By 2030, equity research had stopped being a profession. Nobody could say when.

It started with earnings-call summaries. Then the models rebuilt spreadsheets, reconciled filings, tracked competitors, and updated estimates before a human had finished reading the press release. One analyst covered a hundred names, then five hundred. By 2028 an institution could rent the output of a fifty-person sector team for less than one junior associate.

Everyone had expected AI to make analysts better. It made analysis abundant instead, and abundance took the price to zero.

Facts stopped being worth much

For decades, firms paid for a scarce input: human attention applied to financial information. A good semiconductor analyst spent years on supply chains, factory visits, and cycle intuition. That was expensive because it was slow to build and impossible to scale.

By 2030, machines held continuously updated models of every listed company on earth, not spreadsheets but causal maps of customers, suppliers, pricing, capacity, regulation, management credibility, and thousands of relationships between them. Research had been episodic: quarterly prints, investor days, conferences. Now it ran on every price change, job posting, freight movement, patent, and line of code, in real time.

The 60-page initiation report died first. It had existed to compress months of work into an hour of reading. Once the whole analysis could be regenerated on demand, nobody needed the compression. Investors stopped asking “What’s your thesis on Nvidia?” and started asking “What moved in the Nvidia belief distribution in the last six months?” The answer came back as probabilities and return distributions, not prose.

The problem analysts had been solving was the wrong one

The industry assumed the core task was forecasting the company. Machines got good at that quickly. The harder and more valuable question turned out to be: what does the market already believe?

By 2030 every serious system ran two models in parallel. One estimated reality. The other estimated consensus. Alpha was the difference.

The questions changed accordingly. Not “how fast will this grow?” but “what growth rate is the price already assuming, which futures are underpriced, and which consensus assumption would cause the biggest repricing if it broke?”

Fundamental research had become expectations research.

Machines made hypotheses cheap, so the work became killing them

Markets did not become efficient. Simple information such as an earnings beat, a channel check, a valuation screen got absorbed in seconds. But a machine could also produce ten thousand coherent explanations for the same data point: demand weakness, inventory digestion, share loss, macro, a product transition, an accounting quirk. Generating theses was free.

The bottleneck moved to eliminating them.

The best firms stopped asking AI for better theses and built engines to destroy them. A bull agent claims enterprise AI adoption is accelerating; the falsification layer asks what else would have to be true — cloud consumption, hiring, data-center utilization, networking demand, supplier lead times — and goes looking for contradictions. Adoption booming while infrastructure usage sits flat weakens the belief.

The pipeline changed from thesis → evidence → recommendation to hypothesis → predicted observations → adversarial tests → posterior → position size. Research started to look like experimental science.

A single stock was no longer analyzed by one model but by hundreds — bull, bear, historical analogue, accounting, substitution, positioning, supply chain, management credibility, reflexivity — deliberately set against each other, with an arbitration layer deciding which disagreements deserved more work. Firms had built artificial scientific communities. The twelve-person analyst meeting looked antique.

Sector coverage dissolved

The sector analyst went next, not because expertise was useless but because the economy doesn’t respect the org chart. A research department splits semiconductors, utilities, industrials, real estate, and materials. An AI-compute thesis runs through all of them: compute drives electricity, electricity drives transmission, transmission drives transformers, power constraints move data-center sites, sites reprice fiber and water. Once a model could reason across the whole chain, the silo was just friction. Departments reorganized around economic systems — an “AI compute economy model” in which any given stock is one node.

Alpha migrated down those chains. Knowing AI demand was strong was worthless; everyone knew. The money sat several hops out: adoption → inference load → hyperscaler economics → power scarcity → behind-the-meter generation → turbine demand → component shortages → pricing power at a supplier nobody associated with AI. The best trades of 2030 looked bizarre at initiation because the machines had followed the constraint chain further than any human had.

Research also turned active. Instead of waiting for filings, a system estimated the information value of each possible experiment — developer surveys, marketplace rankings, GitHub dependencies, former-customer interviews — and spent budget on the one that would cut uncertainty most. “What should I read?” became “What evidence should I create?”

Then the models started modeling each other

By 2029 a large share of institutional capital was run by AI, so every model had to forecast other models: how systematic strategies would read a services-growth number, where risk engines would cut exposure, how those flows would be reinterpreted. Fundamentals → machine interpretation → positioning → price → interpretation again. The fundamental/quant distinction stopped meaning anything. All investing had become partly game-theoretic.

