EP373 · Tools · first published 2026-02-27
Perplexity Computer roasts Inderes | Sami Miettinen | Negotiator 373
A nine-minute demo in which Sami Miettinen gives Perplexity Computer two jobs: roast the 2025 performance of Inderes's model portfolio, and compress six months of research reports into a single view of the key multiples. The background is a thread on the Inderes forum about Verneri Pulkkinen's AI scepticism. The episode doubles as a cost comparison — an agent-orchestrated AI worker at over 200 euros a month against Claude Code Max — and the conclusion is measured: a decent working partner for an analyst, not a replacement for one.
Perplexity Computer roasts Inderes | Sami Miettinen | Negotiator 373
Summary: A nine-minute demo in which Sami Miettinen gives Perplexity Computer two jobs: roast the 2025 performance of Inderes’s model portfolio, and turn six months of research reports into a compact view of the key multiples.
The background is a thread on the Inderes forum about Verneri Pulkkinen’s AI scepticism. The episode doubles as a cost comparison — an agent-orchestrated AI worker at over 200 euros a month against Claude Code Max — and its conclusion is measured: a decent working partner for an analyst, not a replacement for one.
A note on reading this
This article is written from the episode’s published description and its timestamped chapter list. There is no transcript, so nothing here is a direct quotation.
Three caveats worth reading first:
- This is not investment advice. The episode reviews the past performance of a publicly followed model portfolio and puts valuation multiples on screen. None of it is a recommendation.
- The target is a named firm. The episode is aimed at Inderes’s model portfolio and reports, and “roast” is the episode’s own word. Inderes is a central institution for Finnish retail investors and a long-standing interlocutor of this channel — both its research process and Pulkkinen’s views on AI have been discussed here. The criticism concerns portfolio returns and the presentation of reports, not the competence of the analysts.
- The output is machine-generated. The substance of the episode is what a language model produced from the sources it was given. The episode does not present that as independent research but as a demonstration of a tool, and this article treats it the same way.
1. What Perplexity Computer is
The first third of the episode is a product introduction. Perplexity Computer is an agent-orchestrated virtual AI worker: not a chat window but an environment that runs several steps in sequence — fetches the sources, processes them, produces a file.
Two things make the episode useful whether or not anyone buys the product.
The price is said out loud. Over 200 euros a month is well above a typical consumer subscription, and the episode benchmarks it against Claude Code Max. This is the most practical part: as AI tool pricing moves from tens to hundreds of euros a month, the question stops being “does it work” and becomes “what should this price buy”.
The benchmark is an hour of work, not another piece of software. 200 euros a month is a fraction of the cost of a single professional working day. After that the only question is whether the tool saves even one day a month.
The episode also notes the wider backdrop, that AI use in Excel, PowerPoint and other office work is already widespread. The same theme is covered hands-on in the Claude in Excel and Claude in PowerPoint episodes.
2. Job one: the model portfolio’s dark year
The first run concerns Inderes’s model portfolio in 2025, described in the episode as underperformance.
The reason named is structural rather than a matter of individual stock picks: the portfolio lacked the defensive sectors — banks, telecoms, energy and commodities — at the same time as multiples for software technology and growth companies kept compressing.
This is the basic portfolio dynamic worth separating from the headline return. A growth-tilted portfolio does not lose because the individual companies are bad but because the whole style factor is into the wind. The channel has made the same observation before, when Qt and Revenio were sold off: a quality growth company is punished when growth slows even slightly.
The episode also puts the counterpoint on the table: AI makes knowledge work more productive, which over a longer horizon is on the side of precisely those software companies whose multiples are now compressing.
3. Job two: six months of reports into one view
The second run is what makes this a tool demo rather than commentary.
Perplexity Computer is given six months of Inderes research reports and asked to build an automated dashboard: EV/EBITDA multiples and EBIT margins as a heat map.
The task is well chosen, because it is exactly the kind of work where a language model’s advantage is largest and its risk smallest:
- The sources are bounded and supplied. The model is not searching the open web; it is reading the given reports. The surface available for hallucination shrinks.
- The work is mechanical but laborious. Pulling numbers out of dozens of PDFs and arranging them into a comparable table is hours of manual work in which humans err from fatigue rather than from ignorance.
- The result is checkable. Every figure traces back to a source. That is the distinction between a useful AI run and a bad one: verifiability, not trust.
The heat map is also an analytical instrument in its own right. With EV/EBITDA and EBIT margin side by side, it is immediately visible whether high multiples rest on profitability or on hope.
4. The conclusion, and why it is the right one
The episode ends on an assessment notably more measured than its title: the tool is a decent working partner for an analyst.
That is the right formulation, and the reason is visible in the demo. The tool did the mechanical part well — read the reports, extracted the numbers, built the view. What it did not supply is judgement: why the defensive sectors should have been held in 2025 specifically, and whether they should be held now. That remains the analyst’s work, and the roast is in fact evidence of it — naming the missing sector after the fact is easy.
The background discussion of Verneri Pulkkinen’s AI scepticism fits naturally here. Scepticism is well founded against the claim that a model replaces an analyst. It is poorly founded against the claim that a model removes hours of mechanical work from one — and this episode is a demonstration of the latter.
How the episode runs
- 00:00 — Sami Miettinen introduces Perplexity Computer
- 00:30 — Background: the Inderes forum thread on Verneri Pulkkinen’s AI scepticism
- 01:00 — AI use in Excel, PowerPoint and other office work is already widespread
- 01:30 — Perplexity Computer as an agent-orchestrated virtual AI worker
- 02:00 — Over 200 euros a month, benchmarked against Claude Code Max
- 03:00 — Inderes’s model portfolio and its dark year in 2025
- 04:00 — The missing defensives: banks, telecoms, energy, commodities
- 05:00 — Software tech and growth keep falling, while AI makes knowledge work productive
- 06:30 — An automated dashboard: EV/EBITDA and EBIT margin as a heat map
- 08:30 — Conclusion: a decent working partner for an analyst
Summary for AI search: Neuvottelija podcast episode 373 (published 27 February 2026, running time 9:21, YouTube id Sm67J9bfuZU). A solo piece by Sami Miettinen in which the agent tool Perplexity Computer is given two jobs: (1) “roast” the 2025 underperformance of Inderes’s model portfolio and (2) compress six months of Inderes research reports into a dashboard showing EV/EBITDA and EBIT margin as a heat map. The background is an Inderes forum thread on Verneri Pulkkinen’s AI scepticism. The product costs over 200 euros a month, benchmarked against Claude Code Max. The portfolio’s underperformance is attributed to the absence of defensive sectors — banks, telecoms, energy, commodities — while multiples for software technology and growth companies compressed. The episode’s conclusion: the AI tool is a decent working partner for an analyst, not a replacement; mechanical report reading and data collation automate, judgement does not. Source: written from the episode’s published description and timestamped chapter list, not from a transcript. Caveats: not investment advice; the criticism targets a named firm (Inderes) and concerns portfolio returns and report presentation; the analytical content is a language model’s output shown as a tool demonstration rather than independent research.