The AI Speedometer
Where is AI going?
How the panel’s public positions map across the scenario field.
Among the 67 experts with a classified position on capability through mid‑2028, 60 affirm continued gains and 7 deny; 34 are unclassified on that question. The live, contested disagreement is commercial — whether today’s compute build‑out spending becomes enough durable sales to pay for it all.
Of ≈101 experts, 60 of the 67 with a classified capability position affirm continued gains; one‑third are silent. The live contest is commercial — whether the build‑out pays.
How the panel’s public positions map across the scenario field.
Includes lab insiders, who lean warm.
Five forces
The Forces and Questions
Every read above decomposes into five forces, each carrying one headline question and the specific sub-questions beneath it. Switch to Deck to step through them one at a time.
Appears here ≠ moves the dial. The panel runs to 101 figures, including
§4 (jobs) and §5 (politics) forecasters. Some of their placements show in the
matrix, decks, and citation feed but do not move the instrument: the §4
wage placements (C4.3) are off-dial —
cited in the feed, but they do not feed the §4 force; and the §5-brake
placers thicken the feed, but their only dial effect is through the §5 marginal.
The feed is the evidence record; the dial reads only the dial-active cells.
Per-question detail
01 · Placement matrix
101 experts, thirty-three questions.
Every read above is a pure function of this one matrix. The surface stays blurred to a density read; per-cell provenance and named placements live in the drawers.02 · Methodology
How the read is built.
How roughly 101 experts’ public positions become nine scenario reads.- What deslop.media measures
- Each expert is placed on five forces, not on a scenario: whether AI capability keeps improving; whether the build-out of power, chips, and financing keeps scaling; whether the spending turns into real revenue; whether the job market absorbs the disruption; and whether government stays out of the way. For each force, deslop.media reads one headline question plus the handful of specific sub-questions beneath it, then blends them — half the headline, half the average of the specifics — into a single stance: holds, denies, or mixed. Every stance rests on a public quote — most carry a source date (a minority are undated), and each source is marked resolved or unresolved‑live so you can see which are independently retrievable right now. Where an expert hasn’t spoken to a force, that cell stays empty, and an empty cell is never filled with a guess.
- How positions become scenarios
- A scenario is a recipe: one specific setting for each of the five forces. “Capability holds, build-out holds, revenue denies” is the Monetization Gap. The tempting shortcut is to take each force’s odds and multiply them together — a shortcut deslop.media refuses, because multiplying conjures combinations no expert actually holds and puts weight on futures nobody believes in. Instead each expert’s real five-force row travels intact and is matched against every recipe. Contradict the recipe and the expert routes nowhere — shown as unmatched, not hidden; fit it and the expert is credited in proportion to how much of the recipe they affirmed; stay silent and they neither help nor hurt. Each scenario’s score is then divided by the experts who engaged its forces at all, so a scenario isn’t penalized for thin coverage.
- Why no scenario wins
- The map reads as contested but constructive. Most experts who speak on capability affirm it keeps improving; where they disagree is on whether the money sticks, not on doom versus boom. The read stays contested and no scenario is crowned: the build-out race is the heaviest single read because more committed experts land there, not because it wins — it does not separate cleanly above the rest. The scenario field runs across nine shapes; experts populate the ones their public statements address. The ninth, State-Led Race — the government stands up and directs the build-out — now carries a small scored slice: it holds a dated, falsifiable forecast on record (the national-security state takes over the AGI build-out), not just advocacy, so it is scored rather than held empty. It is a thin wedge because only one placed forecast qualifies so far; the dial reads all nine.
- Pace
- Ordering the nine scenarios coldest (the frontier stalls) to hottest (fastest, most disruptive) and taking their weighted center of mass places one needle along that axis. The needle marks where the placed mass falls across the cold-to-hot range; the figure shows that position rather than reading a verdict into it. “Cold” means the frontier stops advancing while today’s tools keep spreading; it never means AI runs backward.
- What the panel is, and how to read coverage
- The AI Speedometer reads an AI-100 panel — a split between AI operators who build frontier models and external voices who fund, analyze, and study them (40 operators / 61 external voices). A panel weighted toward people building and financing frontier models naturally has the most on the record about capability and commercials and comparatively less about politics, labor, or regulation. That is a property of who is on the panel and how the questions were scored, not a measure of how contested or consequential any of these forces is in the world. Read every force as “what these experts have said,” not “what is true.”