In 2030 that loop broke. Several dominant systems, trained on different data at different firms, had converged on similar internal representations. A macro shock invalidated one shared assumption and thousands of portfolios cut the same exposures at the same moment. Each model was rational. Together they were unstable, and liquidity vanished. It became known as the Model Consensus Crash, and it taught a lesson humans had already learned several times: a truth everyone sees can be the largest risk in the market. Model diversity became as important as model quality.

What the moat turned out to be

Frontier models commoditized. Every fund had strong reasoning engines, agents, and alternative data. Returns stayed widely dispersed anyway, because the edge wasn’t intelligence but the architecture around it. The firms that won had five things:

• Better questions — they investigated what rivals hadn’t framed.

• Better ontologies — they represented economic reality differently.

• Proprietary memory — decades of hypotheses, decisions, errors, and outcomes a machine could learn from.

• Explicit epistemology — written rules for what counts as evidence and when beliefs must change.

• Capital discipline — because being right decides nothing; sizing, liquidity, horizon, and survival still do.

AI commoditized analysis. It didn’t commoditize judgment. It changed what judgment meant.

Humans stayed, but higher up: objective functions, risk governance, ontology design, model diversity, regime-change detection. The job was no longer knowing more facts. It was deciding how a system should form beliefs about a future it can’t know.

Looking back, the scarce resource had kept moving — from information to interpretation, from interpretation to questions, from questions to epistemology, and finally to capital. A perfect forecast still leaves someone to decide how much to own, for how long, and against what.

Epilogue

In 2030 a well-known fund closed its fundamental research department. It had once employed 118 analysts. Six humans remained, none covering stocks. One studied model failures. One designed experiments. One maintained the ontology. One audited causal inference. One governed portfolio risk.

The sixth spent her days on one question: What are we all assuming that might no longer be true?

The boss called her the most valuable analyst the firm had ever employed. She never wrote a research report.

Subscribe now

The most interesting thing I saw was not another impressive AI demo.

It was how quickly AI is becoming ordinary.

Engineers use coding agents constantly. Among the people I spoke with, Codex came up far more often than Claude Code. More importantly, developers who previously defaulted to frontier American models are increasingly experimenting with Chinese open-source models, which is similar to what’s happening across American enterprises.

Chinese companies are no longer asking whether employees should use AI. They are calculating how many employees they still need once they do.

Doubao by ByteDance is everywhere.

Tencent’s WorkBuddy is quickly appearing in actual office workflows.

And conversations about AI have moved surprisingly far beyond productivity. I heard people casually discuss a future where relatively few humans perform economically necessary work, while everyone else receives some form of government-provided basic income.

Meanwhile, outside the AI bubble, China felt strikingly different from the country I visited 3 years ago.

Cities were cleaner, more organized and more digitally integrated. Cameras were everywhere. Infrastructure felt increasingly finished rather than developing.

Yet the economy simultaneously felt intensely deflationary.

Cheap food. Cheap services. Cheap delivery. Constant promotions. Businesses fighting for increasingly reluctant consumers.

Put these observations together and a more interesting possibility emerges:

China may become the world’s first large-scale experiment in AI-driven deflation.

Not because China necessarily has the world’s best proprietary model.

Precisely because it may not need one.

1. The model is already becoming interchangeable

The first surprise was coding.

Globally, Claude Code remains much more widely used than Codex.

That is not what I observed in China.

This is anecdotal, not a Chinese market-share estimate. But there are structural reasons China could diverge.

Anthropic explicitly restricts access by Chinese-controlled companies, and Alibaba reportedly banned internal Claude Code use in July. OpenAI also does not officially support API access from mainland China, so neither ecosystem has frictionless access. Yet the broader direction is obvious: relying permanently on expensive foreign proprietary intelligence is increasingly unattractive when domestic alternatives are improving this quickly.

And the open-model data is much stronger than my anecdote.

That changes the mental model.

The AI race may not converge toward:

one company owns intelligence and charges everyone rent.

It may increasingly look like:

frontier labs discover intelligence, open models compress its price, and applications capture the surplus.

China is unusually well positioned for the second world.

2. Doubao shows what happens when AI stops being a destination

Doubao surprised me even more.

