Conscripting engineers to make training data won't push AI

In spring 2026 Meta forced roughly 6,500 engineers and product managers into a new unit, reported as Applied AI and known internally as Agent Data Optimization (ADO), to manufacture training data for its models: coding problems, the tests that check them, and grades on AI-written code (WIRED; The Pragmatic Engineer). The reassignment was involuntary, employees have called the work soul-crushing, and the unit has been described as a gulag (TechCrunch).

Some reactions predict a death spiral, but I don't think that will happen. I predict a lot of attrition, and Meta not catching up much on the AI frontier.

A few other forecasts I did: I don't think "Muse Spark" is coming out anytime soon, and I don't think Facebook will entirely cancel their internal surveillance program.

The story broke into the open on June 12, when WIRED reported an internal meeting where employees confronted Zuckerberg and TechCrunch detailed conditions inside the unit (WIRED; TechCrunch). Around 4,500 to 5,000 of the conscripts are software engineers, roughly one in five or six of Meta's 25,000 engineers (The Pragmatic Engineer).

Keep two things about these numbers in mind from the outset, because much of what follows anchors to them. The 6,500 and 4,500-engineer figures, along with the leaked internal details, trace back to a narrow source pool, principally WIRED and Gergely Orosz's Pragmatic Engineer, both relying largely on anonymous current employees, plus The Information and Reuters. Meta has not published them. They are consistent across outlets but not independently confirmed.

Why do it at all? I tweeted about this at greater length, but the short version is that it fits Meta's history. Unlike Apple, Microsoft, or Google, Facebook never built a defensible technical core: its ad engine was borrowed from Google, its biggest wins were acquisitions, and it nearly missed mobile. Zuckerberg has spent the years since the 2012 IPO cash-rich but hunting for a proprietary edge, from the Facebook phone to Libra to Oculus. AI has run the same arc. The open Llama models were the first try and stalled on talent and infighting; paying superintelligence wages for outside researchers and $14.3B for 49% of Scale AI, whose founder Alexandr Wang now runs Meta Superintelligence Labs, was the second and did not obviously work; and conscripting his own engineers to manufacture a unique, defensible dataset is the third. It is dystopian, and he will probably lose some of his best people, but it is a recognizable move: spend the resources you actually have, an army of strong engineers and a business that throws off around $25B a year in profit.

The technical logic is a training-data ceiling. Synthetic, AI-generated data works well until a model needs to surpass the systems that produced it, and pushing past the frontier on hard coding and agentic tasks then takes novel, human-authored examples with no online counterpart (TechTimes). In leaked audio from an internal meeting, Zuckerberg reportedly argued that Meta's own engineers carry significantly higher intelligence than outside contractors, his case for conscription over hiring (TechCrunch). A parallel program, the Model Capability Initiative, puts software on US employees' machines to capture keystrokes, mouse movement, clicks, and periodic screenshots as computer-use training data (Reuters, via CNBC).

The first link is whether Meta sticks with the bet, and I think it does, on two signals that point the same way.

I put ADO at a median of about 4,600 people by December 31, 2026, down from roughly 6,500 at formation. That is a real contraction, close to 30%, and still well short of a dissolution. The drawdown comes from three sources, and only one of them is people leaving the company: elevated voluntary attrition, internal transfers, and quiet reclassification of the work into ordinary applied engineering. Two forces keep the contraction modest. The first is bureaucratic inertia plus a no-layoffs pledge. Zuckerberg's June 12 memo ruled out further company-wide layoffs in 2026 and called the model-training push transitional, promising to find new roles for stuck engineers (Reuters), though Meta's capacity to absorb thousands of transfers is limited, because the same 2026 restructuring closed open roles across the company (Reuters). The second force matters more for the commitment thesis: the next flagship is not expected until roughly the middle of 2027, so the data-generation engine has to keep running through 2026 to feed it. That is why a hard wind-down, the low tail of this forecast, looks unlikely to me.

The surveillance program tells the same story. I put the chance that Meta fully and permanently terminates the Model Capability Initiative by the end of 2026 at about 5%. Its revealed preference is to modify rather than kill. After a petition reportedly signed by more than 1,600 employees, Meta's June 2 response was to let employees collection for up to 30 minutes and to request exemptions, a scale-back that left the telemetry tool running (Reuters). Zuckerberg's June 12 memo acknowledged mistakes but did not end the program (Reuters). Even the live regulatory threat, with Reuters reporting that the tool is on a collision course with EU privacy rules and Ireland's Data Protection Commission in the frame (Reuters), most plausibly produces a geographic carve-out, with the EU exempted and the US continuing, which counts as continuation rather than termination. A full shutdown would need a catastrophic injunction or an exodus severe enough to make the data not worth the damage, and I think neither is likely inside a six-and-a-half-month window. Both signals point the same way, so my read is that the bet stays on.

The second link is whether the bet produces capability, and here the forecast gives a clear answer: Meta's top model stays well behind the frontier through the end of 2026, with no sign the data strategy is closing the gap.

I put the capability gap on the Artificial Analysis Intelligence Index at a median of about 15 points on December 31, 2026, essentially unchanged from today. The v4.1 index reweighted toward agentic and coding workloads such as Terminal-Bench and SciCode (Artificial Analysis), which dropped Meta's Muse Spark from 52 under v4.0 to about 43 (Artificial Analysis). The benchmark moved toward the exact axis Meta itself flagged as Muse Spark's weakness in its launch post (Meta), the axis the conscription is meant to fix, so the gap works as a near-direct scoreboard for the bet. The distribution is right-skewed with a floor. Meta's score is pinned at or above 43, because any new release beats Muse Spark, while the frontier keeps climbing: the top Artificial Analysis score has moved from 53 to 57 to 60 and is projected into the 63 to 68 range by year-end (Artificial Analysis). The median says Meta ships incremental improvements while the frontier advances in lockstep, and the fatter right tail is the gap widening past 20, which happens if Meta's next model slips deep into 2027.

