From Memory-Price Forecasts to the Future of Analyst Software — Part 2 of 3

A failed forecast leaves a useful research asset

The project began with a concrete goal: gather historical NAND and DRAM memory prices, fine-tune a time-series foundation model, and build an internal Vercel dashboard for SemiAnalysis. The researcher selected a model that supported covariates—additional inputs alongside the price history. Agents found data sources, considered imperfect price data, helped select an hourly rented GPU, and supported fine-tuning and evaluation.

The workflow reflects the preference that opens this part of the conversation: small skills, explicit goals, and tasks bounded by one context window. Elaborate hooks and long chains of agent interactions had proved unreliable. The researcher instead uses focused skills to bring in relevant context, then sets a goal the agent can finish within the session.

The forecast itself looked unlikely to work. The researcher tentatively blamed market regimes: periods in which different relationships govern prices. A macroeconomic shift can reverse which strategies work, a problem the other participant argues that many language-model stock-trading games overlook. In memory markets, regimes can be nearly unique events. A long price history may therefore contain very few comparable situations from which to learn, leaving analysts with heuristics and checklists rather than confident predictions. The experiment supplies no numerical evaluation of that diagnosis.

The weekend still produced assembled price histories from public sources, paid APIs, and enhanced data sources, alongside inputs such as consumer sentiment. The next goal became explaining regimes rather than predicting prices: a shareable dashboard covering each cycle’s beginning, middle, end, and drivers. That dashboard was still planned. The researcher reports completing the information gathering in one or two days, contrasting it with earlier manual work reading annual reports, assembling spreadsheets, and constructing historical narratives.

Expert amplification creates an apprenticeship problem

Fast research raises a harder question: does the person receiving the report know enough to criticize its reasoning? The stakes rise when analysis goes to clients. The project owner acknowledges frequent mistakes and treats AI as a junior analyst under experienced supervision. It performs laborious information gathering; the human integrates the findings, evaluates scenarios, and decides what they mean.

That division of labor creates an apprenticeship problem. Repeated research helped the experienced analyst learn which patterns mattered and which analyses they could perform reliably. They do not yet see equivalent higher-level learning in their agents. Meanwhile, automating the work that trains juniors may weaken future reviewers’ judgment. The researcher is considering tighter limits on junior staff’s AI use so they still have to demonstrate independent thinking.

The researcher calls the judgment-intensive finish the “last 5%”—a figure of speech, not a measured share of effort. It includes revising assumptions and deciding what the analysis should ultimately say. Unattended agent iterations can instead compound noise and degrade context. For now, the powerful arrangement is a human directing agents and recognizing when their work has gone wrong.

Parallel attempts move thinking into review

This workflow changes the researcher’s expectations for artificial general intelligence (AGI), understood here as the ability to automate or substantially change common information jobs. Recurring quarterly analysis—finding interesting developments in a body of data—looks exposed to a thoughtfully designed agent system. That is a judgment about potential, not a measured comparison with entry-level analysts. The same speaker says their system is not AGI today: it makes errors, skips skills, and appears to deteriorate as context accumulates. A human still checks and finishes its work.

Adoption lags that sense of possibility. The researcher has encouraged New York analysts to try Claude Code, while the participants distinguish widespread daily AI use in their circles from deeper adoption of agents. The researcher predicts that Claude Code or CoWork could become baseline information-work tools within roughly 24 months, extrapolating from personal utility and competitive pressure. The entry point can be mundane: in a secondhand anecdote, a portfolio manager became enthusiastic about CoWork after using it to read email.

One mechanism behind the enthusiasm is rapid parallel attempts. An agent can be less capable individually yet still help because it starts quickly and works alongside other attempts. The human compares several approaches instead of spending the available time producing one. Review becomes a principal site of thinking.

But checking an answer is narrower than exercising judgment. In the researcher’s hypothetical example, an analyst tells an agent to prioritize valuation, then encounters evidence of fraud. Fraud can override the valuation instruction. Knowing when one consideration overrides another requires an internal model developed through experience. The participants worry that people who start directly in review will miss that learning. Repeated prompting can also become compulsive; sometimes the next useful action is to stop and think.

