The task of the economist will then be to figure out what out of the nearly infinite pile of words they have immediate access to is worth reading and thinking about, and how to tag the conclusions from one’s own reasoning so that they then get noticed by the people who could benefit from them.
It is going to be a wild ride.
One marker of the way things are going: Two days ago there was an email from Jason Furman:
Jason Furman: Announcing Econvitals ‘Ilaunched Econvitals, a new macroeconomics system with a[website]…[Fedspeak tracker]and a[fiscal simulator]; and automated posts on[X],[Bluesky],[Threads]and[briefings@econvitals.org]…
Indeed: <https://econvitals.org/> and <https://jasonfurmanecon.substack.com/>:
[Econvitals Underlying Unit Labor Costs]….
How it is made## This measure is intended to reflect the underlying growth rate of unit labor costs: (1) it is based on multiple measures of wage growth, (2) it is based on trend productivity growth for the cycle not the much more volatile actual productivity growth over the last quarter or year and (3) it is expressed in units that match PCE price inflation…. five published wage gauges… on a PCE-equivalent footing, then averaged with weights that favor the steadier gauges….
Caveats## A house composite, not a BLS series….
Sources## BLS (Employment Cost Index, average hourly earnings, compensation per hour, weekly earnings, productivity); Federal Reserve Bank of Atlanta (Wage Growth Tracker); BEA (PCE price index).
Entry revised 2026-10-22
<[https://econvitals.org/methods/#wage-tracker]>
What is this thing?:
anthropic/claude:What Would It Have Taken to Build Econvitals Before AI? <[https://econvitals.org/about/before-ai/]>: ‘The code… [is] a data layer that pulls about 110 sources… a library of about 790 charts, with a system that redraws them… [LLMs writing] daily, weekly and monthly [email] briefings; the site itself, including the Lab and the tools; Fedspeak, which tracks and scores what Fed officials say; and the GDP and PCE estimates and fiscal projections. About a fifth of the code was automated tests….[A human software] developer produces only about 10 to 50 finished lines [of code] a day. At 25 a day, 484,000 lines is about 84 person-years [of work]… with a plausible range of 50 to 150…
How useful is this likely to be? Is this likely to absorb my attention in a good way by containing data transformations and averages that I would spend time calculating myself from FRED <http://fred.stlouisfed.org/>?
Jason has limited ambitions for EconVitals. He does not think it will replace him, at least not yet:
Jason Furman: The Fed Is Off Its Own Reaction Function. It Should Tell Us Why <[https://econvitals.org/posts/2026-10-26-fed-reaction-function/]>: ‘Welcome to the first occasional post from me, Jason, a human. Almost everything on Econvitals is an automated briefing, but these posts will all be written by me, as are the chart(s) of the day and the opening paragraph of the Morning Note. I’m not sure how often I’ll write these posts (thus the nonspecific “occasional”). My plan is to make them always data oriented but divided between the descriptive and the explanatory with a touch of opinion…. These posts are all copyedited and fact-checked by AI, and all the data analysis is done interactively with AI. Oh, and I suck at titles, so the AI may do those as well…
And how much work will it be for Jason and company to keep this thing from drifting into irrelevance or error? Jason may say that hallucination is a solved problem as long as you gate numerical claims through deterministic scripts and as long as you have three different agents watching one another. But he is, I think, wrong. These things do drift without careful attention.
Thus there are lots of open questions here.
But this should be fun.
What Jason is doing here is very praiseworthy. And it is a marker of the transformations in how we are going to access information that are coming for us with the arrival of the attention info-bio tech mode of production.
From my point, the most interesting thing here is that Jason really does believe that the whole system will be a net time and accuracy saver as he produces things like this:
Jason Furman: The Fed Is Off Its Own Reaction Function. It Should Tell Us Why <[https://econvitals.org/posts/2026-10-26-fed-reaction-function/]>: ‘The Federal Reserve’s Federal Open Market Committee (FOMC)…Summary of Economic Projections)… have… been a very good guide… [When] it ends up raising rates by less than the dots said it would… in most of those cases that was… [a] response to inflation coming in lower… or unemployment higher….[But in] the last few years… the Fed has deviated…. Is the Fed badly off course, or is there something different, and more obviously transitory, about this inflation? I actually think the Fed has a good answer to why it is so far off its reaction function—but it should be more explicit about giving us that answer….