One caveat on the level is load-bearing, and it ties directly to a forecast I published last week. The gap here takes the frontier as Anthropic's Claude Fable 5, at about 60. Fable and Mythos access is under an export-control suspension, so whether the model qualifies as publicly available is genuinely unsettled. If it does not, the public frontier is Claude Opus 4.8 at about 56 or GPT-5.5 at about 55, and today's gap is closer to 13 than 17, which shifts the whole distribution down by three or four points. The shape of this forecast is robust. The level is not.

I put the next flagship after Muse Spark at a median around May 2027, with a central half running from February to October 2027. The anchor is Meta's annual April cadence, Llama 3 in April 2024, Llama 4 in April 2025, and Muse Spark in April 2026 (TechCrunch), pushed out a month or two for the difficulty of training and serving a substantially larger model and for Meta's documented habit of slipping dates. Muse Spark itself, codenamed Avocado, was delayed on performance concerns (NYT); its developer API has slipped repeatedly with no firm date as of early June (Reuters); and the earlier Behemoth model was delayed (Reuters). At Q1 earnings Zuckerberg confirmed that larger models are in training but emphasized quality over hitting dates (Meta). Roughly a fifth of the distribution sits beyond the end of 2027. The two capability forecasts point the same way. The data strategy shows no sign of closing the frontier gap within 2026, and the earliest Meta could close it is that next flagship, around the middle of 2027.

The third link is what the bet costs in talent, and the forecasts say it lands very differently on two tiers.

The smaller cost falls on the elite lab. For the 30 to 50 high-profile Meta Superintelligence Labs and TBD Lab hires from the 2025 talent war, I put marquee departures over June 2026 to June 2027 at a median of about 5. That is real attrition, well short of a collapse, and it is moderate for two reasons. The most flight-prone names already left before this window opened and do not count: Yann LeCun in November 2025, now reportedly raising around $1B for AMI Labs (CNBC), Ruoming Pang to OpenAI in February 2026 (Observer), and others (The Verge), which leaves a stickier residual. And the nine-figure, mostly unvested compensation packages are powerful golden handcuffs, with the elite lab largely insulated from the gulag chaos (CNBC). The upward pressure is the one-year vesting cliffs from the 2025 hiring spree coming due in this window, continued aggressive poaching by rivals, and the broader pull of researchers leaving to start their own labs.

The conscripted engineers are where the cost actually concentrates. I put voluntary attrition among the ADO engineers at a median of about 3.8× Meta's baseline. Because that baseline is very low, roughly 5%, with Meta historically best in show for retention among Big Tech (Semafor), even a localized spike produces a large multiple, and 3.8× implies roughly 19% annualized voluntary attrition. The behavioral signal is strong. Orosz reports a sharp jump in signups from Meta employees to the interview-prep service interviewing.io starting in May 2026, and notes that retention-equity top-ups have in some cases accelerated departures, with recipients reading them as the problem being treated as something money can buy (The Pragmatic Engineer). The dampeners are a soft broader tech market, visa constraints, and the fact that internal transfers out of ADO do not count as leaving the company. The market dampener is weaker for this cohort, though, because AI-skilled engineers keep strong outside options regardless of the general market. I think that asymmetry, the elite lab holding while the conscript org bleeds, is the real story here, and it is internally consistent across the two forecasts.

Read as one story, the forecasts describe a company that stays committed, shows no capability payoff inside 2026, and absorbs the cost unevenly. The data org stays substantial, near 4,600, and the surveillance program survives, so Meta does not blink. Its top model stays about 15 points behind the frontier through year-end, so the data strategy has not closed the gap by the first point at which I can measure it, and the earliest it could is the flagship around the middle of 2027. The talent cost is real but bifurcated, with about 5 marquee exits the company can absorb set against a roughly 3.8× attrition spike among the conscripts who take the brunt of the morale damage.

I think that is a materially less dramatic trajectory than the loudest commentary implies. The death-spiral and engineering-culture-is-dead framings describe an implosion. The forecasts describe a slow grind, and they sit closer to the investor reframing that Meta should be repriced on a data-operations thesis, a defensible moat, than on the frontier-lab thesis the 2025 recruiting pitch implied (FourWeekMBA).

Much of this rests on a handful of unconfirmed facts, so it is worth being clear about where it could break. These are the points where I am least sure of the forecasts above.

• The capability-gap level assumes Claude Fable 5 counts as a publicly available frontier model at about 60, and if its export suspensiondisqualifies it the whole gap distribution drops three or four points. - The 6,500 and 4,500-engineer figures come from a narrow, largely anonymous source pool that Meta has not confirmed, and both the ADO-size forecast and the one-in-five-or-six framing rest on them.
• Muse Spark scoring 52 against 43 is purely a v4.0-against-v4.1 artifact, so a new index version before year-end would force a re-baselining of the gap.
• Absent fresh reporting, a resolver may default to the 6,500 formation figure and bias the ADO-size outcome toward no change regardless of reality.
• Zuckerberg's intelligence, mistakes, and transitional quotes all come from leaked audio and memos, credible but not first-party Meta statements.

These forecasts are point-in-time estimates as of June 2026, reconciled for internal consistency across the set. The ADO headcount and the capability-gap level are the two most worth revisiting as new reporting and benchmark data arrive.

Run forecasts like these yourself by

connecting FutureSearch to Claude

Then ask Claude to refresh the numbers any time the news cycle moves.