It is not merely a popular Chinese chatbot. It is becoming a default interface for ordinary people who would never describe themselves as AI users.

That is the important transition.

AI adoption becomes economically important when people stop thinking about whether something is AI at all.

The interface disappears into everyday behavior.

China may reach this state unusually quickly because ByteDance, Tencent and Alibaba already own enormous distribution surfaces.

The winner therefore may not be the company with the best benchmark.

It may be the company that can make intelligence ambient.

3. WorkBuddy may be more important than another model release

Tencent WorkBuddy is an even more interesting signal.

It launched only this year, yet QuestMobile data show its active user base rising more than 115% over three months. Another industry report put June PC visits at roughly 21 million, leading China’s AI-native office-agent category.

Why does this matter?

Because the economic transition from chatbot to agent changes what AI competes with.

The unit of competition shifts from:

cost per token

to:

cost per completed job.

Once that happens, the denominator investors should watch is no longer AI users.

It is human hours displaced per dollar of inference.

4. Chinese companies appear to understand this faster than many investors

Several people I spoke with described companies explicitly trying to reduce headcount through AI.

Public evidence suggests this is not isolated.

Media recently documented what it called “quiet layoffs” across Chinese companies as employers adopt AI, particularly in technology, advertising, entertainment and other white-collar functions. Firms are often reducing hiring or trimming teams gradually rather than announcing giant AI layoffs.

This distinction matters.

AI does not need to fire 30% of workers tomorrow to transform labor economics.

A company with 100 workers that previously expected to become a 150-person company may simply remain at 100.

That never appears in a layoff announcement.

But economically, 50 jobs disappeared.

This may become one of the largest measurement problems of the AI era.

5. Deflation may accelerate AI rather than merely coexist with it

Consumer goods prices fell month over month. Pork prices were down 13.3% year over year. Housing remained weak. July retail sales grew only 0.6%. New-home prices were still falling year over year.

The lived economy feels like one of permanent price competition.

And that creates an unusual AI adoption loop:

weak demand → margin pressure → automation → lower labor requirements → lower costs → more price competition

This is almost the opposite of the current American AI story.

In America, AI currently looks inflationary in the physical economy:

more data centers,

more electricity,

more GPUs,

higher engineering salaries,

more capital expenditure.

In China, AI may increasingly look deflationary in the service economy:

Same technological shock.

Different macro transmission mechanism.

That distinction may matter enormously for investors.

6. China’s physical infrastructure hints at another advantage: deployment capacity

The other thing that struck me was simply how developed much of urban China now feels.

Clean streets.

Dense transportation infrastructure.

Fast delivery.

Digitized payments.

Highly organized public spaces.

And China has formal national rules governing public-security video systems and facial-recognition deployment, which itself illustrates how embedded this infrastructure has become.

There are legitimate privacy questions here. But from an AI-adoption perspective, the broader observation matters:

China is extremely good at deploying systems once technology becomes cheap enough.

Western analysis often focuses on invention.

China’s comparative advantage may increasingly be diffusion.

EVs demonstrated this.

Solar demonstrated this.

Mobile payments demonstrated this.

E-commerce demonstrated this.

AI may follow the same pattern.

The State Council has explicitly targeted more than 70% penetration of next-generation intelligent terminals and agents by 2027 and more than 90% by 2030.

Those targets may or may not be measured cleanly.

The direction is unmistakable.

7. The most unsettling conversations were not about models

Several people independently brought up a future in which AI performs most economically necessary work and governments provide income to the rest of the population.

I would not interpret this as evidence that China is preparing universal basic income.

It is not.

But the conversation itself is a signal.

The social imagination has moved.

Three years ago, most AI discussions were about whether chatbots would become useful.

Now ordinary conversations are reaching questions like:

What if society simply needs much less human labor?

That does not mean the answer will be UBI.

It means AI has moved from a technology narrative into a political-economy narrative.

That transition may happen sooner than markets expect.

The investment implication

My biggest takeaway from China is therefore not “buy Chinese AI.”

It is something more uncomfortable.

We may be overestimating how much value ultimately remains in the model layer while underestimating how quickly cheap intelligence changes the rest of the economy.

China provides an extreme test case because several forces are colliding simultaneously:

abundant open models,

rapidly improving domestic AI,

massive consumer distribution,

corporate obsession with cost,

weak pricing power,

high deployment capacity,

and a government explicitly pushing AI into the economy.