[Last] December the Fed expected
[core PCE inflation]of 2.5 percent in 2026, and its September projection is 3.4 percent. Under a Taylor rule reaction function that revision alone would justify a 1.35 percentage point hike in the federal funds rate…. That is way more than [we see].… That shortfall of 1.15 percentage points is unusually large….[Perhaps] the Fed is wiser than the formula. It understands, for example, that some of the inflation is the transitory result of tariffs and the Iran conflict…. I confess, I began this exercise thinking the Fed was way off base—but after quantifying it I end up more sympathetic to its approach…. I like a Taylor-style reaction function even better where the Fed articulates a path and explains that it will deviate from it based on how inflation and unemployment deviate from its forecast…. The people on the FOMC—and the staff who support them—know a lot more than the formula knows, and they should use that information in setting policy. But they should also clearly convey it…
And things like this <https://econvitals.org/tools/fedspeak/>:
which is potentially quite useful, but would have been much more work than it has value before the days of the LLM.
Ok. What do I have to say about all this? I have two things:
First: Measure inflation with the Dallas Fed’s trimmed-mean or the Cleveland Fed’s median PCE, and Jason says that the Fed “sits almost exactly on its reaction function.” Between July 2025 and July 2026, core PCE inflation rose from 2.8 to 3.0 percent. Over the same twelve months the trimmed mean fell from 2.8 to 2.3 percent, and the median fell from 3.5 to 2.7. So the headline question, “Is the Fed off course?”, can be rephrased as: “Are outlier price changes information or noise?” Suppose the big moves in a handful of categories are noise. Then the FOMC is doing what it has done since 2012, and doing it well. Suppose instead they are information, the leading edge of a broader repricing. Then the Fed is about 1.15 percentage points too loose, which puts it among the most dovish readings in its own record.
That is a discussion worth having in public.
I think honest participants should then come out of it uncomfortable rather than confident. Over 2012-2026 the record does not clearly prefer one way of measuring underlying inflation. The choice is a judgment call, and it is the judgment the FOMC is implicitly making right now.
The second thing I would highlight is Table A3. Fit the Fed’s moves to revisions in its own forecasts, and the weights come out at 1.57 on inflation and 0.99 on current unemployment. That is almost exactly John Taylor’s original 1.5 and 1. Fit the same moves to realized data instead, and the weight on trailing core inflation is negative: -0.85 over 2012-2026 and -0.23 since 1990. The FOMC has been reacting to its own forecast of inflation, not to the inflation of the past twelve months.
I approve of Jason’s wish that the Fed explain its deviations from its own reaction function, and Waller’s October speech made a good start. Waller lays out three options:
The Fed can say nothing, which risks markets pricing in anywhere from zero to five hikes.
The Fed can give strong forward guidance, which locks in a path the data may not justify.
The Fed can take the “signaling option” in between: signal roughly how far rates are likely to go over some interval, and leave the pace dependent on the data. He sees the SEP playing that role. In September, sixteen of eighteen participants penciled in at least one more hike this year.
But Waller’s signal tells us where the Fed is heading. Jason is asking why. The natural next step would have two parts. First, publish the rule-implied rate at every meeting, not just twice a year in the Monetary Policy Report. Second, state how much of the inflation surprise the committee is discounting, and on what grounds. I approve of Jason’s draft language for the FOMC:
We are discounting about half a percentage point of inflation as due to transitory factors that we expect to unwind or reverse over the coming year. If we are wrong in that expectation we will make up for it with further rate hikes…
That final sentence is the crucial one. It turns a judgment into a commitment to trigger rate increases if the data disagree with the current expectation.
Second: Economic analysis has three steps. The first is “here is the number.” The second is “here is the proper context for the number.” The third is “here is why this number, rather than the hundred others, is the one that matters today”. anthropic/claude and EconVitals do the firsr two. But MAMLMs are not good at the third step. Knowing which number matters today depends on what people are arguing about, on what the FOMC is worried about, and on what has quietly changed since last quarter.
So where does this leave the economist? The economist’s future job is less producing analysis than curating it. When a competent macro briefing costs almost nothing to produce, the scarce resources become attention and trust. That means the economist’s job is deciding what in the nearly infinite pile is worth reading. It also means tagging your own conclusions so they reach the people who can use them. Econvitals is valuable not because it has 500,000 lines of code. Lots of people will soon have 500,000 lines of code. It is valuable because Jason Furman’s name is on it, and his reputation is on the line every time the Morning Note goes out.
This is the attention info-bio tech mode of production arriving, and Jason has just given us an early view of it. What he is doing is very praiseworthy. It will be fascinating to see whether, a year from now, the system is still measuring what it says it measures. I hope it is. I expect keeping it that way will take more of his time than he thinks.
##attention-info-bio-tech-mode-of-production
##macro-outlook
##mamlms
##tuesday-trends-turmoil
#500000-lines-of-code-to-try-to-track-the-economy
#econvitals
#jason-furman
#vibe-coding
#federal-reserve
#fed-reaction-function
#taylor-rule
#monetary-policy
#core-pce-inflation
#trimmed-mean-pce
#transitory-inflation
#fomc-communication
#economic-data
#model-drift
#economics-profession