Reflection across sessions and an operational AGI threshold

A proposed response to missing higher-level learning is a nightly batch job that reviews the day’s activity and extracts lessons. The participants discuss `Heartbeat.md` as a pattern for periodic agent reflection. They see promise in the recurring process despite the surrounding system’s tendency to garble context and miss initial instructions.

A hypothetical customer-service agent makes the idea concrete. It would specialize in one kind of work, carry lessons across sessions, retrieve past interactions, and verify its work to reduce errors. Its advantage could be access to more history than one human can retain, applied within a narrow remit. This is a proposed architecture for a larger service, not a demonstrated autonomous business.

The conversation then introduces GDPVal, a benchmark comparing model outputs with expert work on white-collar tasks. One participant describes 50% as parity with an industry expert and says models referred to as “5.2” and “Opus 4.5” have crossed that threshold, with performance now in the seventy-something range. Exact per-model scores and evaluation settings are unspecified; the comparison concerns those tasks.

Both participants accept expert-level white-collar task performance as an operational definition of AGI. Requiring superintelligence would change the target. One remains unconvinced by stronger visions of superintelligence while staying open to changing their mind; the other says their timelines have moved substantially earlier. That optimism coexists with the practical need to supervise today’s systems.

More information work could sell for less

The economic analogy begins with agriculture: technology can greatly reduce the share of workers needed in an established occupation. Both participants expect humans to invent new work, perhaps another category beyond manufacturing, white-collar work, and mature finance. They do not expect an effortless transition. One warns that a five- to ten-year adjustment could be socially disruptive and observes that, in their experience so far, people work harder rather than gain leisure.

A deliberately speculative hypothesis follows: AI could expand the quantity of information work while reducing its price. More analysis would not necessarily translate into proportionately more economic value. One participant worries that this could challenge how gross domestic product (GDP) captures AI’s effects. The other imagines a severe adjustment—a possible “Great Depression of AI”—before society learns how to use the expanded capacity.

Useful infrastructure can still be financed at the wrong price

For infrastructure investment, one participant favors railroads over the internet as the analogy. They argue that AI has already exceeded the internet buildout in absolute cumulative spending, though the comparison lacks a common period and inflation base. Their railroad account emphasizes roughly 45 years containing three boom–bust cycles. They expect AI’s cycles to move faster, partly because information travels faster, but doubt there will be just one rise and collapse. A transformative buildout can survive several financing busts.

Comparing investment intensity requires consistent denominators and timing: - Railroad investment: after a rough recollection about GDP, the speaker gives figures of 4.8% of gross national product and 25% of gross fixed capital investment, without specifying the historical period. - Stargate spending: one participant compares it with 2% of US GDP; the response immediately notes that the spending would span more than one year. These figures do not establish an annual, like-for-like comparison.

The failure mechanism is clearer: conviction that infrastructure must be built can overwhelm price discipline. Investors recognize the cost only afterward. The speaker frames AI investment as demand still being discovered while suppliers deploy capacity as quickly as possible. They expect supply and demand eventually to meet, but do not know when and reject assuming indefinite growth.

Claude Code provides a personal example of growing demand. One participant reports outgrowing the Max plan and moving to API use, estimating the tool’s value to them at $20,000–$30,000 a year or more. That is subjective value, not annual spending or measured return on investment. Managing agents is itself a skill, and early difficulty is part of learning a new tool. The participants compare the search for agent interfaces and operating practices with the unsettled conventions of the early internet.

Replace the analyst’s interface; retain trusted data

The interface argument uses a programmer’s analogy: Excel and Bloomberg function as integrated development environments for analysts. They organize daily work but require users to learn application-specific functions, keys, and menus. One participant predicts that an agent could retrieve information, analyze it, and assemble a better interface for the task. They expect software development to move first and acknowledge that finance has not reached that maturity. The prediction assumes reliable retrieval and analysis.

A participant whose first startup attempted to compete with Bloomberg locates its core value in messaging, journalism, and data feeds. The researcher nevertheless describes an effort at SemiAnalysis to replace Bloomberg with the FactSet API plus Claude Code: trusted, verifiable data feeding an agent that performs analysis and produces charts.