If intelligence keeps getting cheaper, these forces reinforce one another.

What does an economy look like when intelligence becomes abundant before demand does?

China may give us the answer earlier than anywhere else.

Subscribe now

AI is rapidly commoditizing the mechanical part of investment research.

Reading filings.
Comparing earnings calls.
Building models.
Searching supply chains.
Testing hypotheses.

These used to consume most of an analyst’s time.

Increasingly, they do not.

That changes where investment edge comes from.

The scarce resource is moving from answer production to question selection.

And that suggests a very different architecture for AI-native fundamental investing.

1. Semantic Radar: Find What You Didn’t Know to Look For

Traditional research begins with a company.

Let’s research Nvidia.

A better system begins with a change.

Imagine AI continuously mapping relationships across:

suppliers → products → customers → capacity → pricing → inventories → capex → earnings.

Its job is not to summarize everything.

Its job is to notice broken relationships.

Subscribe now

Read more

Yet Anthropic, with rumored ARR over $70B in late July and reaching $100B by year end, still seem to be growing rapidly while open-source models take share.

That sounds contradictory, only if we assume AI demand is fixed.

The truth is that enterprise AI usage is expanding so quickly that frontier labs can lose share of total inference and still grow in absolute dollars: Cheaper models are increasingly used for routine workloads, while frontier models remain preferred for complex coding, reasoning, and high-stakes tasks.

I discussed how I combined closed and open-source models to lower the bills for my AI native fund in this post.

But the most important question now is:

What happens when enterprises become good at deciding which tasks actually need frontier intelligence?

The architecture is changing

The first generation of AI applications was simple:

Choose a model. Send everything to it.

The next gen…

Read more

I didn’t short SaaS luckily, however I missed a great opportunity earlier in the year to buy SaaS stocks at historical low valuation.

My old thesis was simple:

AI agents → fewer employees → fewer SaaS seats → lower software revenue.

The first two steps may eventually be right.
The mistake was assuming the shift happens quickly and the rest automatically follows.

What I missed

AI makes execution cheaper.

But when the cost of producing work falls, the amount of work can explode.

More code creates more reviews, tickets and dependencies.

More agents create more workflows, permissions and exceptions.

More automated output creates greater need for context, governance and coordination.

So the important variable may not be:

How many humans use the software?

It may be:

How much economic activity flows through the platform?

Atlassian made this especially clear to me.

The obvious bear thesis was that AI coding tools would reduce develope…

Read more

Heading into June, I was almost all-in on semiconductors and AI. I believed they were the alpha engine of the world, and I’d added aggressively at the late-March lows and into early April as I viewed that the US Iran war would ease.

Then, through late May and early June, I started to feel uneasy about the frenzy and trimmed my semi/AI exposure significantly. The result is that my portfolio was down 3.5% from the peak versus 15-50% drawdowns for many semi stocks.

Three things set off my alarm:

1. Sentiment went manic. Jensen Huang offhandedly called Marvell “a great company,” and the next day Marvel stock jumped 20%. People were building websites to track his Korea trip in real time: every company he so much as visited would pop.

2. Valuation hit a historical extreme. Semi multiples were at all-time highs and stretched far above every moving average - the chart had gone vertical, shooting straight up from the late-March low like a parabola. Parabolas don’t resolve sideways.

3. The narrative quietly cracked. Silicon Valley started talking about token efficiency and pivoting away from “tokenmaxxing”, the first hint that the “spend infinitely on compute” thesis had a ceiling. Crucially, this was still confined to the smaller tech circle; it hadn’t yet transmitted to mainstream Wall Street attention.

I started adding to semi again since yesterday and bought Kioxia, semi ETFs, SharonAI, Maxlinear and others, after I did historical research on memory stock cycles.

Subscribe now

Read more

I built an investment loop consisting of AI analysts and an AI portfolio manager, then ran the process independently through Claude Code and Codex. Both systems reached a bullish conclusion. Meta was deeply unpopular, trading near its lowest valuation multiple in three years, while both LLMs saw an asymmetric payoff: much of the AI-spending risk was already reflected in the stock, but the potential improvement in was not.

Their conclusion was essentially:

The bull case is more likely to win on direction, even if the precise path remains uncertain.