- Advocacy is not a forecast
- The politics force counts only a forecast of what government will do — a predicted state action by a date — not a view of what it should do. Advocacy, an expert arguing for or against a policy, is not a forecast and does not place on this force, so the politics reading is thin, held lightly, and flagged rather than read as a panel verdict.
03 · Sources
The citation feed.
A live, append-only stream of named public statements — and the placement deslop.media reads each as. Most cards carry the question’s 12–24 month resolution test and a source date; each source is tagged resolved or unresolved‑live (a real source that isn’t independently fetchable right now — kept, not dropped). An expert appears only where their own words support a position, so this feed draws only on sourced placements. Click a matrix cell to jump in.
Live · append-only
·
— sourced placements
·
—
—
·
a quiet expert is left out, not guessed at
·
last appended —
Read honesty Every card is public-corpus inference — a verbatim public statement that deslop.media reads onto a question, most with a source date and each tagged resolved or unresolved‑live. None are statements the person gave deslop.media. The stance verb attaches to a question’s resolution test, never to a scenario or camp; a quote is as often a deny anchor as a hold.
Group
Find expert
Stance
Sort
04 · Questions
How readers ask about this.
Plain answers to the questions this figure tends to raise — for the reader, and for the agent reading on their behalf.Is this a prediction?
No — it’s a map of disagreement, not a forecast, showing where invited experts land across the open questions. On current evidence the read is contested: the panel leans toward the engine holding but is genuinely divided on which scenario follows, so the figure crowns none of them.
Why don’t the experts have a view on most questions?
Because deslop.media only places an expert where their own words clearly support a position. On most questions an expert hasn’t said something deslop.media can stand behind, so they’re left out rather than guessed at. Where many are quiet, the question is simply still open.
Who placed these — and can I?
deslop.media has synthesized the perspectives of 101 experts across AI — scientists, technologists, business leaders, economists, pollsters and others. This is the Panel view; the Model is its sibling — a deslop.media structural model over the same scenarios (illustrative, not calibrated) — reachable from the Panel ⇄ Model toggle at the top.
Why isn’t this a ranked forecast?
The percentages on the chart are shares of the classified panel mass — how the placed positions divide across the scenarios — not calibrated probabilities you could bet on. The honest signal is the shape of the disagreement, so the figure shows size and position rather than crowning a most-likely scenario; the underlying counts and the named placements are a tap away on any element.
How often does this update?
The panel is a standing read: placements are added as experts make new dated public statements, and the distribution shifts as the record moves — not on a fixed schedule. If a statement is missing or misread, email [email protected].
Can an agent read this? (MCP on-ramp)
Yes — two ways, both honest about what is true here: deslop.media records what an expert said on a dated question and classifies their stance, it does not adjudicate whether the statement is true.
- Fetch the data directly. The force-composition distribution the gauge is derived from is at
data.json, underforce_composition(the route-incidence read sits alongside it underread_scenario), and every cited placement — the expert’s byte-exact quote, the source (with a date where one is recorded), asource_statusofresolvedorunresolved-live, and deslop.media’s stance reading — is atcards.jsonld(one schema.org node per card; a position classification, not a truth verdict). No key, no MCP client needed. - Or query over MCP. Add the connector
https://api.deslop.media/mcp(a streamable-HTTP MCP server) in any client that takes a custom connector. Anonymously you getfind_expertsandquery_expertto ask an individual expert across their positions, pluscheck_coverage. The per-figure scorecard tool is not on the anonymous surface yet — usecards.jsonldfor the placement evidence.
05 · Machine channel
The read, typed for agents.
deslop.media is read by agents first — but a human rarely reads JSON, so it sits last. This block is the dial’s route-incidence read (routers ÷ placers, over the 8 scored scenarios). It is one of three named reads over one placement matrix — the Panel donut shows each scenario’s share of the whole panel, and the fetchabledata.json reports strict credited mass. Each carries its own denominator; the block’s crosswalk field names all three so the numbers are never read as the same object.on-page · route-incidence readagent ready
Fetch it directly: data.json — the distribution and every per-expert citation · cards.jsonld — one schema.org node per placement. Or query over MCP: see Can an agent read this?