The replacement decision depends on what the user needs: - Analysts need data and analytical tools. Trusted data access, analysis, and chart production may justify moving away from the terminal. - Traders also need the information network. The participants describe deals conducted through Bloomberg’s messaging system, referred to as IB, and tracked by regulators. Those relationships make replacement harder. A better charting workflow does not reproduce the network through which a trade happens.

Coding strength and research usefulness can diverge

One participant says Codex 5.3 changed their assessment of the coding competition. Its coding performance made their earlier emphasis on Anthropic’s token efficiency less decisive. They still prefer Opus 4.6 for general information work, using code as the medium for research and rubric creation rather than treating software as the final product. These are personal assessments without accompanying benchmark comparisons.

A website-research request illustrates the difference. The participant wanted Codex to ingest website contents; it tried to build scraping software. They hypothesize that coding-focused reinforcement learning encourages this behavior. Turning research into a software project can be the wrong abstraction, even when the coding is capable. Their uncertainty about Codex’s web-search access does not establish that search was unavailable.

The participant remains optimistic that combining a new pretrained model with Codex 5.3’s reinforcement-learning stack could change the competitive balance. That is a forecast about a future combination, not a demonstrated result.

The adoption discussion exposes three weak proxies for overall usage: - Commit attribution misses unsigned activity. One participant acknowledges that an earlier comparison undercounted Codex because it did not sign off every commit; a request-based comparison looked closer. - Extension installations cover only one channel. VS Code extension usage misses Claude Code’s command-line interface and potentially other installers. - Personal networks are selective. Unfamiliarity with GitHub Copilot users is challenged as a San Francisco social-bubble effect.

A reported doubling of Codex users from January to February raises a separate question: where did the growth come from? One participant argues that ChatGPT promotional modals could move substantial traffic. The other infers little contribution from a secondhand account of insufficient investment in the web experience. Neither absolute user counts nor acquisition attribution resolve the disagreement. Later praise for the Codex Mac app retains the earlier task distinction: strong personal results for coding, less success with non-coding work.

Multiple providers offer choice; middleware needs a durable role

Conductor offers a shared application for using Codex and Claude Code on different tasks. Its proposed advantage is provider choice when first-party apps favor their own tools. The counterargument concerns business durability: an intermediary above fast-growing, well-funded platforms needs a compelling reason to survive if those platforms absorb its features. The skeptical participant has not used Conductor personally, but finds value in having coding agents review each other’s work.

The choice resembles the tradeoff between multiple cloud providers and one: - Use several providers to choose the best tool for each task. - Commit to one provider to use its integrations and smaller features more deeply. The participants see no universal answer. They also caution against treating AI competition as winner-take-all: a small share can support a substantial business, and the second-largest competitor can still be very large.

Microsoft’s dilemma: grow Azure or rebuild software defenses

The closing exchange asks Doug O’Laughlin about a characterization that he thinks Microsoft is out of AI. He rejects it. His argument is that Microsoft has unusually much to lose: its horizontal software business provides the interfaces humans use for information work. Excel, PowerPoint, and email face the same potential displacement discussed through the analyst workflow.

O’Laughlin acknowledges that Azure keeps Microsoft in the race. His concern is that an infrastructure customer can become a stronger software competitor. Supplying OpenAI with GPUs earns infrastructure revenue while strengthening a partner that could challenge Microsoft’s application business.

The discussion frames two strategic responses: - Concentrate on Azure infrastructure: accept a role analogous to telecom’s commodity “dumb pipes,” with Charter as the comparison. - Reinvest in product defenses: build competing features and restore a defensible software business. Executing either path poorly, the speaker argues, could leave Microsoft in an even worse position.

The final tension is capital allocation. Rebuilding product defenses could require diverting investment from Azure, while expectations for future Azure revenue weigh heavily in Microsoft’s valuation. The speaker speculates that Microsoft excluding Azure would command a low valuation multiple, without offering a calculation. The unresolved question is how Microsoft can fund AI’s growth while defending the software that growth may displace.