Fast-forward to today. The market is beginning to reconsider Meta as the perceived return on its AI capital spending improves. More importantly, Meta is no longer presenting AI purely as an internal tool for better advertising and lower costs. It has begun charging developers for model access and is increasingly being viewed as a potential supplier of AI models, chips, compute, and infrastruc…

Read more

This is because that I’ve been practicing token efficiency for a long time. Working on my own without the luxury of an unlimited research budget, I had to be deliberate from day one.

This coincides with the most recent “Token Efficiency” trend in Silicon Valley.

A month ago, “Tokenmaxxing” was the trend. Today, Token Efficiency is what every US tech firm wants to talk about. The rise of GLM 5.2 accelerated this trend. Btw I wrote about GLM 5.2 last week citing its impressive performance in research.

Here are four Token Efficiency measures I rely on the most.

Subscribe now

Read more

Although GLM 5.2 was praised for its coding capabilities, it also did well for investment related tasks. I did a quick experiment using GLM 5.2 and was impressed.

Subscribe now

I gave it a simple prompt for stock idea generation and it did advanced reasoning + web search and came back with over 10 less visible, high-interest stock targets for further research and also clear explanations about why market gap existed.

Actually GLM’s parent company went public in Hong Kong stock market in Jan 2026 and has gone up 1745% in less than 6 months.

This is the prompt I gave GLM 5.2 and the output it generated.

Prompt:

Read more

As a mom with two kids, 10+ years of investing experiences and paranoid passion for vibe coding, I did my latest experiment: SpaceX for Kids - education wrapped in mini-games and mini-courses.

Read more

I’ve already built 30+ AI fundamental stock analysts that do various investing tasks on their own.

For the past five months I’ve been building multiple AI PM systems that pick stocks autonomously. I’ve also built them with different investing styles, mimicking different top investors or funds. And on top of AI PMs, I built an AI asset allocator that allocates virtual capital across AI PMs - lots of fun adventures.

Most importantly, they’re not AI quant investing systems. But fundamental ones. They read, form opinions about businesses, decide what’s worth owning, and trade. No human in the loop.

Here’s what I learnt.

Subscribe now

Firstly, quant was always the easy thing to automate, because computers already think in rules and numbers. Fundamental investing is the hard thing. It’s judgment. It’s deciding whether a business is good and whether the price is wrong - all questions with no clean answer. Until recently no machine could try systematically. That’s …

Read more

Anthropic released a new cool function of Dynamic Workflows yesterday. Basically you can now launch 50, 100 or more agents inside Claude Code to work simultaneously. It reminded me of the Wide Research function by Manus, which I wrote about this Jan.

The AI-Empowered Investor

Using Manus Wide Research for Investors

You might have heard of Manus, the AI agent application recently acquired by Meta. It reached 100mn ARR and 2bn valuation within a year for a good reason: Many people spent hundreds of dollars or more on it every month…

Read more

8 months ago · 4 likes · Alphaguru.ai

This is a task I use Claude Workflows to do: Years ago when I started working in the investment industry, I spent weeks of time to just gather genuine organic growth data for companies by manually cleaning and calculating data from filings and earnings call transcripts. Now I have AI.

Prompt for Claude Code:

Create a workflow - I want to gather real organic growth excluding FX effect for top 50 S&P companies for the past 5 years. Use existing sec filings to do that.

This is an overview of the workflow agents - 50 of them working together.

In 8 mins Claude Code returned this:

Workflow complete — 100 agents, 4.17M tokens, ~8 min. The verifier surfaced an important data-quality issue (NVDA’s reported …

Read more

I did a fun yet powerful experiment: I asked Codex to act as a portfolio manager and Codex is tasked to use Claude Code as the stock analyst to find undervalued AI beneficiaries.

I gave Codex (5.5 Extra High) a simple prompt and it ran fully autonomously for about 55mins. During the 55mins, Codex controlled Claude Code to do the research, then Codex reviewed Claude Code’s output and generated the final report with top 10 stock picks.

