered agents remember relationship history and let that shape how they respond emotionally. The team tested their prototype 'Auri' on a messaging platform over 21 days, reporting improvements in perceived naturalness and emotional coherence, though specific metrics weren't disclosed in the abstract.
Why it matters: This is early academic work, but it signals where AI companions and customer-facing agents are headed: systems that don't just remember facts but develop something like relational continuity—potentially useful for healthcare support, coaching apps, or any context where emotional consistency matters.