Prompt for Codex:

You are a superior fundamental stock portfolio manager. Now you want your analyst to find the best candidates for undervalued, misunderstood and overlooked stocks that benefit from AI in nonlinear way but not recognized enough yet by the market. You will direct claude code as your analyst to do all research, you will review its findings, challenge it, push it to improve and do more work, until you feel satisfied with its work. you’ll not do the work directly. understood? in this folder. The majo…

Read more

One simple prompt:

Use 3 agents to act as 3 stock investment analysts to work on one stock analysis report - one bull analyst, one bear analyst, one portfolio manager - each agent will have own chrome tab to conduct deep research to provide different perspectives for the stock. They’ll then work on a same tab to present their finding in slack, with back and forth, like deep serious thoughtful investor discussions. Finally portfolio manager will make decision whether to buy or sell the stock.

Then the portfolio manager will also create a html report to summarize all discussions and rationale for final decision.

I want see everything including slack discussions and the final html report. The stock is Meta. Each agent will behave like the best superior stock analyst with superior judgement, deep thinking and long term growth investing style. Understand?

In about 15mins,…

Read more

Here’s what I’ve learned.

Subscribe now

1. Codex is a Much Better OpenClaw

Codex can do almost everything OpenClaw does, while it’s 10x easier and safer to use Codex then OpenClaw. Also Codex’s performance is much stronger than OpenClaw.

2. Codex hallucinates less than Claude Code

Codex is more deliberate and reliable. It follows prompts precisely, stays on mission, and produces cleaner, higher-quality output with fewer bugs.

Claude Code is faster and more “creative,” but it can feel rushed and it often ignores custom instructions, including the rules I’ve spelled out in .md files.

The two complement each other nicely. For a major task or project, I’ll ask Claude Code to brainstorm and plan, then hand the implementation off to Codex.

3. Use Codex as a counterparty to Claude Code

When I’m vibe coding or running investment ana…

Read more

During the entire process, AI enabled me to conduct deep-dive research and form conviction without any external resources.

Subscribe now

The Turnaround Typology

I started with a typology. In July 2025, I had established something called the Phoenix framework, a screen for turnarounds.

Four conditions, all required: stock down a lot, an under-appreciated catalyst, a financial inflection in motion, and a valuation in the bottom decile.

Intel surfaced from the screen.

Read more

Semiconductors are the most important industry of this decade.

But in the past, two things made the sector uniquely difficult for anyone without dedicated semiconductor experts or expensive paid research.

First, a chip passes through over a dozen of layers of specialization before it ships — each with its own oligopoly, its own roadmap, its own bottleneck risk. Judging which company will outperform another in three years often requires reasoning across many of those layers at once. The technical entry barrier is legit.

Second, the stocks are cyclical. The group drops 30%~40% percent every three or four years, and each time a new cohort of investors learns the lesson for the first time.

Without institutional research access nor experts resources, I have been trying to use AI to shift where the research bottleneck is, and that shift seems to be useful.

• Today Vicor reached ATH, becoming another good semi pick I did using AI. I bought it first i…

Read more

I wrote a bull case on Google last June. The thesis worked. The stock doubled.

But I’ve recently exited Google position due to one major reason: not because ads is being disrupted by AI, but the AI game changed faster than Google adapted.

Subscribe now

Read more

One of the less recognized capabilities for AI to generate alpha is:

Find where the second-order demand hasn’t been fully priced yet.

Today I’m writing about a concrete use case for AI to identify stocks before the market fully recognized their potential.

And without privilege, expensive WallStreet research or expert calls.

The target is Lumentum, up 4x since AI helped me identify it last September.

Subscribe now

The Setup: September, Before the Crowd Fully Spread Out

Back in September, I wasn’t looking for a stock.

I was asking AI a different question:

“If AI keeps scaling, what must exist for it to work?”

Read more

Subscribe now

Path 1: Operate Inside Claude Code/Claude CoWork

This is the “Claude Code/CoWork as operating system” approach. You don’t build any permanent systems. You just work inside Claude.

Every morning, you open your terminal or Claude app, and start talking to Claude Code/CoWork the way you’d talk to one dedicated investment analyst sitting next to you.

You dispatch various tasks to Claude Code/CoWork: Scrape this earnings transcript. Compare these two companies’ gross margin trajectories over the last eight quarters. Summarize what sell-side is saying about this name. Run a DCF with these assumptions and show me how sensitive the output is to revenue growth.

Claude Code/CoWork fetches data, writes scripts on the fly, runs them, and hands you the output. When you need a chart, it makes one. When you need a table, it builds one. When you need to read a 200-page 10-K, it reads it an…

Read more